Category: Semiconductor Stocks

Arm: Data Center Royalties Double YoY, Riding Grace Blackwell, Vera Rubin Growth

Posted on by io-fund

AI’s need for high-performance, energy-efficient chips creates a long-term tailwind for Arm, as the company’s heterogenous CPU architectures are seeing rapid adoption in data center applications.  

The company’s license and royalty revenue model had centered around its v9 architecture, as it commanded double the royalty of v8 at ~5%, and is featured in “virtually all high-end data center chips” and a majority of smartphones. For example, Arm’s Neoverse V2 (based on v9) powers Nvidia’s Grace CPU on its Grace Hopper and Grace Blackwell platforms, along with Amazon’s Graviton4 CPUs, Google’s Axion CPUs, and more.  

Arm is also pushing ahead with its Compute Subsystems (CSS) platform to help accelerate time to market for complex chip designs, such as Microsoft’s newest Azure Cobalt 200 CPU rolling out through 2026. CSS notably carries double the royalty rate as v9, which management placed at roughly 10%, providing another lever for growth as Arm continues its trend of doubling royalty rates per each architecture generation. 

Arm’s exact data center revenue is unclear, though data center royalties doubled YoY in fiscal Q2, likely driven by the continuing ramp of Nvidia’s Grace Blackwell platforms. For another view on data center-linked revenue, Arm’s management explained that it is reasonable to assume cloud and networking would reach 15% to 20% share of royalty revenue for the fiscal year, up from ~10% last year, or potentially up to $525 million.  

Down to the fundamentals, Arm’s revenue growth accelerated 22 points in fiscal Q2 to nearly 35% while margins expanded and FCF surged, though revenue growth is expected to normalize to 25% in Q3. While growth pales in comparison to key customers such as Nvidia, forward estimates are lower than current growth rates with analysts projecting Arm to grow at a ~20% CAGR over the next few years. However, there are a handful of tailwinds that could propel revenue growth to exceed current estimates, closer to the 25-27% CAGR range.  

Arm’s Edge vs x86 Lies in Energy Efficiency, Cost 

Arm is seeing rapid growth in the data center/server CPU market, with the company forecasting its server CPU share at the top hyperscalers to reach as much as 50% by the end of 2025, up from just 15% in 2024. These strong share gains versus x86-based chips from AMD and Intel primarily stem from Arm’s lower power consumption and price performance advantages.  

Arm’s designs are built on a Reduced Instruction Set Computing (RISC) architecture, which means the processors feature smaller, optimized instruction sets that allow the CPU to execute tasks more rapidly. This is in contrast to x86 processors, based on a Complex Instruction Set Computing (CISC) architecture, which allows the processors to complete more complex tasks with fewer instructions, leading to higher power consumption.  

RISC-based processors that Arm offers feature simpler hardware designs, accelerating the deployment process, offering lower per-chip cost and better performance per-watt. Arm also provides a foundation for ‘heterogeneous compute’ platforms, which integrate CPUs, GPUs, and NPUs to facilitate lower power and improved efficiency by allocating workloads to the most suitable processor in the platform. 

Arm’s rapid growth in AI data centers stems from this performance and efficiency advantage, as data centers are being designed and optimized for performance-per-watt as GPU racks get increasingly powerful with each generation. We had covered this in the summer of 2024, AI Power Consumption: Rapidly Becoming Mission-Critical, noting that performance and efficiency will be front of mind as the industry scales towards one million GPU clusters. We had also covered Arm’s growing data center tailwinds and support of next-gen AI chips in March 2024 in our free newsletter, Arm Stock: AI Chip Favorite Is Overpriced. 

Amazon, Google and Microsoft are all designing and deploying custom Arm-based CPUs for these significant performance and efficiency advantages – testing by Signal65 showed that AWS’ Graviton4 CPU consistently outperformed AMD and Intel chips across a variety of workloads. Amazon announced in early December that for the third year in a row, its Arm-based Graviton CPUs accounted for more than half of the CPU capacity it added. 

Source: Signal65 

Google says its Axion CPUs can offer up to 65% better price-performance and 60% better energy efficiency versus x86 alternatives, excelling at matrix-heavy inference workloads to offer a “compelling CPU-based ML inference platform, alongside GPUs and TPUs.”  

Microsoft’s Cobalt 100 CPUs boast ~48-53% better performance and ~91-99% better price-performance on real-time data processing and caching, and web infrastructure and networking workloads versus AMD’s Genoa instances in benchmark testing. Microsoft says its upcoming Cobalt 200 CPU, ramping in 2026, can deliver up to 50% better performance over the Cobalt 100.  

Meta and Arm struck a partnership in October, under which the social media giant will use Arm’s Neoverse platform to optimize its AI ranking and recommendation models. Meta said Neoverse will help it “deliver higher performance and lower power consumption compared to x86 systems” and achieve performance-per-watt parity.

Brief Background on Arm, Revenue Model and Key Products 

Arm offers the most popular CPU architecture in the world with 325 billion chips shipped since inception, of which 31 billion were shipped in FY25. Arm is most dominant in mobile CPUs with >99% market share, followed by automotive at 44%. This dominant share is achieved through its rich software developer ecosystem of 22 million, likely more concentrated in mobile whereas competing x86 is the more popular instruction set on PCs. Cloud compute and networking are smaller end markets at ~20% and 30% but quickly growing on strong AI compute demand.  

Arm’s revenue model is centered on licensing and royalty revenue for its IP, with dozens of different chip designs and platforms for a multitude of applications across smartphone, data center, automotive and other industries. 

Arm’s different licensing models are the following: 

Arm Total Access Agreements (ATA): A type of license where Arm provides a comprehensive package of CPU designs and related technologies for an annual fee. ATA has a fixed term and Arm reserves the right to modify the package by adding or removing specific products. Arm reported 48 ATA licenses as of fiscal Q2, up three QoQ and nine YoY. 

Arm Flexible Access Agreements: This model provides a selection of CPU designs and related technologies for an annual fee, although the latest products are not included like under ATA. Flexible Access customers also need to pay a single-use license fee for specific products if they are included in the final chip design. Arm had 312 Flexible Access licensees in Q2, down one QoQ but up 43 YoY. 

Arm also has Technology Licensing Agreements (TLA) that involve licensing a specific CPU design or technology to the customer for a fixed fee, either for a set term or number of uses; and Architecture License Agreements (ALA) under which customers design their own customized CPU designs using the Arm’s Instruction Set Architecture (ISA). 

Moving to products, Arm offers a wide range of different IP designs and platforms for different end markets – its Neoverse family targets AI/HPC and data center applications, while Cortex primarily targets smartphones and laptops. 

Arm’s Neoverse family includes eight different designs across three lines, Neoverse-N, Neoverse-V and Neoverse-E. N is optimized for maximal performance per watt/per dollar for scale-out applications, DPUs, networking switches and custom ASICs, and E is optimized for maximal throughput. V is optimized for maximal per core performance for HPC and memory-intensive applications, featuring 32 to 128+ cores and drawing 80-350W of power (versus 500W for AMD’s 128 core EPYC 9755 processors). Neoverse has now surpassed 1 billion cores deployed as of this quarter since launching in early 2019. 

Arm’s Cortex family includes more than 46 designs, offering customers flexibility to optimize for performance, power efficiency, throughput or more, for a range of applications from software-defined vehicles, smartphones, edge IoT devices, laptops and more. Arm also offers its Mali and Immortalis designs for mobile and consumer GPUs, as well as its Ethos NPUs for edge AI devices.  

Arm is also pushing further into Compute Subsystems (CSS), which are pre-integrated, nearly-finished CPU packages that bundle CPU cores, interconnect, memory, power management and software to reduce design time and accelerate time to market. Arm currently offers three different CSS platforms, Neoverse CSS for data centers, Lumex CSS for smartphones and PCs, and Zena CSS for automotive. Arm signed three new CSS licenses in Q2 to bring its total to 19, adding that demand for CSS exceeds its expectations. 

Powering Nvidia’s Grace, Vera CPUs 

While Arm’s designs underpin the major hyperscalers’ in-house CPU efforts, it also powers Nvidia’s Grace Blackwell and upcoming Vera Rubin platforms via the Grace and Vera CPUs. As a brief reminder, the GB200 and GB300 feature 72 Blackwell chips connected by 36 Grace CPUs, underscoring the importance of Arm’s CPU involvement within the rack.  

The Grace CPU features 72 of Arm’s Neoverse V2 cores connected by Nvidia’s Scalable Coherency Fabric (SCF) to offer 3.2 TB/s of bisection bandwidth, which Nvidia says its double that of traditional CPUs. Grace also delivers ~2x performance per watt and the highest memory bandwidth over other leading servers. In the NVL72 configuration, the Grace CPU helps deliver up to 18x faster data processing with up to a 5x better TCO.  

As a standalone CPU (Grace CPU C1), Grace delivers 1.5x to 3x faster throughput and comparable or faster performance versus x86 instances, with power consumption of just 250W or 500W including memory, versus ~400W and ~900W for x86, per Nvidia.  

Source: Nvidia 

Nvidia’s Vera CPU will feature 88 ‘Olympus’ custom Arm cores with spatial multi-threading, which, according to CEO Jensen Huang, “enables each thread to have the full throughput of a single core, giving the chip the same processing capacity as 176 cores” and enables it to optimize for performance or density at any time. Vera is also the first CPU to support FP8 precision, and features 3X more system memory and >2x memory bandwidth as Grace with less than 50W of memory power consumption, making it ideal for agentic AI, KV-cache management for inference and memory-bound workloads. 

Arm’s Long-Term Growth Centered Around Data Center Opportunities 

Arm’s long term growth opportunities are likely to be focused around AI data center deployments, considering the company’s increasing role in Nvidia’s GPU systems, along with custom CPU deployments at the hyperscalers. However, Arm is by no means a hypergrowth stock and will not experience a hypergrowth trajectory in the same fashion as some of its customers like Nvidia; rather, the growth story will center on maintaining a >20% revenue CAGR and potentially stronger earnings CAGR as data center deployments featuring its chips scale. 

Moving through 2026, Arm has solid visibility into Nvidia-linked growth as GB200/300 racks continue to ramp with Rubin on deck for the second half, backed by Nvidia’s visibility into ~$320 billion in orders for its fiscal 2027. Assuming GB200/300 rack shipments of ~28,000 to 30,000 in 2025, per Morgan Stanley, this would project to more than 1 million to 1.08 million Grace CPUs shipped.  

For 2026, analysts project GB200/300 shipments to rise to 55,000 to 70,000, or roughly doubling or potentially more than doubling YoY. Grace CPUs will match that trajectory, and if this does pan out, it can reasonably be inferred that Arm’s Nvidia-linked revenue could double in 2026. There’s also Nvidia and OpenAI’s agreement to deploy up to 10GW of AI infrastructure, said to be separate from Stargate, with the first GW coming online in the second half of 2026 on Nvidia’s new Vera Rubin platform.  

Some of Arm’s other tailwinds in the data center next year include the ramp of AWS’ Graviton5 CPUs and the rollout of Microsoft’s Cobalt 200 chips, as well as other components on the networking side, including Mellanox’s (Nvidia) BlueField DPUs, AWS Nitro DPUs and platforms using Broadcom’s Tomahawk such as those from Arista.  

Capex signals from Big Tech/hyperscalers (Microsoft, Meta, Alphabet, Amazon, Oracle) remain robust with calendar 2025 spending projected to be around $435 billion, while initial estimates for 2026 capex are around $583 billion, up 34% YoY. On a dollar basis, this points to an initial estimate of ~$148 billion in growth, versus ~$173 billion in 2025, signaling robust AI spending is poised to continue.  

Stargate represents a large long-term opportunity for Arm, serving as the core CPU provider for the project with the potential for other design opportunities in the future. In October, Stargate announced five new sites to bring its total planned capacity up to 7 GW, with Abilene expected to scale to 1.2GW by mid-2026 and two of the new sites able to scale to 1.5GW by early 2027. However, Arm continues to remain tight-lipped about the long-term annual revenue opportunities from Stargate, simply saying in Q2 that the “demand picture for compute is greater” now than it was when Stargate was first announced, and the new sites “expand visibility into future AI capacity.”  

On a broader view, forecasts point to potentially >100GW of AI data center capacity coming online through 2030. For example, McKinsey projects AI training and inference-dedicated capacity to rise from ~44GW in 2025 to more than 155GW by 2030, rising at a nearly 29% CAGR. Putting the pieces together here suggests that there will be tens of GWs of capacity coming online over the next few years that will feature Arm IP, whether it be primarily via Nvidia’s rack-scale deployments or hyperscalers rolling out next-gen custom CPUs or networking growth supporting larger clusters.  

This is one leg of Arm’s key competitive advantage, in that it is now featured in a majority of AI accelerators deployed, along with increasing presence in custom CPUs and networking components, which could be furthered in the future via its acquisition of Ethernet and RDMA startup DreamBig Semiconductor. The other leg is Arm’s price and performance per watt advantage over x86 rivals when power is arising as a key bottleneck, allowing customers to extract more compute per megawatt.  

These factors support the potential for Arm to exceed current estimates for a ~20-21% revenue CAGR over the next few years, potentially to the 25-27% range. This would likely require Arm to exceed estimates by ~$50-70 million per quarter, which Arm has shown is doable under the right conditions, on top of a ~20% growth baseline.  

For example, for fiscal 2027 (Mar ’26 to Mar ’27) revenue to grow above 26% to around $6.25 billion, Arm would need around a $70 million beat on average each quarter (what it delivered in Q2). This is supported by Blackwell and Rubin ramping in full volume combined with Cobalt, Graviton CPUs and new Tomahawk switch platforms ramping, along with other products in smartphones and automotive. Increasing blended royalty rates as v9-based and CSS-based chips shipped also support stronger growth moving through fiscal 2027. 

Modeling off a 20% growth baseline (continuing the trajectory from FY24/25/26), would place fiscal 2028 revenue at ~$7.5 billion, or ~7% ahead of current estimates for $7.02 billion. Similar quarterly outperformance as Rubin and then Rubin Ultra ramp could drive revenue up to $7.8 billion, or ~24.8% YoY.  

Potential Shift into Full Chip Design, SoftBank Ties 

Another outlet for growth could stem from a potential transition from IP licensing into full chip design, such as for standalone complex SoCs (system-on-chip) or chiplets, though details on how Arm would progress into this arena are sparse. Analysts continue to prod for details considering Arm has hinted about this move for a few quarters, but management remains quiet on exactly when and how this move could happen.  

Ross Seymore, Deutsche Bank: “You mentioned about exploring different sorts of go-to-market methodologies, chiplets, etc. When do you expect to give us more color on when that's going to go from exploration to return on investment or the actual strategy?” 

Jason Childs, Arm EVP and CFO: “The way we think about when we announce something, if it were to be something related to full SoCs, it would be once there's tape-out, once there's samples back and once there's actually noncancelable customer orders, when we achieve all 3 of those milestones, that's when we would probably talk about something because this would be a new business and something we haven't done before. So whenever those milestones are achieved, that's when you should expect to hear from us.” Arm CEO Rene Haas also mentioned that developing chiplets or SoCs would require a higher level of operating expenses, such as what is seen in the slight step-up for Q3 (adj opex moving from $648 million in Q2 to $720 million in Q3). Considering tape-out timelines at TSMC are likely around six months at the soonest, it may not be until later in 2026 or 2027 that Arm provides more details.  

This is also tied in to Arm’s parent SoftBank, who is effectively partially funding this effort and is a major customer for Arm with revenue increasing $52 million QoQ from $126 million in fiscal Q1 to $178 million in Q2, or ~16% of revenue. Arm’s management said that this is a “good run rate to assume going forward,” implying SoftBank will contribute around $600 to $700 million annually in license plus design service revenue. 

What this means is that SoftBank is licensing Arm’s IP to work with it on exploring future chip solutions, with design services “being effectively a kind of a funded R&D model,” per Arm. EVP & CFO Jason Child explained that “at some point, probably in the next year or so, you'll hear us talk about what products those might be. But, obviously, that's not just up to us. It's when SoftBank's ready to talk about what these products could look like and what the revenue profile etcetera is. And so, when that would occur, it's likely to assume that there would be somewhat different revenue source, whether it's royalties, or gross revenue from selling a chip. If in fact it's a full SoC, those are all things that are still to be worked out. And, yeah, I would think of that as being, to some extent, cannibalistic of whatever the current license and design services.” 

This move could help Arm unlock more value for its IP from selling chips externally instead of simply collecting a single-digit royalty fee – for example, Nvidia raked in more than $51 billion in data center revenue whereas Arm’s entire royalty revenue was $620 million. Even if Arm successfully orchestrates a move and can ship a couple hundred million worth of self-developed chips quarterly (below 0.5% of Nvidia and AMD’s combined data center revenue), this could still represent a huge boost to Arm’s revenue generation. 

Financials 

Revenue Accelerates 22 Points in FQ2 

Arm’s revenue growth accelerated more than 22 points from 12.1% YoY in fiscal Q1 to 34.5% YoY in fiscal Q2 to $1.13 billion, while QoQ growth rebounded from (15.1%) to 7.8% QoQ. Growth has been lumpy historically. 

Royalty revenue increased 21% to a record $620 million, with the largest contributors to growth being smartphones with higher royalty rates and data center, with royalty revenue doubling YoY.  However, this marked a slight deceleration from 25% YoY growth in the prior quarter. Licensing revenue rose 56% YoY to $515 million on normal timing fluctuations, accelerating from (1%) growth in the prior quarter.   

For Q3, Arm guided for revenue of $1.225 billion at midpoint, though this represents a deceleration to 24.6% YoY and 7.9% QoQ growth. Royalty revenue is guided to be up just over 20% YoY, maintaining Q2’s growth or a marginal acceleration, while license revenue is guided to be up 25-30% YoY. 

AI Revenue  

Arm does not provide specifics into its data center revenue contributions, but as noted above, data center royalties doubled YoY on continued deployment of Arm-based chips at hyperscalers. Data center Neoverse royalties more than doubled YoY, and Arm expects to reach 50% share in of CPUs deployed by hyperscalers by the end of 2025.   

For another view, Arm’s management explained that it is reasonable to assume cloud and networking would reach 15% to 20% share of royalty revenue for the fiscal year, up from ~10% last year. Assuming Q3 and Q4 see royalty revenue rise ~20% YoY, this could project cloud and networking’s contribution for fiscal 2026 to ~$394 to $525 million.  

Key Metrics  

Arm’s key metrics were mixed in Q2, with annualized contract value (ACV), normalizing license revenue, showing strong growth yet RPO declined. ACV increased 5% QoQ and 28% YoY to $1.6 billion, its second quarter of 28% YoY growth and a strong acceleration from the low/mid-teens previously.

However, RPO declined (6%) YoY but was up 1% QoQ to $2.25 billion, reversing from a 3% increase in Q1. RPO growth has struggled over the prior five quarters, with Arm reporting YoY declines in four of these five. Arm expects to recognize ~29% of RPO as revenue over the next 12 months, or ~$651 million.  

Margins 

Arm saw strong margin expansion down the line, with operating and net margin expanding at a much larger degree than gross margin in Q2, signaling that adoption of its higher margin v9 and CSS platforms is translating to bottom line strength.  

  • GAAP gross margin was 97.4% in Q2, up 1.2 points YoY and 0.2 points QoQ.   
  • GAAP operating margin was 14.4%, up 6.8 points YoY and 3.6 points QoQ. Adjusted operating margin was 41.1%, up 2.5 points YoY and 2 points QoQ; for Q3, adjusted operating margin is implied to be ~39.4% at midpoint assuming gross margin is flat QoQ.    
  • GAAP net margin was 21%, up 8.3 points YoY and 8.7 points QoQ.  

Earnings 

Arm delivered strong GAAP EPS growth in Q2 as margins expanded down the line, while adjusted EPS growth was more muted but solid nonetheless.   

GAAP EPS was $0.22 in Q2, up 120% YoY and more than 66% ahead of estimates for $0.13. Adjusted EPS was $0.39, nearly 18% ahead of estimates for $0.33 and representing growth of 30% YoY.   

For Q3, Arm guided for adjusted EPS to be $0.41, +/- $0.04, for YoY growth of just 5%. Q4 is estimated to see growth of just 2.7% YoY to $0.56, before reaccelerating to >29% YoY growth in each quarter of fiscal 2027.  

For fiscal 2026, Arm is expected to earn adjusted EPS of $1.72, up 5.4% YoY, before accelerating to 32.2% growth in fiscal 2027 to $2.27. 

Cash 

Cash flows improved substantially on a YoY basis, and Arm’s balance sheet remains robust and debt-free.  

  • Operating cash flow was $567 million for a 50% margin, up from a 0.7% margin in the year ago quarter and a 31.5% margin in the prior quarter.  
  • Adjusted free cash flow was $411 million for a 36.2% margin, a significant increase from (7.7%) in the year ago quarter and 14.2% in the prior quarter.   

Cash and equivalents totaled $3.26 billion and debt was zero.  

Notable Risks

Arm has a handful of key risks, notably its premium valuation compared to other leading AI chipmakers despite lagging on growth metrics, and that the AI buildout will more directly benefit the primary AI data center beneficiaries while AI while see barely a fraction of AI spending. This premium valuation versus its customers is not new to the story, as we had covered this in August 2024 in our newsletter, Arm Stock: Buy Its Customers, Not The Stock. 

Arm trades at 25.3x forward PS with revenue growth expected to be ~21%, whereas Nvidia trades at a 21.4x multiple with growth projected to be 3x the rate of Arm’s at 63%. Broadcom also trades at 17x with AI revenue likely more than doubling this year to more than $40 billion, or more than 8X Arm’s projected annual revenue. However, it is important to note that Arm is trading at the lower end of its valuation range since its IPO, having traded as high as 50x forward PS and as low as 16x (for an average of nearly 32x).  

This valuation premium is matched on the bottom line, with Arm trading at a 67.2x multiple, versus both Nvidia and Broadcom at 40x and 34x respectively. This premium valuation presents risks considering Arm again is growing much slower than its peers, with EPS growth projected to be 5% for Arm versus 57% and 49% for Nvidia and Broadcom. 

Even with Arm increasing royalty rates by 2x with each new architecture, from 2.5% with v8 to 5% with v9 and now to 10% with CSS, Arm’s growth may continue to lag that of peers as the AI buildout progresses, and it may have to take the leap into design to capture more incremental revenue and accelerate growth significantly. 

The smartphone market will be key to watch throughout this year as rising memory prices are expected to impact growth, with IDC projecting the market to decline (0.9%) in 2026, revised from a prior view for 1.2% growth, and other groups forecasting a decline of more than (2%) YoY. Considering smartphones contributed ~45% of royalty revenue in fiscal 2025, data center growth may not be enough to offset a soft smartphone market this year. 

Arm also faces a higher degree of related-party risk from SoftBank, with analysts from BofA believing that SoftBank could account for 25-30% of licensing revenue, and that fiscal 2026 licensing revenue could decline (5%) YoY when excluding SoftBank.  

China exposure presents a risk, with the geography contributing approximately ~22% of revenue in Q2; for the first half of fiscal 2026, China accounted for 21% of revenue, up 3 points versus the same period in fiscal 2025. Arm did say that the “demand in China looks to be as strong as we've ever seen” and it recorded one of its largest license deals in the quarter, though China is openly supporting RISC-V. This new architecture is Arm’s open-source competitor, which emphasizes register access over direct memory access, which may be more suitable for parallel processing. While it is unlikely that RISC-V overtakes Arm in the near-term, it could become a serious contender in future years and a headwind in a major market, given Chinese firms such as Alibaba and others have launched RISC-V CPUs and server CPUs this year.  

Conclusion 

Arm’s presence in the data center is sharply rising as it is powering some of the most important AI platforms currently (and soon to be) shipping, including Nvidia’s Grace Blackwell and Vera Rubin and the hyperscalers’ custom CPU efforts. The ramp of these platforms through 2026 and 2027, combined with strong AI capex trends and a focus on performance per watt as power emerges as a key bottleneck can drive strong growth for Arm over the next few years.  

However, the major downside to Arm’s model is that the company only sees a small percentage of the end market value it creates, and at times it can be better to own Arm’s customers instead of Arm in the midst of these strong trends. For example, mobile handsets created a $200+ billion segment for Apple yet only resulted in (roughly) $3 billion for Arm. The deployment of hundreds of GW of AI data center capacity could require $3 trillion to as much as $7 trillion in spending, yet Arm is only currently expected to scale from less than $5 billion in revenue to almost $12 billion in annual revenue by 2030, barely seeing a fraction of this growth.  

For now, we are passing on Arm yet will certainly reconsider if the company pivots toward design.

Damien Robbins, Equity Analyst at I/O Fund contributed to this analysis.

Every Thursday at 4:30 pm Eastern, the I/O Fund team holds a webinar for premium members to discuss how to navigate the broad market, as well as various stock and crypto entries and exits. Beth Kindig offers weekly deep dives including lesser-known cryptocurrencies and AI stocks, plus the team offers trade alerts. The I/O Fund team is one of the only audited portfolios available to individual investors. If you’d like to subscribe to the Advanced Market Signals plan, email us at premium@io-fund.compremium@io-fund.com.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

I/O Fund Portfolio & Must-Read Theses

Posted on by io-fund

Below are our current positions and corresponding theses. In most cases, we have written about the stock many times. What is listed below are the most pertinent analysis for becoming acquainted with the stocks we currently hold. If you want to read more, please use our search bar and search by stock name to pull up more archived articles.

This list will be updated and refreshed when positions are added or removed. Please check back often for updates!

Audited Returns

Quarterly Webinars and Analysis

Nvidia

Astera Labs

Alphabet 

Applied Optoelectronics

Arm

AMD

Bloom Energy

Broadcom

Coherent

GE Vernova

Lumentum

Micron

Palantir

SanDisk

Last updated on 06/18/2026Last updated on 06/18/2026

I/O Fund Portfolio & Must-Read Theses

Posted on by io-fund

Below are our current positions and corresponding theses. In most cases, we have written about the stock many times. What is listed below are the most pertinent analysis for becoming acquainted with the stocks we currently hold. If you want to read more, please use our search bar by also ticking Pro in the filters and search by stock name to pull up more archived articles.

This list will be updated and refreshed when positions are added or removed. Please check back often for updates!

Audited Returns

Quarterly Webinars and Analysis

Nvidia

Astera Labs

Alphabet 

Applied Optoelectronics

Arm

AMD

Bloom Energy

Broadcom

Coherent

GE Vernova

Lumentum

Micron

Palantir

SanDisk

Last updated on 06/18/2026Last updated on 06/18/2026

AMD Q3: The Catalyst is Expected in H2 2026, Could Ramp Sooner

Posted on by io-fund

AMD’s bigger moment was intra-quarter when the OpenAI deal was announced, as it’s a clear signal the company is able to grab the attention of AI’s leading development firm. According to Lisa Su, the 6GW deal is expected to amount to “generate well over $100 billion in revenue over the next few years.”  

However, as noted in July’s Top 15 Report, “the risk to AMD is primarily in Q2’s data center growth decline, and how quickly the company can ramp its MI355s and subsequent MI400s while in the midst of Nvidia’s large shadow – will we see a solid surprise arrive in Q3, Q4 or even into next year? My best guess is the most meaningful AMD moment is not likely to occur during Blackwell’s NVL72s release – I think 2025 belongs to Nvidia and somewhere between 2026-2027 we switch it up.” 

After putting the loss of China revenue in the rear-view mirror, the data center segment sharply rebounded this quarter, up 22% YoY and 34% QoQ for revenue of $4.34 billion. However, we aren’t quite there yet in terms of a strong inflection as it was stated data center would grow 4% QoQ with strong growth server (a nod toward CPUs instead of GPUs). 

AMD is a stock where I’ve been intentional about managing expectations. The upside is compelling — as the second place in data center GPUs is wide open. Yet for those who have followed our coverage, the timing has always been key: meaningful execution in AI accelerators is not expected to materialize until the second half of 2026. In other words, the long-term opportunity is substantial, but patience remains part of the thesis.

MI400s Arriving in H2 2026 

As stated in the most recent Top 15 AI Stocks report: “The MI400 series will be the start of rack-scale systems for AMD, starting with Helios, which will connect up to 72 GPUs similar to Nvidia’s NVL72 systems […] AMD stated in their last earnings report they have an ambitious goal of reaching tens of billions in MI400 sales. Investors should take note that management is specifically calling out the MI400 for this, arriving in H2 2026. The readthrough is that OpenAI is an early validator that the MI400s have serious chops, and where OpenAI goes, the rest of the AI market tends to follow.” 

I’m quoting all of my previous comments so the timing is crystal clear. Here was the update this evening in terms of timing – aligned with my current understanding: “But given what we see today, we see a very good demand environment into 2026, so we would expect that MI355 continue to ramp in the first half of '26. And then, as we mentioned, MI450 Series comes online in the second half of 2026, and we would expect a sharper ramp as we go into the second half of 2026 of our data center AI business.” 

Last month, more information was shared on AMD’s Helios systems, primarily that Meta’s Open Rack Wide specifications were met, which refers to improvements for power, cooling and serviceability. The end result is ecosystem validation that AMD can offer an open standard for AI infrastructure based on Meta’s data center rack designs.  

One key area where Helios stands out is memory — the platform offers roughly 50% more total memory capacity compared to Nvidia’s Vera Rubin rack architecture. That said, if you’ve followed AMD’s AI story as closely as I have (and I know many of you have), then the most important leap in this generation of GPUs is not found in Helios specs or even this quarter’s commentary. Rather, it’s in the demand signals. For the first time, some of the most influential AI customers — including OpenAI, Oracle, and Meta — are preparing to deploy the MI400 Series in meaningful volume. That level of hyperscaler commitment is something AMD hasn’t enjoyed in prior GPU generations (MI300s), and it represents an important shift in the company’s competitive positioning. 

AMD Signs 6GW Partnership with OpenAI 

In early October, AMD signed a landmark deal with OpenAI to supply the ChatGPT parent with 6GW worth of GPUs, starting with 1GW worth of AMD’s MI450 GPUs in the second half of 2026. AMD said the deal would be worth “tens of billions” but declined to provide exact specifics, yet analysts have chalked the deal as worth potentially upwards of $100 billion at the full 6GW scale. In conjunction with the deal, AMD is issuing a warrant to OpenAI to purchase up to 160 million shares. This enables OpenAI to take up to a 10% stake in the chipmaker, though vesting will not begin until the first 1GW deployment and is tied to certain share price targets. 

The deal is expected to provide significant upside potential for both revenue and earnings by 2030, per BofA’s estimates, even assuming a significant discount to Nvidia’s GPUs on a GW basis – AMD’s opportunity per GW is pegged at $17.5 billion, compared to $25 billion-plus for Nvidia’s Blackwell Ultra GPUs.  

BofA’s scenario analysis projects AMD’s total revenue as high as $63.1 billion by calendar 2027 assuming one full GW is deployed, a rather quick timeline considering first shipments are not expected to commence until the second half of next year.  

Under this assumption, BofA projects AMD’s earnings power as high as $10.15, a 35% uplift to consensus estimates at the time for $7.50. By calendar 2030, deployments are expected to culminate with 2GW, or ~$35 billion assuming the opportunity per GW remains flat at $17.5 billion; this could result in EPS as high as $15.80, per BofA, 47% above consensus prior to the deal.  

However, considering that next-gen GPUs continue to command higher prices (such as Nvidia’s Rubin and Rubin Ultra moving to $30-35 billion per GW), AMD may also be able to charge a higher premium for its GPUs and still maintain a significant price-performance advantage to Nvidia. Thus, assuming a mid-$20 billion per GW opportunity by 2030, AMD could see $45 billion-plus with 2GW delivered in the final tranches. Given consensus estimates were ~$65 billion prior to the deal, this would project revenue potentially at $110 billion by 2030. 

Our estimates are aligned with Lisa Su’s commentary, where she stated in the opening remarks that “We expect this partnership will significantly accelerate our data center AI business with the potential to generate well over $100 billion in revenue over the next few years.” 

One major question surrounding the deal is OpenAI’s spending spree, and how the company will not only fund this, but its other GPU and cloud computing deals it has signed over the last month. In September, OpenAI had already projected cash burn at $115 billion through 2029, yet has signed $1.4 trillion worth of deals with Nvidia, Oracle, Azure, AWS and others.  

Commentary for AI Growth in FY26-FY27 

The last guide that AMD provided on GPUs was $6.5 billion in revenue by the time we exit this year. Management is hinting they will see “tens of billions” in their AI business by 2027. If we assume this means a minimum of $20B (perhaps more) then it coincides with roughly minimum 200% growth in AMD’s AI business over a two-year time span.  

“In summary, our AI business is entering a new phase of growth and is on a clear trajectory towards tens of billions in annual revenue in 2027, driven by our leadership rack scale solutions, expanding customer adoption and an increasing number of large-scale global deployments. I look forward to providing more details on our data center AI growth plans at our Financial Analyst Day next week.” 

Current Consensus Estimates Show Mismatch In FY28-29 

Though it is still uncertain as to how the OpenAI deal will ramp with the subsequent 5GW and timing for those deployments, consensus estimates still show a mismatch in FY28-29, with YoY growth pegged at <2%.  

Some of this stems from the inherent difficulty from projecting 3+ years into the future (and a much smaller # of analysts projecting long-term, dropping from 32 in FY28 to 5 in FY29). However, running off the assumption that 6GW is worth >$110 billion with the opportunity per GW rising from $17.5 billion to mid-$20 billion over the course of the deal, there is a >$30 billion mismatch in forward estimates.  

For example, post-deal, estimates for FY27 have risen 22.5%, FY28 by 37% and FY29 by 22.5%. On a dollar basis, FY27 has risen by nearly $11 billion, FY28 by $14 billion, and FY29 by $11.5 billion. Adding in the $2 billion jump in FY26 and a ~$48 billion jump in FY30 (limited data but initial consensus at approx. $65 billion), the total increase in estimates amounts to ~$86.5 billion.  

Based on rough back-of-the-napkin math for ~$110 billion in the total opportunity, $23.5 billion is unaccounted for, likely landing in the FY27-29 time frame as deployments ramp. Thus, there could exist future upside to revenue estimates later in the decade as the pace and timing of the ramp becomes more clear.  

Oracle to Deploy 50K MI450 GPUs with Expansion Potential 

Oracle has emerged as another large, public backer for AMD’s upcoming MI450 GPUs, with the company announcing on October 14 that it would be deploying an initial 50,000 GPU cluster starting in the second half of 2026, with room to expand in 2027 and beyond. This builds on an existing planned deployment of a zetta-scale cluster of 131,072 MI355X GPUs announced earlier this summer. 

This deployment is expected to carry an all-in cost of $3.5 billion to $4 billion to Oracle for ~700 72-GPU racks, including storage and networking, or nearly $5.4 million per rack at the midpoint. For comparison, Nvidia’s GB300’s are estimated to carry an all-in cost of $80,000 per GPU, or ~$5.6 million per rack. 

Oracle Cloud Infrastructure executives said that they believe “customers are going to take up AMD very, very well — especially in the inferencing space.” This is where AMD is packing a punch with 31.1TB of HBM content in the MI450’s Helios rack, 1.5x more than the GB300 NVL72, to significantly increase bandwidth and throughput for inference tasks. The I/O Fund was early to discuss this angle in the analysis “AMD vs Nvidia” 

AMD expressed confidence in delivering for Oracle in future years, as well: “Oracle announced they will also be a lead launch partner for the MI450 Series, deploying tens of thousands of MI450 GPUs across Oracle Cloud Infrastructure beginning in 2026 and expanding through 2027 and beyond.” 

Q3 Revenue Grew by 36% 

AMD’s Q3 revenue grew by 35.6% YoY and 20.3% QoQ to a record $9.25 billion, beating estimates by 5.7%. The revenue growth accelerated by 400 basis points from the 31.6% growth reported in Q2, reflecting strong momentum across the data center AI, server and PC businesses. The strong sequential revenue growth was primarily driven by growth in the data center, client & gaming segment, as well as modest growth in the embedded segment. 

The company also guided for a strong Q4 revenue of $9.6 billion at the midpoint, representing a YoY growth of 25.4% and 3.8% sequentially. It beat the analyst's estimates by 4.3%. The revenue growth will be primarily driven by strong double-digit growth in the data center and client & gaming segments. Similarly, to the last quarter, the revenue guidance does not include any MI308 chip sales to China. However, this time management indicated that MI308 chip sales could be coming soon.  

“So look, it's still a pretty dynamic situation with MI308. So that's the reason that we did not include any MI308 revenue in the Q4 guide. We have received some licenses for MI308, so we're appreciative of the administration supporting some licenses for MI308. We're still working with our customers on the demand environment and sort of what the overall opportunity is. And so we'll be able to update that more in the next couple of months.” 

Analysts expect revenue to grow 19.4% YoY to $8.88 billion in Q1 and then accelerate to 24.6% growth to $9.58 billion in Q2 2026. Looking forward, analysts expect revenue to grow 27.9% YoY to $42.33 billion in 2026 and accelerate 8.5 percentage points to 36.4% YoY growth to $57.72 billion in 2027. 

Data Center Segment Grew by 34% QoQ 

Data Center revenue rebounded strongly in Q3 as it grew by 22% YoY and 34% QoQ to a record $4.3 billion. The strong growth was primarily driven by the ramp of the Instinct MI350 Series GPUs and server share gains. Server CPU revenue reached an all-time high as adoption of 5th Gen EPYC Turin processors accelerated rapidly, accounting for nearly half of overall EPYC revenue in the quarter. The sales of prior generation EPYC processors also continued to be strong. 

The company also reported record sales as hyperscalers expanded EPYC CPU deployments to power both their own first-party services and public cloud offerings. Hyperscalers launched more than 160 EPYC-powered instances in the quarter. Currently, there are more than 1,350 public EPYC cloud instances available globally, up by about 50% YoY. 

Management expects cloud demand to remain very strong as hyperscalers are significantly increasing their general-purpose compute capacity as they scale their AI workloads. Many customers are now planning substantially larger CPU buildouts in the coming quarters to support the strong AI demand. Also, enterprise demand is very strong as the EPYC server sell-through increased sharply YoY and sequentially, reflecting accelerating enterprise adoption. 

AMD’s Instinct GPU business continues to accelerate. It is witnessing a sharp ramp of MI350 GPU sales and broader MI300 deployments. Multiple MI350 Series deployments are underway with large cloud and AI providers, with additional large-scale rollouts on track to ramp up over the coming quarters. 

Management was quite optimistic about future AI business growth. “Looking ahead, our data center AI business is entering its next phase of growth with customer momentum building rapidly ahead of the launch of our next-gen MI400 Series accelerators and Helios rack-scale solutions in 2026.” 

Client and Gaming Segment Grew by 73% YoY 

The client and gaming segment grew by 73% YoY and 12% QoQ to $4.05 billion. The strong growth was primarily driven by the acceleration in the Ryzen portfolio. It was stated: 

“Our PC processor business is performing exceptionally well with record quarterly sales as the strong demand environment and breadth of our leadership Ryzen portfolio accelerates growth. Desktop CPU sales reached an all-time high with record channel sell-in and sell-out led by robust demand for our Ryzen 9000 processors which deliver unmatched performance across gaming, productivity and content creation applications. OEM sell-through of Ryzen-powered notebooks also increased sharply in the quarter reflecting sustained end customer pull for premium gaming and commercial AMD PCs.” 

The gaming revenue grew by 181% YoY and 16% QoQ to $1.3 billion. The strong growth was driven by higher semi-custom revenue and strong demand for the Radeon GPUs. Management stated, “Semi-custom revenue increased as Sony and Microsoft prepare for the upcoming holiday sales period. In gaming graphics, revenue and channel sell-out grew significantly driven by the performance per dollar leadership of the Radeon 9000 family.” 

Embedded revenue was down (8%) YoY and up 4% sequentially to $857 million. Revenue increased sequentially as the demand environment strengthened across multiple markets. 

Margins 

The company’s profits are growing. However, margins are negatively impacted by higher operating expenses to support strong future AI opportunities.  

  • The company’s Q3 gross profits grew by 40% YoY and 56% QoQ to $4.78 billion. The gross margin was 52%, up 200 basis points YoY primarily driven by a higher profitable product mix. The adjusted gross margin was 54%, in-line with management guidance. Management has guided an adjusted gross margin of 54.5% for the fourth quarter. 
  • Operating income was up 75% YoY and up 1048% QoQ to $1.27 billion. The operating margin improved by 300 basis points YoY to 14%. Adjusted operating margin was down by 100 basis points YoY to 24% and missed the management guidance of 25% as the adjusted operating expenses increased by 42% YoY to support the significant AI opportunities and go-to-market activities for revenue growth. Management has guided an adjusted operating margin of 25% for the fourth quarter. 
  • Net income was up 61% YoY to $1.24 billion or 13% of revenue, up 200 basis points YoY. The adjusted net income was up 31% YoY to $1.97 billion or 21% of revenue, down 100 basis points YoY.

Adjusted EPS Grew by 30% YoY 

The company’s GAAP EPS grew by 59.6% YoY to $0.75, beating estimates by 10%. Adjusted EPS rose by 30.4% YoY to $1.20, beating estimates by 2.4%. 

Analysts expect adjusted EPS to grow by 22.3% YoY to $1.33 in Q4 and accelerate to 26.4% growth in Q1 and 183.2% YoY growth in Q2 to $1.36. Looking forward, they expect the adjusted EPS to grow by 61% YoY to $6.35 in 2026 and 45.7% YoY to $9.25 in 2027.

Cash Flow and Balance Sheet 

The company’s cash flows are growing primarily driven by higher revenue and profits.  

  • Q3 operating cash flows grew by 185% YoY to $1.79 billion or 19% of revenue, up 10 percentage points YoY. 
  • Q3 free cash flows grew by 208% YoY to $1.53 billion or 17% of revenue, up 10 percentage points YoY. 
  • The company had cash and short-term investments of $7.24 billion at the end of the quarter, up from $5.87 billion in the previous quarter. While debt remained the same at $3.22 billion. 
  • Inventories increased by 10% sequentially to $7.3 billion. 

Conclusion 

For the far majority of stocks, we would not have a placeholder in the I/O Fund portfolio this far ahead of execution. AMD is unique because the data center GPU market desperately needs a second-place contender. Investors may appreciate Nvidia’s pricing power, but hyperscalers and companies like OpenAI do not; they’d like to see more competition and optionality including lower prices. That is why we are seeing Meta work alongside AMD to bring Helios to market. There are many investment opportunities in AI across AI networking, AI energy, AI software, AI data layer and more – but none compare to the sheer size and strategic importance of GPUs, particularly when there are so few players competing for that share. That scarcity dynamic is precisely why AMD remains a special case in our portfolio. 

Equity Analyst Royston Roche contributed to this analysis. 

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

Astera Labs Q3 Earnings: Blowout Report Meets UALink Uncertainty  

Posted on by io-fund

Astera Labs reported another impressive quarter in Q3 with revenue maintaining 20% QoQ growth, beating estimates by 11.7%, while GAAP operating and net margins showed solid QoQ expansion. Fiscal 2025 revenue is expected to be around $831 million at midpoint, $55 million above current estimates for $776 million and pointing to approximately 110% YoY growth. 

There are two items weighing on Astera’s report this evening. The first is the guide for Q4 of $249M at the midpoint, which implies 76.5% YoY growth yet 8% QoQ growth. This is a slowdown from the 20% QoQ growth this quarter.  

However, if we zoom out, we see this is actually a large beat as Q4 was slated to report $216.5M equaling a 15% beat. Also consider Astera has seen significant beats every quarter even after raising guidance, thus the conservatism we’ve seen thus far could continue. When looking at the bottom line, we see % surprise has been exceptional – don't believe I’ve seen this consistency and magnitude of bottom line beat before. 

Fundamentals aside, Astera’s stock has recently been haunted by Ethernet for Scale Up Networking (ESUN). ESUN was announced mid-October following the Open Compute Project conference with support from AI chip partners and strong Ethernet players, including Nvidia, AMD, Broadcom, Arista, Cisco, HPE, plus a few hyperscalers such as Meta, Microsoft and AI heavyweight OpenAI. The week of the announcement, Astera dropped as much as 33% from an October high of $225 to $154.  

Astera’s product line has greatly benefited from the boom in the PCIe networking protocol, and the readthrough is that PCIe will see competitive pressure from Ethernet.  

There is a lot to unpack from the earnings report, yet the PCIe versus ESUN debate takes precedence. As you can expect, the Financials section shows a trend very much in play, but we have to address the boogieman first before we can assess Astera’s ability to continue its remarkable growth trajectory.  

UALink versus Ethernet Scale-Up Networking (ESUN) 

UALink is an open-source alternative to proprietary interconnect protocols such as Nvidia’s NVLink. Astera offers a portfolio of UA linked connectivity solutions, including AI fabrics for signaling conditioning and other I/O components. The expectation is that PCIe’s low latency would carry UALink as a solid alternative to Nvidia’s NVLink with proliferation expected around 2027.  

In the past, Astera has counted ten customers as leveraging PCIe in the short term and a combination of PCIe and UALink in the midterm before “transitioning perhaps to a broader UA Link deployment in 2027 and later.” 

The market is concerned because ESUN is now a third viable option and one that comes with a sense of familiarity given its Ethernet based, has large backers and perhaps most importantly – can exceed UALink when it comes to time to market. As you’ve likely picked up on with our energy coverage, time to market is everything right now. The ESUN consortium is banking on Ethernet moving quickly for scale-up networking.  

As it stands now, PCIe is optimized for scale-up networks due to low latency and speed measured in nanoseconds whereas Ethernet is best for scale-out networks. However, ESUN is proposing an Ethernet solution for scale-up with the press release stating: “ESUN is a new workstream collaboration designed as an open technical forum to advance Ethernet in the rapidly growing scale-up domain for AI systems.” 

There are quite a few details to consider in terms of how this plays out, yet the most likely outcome (from where I stand today) is that both are needed. UALink has specific benefits that will be tough for Ethernet to displace. Here is the technical description: “The [UALink] specification enables load, store, and atomic operations between 100s of GPUs while optimizing the protocol stack to minimize end-to-end latency, reducing valuable die area on GPUs and switches, and reducing interconnect and switching power consumption. UALink will support state-of-the-art up to 200Gbps per lane (equivalent to Ethernet) to provide the high bandwidth required between GPUs while also keeping latency in the 100s of nanoseconds (vs. multiple microseconds for Ethernet)” 

It goes back to this last part, which is the 100s of nanoseconds for UALink versus microseconds for Ethernet (as it stands today) that will require a leap in product design and successful deployment in order to displace UALink and the PCIe protocol (where Astera’s solutions fit in). Market participants are doing what many time-strapped investors do – scanning and seeing the large backers for ESUN and assuming it means UALink will not be successful. There is a time to market issue for UALink, yet in the meantime, PCIe remains a strong choice for fast, scale-up systems. PCIe is deployable right now for scale-up pods and CXL is also a strong choice for memory pool connectivity (Astera participates in all of this). 

Quick Takeaway: ESUN is attempting to make Ethernet work for scale-up whereas UALink was built from scratch for scale-up. The primary benefit ESUN offers is to move quicker than UALink (as discussed above, ALAB is saying it’ll be 2027 for UALink to be fully deployed). However, in the meantime, Astera’s PCIe solutions are in high demand and deployable now. Even if ESUN moves faster commercially, there is a performance gap that helps to ensure that Astera’s positioning with PCIe/CXL remains intact. That performance gap is best described as the low latency required for what are the most in-demand AI workloads today – those that require memory pooling and GPU-to-GPU communication.  

Commentary on the Earnings Call regarding UALink and ESUN 

The first question on the call pertained to this concern with an analyst inquiring about any changes in management’s outlook given the recent developments around Ethernet. It’s important to note the analyst acquiesced that it could take 18-24 months for a new protocol to be spec’d out.   

Here is what was stated – providing a longer quote given this topic has caused the stock to selloff twice. 

“We continue to see our market opportunity grow for our scale-up products, particularly this Scorpio X product like you noted. Scale up, as you can imagine, it's a very large market. We estimate it to be in tens of billions of dollars like you correctly noted, some of these design wins take last over multiple generations, simply because of the investment that goes into developing the software and the hardware required for killer topologies.  

For us, if you think about our business today, we are getting ready to ramp into production with our PCIe-based scale-up solutions, it's been extremely popular. There are several customers that are using PCIe like protocols for scale. A new entrant was Qualcomm that publicly announced their new AI 200 inference rack that feature PCIe-based scale-up. For Astera, we have engaged with over 10 AI platform providers. And we expect that these design wins and engagements that we have will continue to ramp.  

In fact, we expect this to go 2029 just based on some of the multi-generation nature of these design wins. For us, UALink is also a very meaningfully additive opportunity as customers start adopting it, just based on the higher data rate support the spec has been around, like you noted, for over a year now in terms of the consortium being formed. The spec is stable, the ecosystem is forming, silicon development is in full gear. And many of these customers, we have currently engaged with RFPs and RFQ. So the momentum is really built up very nicely and continues to grow. So we do expect meaningful revenue from UALink to start coming in 2027.” 

Notably, there is already a hint that UALink could hurry along to become available in H2 2026: “We continue to expect a portfolio of UALink solutions to be available to customers in the second half of 2026 with early revenues generated in 2027.” 

Scorpio P-Series is Ramping Now, and X-Series will Ramp Early 2026 

The more immediate catalyst for Astera as we look toward Q4 and into early 2026 are the Scorpio products, which we’ve covered in the past.  

Scorpio P-Series represents 10% of revenue now, yet management stated it will quickly double to exit the year at 20% of revenue. From there, management has implied Scorpio X will exceed Scorpio P’s revenue percentage. Net-net, that means Scorpio will reach 50% of revenue sometime in H1 2026 up from effectively 0% of revenue in H1 2025. 

Here is a quick refresher on these two solutions: 

The Scorpio P-Series is a small chip that connects the CPU, GPU, NIC and NVMe storage. Rather than building a large switch, the company built a smaller device that is more efficient for high-speed signals to help feed GPUs with data. The fewer ports and smaller switch decrease complexity in a bid to compete against Broadcom with twice the lane count.  

The X-Series is for back-end networking in GPU-to-GPU configurations (and custom silicon configurations), and will offer a higher port count. Astera is essentially building something similar to Nvidia’s NVSwitch with the X-Series, but for PCIe-enabled GPUs and ASICs. Per an earnings call earlier this year: “And this one, like Mike noted, it's a greenfield use case, meaning if you keep Nvidia and NV Switch aside, everyone else is starting to build configurations that are obviously going to need some kind of a switching functionality, which is what we are addressing with our X Series device.”   

The X-Series improves efficiency for ever-increasing AI cluster sizes. The majority of AI clusters are in the tens of thousands GPUs, but are expected to go to the hundreds of thousands (already has with X and some other Big Tech companies), and will see AI clusters with millions of GPUs over the next couple of years.  

In an effort to identify a catalyst that can sustain Astera’s exceptional growth, it would be this product that does so. The X-series is used to interconnect GPUs for higher GPU utilization, resulting in higher ASPs. Per previous commentary: “So to that standpoint, X-Series does bring in a lot more value, and therefore, you can assume that the ASPs tend to be significantly higher. And that's — again, there are different — the X-Series is not one device, to be very clear, there are multiple part numbers. So there would be situations where maybe one part number is not at the same level as P-Series. But in general, you can just look at it from a per lane standpoint or per port standpoint, and look at the value delivered.  And on that basis, the X-Series will always be a much more valuable, much more higher ASP product than a P-Series.” 

Notably, Astera maintains their largest opportunity for the X-Series is on the custom silicon side although they foresee hyperscalers wanting to customize their racks in a way that prevents vendor lock-in from both Nvidia and Broadcom.   

“So these are fabric switches that are used to interconnect multiple accelerators together. So to that standpoint, a, it's not only a significant dollar opportunity because the ASP of this product tends to be high. But these are also products that are turning out to be anchor sockets for us. If you think of an AI rack being built, you have the accelerators and then you have the fabric that interconnects the accelerators.    

So what we are transitioning and what we're excited about is that the Scorpio X device is now translating to be an anchor socket. Think of it as like a mothership around which we are able to now add a lot more products that go along with it, whether it's the silicon level products or module or other form factors that we're considering.” 

The longer refresher on Scorpio P-Series and Scorpio X-Series is necessary because the primary catalyst we identified earlier this year has not even ramped yet. Scorpio P-Series only began shipping this quarter and Scorpio X-Series will begin to ship next year. Here’s the most recent update on when these are shipping: “Scorpio P-Series continued its initial volume ramp at our lead customer, and we are excited that our P-Series revenue will further broaden with recent new design wins across a variety of AI platforms at multiple hyperscaler customers. Scorpio X-Series is shipping in preproduction quantities with a volume ramp expected throughout 2026.” 

Management confirmed they continue to expect a strong ASP uplift for the upcoming Scorpio X-Series ramp in 2026: “Looking ahead, we are gearing up for Scorpio X-Series to shift to high-volume production over the course of 2026. With this ramp of Scorpio X-Series for scale-up connectivity topologies next year, we expect our overall dollar content opportunity per AI accelerator to significantly increase, representing another step-up from a baseline revenue standpoint.” 

To future proof Scorpio solutions, Astera Labs is acquiring xScale Photonics as the market is expected to rely more heavily on photonics (as opposed to copper) a few years from now with management stating: “However, as data rates increase and scale-up domains go beyond 1 rack, clearly, at some point, you will need optical interconnects for scale-up. And there is already a big market for optical interconnects at a data center scale.” The result will be photonic solutions with higher data rates for Astera’s Scorpio solutions. 

Financials 

Revenue Beats by Nearly 12% 

Astera delivered a strong beat on the top line in Q3, with revenue up 103.9% YoY and 20.1% QoQ to $230.6 million, beating the $206.4 million estimate by 11.7%. This maintained Q2’s sequential growth rate of 20%, though YoY decelerated by ~46 points as the company begins to lap tougher comps on a dollar basis. Management said the strong growth was driven by new AI platform ramps featuring multiple products and “robust demand” across its signal conditioning, smart cable module (SCM), and switch fabric portfolios. 

For Q4, Astera guided for $245 million to $253 million in revenue, coming in well ahead of estimates for $216.5 million and pointing to YoY growth of 77% and QoQ growth of 8%, driven by continued PCIe 6 momentum and robust growth from Taurus Ethernet SCMs. This would technically mark the company’s first <100% growth quarter since the end of 2024.  

While Astera did not provide a full-year guide, extrapolating from the midpoint of Q4’s guide projects FY25 revenue at ~$831 million, ahead of estimates for $776 million and corresponding to nearly 110% YoY growth. This will likely force FY26 revenue revisions to move at least $100 million higher from the current $1.04 billion assuming consensus continues to project 34-35% YoY growth. 

GAAP Operating Margin Expands ~32 Points YoY to 24% 

While the revenue beat and raise is certainly welcomed, the improvement in GAAP margins down the line was impressive, with GAAP operating margin expanding to 24%.  

  • GAAP gross margin was 76.2%, ahead of guidance for 75%. This marked a marginal 0.4 point sequential improvement but a 1.5 point YoY contraction. Adjusted gross margin was 76.4%. 
  • GAAP operating margin was 24.0%, well ahead of guidance for 17.9%, and expanding 3.3 points QoQ and nearly 32 points YoY. This YoY expansion from (7.9%) in Q3 ’24 is quite impressive considering the company was reporting triple-digit revenue growth in each quarter; this also reinforces that the company is comfortably GAAP profitable. Adjusted operating margin was 41.7%, up 2.5 points QoQ and 9.3 points YoY. 
  • GAAP net margin was 39.5%, up nearly 13 points QoQ and more than 46 points YoY. Adjusted net margin was 38.3%, down 2.4 points QoQ but up 2.7 points YoY. 

For Q4, Astera guided for slight sequential moderation in margins down the line, with gross margin guidance at 75%, in line with prior quarter guidance. GAAP operating margin was guided to be 22.2% at midpoint, down 1.8 points QoQ but still up more than 22 points YoY. Adjusted operating margin was guided at 39.9%, down 1.8 points QoQ but up 5.6 points YoY. 

GAAP EPS Beats by 92% 

With the strong expansion in GAAP net margin, Astera delivered a 92.3% beat on GAAP EPS, reporting $0.50 in Q3 versus the $0.26 estimate. Adjusted EPS was $0.49, up 113% YoY and solidly ahead of the $0.39 estimate. 

For Q4, Astera guided for $0.20 in GAAP EPS, below the $0.26 estimate due to a 45% income tax rate. Adjusted EPS was guided at $0.51, up 38% YoY. This guidance would bring FY25 GAAP EPS to $1.17 (versus estimates for $0.96) and adjusted EPS to $1.77 (versus estimates for $1.58).  

Cash and Balance Sheet 

Cash flow margins moderated quite sharply on both a QoQ and YoY basis, though cash flow generation remained decently strong with an OCF margin of 33.9%.  

  • Q3 operating cash flow was approximately $78.1 million for a 33.9% margin, down from a 56.2% margin in the year ago quarter and a 70.5% margin in Q2.  
  • Q3 free cash flow was approximately $65.8 million for a 28.5% margin, down from a 41.4% margin in the year ago quarter and 69.5% in Q2. 
  • Cash and equivalents totaled $1.13 billion and debt remained zero. 
  • Accounts receivable were $42.9 million, rebounding from $24.3 million last quarter but remaining lower than the $69.8 million from Q1. Inventories moderated slightly to $51.7 million from $58.6 million in Q2.  

Conclusion: 

We do a lot of checks and balances at the I/O Fund to figure out if a stock is seeing unexpected headwinds. We look for catalysts (Scorpio solutions), we double check our product understanding (PCIe has unique benefits over Ethernet that will be hard to completely displace especially for memory pooling and GPU-to-GPU communication), we do a thorough checklist of the fundamentals with Astera crushing on the top line and even more so on the bottom line.  

To me, ESUN signals that Ethernet vendors are a little afraid of being left out of the scale-up party. The most probable outcome is that both ESUN and UALink coexist to address different layers of an AI cluster and as part of a broader shift away from proprietary networking. Even in a downside scenario where Ethernet moves faster than expected, we probably have an 18-month window before those dynamics materially affect deployments. It won’t be a boring 18 months either, rather it’ll be pretty exciting in terms of the importance of AI networking.  

With that said, Astera is a hypergrowth stock with an AI valuation. We are always prudent with these stocks by following risk management plans. We saw this with Palantir this week, reminding us that price action doesn’t always follow fundamentals in the near term. Our antennas are up, not because of concerns specific to Astera, but because prudent positioning is part of being an investor in high-growth AI names. 

Equity Analyst Damien Robbins contributed to this article.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Readings:Recommended Readings:

AMD Q3: The Catalyst is Expected in H2 2026, Could Ramp Sooner

Reddit Q3: Setting a High Bar with Top Line Strength and 10-point Sequential Margin Expansion

Taiwan Semiconductor Q2 Earnings: FY25 Guidance Raised on Strong AI Demand 

Posted on by io-fund

Taiwan Semiconductor reported earnings last week, providing the first glimpse into the AI semiconductor industry in the second quarter. Riding strong AI chip demand, TSMC boosted its full year revenue growth guidance, yet there are still some lingering doubts about the chipmaker’s growth in Q4.  

HPC revenue continues to accelerate, reaching a record at $18 billion in Q2. Net profit reached a record in Q2 as margins outperformed. With that said, TSMC signaled significant margin pressures in Q3 from FX and ramping overseas fabs.

TSMC boosts FY revenue by 5 points 

Q2 revenue increased 44.4% YoY and 17.8% QoQ in USD to $30.07 billion, well above TSMC’s guidance for $28.4 billion to $29.2 billion in revenue. This was driven by strong demand for AI accelerators built on TSMC’s 3nm and 5nm nodes.  

For Q3, TSMC guided revenue of $31.8 billion to $33 billion. At midpoint of $32.4 billion, this represents YoY growth of 37.8% and QoQ growth of 7.8%. FX is playing a role here in this high-37% guide, with TSMC noting that revenue growth in NT$ is expected to be negatively impacted by 6.6 points based on current exchange rates, or ~31.2% YoY.  

For the full-year, TSMC boosted its revenue growth forecast from mid-20% YoY to close to 30% YoY, driven by robust AI and HPC demand.  

As it stands, TSMC’s growth in the first three quarters is far above the close to 30% guide, suggesting Q4 could see growth in the single digits. Management said they remain more conservative on Q4 at the moment. This is because although Q4 is typically seasonally strong for consumer electronics and smartphones, there is risk that tariffs put a damper on growth. 

Management added that they have not seen any changes in customer behavior, but they are well aware of uncertainties from tariffs on consumer and price-sensitive end-markets. They added that Chinese rebate programs are stimulating near-term demand upside, but this is expected to phase out rather quickly and only create a mild recovery in non-AI demand this year. 

HPC revenue rises 14% QoQ and up 19.2% CC 

TSMC continues to benefit from robust AI accelerator demand, with HPC now accounting for three-fifths of the chipmaker’s revenue. TSMC stated that HPC revenue rose 14% QoQ in NT$ in Q2, though this increase was more pronounced on a US$ basis due to FX. TSMC’s revenue is recognized in US$, so every 1% appreciation of the NT$ adversely impacts NT$ reported revenue by ~1%.  

Therefore, on a US$ basis, HPC revenue rose 19.2% QoQ to ~$18 billion, or up nearly $3 billion from Q1 in constant currency — its largest growth on record. The pace of acceleration in HPC revenue has been astonishingly quick, as the segment is now 2.5x larger than it was two years ago at $6.9 billion. HPC’s share of revenue also has increased 8 points YoY to 60%. 

Smartphone revenue increased 7% QoQ to account for 27% of revenue in Q2, notably the strongest sequential increase for the segment since 2022.  

IoT revenue increased 14% sequentially to account for 5% of revenue, while automotive was flat and also accounted for 5% of revenue. DCE increased 30% sequentially to account for 1%, while other segments rose 14% sequentially to 2% of revenue.  

Revenue by Technology 

TSMC’s advanced nodes – 3nm, 5nm, 7nm, and soon, 2nm – contribute the majority of revenue at 74% in Q2, fueled by AI accelerators and smartphones. 3nm ticked back up to 24% of revenue in Q2, while 5nm held flat at 36% of revenue, supported by Nvidia’s Blackwell GPUs. 7nm saw its contribution shrink one point to 14%, while mature nodes also shrunk one point to 26%. 

TSMC shared some more details about its upcoming advanced nodes, stating that it remained on track for volume production on its 2nm node beginning in the second half of 2025, with a ramp profile similar to the 3nm node. Management also stated that they “expect the number of new tape-outs for 2nm technology in the first 2 years to be higher than both 3nm and 5nm,” driven by HPC and smartphone. 

For the A16 node (1.6nm), TSMC said that it remains on track for volume production in the second half of 2026, believing this node will be best suited for “HPC applications with complex signal routes and dense power delivery networks.” 

Margins guided to decline due to FX headwinds in Q3 

Despite some FX headwinds to gross margin, TSMC’s operating margin outperformed, driving profit to a record level in Q2. However, Q3’s guide showed increasing FX headwinds and sharper sequential impact on margins.  

  • Gross margin was 58.6% in Q2, at the high end of the guided range for 57% to 59%. Gross margin declined sequentially from 58.8%, with a 2.2 point headwind from FX and a 1 point headwind ramping overseas fabs offset by higher capacity utilization. 
  • Operating margin was 49.6%, increasing 1.1 points sequentially from operating leverage, above guidance for 47% to 49%. Operating margin was up more than 7 points YoY. 
  • Net margin was 42.7%, down slightly from 43.1% in the prior quarter but up nearly 6 points YoY.  

For Q3, margins will decline due to FX: 

  • TSMC guided gross margin to decline sequentially to 55.5% to 57.5%, or 2.1 points to 56.5% at midpoint. This is again due to continued FX headwinds, with approx. 2.6 points from unfavorable FX and overseas fab ramp in Kumamoto and Arizona.  
  • Operating margin was guided to decline sequentially to 45.5% to 47.5%, or 3.1 points at midpoint. This would be the lowest level since Q2 2024. 

EPS increased 67% YoY, up from 54% 

TSMC delivered record profit in Q2, rising 61% YoY to NT$398.3 billion, or ~$12.8 billion. Adjusted EPS of $2.47 beat estimates for $2.31 and increased nearly 67% YoY, accelerating from recent growth in the 50% range. However, EPS growth is forecast to decelerate rather rapidly through Q4, with TSMC barely maintaining double-digit growth.  

For 2025, adjusted EPS is expected to increase 35.2% YoY to $9.52, up from 31.5% growth two months ago.  

Cash Flows dip yet Balance Sheet is Healthy 

Cash flow margins dipped by a larger margin sequentially, and TSMC’s balance sheet remained healthy.  

  • Operating cash flow was $16.2 billion for a 53.8% margin, down from a 74.5% margin in Q1 and a 56.1% margin in the year ago quarter. 
  • Free cash flow was $6.5 billion for a 21.7% margin, down from a 35.1% margin in Q1 and a 25.5% margin in the year ago quarter. 
  • Cash, equivalents and marketable securities totaled $90.4 billion, while debt totaled $32.3 billion. 
  • Capex rose more than 51% YoY to $9.6 billion, slowing from a 74% pace in Q1. TSMC maintained its full year capex guide at $38 billion to $42 billion.  
  • Inventories were $10.43 billion, up from $8.83 billion in Q1; however, the sequential increase looks to have been impacted by FX, as inventories in NT$ were up less than 4% QoQ. 

Earnings call Q&A  

While management offered little to no clarity on long-term AI growth or CoWoS capacity, they discussed long-term diversification of advanced node manufacturing to the US. Management also offered insights into advanced node capacity that signal Nvidia’s growth could remain strong come 2H. 

Arizona Expansion 

TSMC provided an update on its global expansion plans, which is important considering onshoring US manufacturing helps reduce geopolitical risk from China for the AI server supply chain. Management shared that they are expecting to bring 30% of their 2nm manufacturing to the US in Arizona. 

As a result, TSMC is accelerating and expanding its presence in Arizona with its recent $165 billion investment plan, for six fabs, two advanced packaging fabs, and a major R&D facility to meet high multi-year demand from customers. The second fab in Arizona, utilizing 3nm tech, has finished construction with TSMC aiming to speed up volume production by several quarters. The third fab, offering 2nm and A16 advanced nodes, is under construction.  

Management added that “despite the higher cost of overseas fabs, we will leverage our increasing size in Arizona and work on our operations to improve the cost structure,” to help minimize gross margin dilution impacts. This is important considering TSMC is forecasting increasing margin dilution from its growing overseas presence, widening from “2% to 3% every year in the early stages and widen to 3% to 4% in the later stages” over the next five years. 

Nvidia H20 Impact 

Morgan Stanley’s Charlie Chan questioned about Nvidia’s approval to resume shipping its H20 GPU to China, and if unlocking the Chinese market again would help TSMC raise its AI accelerator growth CAGR upwards from the mid-40% range.  

C.C. Wei was rather tight-lipped about the potential impact, given that shipments have (likely) not yet resumed, saying that it is too early to provide an estimate on how this would impact growth. Wei explained that TSMC is not yet ready to increase its forecast, and “another quarter probably will be more appropriate to answer your question,” hinting that the AI accelerator CAGR may be updated in Q3.  

However, it is expected that a majority of the H20s to be sold will be from existing inventory in the Taiwanese supply chain, meaning chips already built and revenue already booked. Therefore, it’s hard to see how TSMC could meaningfully increase its CAGR for the next four years based on just the H20. 

In terms of China, however, the bigger opportunity here for TSMC may stem from Nvidia’s China-specific Blackwell B30 GPU, which is estimated to see shipments of up to 1.2 million units, or ~20% more than the total estimated H20 inventory. The B30 is expected to hit the market in Q4, following the H20’s resumption largely in Q3. 

Advanced Node Capacity  

Goldman Sachs’ Bruce Lu asked management about advanced node capacity, and supply-demand imbalances as more AI chips begin to shift to the 3nm node. While TSMC did not offer much beyond capacity being tight, one comment suggested Nvidia’s demand remains very strong. 

C.C Wei explained that TSMC’s 5nm capacity is “very tight,” while 3nm capacity is even tighter and will continue to remain tight for a couple of years. Wei explained that TSMC can quickly retool advanced node fabs, such as 7nm to 5nm, 5nm to 3nm, etc, and keep utilization high to help meet demand. He would not commit to saying that demand would outpace supply, but that was implied given his comments of trying to “narrow the gap” between supply and demand.  

However, one of the more important comments here was Wei stating that TSMC is using 7nm capacity to support 5nm demand, which provides another piece of evidence alongside surging HPC revenue that Nvidia’s Blackwell ramp is accelerating rapidly. Blackwell is built on TSMC’s N4P process, a subfamily of its 5nm node, offering higher performance, better power efficiency and higher transistor density.  

Nvidia CEO Jensen Huang had stated that Nvidia was shipping nearly 1,000 racks per week to hyperscalers in May and expecting to ramp further, and this comment from TSMC supports lasting 5nm demand, likely from Nvidia given its high share of CoWoS (and manufacturing) capacity.  

Conclusion 

TSMC’s earnings provided more evidence that AI GPU demand remains strong, particularly for Nvidia, with HPC revenue rising to a fresh record with its largest QoQ increase, and commentary for 7nm capacity helping meet high 5nm demand. Supported by this robust AI and HPC demand, TSMC boosted its full-year guidance from mid-20% revenue growth to close to 30% growth, despite lingering concerns of tariff-related weakness come Q4.

Damien Robbins, Equity Analyst at I/O Fund contributed to this analysis.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

Taiwan Semiconductor Stock: AI Growth Amid Geopolitical Risk 

Posted on by io-fund

Despite their leadership, AI stocks like Taiwan Semiconductor and Nvidia are flat year-to-date and trading at similar levels as June 2024. Clearly, the AI trade is not as straightforward as it might seem. Taiwan Semiconductor, in particular, sits at the center of geopolitical tensions — yet those tensions tend to surround companies with deep IP in the AI economy. What makes this economy so distinct is not just the extraordinary commercial demand, but also its rare, historical role in shaping global alliances (and adversaries). 

Investors are confronted almost daily with friction between the U.S. and China — and at the center of it all is one stock: Taiwan Semiconductor (TSMC). While enthusiasm around AI demand remains strong, assuming it will simply override geopolitical headwinds is overly optimistic. Onshoring a supply chain like TSMC’s takes years, yet markets can react to a negative headline in seconds. 

Headlines aside, the bigger picture is that TSMC is deepening its moat with advanced nodes, such as N2 and A16. The company already powers tens of trillions in market cap on the stock market when you consider Apple, Nvidia, Broadcom, Amazon, AMD and Google are customers of TSMC. Essentially, all mega cap stocks have an AI strategy spanning merchant GPUs and custom silicon, and of course, software – yet the common denominator to these strategies is they all funnel into TSMC. 

The problem my firm helps with is this — how does an investor ride out the inherent cyclical nature of semiconductors given the powerful, secular trend of AI? For every stock that becomes a multi-generational winner, there are dozens that never reclaim their all-time highs. The cloud sector for example, is becoming an all-time high graveyard with once-upon-a-time Wall Street darlings trading meaningfully below their ATHs for over three years now. 

TSMC will very likely push beyond its ATH yet returns can increase meaningfully if an investor has the guts to buy during a steep selloff. Other times, that selloff isn’t coming and it’s best to buy before a breakout. We answer these complex questions in the analysis below. 

TSMC’s Advanced Nodes have Created a Competitive Moat 

The most advanced node shipping today is the 3nm, offering 15% better performance than the 5nm process when power level and transistors are equal. The die sizes are an estimated 42% smaller than the 5nm and TSMC also states the 3nm process can lower power consumption by as much as 30%.  

Power efficiency is a major advantage, helping to deepen TSMC’s moat. Samsung was first to introduce 3nm process chips in 2022 yet has not been as competitive on yield and power efficiency at a roughly 10% to 20% difference compared to TSMC. The moat is visibly seen in TSM’s pricing power with the dominant foundry charging 25% more for its 3nm process compared to its 5nm process, and customers are willing to forego Samsung to pay the higher pricing.  

Last year, companies such as Apple, Nvidia, AMD and Intel committed to working with TSMC for its 3nm process, and eventually Google and Qualcomm left Samsung “after careful consideration” to also secure a partnership with TSMC.  

This was an important moment for TSMC to complete its near-monopoly in advanced nodes as Google had been outsourcing its Tensor processors to Samsung’s foundry for four generations, before moving to TSMC for the fifth generation. Qualcomm also switched to TSMC from Samsung for the Snapdragon 8 Gen 4 series. 

To attract these large customers with different end markets, TSMC offers a few 3nm processes, such as the N3E, N3P and N3X. This allows a company like Apple to customize the 3nm chips differently than AI chips for hyperscalers. N3E is the baseline for IP design with 18% increased performance and 34% power reduction, N3P has higher performance and lower power consumption, whereas the N3X will offer high-performance computing very high performance but with higher power leakage. 

To illustrate the near monopoly that TSMC has over other foundries, consider that its market share stands at 67.1%, up 2.4% QoQ in Q4. Meanwhile, second-place Samsung was at 8.1% down from 9.1% for a lead of 59 points.  

When comparing revenue, TSMC reported $26.85 billion in Q4 for a 14.1% increase compared to Samsung’s $3.26 billion, which declined 100 basis points to 8.1%.  

In the latest quarter, advanced nodes below 7nm drove 73% of wafer revenue with 3nm contributing 22% of revenue and 5nm representing 36% of revenue. Nvidia is not on the 3nm process yet for its Blackwell shipments, thus 5nm is outsized in terms of its market share.

TSMC 3nm revenue surpasses 20%, up from 9% year-over-year

Pictured Above: 3nm revenue for TSMC has ramped quickly, up from 9% in the year ago quarter and in its third consecutive contributing >20% of revenue. 

TSMC’s 2nm Nanosheet Transistors (Gate All-Around) 

Looking ahead, 2nm is expected to see volume production in the second half of 2025 with a more advanced iteration called N2P scheduled for volume production in the second half of 2026. The 2nm marks a new era in TSMC's transistor architectureas N3 relied on FinFET while gate-all-around (GAA) is being introduced for N2. As the name implies, the gate is wrapped around on all sides compared to FinFET which had a gate wrapped on three sides. By having the gate wrap “all-around,” a greater surface is created for better electrostatic control and to also reduce leakage.  

For TSMC, the 2nm will feature NanoFlex technology, which is similar to FinFlex to where designers can use cells from different libraries. However, due to the new gate-all-around (GAA) nanosheet transistors, there are additional benefits, such as customizing the width and height of cells. For example, GAA can uniquely widen the channels for a performance boost, or there is an option to narrow the channel to optimize power cost. The goal is to increase the performance-per-watt to enable higher levels of output and efficiency.  

According to management on the earnings call: “N2 will deliver full-node performance and power benefits with 10% to 15% speed improvement at the same power or 20% to 30% power improvement at the same speed and more than 15% chip density increase as compared with N3E.” 

Similar to the 3nm, there will be a few variants of the 2nm chip for customers to optimize performance with power requirements. The first two years of the 2nm ramp is outpacing the 3nm and 5nm ramp, signaling good things to come for TSMC. 

TSMC Stock will Close out the Decade with Pricing Power 

As a growth investor, it certainly doesn’t hurt to keep an eye on the horizon. A16 is the 1.6nm process node that will emphasize backside power delivery. Our firm first covered this topic last year in the analysis: “Taiwan Semiconductor Stock: April Sales Soar From Advanced Nodes" stating the Angstrom era will translate to “future process generations where the nodes are not smaller necessarily, rather the transistors they’re built with will be improved upon.”  

For the A16, the Super Power Rail (SPR) backside delivery will offer a redesign to where power routing is moved from the front to the back, which allows for the signaling on the front side to have lower latency. By reducing voltage drop, SPR becomes attractive for AI workloads since multiple cores are operating at high speeds with complex signal routes and dense power requirements.  

Intel’s PowerVia is first to market with the backside power delivery design, yet TSMC’s design will likely result in higher yields and volume production. TSMC also connects the backside power delivery to each transistor’s source and drain, which is more expensive yet also more efficient compared to Intel’s approach. 

The A14 is due out in 2028 and will offer a significant breakthrough in performance while offering up to 25% to 30% lower power consumption, with increased density of 20% to 23%. There will be an A14 variant that offers backside power delivery in 2029. It's expected that A14 will help to drive forward edge AI due to a combination of speed improvements and power reduction. Pricing for the A14 is expected to increase from $30,000 per wafer for the 2nm process to $45,000 per wafer as we close out the decade.  

TSMC to Grow Revenue Mid-20%; AI Accelerator Revenue will Double 

The company is off to a good start for the year with revenue growth of 35.3% YoY while guiding for an acceleration to the 38% range in Q2.  Revenue was down (5.1%) sequentially, impacted by smartphone seasonality, partially offset by AI-related demand growth. However, the Q2 guide represents 13% QoQ growth with revenue between $28.4-29.2 billion.  

TSMC offers monthly reports with April starting Q2 off strong as monthly revenue surged 48.1% YoY and 22% MoM to ~$11.55 billion, with Bloomberg stating the outperformance could be due to a rush in pre-tariff ordering, although certainly 3nm and 5nm demand helped as well.  

TSMC earnings: strong H1 growth, slower H2 outlook

TSMC earnings show strong growth in H1 followed by lower growth in H2. 

This year, IDC is forecasting Foundry 2.0 will grow by 11% compared to 6% last year. Foundry 2.0 describes a broader range of foundry technologies, with the foundry segment expected to grow 18% down from 20% last year.  

Regardless of which way you dice it, TSMC is guiding for above industry growth, stating in the most recent quarter: “we continue to expect our full-year 2025 revenue to increase by close to mid-20s percent in U.S. dollar term.”  

Of this, AI accelerator revenue is expected to double in 2025 and management also forecasts AI to grow at a mid-40% CAGR for five years from 2024: “Based on our planning framework, we are confident that our revenue growth from AI accelerators will approach a mid-40s percentage CAGR for the next five years period starting from 2024.” 

Slower Growth Up Ahead with H1 > H2 

An area of concern is that TSMC is guiding a slowdown in the second half of the year, given the mid-20% revenue growth is below Q1/Q2 revenue growth of 35% to 38%.  

There was a question on the call about this from analyst Charlie Chan asking: “And also based on your full-year guidance, so called the mid-20%, it seems like second half recovery will be very, very gradual or flattish. So I'm wondering if you're already bake in kind of consumer tech demand impact. And if a tariff have some kind of turnaround, right, meaning, for example, major smartphone brands whether there's a chance for you to revise your full-year revenue guidance? Thank you.”

Management answered the H2 weak guide is due to uncertainty and tariffs: “Charlie, as we also said in the prepared remarks, there are uncertainties and potential risk from tariffs exist.”

Analysts are a bit concerned about the full-year guide given the risks key customers are facing from April’s tariff shocks. JPMorgan analysts say TSMC “could pare [its forecast] slightly to target low- to mid-20%” sales growth, while Deutsche Bank analysts raised the concern that the chipmaker “may also withdraw its guidance as customers adjust to tariffs.”    

Management also stated that things are more “balanced now” — meaning demand is not overwhelming supply like it once did: “Brett, three months ago. Now I can tell you that three months ago, we are barely – we just cannot supply enough wafer to our customer. And now it's a little bit balanced, but still the demand is very strong. And you are right. Other than China, the demand is still very strong, especially in U.S. And so we are confident that we are going to double our AI revenue this year.” 

Echoing these comments, if we look at the segments listed below, we can see that smartphones are reporting higher seasonal weakness than last year. 

TSMC Reports Strength in HPC offset by Smartphones 

HPC Revenue rose 7% QoQ 

TSMC continues to ride AI accelerator tailwinds, evident in its rising HPC revenue and mix. HPC revenue rose 7% QoQ in Q1, surpassing $15 billion for the first time. HPC accounted for 59% of TSMC’s revenue, expanding from 53% of revenue last quarter.

Top tech firms drive TSMC AI chip growth via HPC segment

Pictured above: Major tech companies choose TSMC for AI chips, visible in its HPC segment 

Management stated that they “continue to observe robust AI-related demand from our customers,” and reaffirmed that AI accelerator (GPU + ASIC + HBM) revenue is expected to double YoY in 2025. As stated, management also confidently forecast AI accelerator revenue to grow at a mid-40% CAGR over the next five years starting in 2024. 

Smartphones Declined 22% QoQ: 

Smartphone revenue declined (22%) QoQ due to seasonal trends, accounting for 28% of revenue in Q1. This was larger than last year’s (16%) seasonal decline.

TSMC sees increased seasonal weakness in smartphone segment

TSMC is reporting higher seasonal weakness in the smartphone segment compared to last year. 

IoT, Auto and Other: 

IoT revenue declined (9%) QoQ to account for 5% of revenue, while Automotive revenue increased 14% QoQ to also account for 5% of revenue. Digital Consumer Electronics increased 8% QoQ to account for 1% of revenue, while Other revenue rose 20% QoQ to account for 2% of revenue. 

Gross Margin to See 3% to 4% Headwind  

Margins came in at the higher end of guidance in Q1, with TSMC seeing continuing strength in Q2.  

  • Gross margin was 58.8%, at the high end of management’s guided range for 57-59%, dipping slightly sequentially from the January earthquake impacts and the ramp of the Kumamoto fab. On a YoY basis, gross margin expanded 5.7 points. 
  • Operating margin was 48.5%, at the high end of the guided 46.5-48.5% range.  
  • Net margin was 43.1%, flat with Q4 and up 3.1 points YoY. 

TSMC delivered nearly 54% YoY growth in EPS in Q1 as it delivered a slight beat to $2.12, its third straight quarter with EPS growth above 50% YoY. This growth also reflects TSMC’s operating leverage, outpacing revenue growth in the mid to high-30% range.

TSMC Q1 EPS up 53.6%, but growth expected to plateau later in the year

In Q1, TSMC reported strength on the bottom line with EPS growth of 53.6% although EPS will face tough comps with growth plateau’ing toward the end of the year. 

For Q2, EPS growth is expected to maintain this >50% growth rate to $2.24, before decelerating rather sharply to the mid-single digits by Q4 as it begins to lap these more difficult 50% growth comps. 

FY25 EPS is currently expected to increase 31.5% YoY to $9.26, before decelerating to 15.2% growth in FY26 to $10.66. 

For Q2, TSMC guided for similar gross and operating margin ranges, but flagged some headwinds from the fabs buildout. However, a larger headwind exists – FX. TSMC’s guidance below assumes an exchange rate of US$1 to NT$32.5, yet the current rate sits at US$1 to $NT30.1, down nearly 8% from the guided level.  

  • Gross margin is forecast at 57-59%, down 0.8 points sequentially at midpoint as dilutive impacts from ramping the Arizona fab kicks in. Management added that overseas fab impacts are expected to grow more pronounced throughout the year, forecasting 2-3% dilutive impact for the full year from Arizona and Kumamoto. 
  • Operating margin is forecast at 47-49%, down 0.5 points sequentially at midpoint. 

Over the next five years, management sees the dilutive impact from ramping its overseas fabs widening, projecting it to start at 2-3% each year in the early ramp stages before widening to 3-4% each year. Despite this, TSMC remains confident in its ability to keep long-term gross margins at 53% or higher.  

$100B Investment Announced for Arizona Fabs 

In the recent quarter, the company’s cash flow increased 37% YoY to $19.0 billion, for a 74.5% margin, a slight YoY expansion. Capex was $10.1 billion in Q1, down more than 10% QoQ but up more than 74% YoY. Cash and equivalents rose $7.5 billion sequentially to $81.4 billion, while debt is $30.4 billion. 

In March, TSMC announced a new $100 billion investmentto expand manufacturing in the United States. The $100 billion will go toward building new fabs in Arizona bringing TSMC’s total investment in the United States to $165 billion. Once the new fabs are built, 30% of TSMC’s advanced nodes capacity will be located in Arizona.  

TSMC’s current Arizona fabs began producing chips this year, with Apple being the first to receive chips on the 4nm/5nm process and Nvidia receiving chips later this year. In addition, it’s been reportedthe 2nm process is seeing a 90% yield for memory products in the newer Arizona fab.  

Notably, due to rising costs, there are rumors that TSMC will raise prices from the Arizona fab by 30%. 

Subscribe for Full Access to the Article 

Below the Paywall is the Following information  

  • Our analysis on the stock’s valuation and if the stock is a “buy” or a “hold” given the stock faces immense demand from the AI economy yet must also weather geopolitical tensions. 
  • The I/O Fund’s trading plan for TSMC including never-before published buy targets over a 12 to 18-month time frame.
    Sign Up to Continue ReadingSign Up to Continue Reading

Paid subscribers, click here to view the full articleclick here to view the full article

Not ready to subscribe but want more thoughtful analysis from a top-performing team in tech? Every week, we publish free research. 👉 Sign up here.

For any account-related issues, please contact our customer support team at premium@io-fund.com.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

Nvidia Stock Faces a Choppy Q2, But Tailwinds Build for H2 Acceleration

Posted on by io-fund

When you’re Babe Ruth, the crowd expects to see a homerun. Hit a single or a double and the fans go home disappointed. Nvidia has continued to report exceptional earnings results, yet Nvidia stock is competing with itself at this point. 

Next week, Nvidia will report fiscal Q1 earnings, and the market has become accustomed to the company reporting a string of homeruns and grand slams. While Q1 results will be propped up by China stockpiling the H20s, the outlook for Q2 is the choppiest the company has faced in two years. This is because Nvidia has a narrower path than usual to impress investors as the Hopper generation demand is waning while Blackwell is (finally) shipping but not at the levels originally expected.  

Last quarter, I published the analysis: “Nvidia Suppliers Send Mixed Signals for Delays on GB200 Systems – What it Means for Nvidia Stock.” which stated “Given market jitters around DeepSeek, which turned out to be a non-issue, something more material related to the GB200s, such as growth slowing below expectations at the start of the new fiscal year, could send the stock below $100 — which we would see as a buying opportunity […] ultimately, my firm trimmed our Nvidia position (to a 10% allocation) and will happily buy lower should the assumptions in this analysis materialize. Nvidia remains the stock of the decade; however, stock returns – and product launches — are not perfectly linear.” –February 2025 

That analysis played out. We were able to buy Nvidia at $87, issuing a real-time trade alert that has returned 53% on that tranche since early April. We also added several key Nvidia suppliers that have moved sharply higher. 

With a strong seven-year track record on this name, I felt it was important to share my perspective as Nvidia Week kicks off on Wall Street. 

H20 Export Ban Will Result in $5.5B Inventory Loss and Cost $15B in Revenue 

China has dominated the headlines over the past quarter and exactly how Nvidia plans to overcome geopolitical tensions will be a primary focus in the upcoming call. Export controls have been in place for years, hence the H800 and the H20 GPU variants, which were designed to be less powerful GPUs to comply with export restrictions. Yet, the license restrictions were changed in April, resulting in a $5.5 billion inventory loss for Nvidia. More recently, Jensen Huang clarified it would be $15 billion in revenue stating the inventory will have to be discarded.  

Here is a brief summary of the USA-China GPU licensing restrictions: 

  • In 2022, the Biden administration placed export controls on the A100s and H100s due to bandwidth, requiring the bandwidth to be lowered to 400GBps This led to the A800s and H800s. 
  • In October of 2023, performance requirements were accounted for in the export controls, limiting sales of Nvidia’s A800, H800, L40, L40S and RTX 4090 chips. This led to Nvidia creating the H20 GPUs. 
  • In April of 2025, the Trump administration has effectively banned the H20s by denying the export license. This is on the grounds that H20s can be used in supercomputers and offers 20% faster inference than the H100s.  

While the H20 has reduced compute performance compared to the H100s — including fewer Tensor Cores and lower FP8/FP16 throughput — it retains high-speed interconnect capabilities through its support for NVLink and PCIe Gen4, and features 96GB of HBM3 memory with 4.0 TB/s of memory bandwidth. The 96GB of HBM3 exceeds the H100s, which is large enough for LLM models to run in memory and lower costs. The higher HBM3 also translates to the H20 offering fast communication when clustered with other GPUs for a multi-GPU system supercomputer. 

Furthermore, although the H20 performs at roughly 50% of the H100s, it has a power advantage at 400 watts compared to the H100s 700 watts. Partially due to the lower power, while maintaining high-speed bandwidth for inference, means the H20s have remained attractive to Chinese firms especially in the wake of DeepSeek’s R-1 release. 

Selling chips to China for use in supercomputers has been prohibited since 2022. Meanwhile, many industry experts believe Chinese firms were stockpiling the H20s to build a large supercomputer. The Institute for Progress, a nonpartisan group, wrote a long-form explanationof the loopholes that were being used, stating: “The United States is about to make another strategic mistake: Allowing three Chinese firms to receive over $16 billion in orders for NVIDIA H20 chips, amounting to over 1.3 million chips. This order is over six times the size of Colossus, the largest compute cluster in the world. It would more than double China’s entire existing stock of H20 chips. If these chips are delivered, they will dramatically increase Chinese firms’ ability to develop frontier AI models and deploy them at scale.”  

As far as when the stockpiling began, semiconductor Insights analyst, Claus Aasholm, noted back in December that “The downgraded H20 system, which passes the embargo rules for China, is doing incredibly well. With 50% quarter-over-quarter growth, it is currently Nvidia’s most successful product. The H100 business “only” grew 25% QoQ.” 

According to Reuters, analysts had forecast a total of $12 billion in H20 sales for Nvidia's fiscal year ending in January. However, China revenue for the year was significantly higher at $17.1 billion. 

However, this pales in comparison to what Q1 and Q2 were about to report in terms of China revenue.  

Nvidia Q1 Earnings Preview: Loss of China Revenue Will Sting 

In Q1, Chinese tech companies such as Alibaba, ByteDance and Tencent were hurrying to place H20 orders. The Information reported that Chinese Big Tech companies had placed $16 billion worthof H20 chips in the first three months of the year.  

This would represent a sudden surge of roughly 3X growth given previous quarters peaked at $5.5 billion: 

Nvidia’s China Revenue: 

  • Q1 2025 ending April 2024: $2.49 billion 
  • Q2 2025 ending July 2024: $3.67 billion 
  • Q3 2025 ending October 2024: $5.42 billion 
  • Q4 FY25 ending Jan 2025: $5.52 billion 
Financial table: revenue by region for three and nine months ended Oct 27, 2024, and Oct 29, 2023 (in millions).

Pictured Above: Nvidia’s quarterly revenue in China was $5.4 billion for the October quarter and $5.52 in the January quarter (not pictured).for the October quarter and $5.52 in the January quarter (not pictured). 

When Nvidia stated they would see a $5.5 billion inventory charge in Q1, it was suggesting a very high monthly run rate given the export restrictions were only in effect for the remaining three weeks of Q1. According to the SEC filing “First quarter results are expected to include up to approximately $5.5 billion of charges associated with H20 products for inventory, purchase commitments, and related reserves.” 

If you view China revenue on a fiscal year basis, then The Information is suggesting that the first three months of the year resulted in nearly as much revenue from China as all of last year at $17.1 billion.

Nvidia annual geographic revenue breakdown by customer billing location from 2023 to 2025, highlighting growth in U.S., Singapore, Taiwan, and China

Source: Nvidia 10-KSource: Nvidia 10-K

This helps to illustrate Nvidia was filling a lot of Chinese orders very suddenly a lot of Chinese orders very suddenly before the government intervened. This is further supported by the $16B figure from The Information as revenue of that magnitude from China is not seen in prior quarters.  

Nvidia Q2 Earnings Guide Likely to be Impacted 

Nvidia’s beats have become narrower over the past few quarters. Although it’s not clear what the impact will be in Q2, we have some indication from semiconductor peer AMD that the export ban of the MI308 will be mostly felt in Q2. It's logical to assume Nvidia will disclose something similar in their earnings call.  

In the most recent quarter, Nvidia beat by $1.2 billion. At the start of the AI surge, Nvidia beat by $2.4 billion, and had initially raised guidance by $2 billion for a total upward surprise of $4.4 billion if you generously combine the May guidance with the August results. 

Nvidia quarterly revenue estimates versus actual results from FQ2 2024 to FQ4 2025, highlighting consistent earnings beats and surprise percentages

If we look at fiscal year estimates of $200B and we take the $15B at face value — meaning there is no other impact further out into Q2 and Q3 (there very well could be if we assume $16B were rushed orders, yet the quarterly run rate was $5.5B, which is a floor for the other quarters) — then that’s 7.5% of revenue. This is not a reason to run for the hills, but it’s certainly revenue that has to be absorbed by other SKUs with strong potential Q1/Q2 is where the bulk of the impact is felt.  

Analysts are preparing for lower QoQ growth with Q2 revenue $2.7B higher than Q1 compared to QoQ increase of $5.6 billion in Q3. Part of this is that Blackwell is (finally) ramping but not at the volume originally expected for this quarter.  

Blackwell Revenue May Absorb China Losses 

As the Blackwell vs Hopper GPU transition unfolds, investors are watching to see if Blackwell can make up for declining demand for the aging Hopper architecture as the H20 impact dissipates. 

Nvidia said in Q4 that it delivered revenue of $11 billion for its Blackwell GPU, marking the fastest product ramp in its history. However, there was not a strong ramp in the GB200 NVL72 racks in Jan-March, with delivery estimates finally increasing in April. 

Data from Morgan Stanley places Jan-March GB200 NVL72 rack shipments at ~1,000, while April shipments were estimated to rise sharply to ~1,500, with the majority coming from Foxconn. This aligns with fiscal Q2 GB200 NVL72 rack estimates of 4-5K, up 3-4x sequentially.  

A screenshot of a tweet by Beth Kindig stating that Morgan Stanley estimates Nvidia's GB200 NVL72 rack shipments reached 1,500 in April, up from 1,000 in all of Q1. The tweet was posted on May 17, 2025, at 11:38 AM and has 197.3K views.

Source: Beth_Kindig TwitterBeth_Kindig Twitter 

At the midpoint of those estimates, GB200 NVL72 rack revenue would calculate out to ~$13.5 billion at a $3 million ASP, complemented by HGX B200 shipments and other Blackwell products, which likely contributed 70%+ of the $11 billion from Q4.  

This means even though Q2 could see a bigger impact from China that Blackwell sales may be able to help absorb these losses. The only hitch is that suppliers are not fully in agreement with that takeaway. 

Subscribe for Full Access to the Article 

Below the Paywall is the Following information 

  • What AI-Related Suppliers and Competitors are Saying Ahead of Nvidia Earnings, including an important supplier correlation that broke down in November and worsened in April. 
  • The timing of when Big Tech capex will hit peak growth and what this means for Nvidia’s stock 
  • The I/O Fund’s trading plan for Nvidia including never-before published buy targets over a 12 to 18-month time frame.

IT’S OUR BIGGEST SALE OF THE YEAR — GET $275 OFF ADVANCED FOR FOUR DAYS ONLY. Expires May 27th at Midnight. 👉 Claim Offer Here

Sign Up to Continue ReadingSign Up to Continue Reading

Paid subscribers, click here to view the full articleclick here to view the full article

The last time we published our Nvidia price targets in Where I Plan to Buy Nvidia Stock Next, it was to say the stock would trade below $100. The stock topped five days later at $149. This buy plan was perfectly timed before the DeepSeek selloffs. We then reiterated this price target again the very week when Nvidia’s stock was at $138 and traded below $90 a little over a month later. Real-time trade alerts are sent to Members, including when we snagged shares as low as $87. When you layer these granular details on top of our first entry being $3.15 in 2018 — you will not want to miss the next buy plan our firm is putting into place.you will not want to miss the next buy plan our firm is putting into place.

Please note: The I/O Fund conducts research and draws conclusions for the company’s portfolio. We then share that information with our readers and offer real-time trade notifications. This is not a guarantee of a stock’s performance and it is not financial advice. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis. Beth Kindig and the I/O Fund own shares in NVDA at the time of writing and may own stocks pictured in the charts.

Recommended Reading:

Potential Upstream Beneficiary of Nvidia’s CPO Push Come 2026

Posted on by io-fund

There is a small-cap company that is emerging as a potential beneficiary of an upcoming shift to co-packaged optics over the next two years. Nvidia introduced its first co-packaged optics (CPO) switches with integrated silicon photonics at GTC 2025, addressing the need of providing more bandwidth and faster speeds while reducing energy consumption in exascale GPU clusters.  

Nvidia is utilizing TSMC’s COUPE (Compact Universal Photonic Engine) for its Quantum-X and Spectrum-X switches, and this company’s wafer-level optics tech (WLO) is expected to be an integral part of the CPO supply chain for TSMC’s COUPE.  

For 2025, this company is expecting little to no revenue from CPO, though it expects the business to ramp quite rapidly come 2026 and beyond, due to the advantages CPO offers in terms of bandwidth and its upstream placement in the supply chain serving a critical need for COUPE.  

While this analysis will focus on this stock’s opportunities in wafer-level optics, which has the potential to quickly become a significant driver for revenue growth from the emergence of co-packaged optics, it’s important to note that this segment remains quite small presently while automotive remains the core growth driver for the company.  

Every Thursday at 4:30 pm Eastern, the I/O Fund team holds a webinar for premium members to discuss how to navigate the broad market, as well as various stock and crypto entries and exits. Beth Kindig offers weekly deep dives including lesser-known cryptocurrencies and AI stocks, plus the team offers trade alerts. 📈 The I/O Fund team is one of the only audited portfolios available to individual investors. 📊 If you’d like to subscribe to the Advanced Market Signals plan, email us at premium@io-fund.com. 📧4:30 pm Eastern, the I/O Fund team holds a webinar for premium members to discuss how to navigate the broad market, as well as various stock and crypto entries and exits. Beth Kindig offers weekly deep dives including lesser-known cryptocurrencies and AI stocks, plus the team offers trade alerts. 📈 The I/O Fund team is one of the only audited portfolios available to individual investors. 📊 If you’d like to subscribe to the Advanced Market Signals plan, email us at premium@io-fund.com. emium@io-fund.com. 📧

Nvidia Unveils New CPO Switches at GTC 

The AI industry has been rapidly progressing towards million-plus GPU clusters, with Broadcom outlining late last year that it believes its three hyperscale customers are aiming to each have 1-million GPU clusters by 2027, which will require new scale-up networking architectures and components. 

Co-packaged optics (CPOs) are seen as one possible way to break past the cost and power bottlenecks of current pluggable optical technologies and facilitate the development of larger-scale AI clusters. CPOs place optics directly on the switch die which decreases the distance and thus insertion loss, while improving power efficiency due to that shorter distance. 

Source: Cisco 

Nvidia’s two new switches unveiled at GTC, the Quantum-X Photonics InfiniBand and Spectrum-X Photonics Ethernet, were designed with that million-plus GPU cluster scale goal in mind. Nvidia says that by replacing pluggable optics with silicon photonics on the package, it can “deliver 3.5x more power efficiency, 63x greater signal integrity, 10x better network resiliency at scale and 1.3x faster deployment compared with traditional methods.” In simple terms, the switches are more reliable, more efficient, consume less power, and are quicker to deploy than current networking tech. 

Nvidia VP Ian Buck stated at GTC that the CPO switches help reduce power consumption by eliminating the need for external lasers and pluggable transceivers to achieve a significant reduction in power from 39 watts to 9 watts. Buck explained that this “gives you that benefit from going from 39 watts of power down to only 9 watts of power for the same number of ports, and that's huge. It doesn't sound like 39 sounds a lot. But if you get 400,000 GPUs in an AI supercomputer, there's like 24 megawatts of lasers like so that's a lot of laser light that could be optimized and made more efficient.” 

24 MW may not sound like much, but given that the GB300’s are expected to draw up to 140 kW per rack, that’s the equivalent of nearly 172 GB300 racks (or the equivalent of approx. 12,400 individual GPUs) in power savings from the shift to CPO. Thus, by shrinking the bill of materials by eliminating pluggable transceivers while simultaneously providing a much more power-efficient solution, Nvidia can further reduce TCO for its customers. 

Nvidia Expects CPO Switch Availability as Soon as 2H 2025 

Similar to its GPU roadmap, Nvidia is moving quickly when it comes to its CPO switches, expecting availability for its Quantum-X Photonics InfiniBand switches as soon as the second half of this year, while its Spectrum-X Photonics Ethernet switches are expected to be released in 2026.  

TSMC’s COUPE Paves Way for CPO Adoption 

Nvidia’s new switches rely on TSMC’s silicon photonics platform called COUPE (compact universal photonic engine).  

TSMC first published research on this new platform in 2021, as a team of researchers discussed a primary challenge of mass adoption of silicon photonics, which was at the time the lack of an integration platform that could meet a range of power, performance and cost needs. TSMC’s researchers proposed COUPE, which they said had the ability to meet “the most demanding system requirements and pave the way for [silicon photonics]-based wafer level system integration (WLSI) for high performance computing applications.”  

COUPE combines a photonic integrated circuit (PIC) with a 65nm electronic integrated circuit (EIC) via its SoIC-X advanced packaging tech. TSMC says that utilizing SoIC-X “enables the lowest impedance at the die-to-die interface and therefore the highest energy efficiency.” COUPE is also easily integrated with merchant ASICs and GPUs to form a co-packaged structure.  

TSMC laid out a tentative timeline at the 2024 North American Technology Symposium in April 2024. At the symposium, TSMC laid forth a plan to enable COUPE in pluggable optics in 2025, and COUPE on substrate in a CoWoS-based CPO in 2026, which it said would offer 2x reduction in power and 10x reduction in latency. TSMC mentioned that it was exploring COUPE on CoWoS interposers for further power reduction.  

Recently, TSMC has been rather tight-lipped about COUPE and co-packaged optics — in Q4’s earnings call, CEO C.C. Wei was questioned about CPO and how it would facilitate this shift in the supply chain. Wei answered that for “big volume, I don’t think it will be in this year, or probably we have to wait for one or one and a half year to see that contribution or the volume production. The initial results are quite good, no doubt about it.” 

SoIC capacity is expected to rise at a rapid rate, though there have not been any major updates recently. At the beginning of 2024, TSMC’s SoIC capacity was estimated to reach 5,000 to 6,000 wafers/month by the end of the year, a 150% to 200% YoY increase. 2025’s capacity at the time was expected to reach 10,000 wafers/month, nearly doubling YoY. A recent update from TrendForce reaffirmed the 10,000/month capacity for 2025 and outlined the possibility of another doubling of capacity in 2026. 

It had been reported in May 2024 that TSMC was aiming to expand its SoIC capacity at a 100% CAGR from 2023 through 2026, correlating to an “eight-fold” increase in capacity from 2023’s levels by 2026. 

FOCI’s Role in CPO with TSMC 

While TSMC is taking a more back-seat view to CPO’s timelines, suppliers are expecting a more accelerated ramp, with FOCI and Browave both seeing initial shipments in 2025 and a rapid ramp through 2026. FOCI is viewed as one of TSMC’s closest collaborators on CPO.  

In its 2023 annual report, FOCI stated that “CPO technology has just entered the market, and the production cost is still high,” but with the “explosion of high-speed transmission demand, it is expected that CPO technology will be a necessary technology that cannot be ignored and will enter the market in large quantities after 2025.”  

According to DigiTimes, FOCI anticipates that CPO fiber array products will see some smaller scale shipments in 2025, leveraging its ReLFACon (Reflowable Lensed Fiber Array Connector) products for optical switches and HPC/AI use cases. Mass production of CPO components is expected in 2026.  

FOCI is rumored to be the sole supplier of external fiber array units (FAUs) to TSMC’s first and second-gen COUPE via its ReLFACon product. FAUs are a critical part of CPO tech — the FAUs ensure smooth and efficient coupling of optical fibers to the silicon photonics engines, which enables high data throughput and reduces latency. In COUPE, the FAUs are used to align optical fibers on the PIC; COUPE 2.0 introduces a broad-band coupler that integrates with the FAUs to minimize insertion loss and extend alignment tolerance, which can lower manufacturing costs at the alignment stage. 

FOCI is not the only supplier preparing for strong growth come 2026 — Browave (which also holds a 4.6% stake in FOCI) is said to be fast-tracking its CPO development, as one of the named partners for Nvidia’s switches at GTC. Browave “expects to complete CPO validation by late 2025,” and is preparing for a “breakout” 2026, per DigiTimes.  

Explosive Growth Forecasted for CPO Market 

The CPO market is relatively new in that 2025 is expected to be the first year when growth surfaces before accelerating through 2026. Morgan Stanley estimated the CPO market size at just $8 million in 2023, projecting a 172% CAGR through 2030 in its base case scenario to reach $9.3 billion.  

This base case forecast has a few key assumptions: 1) Nvidia is the first to adopt CPO in its Rubin rack systems in 2026, with Rubin’s shipment volume reaching 200,000 units in 2026 and 700,000 in 2027; 2) other vendors including Cisco, Broadcom and Marvell begin shipments in 2027. 

Morgan Stanley also outlined a bull case scenario, projecting growth at a 210% CAGR, resulting in a significantly larger end market of $23 billion by 2030. This relies on much more optimistic assumptions: 1) Rubin shipments of 500,000 in 2026 and 1.75 million in 2027; 2) broader adoption of CPO by more chip manufacturers; 3) CPO yield rates greatly improved. 

However, Morgan Stanley’s bear case scenario sees the CPO market only $1.3 billion by 2030, on the assumptions that CPO yield issues cause shipment delays and thus a lower customer willingness to adopt the technology.

CPO One of 3 Growth Opportunities Ahead for Himax  

Himax is expected to be a critical player in this CPO push due to its partnership with FOCI, and sees CPO and WLO as one of three different growth opportunities ahead (with the other two being automotive OLED and ultralow power WiseEye AI sensing chips). For Himax, WLO presents a rather large opportunity over the next six to twelve quarters as Nvidia progresses with CPO switches to enable the scaling of data centers to millions of GPUs.  

In June 2024, Himax deepened its strategic partnership with FOCI and acquired a 5.3% equity stake in the firm for $16 million. Himax explained that the partnership leverages and combines WFO’s expertise with FOCI’s advanced ReLFACon solution to develop linear pluggable optics (LPO) and co-packaged optics solutions for AI and HPC chips that “demand enhanced bandwidth, improved data rate, minimized signal loss, reduced latency, and lower energy consumption.”  

Himax says that its WLO technology plays a crucial role in CPO by “providing essential optical coupling capability” as a core part of the CPO solution. Himax is said to be a sole supplier of micro-lens array units to FOCI’s FAUs for first and second-gen COUPE. 

Himax Outlines CPO’s Value Proposition, Ramp Profile 

Shortly after this investment, in August’s Q2 earnings call, management noted that they expect WLO to play an “even more decisive role in the next-generation optical technology landscape,” due to its versatility, precision and “small form factor characteristics that are not feasible with alternative technologies.” They stated that they believe LPO and CPO tech can “generate new, long-lasting revenue streams for Himax.” 

Management added that they are working closely with FOCI to align with multi-year roadmaps of foundry partners and AI chip customers with an effort to meet their near-term production targets. It is assumed that this comment is referring to TSMC’s COUPE and Nvidia’s CPO switches.  

 Q2’s analyst Q&A highlighted that this is not simply an R&D project, but rather one that is progressing quickly towards mass production: 

Q: Donnie Teng, Nomura 

My second question is regarding to the CPO. So would you maybe elaborate more on, you know, what’s the timeline of the CPO product. When should we expect to see some, you know, small volume contribution? And how confident you are to ramp up this business in the mid to long term?  

A: Jordan Wu, Himax CEO 

“First on precise timeline, I’m bound by NDA of my partner and customer, so I’m afraid I cannot give you very, very specific date or timetable. But I can tell you, what we are working on right now, the design is targeting for mass production. It is certainly not a R&D concept. …. Actually we’re way past the stage … and we are now pushing towards mass production ASAP. That’s what I can tell you. In fact, we expect to see some small but very early result hopefully by the end of this year but that’s minimal. But the next year, you know, if everything goes as planned, there will be steady ramping. And the confidence level mid to long term, I would say very confident. 

… So I would say, you know, everybody in the ecosystem is very keen to making sure that this happens ASAP. … And we have a roadmap together with partners, our customers, to really pretty dramatically expand the transmission bandwidth very substantially. I’m talking about by multiple times over the next few years. And you know, some of these projects are already in experimental stage in the earlier experimental stage or more mature experimental stage.” 

Himax confirmed in its 2024 annual report that small-scale production of its first-gen CPO solution was already underway by the end of 2024.  

Himax Provides Major Update in Q4, Lays Out CPO Revenue Opportunity 

Himax provided a major update in Q4 regarding CPO, as it continues to progress through small-scale production of its solution for LPOs and developing its platform for CPO architectures with FOCI. 

Management stated that the long-term prospect of CPO remains unchanged despite the market’s AI jitters from DeepSeek, and this was “evident by the significant increase in [the] customer’s recent trial production volume forecast, indicating an accelerated timeline for CPO technology to enter mass production.” They also expect sample volume increases over the next few quarters. 

Most importantly, Himax’s management outlined what CPO’s revenue contribution could look like in the future, with CEO Jordan Wu emphasizing that Himax is now “more optimistic than ever” about the outlook for WLO and CPO products, which they believe are “poised to generate significant growth opportunities and become a major revenue and profit contributor in the years ahead.” 

Here’s what Wu said about 2025, 2026 and the revenue opportunity in dollar terms: 

“2025 will be a year for engineering validation with only sample shipments for us. So while the sample shipment will accelerate quarter by quarter, the revenue contribution will be rather limited, if you compare that with our total revenue. The fourth quarter, which presumably will be the peak of this year, the revenue is set to be in millions of dollars based on current forecast. But again, it's still quite small compared to our total size and things can still change. It's still early.sample shipment will accelerate quarter by quarter, the revenue contribution will be rather limited, if you compare that with our total revenue. The fourth quarter, which presumably will be the peak of this year, the revenue is set to be in millions of dollars based on current forecast. But again, it's still quite small compared to our total size and things can still change. It's still early. 

Now, in all likelihood, mass production will commence in 2026, but we don't know how, and certainly we won't comment on exactly when, in 2026. It's probably still more than a year from now. There are still many unknowns, like how many customers, how many projects, or their ramping curve, etc. And therefore, while 2026 is likely to be the first year of mass production, it is still early and it's still difficult to give a revenue indication for the year at this point.in all likelihood, mass production will commence in 2026, but we don't know how, and certainly we won't comment on exactly when, in 2026. It's probably still more than a year from now. There are still many unknowns, like how many customers, how many projects, or their ramping curve, etc. And therefore, while 2026 is likely to be the first year of mass production, it is still early and it's still difficult to give a revenue indication for the year at this point. 

Now, if we look further ahead and ignore the exact timing and ramping curve, etc, for the time being, and just try to kind of paint a picture for, I would call it annualized potential revenue for Himax,… when the CPO business reaches an ‘early stage’ of mass production, right … when the industry is perhaps still testing the water, with maybe only premium models equipped with CPO.  

In making the assessment, we have considered leading AI customer total advanced GPU shipment outlook … and the leading foundry's total CoWoS capacity plan, which is very much public information. And we have assumed a very low percentage of CPO attach rate for both, right. … So with such conservative assumptions, I can say that our annualized CPO revenue could still reach hundreds of millions of dollars when we get there. Again, this is early stage, and this is the best I can do in terms of providing a revenue indication.” I can say that our annualized CPO revenue could still reach hundreds of millions of dollars when we get there. Again, this is early stage, and this is the best I can do in terms of providing a revenue indication.” 

Wu believes that the determining factor is not a question of whether CPO adoption will occur, but rather a question of how fast CPO will penetrate the industry due to the benefits it offers to power consumption, bandwidth, and cost. Wu added that he also believes the ultimate demand for the new tech is likely to be far above what Himax can predict at this stage. 

Putting Hundreds of Millions of Revenue in Perspective 

While hundreds of millions of dollars is but a mere splash in the bucket for a company like Nvidia or TSMC when it comes to AI-driven revenue, that opportunity is remarkably large for a company like Himax. 

Himax reported just $907 million in revenue in 2024, or just over a (4%) YoY decline due to global demand weakness and conservative purchasing trends from customers due to market uncertainty. Gross margin improved more than 2.5 points YoY to 30.5%, its first annual expansion since 2021, aided by strong 20% YoY growth in the automotive segment (50% of revenue in Q4) which Himax says enjoys a higher gross margin than its corporate average. 

Himax’s revenue performance has been challenged since 2021, where the pandemic-driven operating environment led to demand outpacing supply in core display driver end markets, leading to 110% YoY automotive revenue growth, 77% YoY tablet IC growth, and 85% YoY smartphone growth. Himax generated $1.55 billion in revenue in 2021, up more than 74% YoY while gross margin nearly doubled to 48.4% due to very favorable price and product mix. Revenue has declined YoY each year since then. 

For an emerging new opportunity that has yet to generate meaningful revenue, WLO holds the potential to grow into the size of 25% to 50% of Himax’s current annual revenue, or multiple times the size of its non-driver business, which has reported revenue between $140 million to $185 million the past four years.  

The CPO-driven opportunity holds remarkable potential to dramatically increase Himax’s topline and growth and put it on a path to quickly reach fresh records for revenue. However, this opportunity also extends down the line as it may provide a substantial boost to Himax’s earnings power.  

Himax is currently operating at around a 30% gross margin and high-single digit operating margin profile. However, this operating margin is driven entirely by Himax’s driver segments, as its non-driver segment had generated widening operating losses from 2021 through 2023 before slightly rebounding in 2024. 

Operating losses for Himax’s non-driver segment widened from ($17.7 million), or an (11%) margin, in 2022 to ($32.1 million), or a (23%) margin, in 2023. However, losses rebounded slightly to ($24.5 million) in 2024, or a (16%) margin. This weakness amplified the rapidly shrinking operating income from drivers, which faced pricing pressures in 2022 and 2023, and weighed down on the segment’s rebound to $93 million in operating income in 2024. 

Himax explained that low sales volumes in non-drivers “led to insufficient revenue to fully cover expenses” over the past few years, but as volume production ramps, such as for products including WLO and WiseEye, Himax expects to generate positive operating income as non-driver products have “higher gross margins as well as higher growth potential” versus drivers. 

Given some of the uncertainties about timing of CPO’s ramp, volumes at mass production, realized selling prices and other factors, Himax has provided no insight into what the margin profile for WLO could look like. Assuming the CPO business ramps as expected and matures into a 40% gross margin, 15% operating margin business at an $500 million dollar scale, it could generate $200 million in gross profit and $75 million in operating income by itself. This alone could represent a ~$0.40 positive impact to EPS.  

If CPO matures into a much higher-margin business due to Himax’s and FOCI’s positioning in the supply playing a core role to help foundry partners meet high demand, say at 60% gross and 30% operating margins (approx. in line with 2021’s corporate operating margin), it could generate $300 million in gross profit and $150 million in operating income, or up to $1 in positive EPS impact.  

At the moment, there is little analyst coverage and limited visibility into Himax’s growth prospects. Current analyst estimates point to a slight YoY decline in EPS in 2025 to $0.44, with revenue rising 8% to $981 million as Himax works its way out of a more challenging macro backdrop. For 2026, EPS is expected to double to $0.88 with revenue growth of 20% to $1.18 billion, though this could be impacted by the timing and pace of CPO’s ramp.  

Himax does have other growth opportunities outside of CPO, with management anticipating strong growth in its WiseEye business in 2025 and continuing momentum in automotive TDDIs, timing controllers and growth in auto OLED. 

Quick Note on Technicals 

In 2019, Himax appears to have put in a major low.  That being said, investors should still expect large swings and the potential for bouts of volatility as this larger uptrend pattern plays out.  

The current drop that we are in is hitting strong support just under $8. If this level can hold, we should see one more push into the $17 region in the coming months, which would complete a large 1st wave.  If this does happen, we would then need a 2nd wave pullback to follow.  This would likely take you back into the $10 – $7 range, before commencing with the larger uptrend pattern. If any bout of volatility takes us under $6.40, then we should expect a drop into the $4 – $2 region before finding a meaningful low.  

While the larger pattern suggests much higher prices for HIMX, based on the overlapping nature of this pattern, investors should expect periods of heightened volatility along the way. The $6.40 price point is a major line in the sand. If this holds and we see a push into the $17 region, investors should also be aware for the potential of a notable pullback before taking off.  

Major Risk: China Sourcing: 

Himax operates via a fabless business model utilizing third-party foundries and OSAT capabilities, and primarily sells products via direct sales teams in core regions Taiwan, China, South Korea and Japan. However, Himax generates more than three-fourths of its revenue from China, which could place it in the cross-hairs of increased geo-political tensions.  

Himax notes that it relies on TSMC, UMC, GlobalFoundries Singapore, Macronix, PSMC, Nexchip and SK Hynix System IC for wafer fabrication, while its WLO production is done in-house in its two facilities in Taiwan. 

China’s revenue share declined slightly from 76.2% in 2023 to 73.4% in 2024, with a majority of this revenue (85%) in the Drivers business. China accounted for nearly 77% of Himax’s Drivers business in 2024, and 63% of its Non-Drivers business. While Himax may not be exposed directly to US-based tariffs due to its concentration in Asia for supply chain and manufacturing, it warned that it expects to “continue to be subject to economic and political events and other developments that affect our customers in Asia,” and if tariffs dent consumer electronics demand either in Asia or globally, Himax could face revenue headwinds as a result.  

The I/O Fund owns a stock that is rumored to be a future CPO supplier to Nvidia (and is a current Nvidia supplier) while being vertically integrated on sourcing. This company’s CEO pointed out this is an area incremental strength and an advantage as they manufacture most of their parts internally (which is quite rare) and can quickly capitalize on a rapid ramp in demand. The company stated that they also will source when the demand requires it, yet the vertical integration on the manufacturing side with primary production in the US, and some in Europe, help avoid heightened tariff risks from China.

Conclusion 

CPO beckons a massive opportunity ahead for Himax due to its positioning upstream in the supply chain for TSMC. Management outlined that the emerging technology could represent hundreds of millions of dollars in annualized revenue assuming a low attach rate as the tech ramps through 2026 and beyond.  

While CPO promises hundreds of millions of potential revenue and possibly a large positive impact to EPS as it matures, Himax’s driver business remains in the driver’s seat for 2025 and 2026 until CPO growth arises, and the segment is still plagued by margin weakness. The driver business is far from 2021’s peak for revenues and margins, and thus Himax’s EPS, and it may struggle to return to 2021’s levels due to the supply chain circumstances and resulting pricing advantages it benefited from at the time.   

Despite that, Himax appears relatively priced ahead of a potential revenue acceleration and development of a multi-million dollar business line. Himax is currently trading at 18x forward earnings, with EPS estimated to double in 2026, and 1.4x forward sales, with prior peaks around 3x. Gross margins are showing strength due to strong automotive sales and strict cost management, while operating income is beginning to recover from a prolonged contraction from headwinds in the driver segment.  

Small caps can have quick and volatile moves, and Himax is no exception. Shares surged 45% in one day in December when it was first named as a potential AI supplier to TSMC, and then rose nearly 100% in seven days in January before plunging nearly -28% during DeepSeek’s market rout and is down nearly 50% from the January high. 

The I/O Fund owns a different NVDA and possible future CPO supplier, sharing this research with our Pro and Advanced subscribers, while discussing potential setups and trading plans in our weekly webinars with Portfolio Manager Knox Ridley. Take advantage of a limited-time offer for $75 off Pro or $100 off Advanced here.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading:

Micron FQ2 Results: Record $1 billion HBM revenue; Mixed Consumer Results

Posted on by io-fund

Micron beat analyst consensus estimates and offered better-than-expected guidance. The company expects strong growth to continue in the next quarter, driven by increased DRAM and NAND shipments. The management also highlighted strong AI data center demand and the ramp-up of HBM products. 

Management was quite optimistic about the opportunity in HBM, by increasing the TAM for calendar year 2025 from over $30 billion to over $35 billion. The management also reiterated that the company’s 2025 HBM production is fully allocated and is in discussion with customers for the calendar year 2026. 

In the past, Micron did not report the AI HBM revenue. Management highlighted that the company reported record data center revenue, with the HBM revenue growing more than 50% sequentially, to over $1 billion in their recent quarter. The HBM shipments were ahead of the company’s projections, demonstrating strong execution of the ongoing ramp. 

The management also highlighted that the company’s HBM3E 8-high is designed into NVIDIA's GB200 system, and the HBM3E 12-high is designed into the GB300. Micron also initiated volume shipments to the third large HBM3E customer in FQ2 and anticipates additional customers over time. At the recent GTC 2025 event, Micron showcased its complete AI memory and storage portfolio to support AI from the data center to the edge. The company’s High-performance Micron 9550 NVMe and Micron 7450 NVMe SSDs are included on the GB200 NVL72 recommended vendor list. They expect multibillion dollars in HBM revenue in FY2025. Micron’s HBM4 is expected to ramp in volume in calendar year 2026 and is 60% faster than HBM3E. 

This quarter, revenue grew by 38.3% YoY yet was down (-7.5%) QoQ to $8.05 billion, beating estimates by 2%. Adjusted EPS grew by 271.4% YoY to $1.56, beating estimates by 9.5%. However, the company’s gross margins were below the guidance due to a higher mix of consumer products and lower NAND prices. 

The FQ3 revenue guidance is $8.8 billion, representing YoY growth of 29.2% and 9.3% QoQ growth at the midpoint, beating estimates by 3.8%. Adjusted EPS guide is $1.57, representing YoY growth of 153.2%, beating estimates by 3.3%. 

Micron’s valuation remains low with a P/E ratio of 22.7 and a forward P/E ratio of 13.6. Its P/S ratio is 3.4 and a forward P/S ratio is 3. A few reasons that Micron’s valuation is lower than AI peers is likely due to memory being more cyclical and Micron seeing fierce competition with SK Hynix and Samsung, which leads to more commoditized pricing. There is China, as well, which represents 12% of revenue – although this is not likely to be a leading reason for the lower valuation. Micron's earnings power also fluctuates wildly at times due to the cyclical nature of the memory market. More recently, the increasing consumer revenue mix and lower NAND prices have also led to analysts reducing estimates, which have also led to lower valuations.  

What Micron will need to answer is if the cyclical nature of the memory market will smooth out as the dollar content of memory is rapidly increasing. AI training and inference rely heavily on high-bandwidth memory (HBM) for the massive memory bandwidth that complex models require. AI servers also use more DRAM and NAND than a traditional server. These are reasons that Micron’s cyclical fundamentals could become more secular as the AI economy is built out. 

Ultimately, we want to keep a close eye on companies that are at the forefront of AI yet have valuations that do not reflect this outsized growth. Below are notes that discuss the most recent earnings report as well as references to our previous coverage. 

Bullish HBM commentary

Management set out an optimistic tone about the opportunity in HBM and increased the TAM to over $35 billion for CY2025, up 17% from the previous estimate provided during FQ1 results. As previously noted, the company’s 2025 HBM production was fully allocated and are further seeing strong demand for the HBM supply in 2026 and are in discussions with the customers on agreements for their calendar 2026 HBM demand. 

“We see strong demand for HBM and have once again increased our HBM TAM estimate for calendar 2025 to over $35 billion. We remain on track to reach HBM share similar to our overall DRAM supply share on a run rate basis in calendar Q4 2025. As previously mentioned, Micron has sold out of our HBM output in calendar 2025. We are seeing strong demand for our HBM supply in 2026 and are in discussions with our customers on agreements for their calendar 2026 HBM demand.” 

They also highlighted that they have initiated volume shipments to the third large HBM3E customer. “We are making good progress on additional platform and customer qualifications with HBM. Micron's HBM3E 8-high is designed into NVIDIA's GB200 system, and our HBM3E 12-high is designed into the GB300. In fiscal Q2, we initiated volume shipments to our third large HBM3E customer and anticipate additional customers over time. We expect multibillion dollars in HBM revenue in fiscal 2025.” 

We covered this in the past when we stated in our December post-earnings update that “Management pointed out they are raising their view of server unit percentage growth for the current year and they anticipate server unit growth to continue in 2025. The CEO also stated that HBM has exceeded their plans due to solid execution on yield and capacity ramps.” We also stated that the next catalyst would be the availability of HBM4 in CY2026. 

Similarly, we highlighted during the September post-earning update that the company delivered strong results due to robust data center/HBM revenue. We informed our members in June 2024 that that HBM is sold out for calendar year 2025 with pricing already contracted for. 

Market Outlook 

Management increased the 2025 DRAM bit shipment forecast from mid-teens percentage range to mid-to high teens percentage range due to the growing adoption of AI in smartphone & PC devices and HBM is also a strong contributor to the bit demand growth.  

“Now turning to our market outlook. Calendar 2024 DRAM bit demand growth was in the high teens, consistent with our prior expectations. Calendar 2024 NAND bit demand growth was approximately 10%, slightly below our previous view of low double digits. We forecast calendar 2025 DRAM bit demand growth in the mid- to high teens percentage range and NAND in the low double-digit percentage range. Over the medium term, we expect industry bit demand growth of mid-teens CAGR for both DRAM and NAND.” 

HBM4E to consume higher silicon 

Management mentioned that HBM4E will consume 4x the amount of silicon compared to D5. “As noted before, HBM3E consumes 3x the amount of silicon compared to D5 to produce the same number of bits. Looking ahead, we expect the trade ratio to increase with HBM4 and then again with HBM4E when we expect it to exceed 4:1. The sustained and significant increase in silicon intensity for the foreseeable future contributes to tightness for industry leading-edge node supply and constraints capacity for non-HBM products.” 

Growth in the second half of CY2025 

Management mentioned that the transition from 8-high offering to 12-high offerings will increase revenue in the second half of the year due to premium pricing. Also, shipping to the third large customer will also lead to a revenue increase in the second half of the year. However, revenue growth rate has peaked for now and will decelerate in the year's second half. 

Q: CJ Muse (Analyst) 

“Very helpful. And a quick follow-up. You revised your HBM industry revenue outlook higher. Curious if there's a framework on how you're thinking about kind of first half versus second half for the industry? 

A: Sanjay Mehrotra (CEO) 

Of course, the revenue in the second half as you go from 8-high to 12-high continues to go up because 12-high will be carrying a certain premium over 8-high. So if we have projected more than $35 billion for calendar year 2025 and a bigger portion of that in second half of calendar '25 versus first half. And more than $35 billion, of course, is the industry TAM for HBM that we have referred to here. 

And I'll just point out that, of course, with respect to HBM, there is expansion of HBM customer base taking place. Micron itself, now we are shipping to a third large customer that we have begun shipping our products to. So that also is contributing to the growth in HBM revenue in the second half as the customer base expands.” 

Financials: 

Revenue growth momentum eases 

FQ2 revenue grew by 38.3% YoY and down (-7.5%) QoQ to $8.05 billion, beating estimates by 2%, driven by strong HBM revenue that grew more than 50% sequentially to over $1 billion.  

FQ2 DRAM revenue grew by 47% YoY to $6.1 billion, representing 76% of total revenue. DRAM revenue was down (- 4%) QoQ, with bit shipments decreasing in the high single-digit percentage range and prices increasing in the mid-single-digit percentage range because of improving portfolio mix. 

The company’s DRAM revenue was down sequentially due to the decrease in bit shipments. However, management expects bit shipment to increase in Q3, which will likely help with the DRAM revenue recovery.   

FQ2 NAND revenue grew by 18% YoY to $1.9 billion, representing 23% of total revenue. NAND revenue was down (-17%) QoQ, with bit shipments modestly higher and prices decreasing in the high-teens percentage range. FQ2 bit NAND bit shipments exceeded management expectations, driven by higher consumer-oriented shipments. Management anticipates growth in DRAM and NAND bit shipments in FQ3. 

  • The company expects strong growth to continue in the next quarter with revenue guidance of $8.8 billion, representing YoY growth of 29.2% and 9.3% QoQ growth at the midpoint, beating estimates by 3.8%. However, revenue growth rate has peaked for now and will decelerate in the calendar year's second half. 
  • Analysts expect revenue to grow 25.3% YoY to $9.71 billion in FQ4 and 23.8% growth in FQ1. 
  • For FY2025 ending August, analysts expect revenue to grow 39.1% YoY to $34.93 billion and 28.6% growth to $44.93 billion in FY2026. 

Strong YoY Margin Expansion, Yet Consumer Weighs on Margins

The company’s margins are improving on a YoY basis driven by operating leverage. However, margins are down sequentially, and the company’s gross margins were below guidance due to a higher mix of consumer products and lower NAND prices. This was partially offset by high-value DRAM portfolio mix. Management expects gross margins to be down sequentially in FQ3 due to higher consumer mix, NAND underutilization charges, and NAND pricing weakness. This was also highlighted by the management during the Wolfe Conference last month. It was further clarified during the earnings call Q&A that gross margins will improve in FQ4. 

Q: Harlan Sur (Analyst) 

"Back in mid-February at an investor conference, I know the team had walked us through the dynamics on a weaker gross margin profile during the May quarter. That's playing out, but you did anticipate an improved gross margin profile beyond this quarter, fiscal Q3. So is that still the case that we should see gross margin improvements maybe starting in fiscal Q4 and potentially beyond? And is that a possible data center and your consumer-related products? Is that across total DRAM and your NAND segments? Any color here would be great.

A: Mark Murphy (CFO) 

Sure, Harlan, this is Mark. I'll take that. So let me just make some comments about the third quarter. It is down sequentially, as we had indicated in the conference. And again, as we said in the conference down primarily due to higher mix of consumer-oriented volumes, lower CQ1 pricing on consumer-oriented markets and industry and can just generally, all that partially offset by higher HBM. We do see — while down, conditions have improved since those public comments. And the updated view is reflected in the guide today. 

Now we're not providing guidance on the fourth quarter. However, we do expect gross margin to be up somewhat. There's always tailwinds and headwinds. As you know, on tailwinds, we do expect market conditions to improve. We do expect HBM and other high-value products to grow and contribute to mix improvement…. So in short, we would expect fourth quarter margins to be up somewhat from third quarter.” 

  • FQ2 gross margin was 36.8% compared to 18.5% in the same period last year and 38.4% in the FQ1. It was lower than the guidance of 37.5% due to a higher consumer mix. Management guidance for the next quarter is 35.5%. 
  • FQ2 adjusted gross margin was 37.9% compared to 20% in the same period last year and 39.5% in the FQ1, missing guidance of 38.5%. Management guidance for next quarter is 36.5%. 
  • FQ2 operating margin was 22% compared to 3.3% in the same period last year and 25% in FQ1, driven by operating leverage. Management guide for the next quarter is 21.1%. 
  • FQ2 adjusted operating margin was 24.9% compared to 3.5% in the same period last year and 27.5% in the FQ1. Management guide for the next quarter is 23.7%. 
  • FQ2 net margin was 19.7% or $1.58 billion compared to 13.6% or $793 million in the same period last year. FQ2 adjusted net margin was 22.1% or $1.78 billion compared to $476 million or 8.2% of revenue in the same period last year.

Adj.EPS growth of 271% 

FQ2 GAAP EPS grew by 98.6% YoY to $1.41, beating estimates by 11.4%. Adjusted EPS grew by 271.4% YoY to $1.56, beating estimates by 9.5% driven by strong operating leverage.   

  • Management adjusted EPS guide is $1.57 for FQ3, representing a YoY growth of 153.2%, beating estimates by 3.3%. 
  • Analysts expect adjusted EPS to grow 73.7% YoY to $2.05 in FQ4 and 41.3% growth in FQ1. 
  • For the FY2025 ending August, analysts expect adjusted EPS to grow 419.6% YoY to $6.76 and 64.7% YoY growth to $11.12 in FY2026. However, due to a higher consumer revenue mix and lower NAND prices, analysts have reduced the estimates significantly for this FY2025 from 783% expected growth in June 2024 to the current expected growth of 420%, with some of the growth being pushed to the FY2026. 

Cash Flow and Balance Sheet 

The company’s cash flow improved, driven by higher profits.  

  • FQ2 operating cash flow margin was 49% or $3.94 billion compared to 20.9% or $1.22 billion in the same period last year. 
  • FQ2 adjusted free cash flow margin was 10.6% or $857 million compared to ($29 million) or (-0.5%) in the same period last year.  
  • FQ2 capex was $4.06 billion, up 193% YoY. The company received government subsidies of $963 million (23.7% of capex) in FQ2 compared to $149 million in the same period last year. Management expects the FQ3 capex to be over $3.0 billion and the FY2025 capex to be about $14 billion. Management mentioned in the earnings call that the majority of the FY2025 capex is to support the multiyear facility investments for DRAM and HBM manufacturing, including the Idaho fab, Singapore HBM advanced packaging facility, and Taiwan DRAM test facility.  
  • Earlier in December last year, the U.S. Department of Commerce finalized the $6.17 billion government subsidies under the Chips Act. It will disburse the funds based on Micron’s completion of project milestones. The $6.17 billion funding package, representing roughly 5% of the total estimated capital expenditure in New York and Idaho, will support the construction of two fabs in Clay, New York, and one in Boise, Idaho. This funding is part of Micron's $125 billion capex plan for both states over the next two decades. 
  • The company had cash and investments of $9.6 billion and debt of $14.36 billion compared to $8.74 billion and $13.8 billion in the previous quarter.  
  • During FQ2, Micron extended the debt maturities through a $1 billion 10-year senior note offering and a $1.7 billion term loan, with proceeds used to pay down notes maturing in 2026 and the previous term loan balance.  
  • The company’s inventory increased to $9.0 billion from $8.71 billion in FQ1 to support the strong AI-demand. Management mentioned that the inventory levels will normalize in a couple of quarters.  

Business Units

Compute and Networking (CNBU)Compute and Networking (CNBU) 

Compute and Networking revenue grew by 109% YoY and 4% QoQ to $4.56 billion. For the third consecutive quarter, CNBU revenue reached a new quarterly record, driven by a more than 50% sequential increase in HBM revenue to over $1 billion. However, the CNBU revenue growth decelerated from 153% and 152% growth in the last two quarters and is likely tied to the Blackwell delays.  

The company’s CEO Sanjay Mehrotra, highlighted in the earnings call, “Our HBM shipments were ahead of our plans, demonstrating strong execution of our ongoing ramp. The combination of our revenue from high-capacity DRAM modules and our industry-leading LPDRAM for the data center also exceeded the $1 billion milestone for the quarter.” 

Mobile (MBU)Mobile (MBU) 

FQ2 Mobile Business Unit revenue was down (-33%) YoY and (-30%) sequentially to $1.07 billion due to inventory corrections.  

Embedded (EBU)Embedded (EBU) 

FQ2 Embedded business unit revenue was down (-8%) YoY and (-3%) QoQ to $1.03 billion. It was lower due to inventory correction in the company’s automotive customers. The CEO highlighted in the earnings call that the customers are nearing the completion of inventory adjustments. “Automotive OEMs, industrial and consumer embedded customers are in the later stages of adjusting their inventory levels. In automotive, which comprises the largest portion of our EBU revenue, memory and storage content per car continues to increase as AI-enabled in-vehicle infotainment systems become more enriched and driver assistance functions become more capable.” 

Storage (SBU)Storage (SBU) 

Revenue for the Storage business unit was $1.4 billion, up 54% YoY and down (-20%) sequentially. The sequential decline in SBU revenue was driven primarily by lower storage investments from data center customers after several quarters of very strong growth and the overall NAND industry pricing. 

Earnings Call Notes:

AI PCs and Mobile

Management expects mid-single-digit unit PC growth in calendar 2025, with a stronger second half. AI PCs require increased DRAM, with a minimum of 16GB of DRAM compared to 12GB content last year. They expect smartphone unit volume growth in calendar 2025 to remain at low single-digit percentages and customer inventories are improving. AI smartphones require DRAM capacities of 12GB or higher compared to 8GB in last year’s models. 

“We expect the PC market to grow mid-single digits in unit terms in calendar 2025, with growth weighted to the second half of calendar 2025. The Windows 10 end of life in October 2025, combined with an aging installed base and the desire amongst customers to ensure that their PC hardware specs can support compelling AI applications in the future, are key catalysts that drive this growth. 

AI PCs required a minimum of 16 gigabytes of DRAM, with many models requiring even higher memory versus the average 12 gigabyte PC content last year. During the quarter, we sampled our 16-gigabit 1 gamma-based D5 products to PC clients. In NAND, we launched our Gen9 based 4600 performance SSDs, the fastest in the world for the client market, and completed qualifications of our 2650 mainstream SSDs at multiple PC OEMs. 

Turning to mobile. Our expectations for smartphone unit volume growth in calendar 2025 remain at low single-digit percentages. Smartphone customer inventory dynamics have played out as anticipated, leading to mobile DRAM and NAND bit shipment growth in our fiscal Q3. AI adoption continues to be a significant driver for increased mobile DRAM demand. 

AI-capable flagship phones increasingly feature DRAM capacities of 12 gigabyte or higher compared to the 8 gigabyte in last year's models. Smartphone OEMs are using Micron's industry-leading 9.6 gigabit per second LP5X DRAM to improve AI performance, delivering up to 20% more tokens per second than those using legacy speed grades on the same SoC.”

Tariffs 

Management mentioned that they plan to pass the costs of tariffs to the customers. China accounted for 12% of FY2024 revenue. “On tariffs, Micron serves as the U.S. importer of record for a very limited volume of products that would be subject to newly announced tariffs on Canada, Mexico and China. We continue to monitor the possibility of future tariffs and are prepared to work with our customers and suppliers to understand future tariff effects and supply chain options that may arise. Where tariffs do have an impact, we intend to pass those costs along to our customers.” 

Valuation:

The company is trading at a P/E ratio of 22.7 and a forward P/E ratio of 13.6. Micron’s P/S ratio is 3.4 and a forward P/S ratio is 3.0. It is trading below its five-year average P/S ratio of 3.7.  A few reasons for Micron’s low valuation are likely due to memory being more cyclical and Micron seeing fierce competition with SK Hynix and Samsung, which leads to more commoditized pricing. More recently, the increasing consumer revenue mix and lower NAND prices have also led to analysts reducing estimates, which have also led to lower valuations. 

Conclusion 

Micron beat the top-line and bottom-line analyst consensus estimates. However, the consumer revenue mix, NAND underutilization charges, and NAND pricing weakness weigh on the gross margins.  

This stock would be a breakout buy rather than a counter-trend entry as we’d like to see the market begin to recognize Micron as being more secular once the HBM-related revenue begins to rival the other more cyclical revenue segments. 

I/O Fund Equity Analyst Royston Roche contributed to this report.

Please note: The I/O Fund conducts research and draws conclusions for the Fund’s positions. We then share that information with our readers. This is not a guarantee of a stock’s performance. Please consult your personal financial advisor before buying any stock in the companies mentioned in this analysis.

Recommended Reading: