Meta is the strongest Mag 7 stock YTD with its peers down as much as 40%, Meta’s (13%) decline since the first of the year is strong on a relative basis and provides some important clues.
It may be easy to lump the Mag 7 together, yet this cohort is different in key aspects that matter. For example, we’ve been cautioning on Tesla’s margins for years, Nvidia continues has an AI lead that appears insurmountable, while the others have open pocketbooks for AI chips without much visibility to offer in terms of how that capex will be paid back and when.
It’s this last piece that Meta has a better handle on. The company is not offering data center infrastructure for developers and R&D departments to create AI applications, who in turn, must find a path to monetization for a virtuous cycle of demand. With what we know today from earnings disclosures, Azure is leading with a $13 billion annual run rate in revenue among the cloud providers (Amazon, Google, Microsoft). Yet, consider that Meta has a new ad product driven by AI that is already on a $20 billion run rate, and this does not include incremental revenue AI is driving across their other ad products, evidenced by an inflection in Meta’s key metric on ad prices.
There are other areas that Meta is quietly leading. The company is releasing a standalone AI app in Q2, powered by Llama 4, with projections for 1 billion monthly active users by year’s end. That’s up from 700 million users today on the embedded AI features it sees in its popular “family of apps” such as Facebook, Whatsapp and Instagram. To compare, Chat-GPT's standalone app has 500 million. While it remains to be seen if Meta AI will reach the company’s lofty goals, there’s certainly a viable path with 3.35 billion currently using its popular apps today.
Lastly, whereas other big tech companies ramped capex 1-2 years ago, Meta is only now moving into large capex spending from $30B in 2023, to $40B last year to $62 billion this year, up 59.4% YoY. This past quarter, capex was up an astonishing 88% YoY. Meta is in the enviable position to increase this capex once they’ve found product market fit within their own platform rather than pre-emptively building for future AI development. I suspect this new trajectory in capex spending means Meta is ready to scale its AI tools on the ad platform side.
Meta’s AI-Powered Ad Tools
Meta is unique in that its customer base has an unusually high incentive to use AI tools. You could argue developers for cloud IaaS also have a high incentive, yet they must in turn monetize to drive scale for their AI apps. Meanwhile, advertisers see an immediate ROI using AI tools and are then incentivized to spend more.
Evidence of this is visible in the stubbornly high ad revenue growth last quarter of 20.6%, which has sustained even as Meta recently lapped tough comp as the average ad price grew 14% YoY.
Advantage+ has $20B Run Rate:
Advantage+ is an automation tool that uses machine learning to match customers to ads. Similar to Google’s Performance Max, the tool uses AI and machine learning to help with optimize bids/budgets, create and find audiences, and to produce ad creatives. Up to this point, Advantage+ has been used for ecommerce and manually turned on, yet Meta is now going to have the tool turned on automatically and roll it out to other verticals soon.
“In this new setup, all campaigns optimizing for sales, app or lead objectives will have Advantage+ turned on from the beginning. This will allow more advertisers to take advantage of the performance Advantage+ offers, while still having the ability to further customize aspects of their campaigns when they need to. We plan to expand to more advertisers in the coming months before fully rolling it out later in the year.”
According to the recent earnings report, over 1 million advertisers used Meta's gen AI tools to create 15 million ads in December 2024.
- Advantage+ shopping campaigns surpassed a $20 billion annual run rate, up 70% YoY.
- Advantage+ users are seeing a return on ad spend (ROAS) improvement of up to 22%.
- Advantage+ Creative is seeing over 4 million advertisers using at least 1 generative AI ad creative tool, up 1 million just six months ago.
Andromeda has 10,000X increase in Model Capacity:
In H2 2024, Meta revealed a new machine learning system built on Nvidia Hopper chips called Andromeda. Per the earnings call: “This more efficient system enabled a 10,000 times increase in the complexity of models we use for ads retrieval, which is the part of the ranking process where we narrow down a pool of tens of millions of ads to the few thousand we consider showing someone.”
Consider that with AI, an advertiser can make 5,000 creatives rather than 5 creatives, changing the game in personalization. Andromeda is aimed at the new frontier of AI advertising, with an emphasis on retrieval as the first step. This first stage in the ad serving process is to retrieve the ads that are most personalized to a user, resulting in thousands of ads before the final ads are selected.
Due to AI, these ads can now be chosen based on learned associations rather than categories or keywords. By offering a 10,0000X increase in model capacity, Andromeda processes large amounts of data to accurately predict which ads will resonate with which users. This leads to more relevance, which in turn, leads to higher conversions. The deep neural network is considered "hierarchal structured” to where it can handle a much larger volume of ads at a much higher accuracy, improving on the previous structure known as “two-tower.”
Meta AI to See 1 Billion Users in 2025
In a recent conference, Chief Product Officer Chris Cox, described Meta’s AI features as Search 2.0. Instead of visiting a search engine, users engage with Meta AI in-app and on the newsfeed. Meta’s standalone app is expected to be announced in late April at the developers conference.
During the conference, Zuckerberg said, “Meta AI differentiates itself in this category by not just offering state-of-the-art AI models, but also unlimited access to those models for free, integrated easily into our different products and apps. So Meta AI is on track to being the most used AI assistant in the world by the end of this year. In fact, it’s probably already there. We’re almost at 500 million monthly actives, and we haven’t even launched in some of the bigger countries yet.” While this version was on Llama 3.2, the new rollout of its standalone Meta AI app will be using Llama 4.0 (detailed below).
Since then, Meta AI has reported 700 million users on a monthly basis while Chat-GPT is at 500 million weekly active users. This is not comparable, however, since Meta AI is not a standalone app yet. The release will be closely watched as Meta will likely push hard on its large social media base to download the app.
Chatbot to Aid in Personalization of Content and Ads:
Meta’s chatbot will also help to aid in further personalization of ads. The data from its chatbot will be stored in memory to help target users.
“We'll be able to remember certain details that people share in one-on-one chats, for example, and use those details to personalize its responses and then really increasing its ability to deliver great content recommendations and enhance really what makes Facebook and Instagram, so valuable for people today.”
CFO Susan Li described the Meta AI roadmap in 2025 during the Q4 2024 conference call, stating something similar as to the strengths of Meta’s app over other AI chatbots in terms of personalization and a large context window:
“And as we look forward to 2025 in our Meta AI road map, we are really focused on doing more to make it feel more personalized. So, I would say some of the most exciting features we're working on include improving sort of the memory dimension of the Meta AI experience. We'll be able to remember certain details that people share in one-on-one chats, for example, and use those details to personalize its responses and then really increase its ability to deliver great content recommendations and enhance what makes Facebook and Instagram so valuable for people today.”
Notably, it’s unlikely Meta AI is monetized this year, although in future years there are plans for a premium tier similar to Chat-GPT.
Llama 4 Models Released Last week
Llama 4 Scout and Llama 4 Maverick were released last week, while Llama Behemoth at 2 trillion parameters is expected to be released soon. These multi-modal LLMs are trained on text, images and video with a mixture of experts (MoE) architecture. MoE distributes a computational load across “multiple experts” (or neural networks) and trains across thousands of GPUs using what is called model and pipeline parallelism. This enables more compute-efficient pretraining, yet the parameters still need to be loaded in RAM, so the memory requirements remain high.
For Llama 4 Maverick, 400B parameters are stored in memory while only 17B parameters are activated during model deployment, which greatly improves inference efficiency by lowering costs and latency.
This is similar to the breakthrough we saw with DeepSeek R1, which was a MoE model that was trained on 671 billion parameters stored in memory, yet when the model is served, only 37 billion parameters are active. The overall result is computational efficiency as only the most relevant parameters are activated for a specific task.
- Llama Maverick can be run on a single H100 DGX system (8 GPUs) or can be run distributed for inference purposes. The inference cost for Llama 4 Maverick is $0.19 to $0.49 per 1 million tokens compared to Chat-GPT4o at $4.38 per million tokens.
- Groq has verified the upper end of this inference cost estimate with Maverick at $0.53. The model leverages 128 experts. According to Meta, Maverick exceeds “comparable models like GPT-4o and Gemini 2.0 on coding, reasoning, multilingual, long-context, and image benchmarks, and it’s competitive with the much larger DeepSeek v3.1 on coding and reasoning.”
- Llama 4 Scout is cheap at $0.13 per million tokens according to Groq and can be deployed on a single H100 GPU, leveraging 16 experts. Scout can remember long threads and documents of up to 10 million tokens. This is the largest context window across LLMs available today.
- Llama 4 Behemoth is a teacher model that is still being trained with 288B active parameters and 2 trillion total parameters, leveraging 16 experts. Llama 4 Behemoth helped train Maverick and Scout and ranks high on internalinternal benchmarks such as MATH-500, MMLU Pro, GPQA Diamond, Multilingual MMU and image reasoning MMMU. According to MetaAccording to Meta, Chat-GPT 4.5 ranks lower on these benchmarks although this will not be officially verifiedwill not be officially verified until Behemoth is released. For reference, an engineer estimates Chat-GPT 4.5 has 5-7 trillion parameters and 600B active parameters.
Recently, Chief Product Officer Chris Cox explained the goals for Llama 4 – it's a long quote but also important to hear what Meta is setting out to achieve with its newest models and also as an open-source leader in LLMs:
“We've finished pretraining the smaller model. The main thing — first, just from an intelligence perspective, we're trying to pack basically, the intelligence of the large Llama 3 series down into really small models, which can then be used with low latency for low-cost on devices on a single host. So basically getting a lot of the intelligence down into the smaller form factor, that's one of the most important things we can do.
The second is just the basis of what's expected in a new model today, which is reasoning. Agentic use cases, just meaning tool use, ability to use a browser, ability to use other tools.
And then the third piece is an omni model. So basically, interacting with image and voice natively. So rather than translating voice into text and then text into LLMs getting text out, turning that back into speech, having speech be native. This is a big deal. I believe it's a huge deal for the interface, the product. The idea that you can talk to the internet and just ask it anything. I think we are still wrapping our heads around how powerful that is.”
Meta Increases Capex 88% YoY in Q4, will increase 59.4% this Fiscal Year
Q4 capex was $14.8 billion, driven by servers, data centers and network infrastructure investments. Meta raised its 2025 capex guidance to $60 billion to $65 billion, with a midpoint of $62.5 billion, up 59.4% YoY, driven by investments in AI infrastructure and its core business.
Meta is also extending the lives of its servers and networking equipment. They will use non-AI and AI servers for longer periods of time before replacing them, which will be 5.5 years, resulting in annual capex savings and depreciation expenses, which is included in its guidance.
In December 2024, Meta announced plans for its 23rd and largest data center in the United States, a 2GW+ AI data center “so large it would cover a significant part of Manhattan” in Richland Parish, Louisiana. The data center plans to outdo Elon Musk’s xAI’s Colossus supercluster of 1 million NVDA GPUs, with plans for 1.3 million GPUs used to train its Llama AI models.
The 4 million square foot campus will be set on 2,250 acres on the former Franklin Farm site. They will build up to nine buildings and plan to bring 1 GW online by the end of 2025. The site will total over 2GW at full buildout as construction is scheduled to continue until 2030.
Meta has a co-location deal with Entergy to develop a 1.5GW natural gas power plant located adjacent to its proposed $10 billion data center. Entergy is investing $6 billion in electric infrastructure, including a 10,000-acre solar farm, three combined-cycle combustion turbines (CCBT) split into two sites totaling 2.26GW and over 100 miles of new transmission lines. Co-location refers to deals where data centers are located next to power plants generating electricity to feed directly to the customer and usually bypassing the electric grid, which would be a back-of-the-meter arrangement.
Meta is seeking approval from Louisiana OSC to begin construction within 10 months. Meta will be matching its electricity use with 100% clean and renewable energy and has pledged to build or acquire 1.5GW of solar power elsewhere in order to offset emissions from the gas plant. Meta will also contribute to a carbon capture and storage project at the Entergy Lake Charles 994MW gas plant. The gas plant is estimated to open between 2028 to 2029.
Incidentally, a report by The Information claimed Meta is in talks to build a $200 billion data center. However, Meta has not confirmed the rumor.
Meta Plans to Increase ASIC Usage
Meta training and inference accelerator (MTIA) is Meta’s custom silicon and is primarily used for inference today for ads and organic content. The goal is to use MTIA for training next year. According to Tom’s Hardware, the “plan is to gradually increase usage if the chip meets performance and power targets."
There have been delays in the past on Meta’s custom silicon program with MTIA originally launching in 2020, yet Meta had to halt the program to buy Nvidia’s GPUs over the past few years with MTIA v1 launching in 2023. This year, MTIA v2 was launched using RISC-V cores that lowers the overall cost as there are no licensing fees. You can read more about RISC-V, the open-source competitor to Arm in previous coverage here.
Regarding DeepSeek, Meta confirmed "that doesn’t mean you need less compute,” going on to explain that a new trend is to apply more compute at inference in order generate a higher level of intelligence and higher quality. “I think that's generally an advantage that we're now going to be able to provide a higher quality of service than others, who don't necessarily have the business model to support it on a sustainable basis.
Financials:
Average Price Per Ad is Inflecting
Q4 revenue rose 20.63% YoY to $48.39 billion, beating consensus analyst estimates of $46.99 billion by $1.39 billion or 2.96%. Full year 2024 revenue rose 22% YoY to $164.5 billion.
Management guided Q1 2025 revenue of $39.5 billion to $41.8 billion, with a midpoint of $40.65 billion representing 11.5% YoY growth at the midpoint. This was in line on a constant currency basis, yet came in under $41.46 billion consensus analyst estimates due to 3% FX headwinds representing 13.72% YoY growth.
Revenue growth in the quarter was driven by 21% YoY revenue growth in its Family of Apps to $47.3 billion, of which ad revenue rose 21% YoY to $46.8 billion, comprising 97.57% of total revenue. Other revenue for its Family of Apps rose 55% to $519 million, driven primarily by WhatsApp Business Platform.
The number of ad impressions served across its services rose 6% YoY, and the average price per ad rose 14% YoY. Ad revenue growth by geographies was led by the Rest of the World at 27%, Asia-Pacific at 23%, Europe at 22% and North America at 16%.
The average price per ad was up 14% YoY – which is the most important metric to watch as pricing should increase with the AI advancements described above. Growth in average price per ad is already inflecting when you consider Q3 was 11% growth and Q4 of last year was only up 2%.
Regarding this inflection, the CFO hinted CPMs (cost per 1,000 impressions) may continue to grow: “Overall, we are seeing healthy cost per action trends for advertisers for whatever is the action that they are optimizing for. And we believe we'll continue to get better at driving conversions for advertisers. And when we do, that will have the effect of continuing to lift CPMs over time, because we're delivering more conversions per impression served, resulting in higher value impressions.”
The Family of Apps average revenue per person (ARPP) rose 15.6% YoY to $14.25, which was a slight YoY improvement from Q4 ARPP growth of 15.4% — so again, consider this growth is being reported on tough comps.
EPS Grows YoY and QoQ
Q4 GAAP EPS rose 50.47% YoY to $8.02, beating consensus analyst estimates for $6.76 by $1.26 or 18.67%. This was a sequential improvement from Q3 GAAP EPS, which rose 37.34% YoY to $6.03, beating consensus analyst estimates of $5.29 by 13.99%.
It’s no secret that AI agents can write excellent code, which can help drive efficiencies at companies by cutting back on the R&D department overhead. According to the earnings call, Meta expects AI to be as good as mid-level engineer, which will positively impact margins and the bottom line:
“I also expect that 2025 will be the year when it becomes possible to build an AI engineering agent that has coding and problem-solving abilities of around a good mid-level engineer. And this is going to be a profound milestone and potentially one of the most important innovations in history, as well as over time, potentially a very large market. Whichever company builds this first I think is going to have a meaningful advantage in deploying it to advance their AI research and shape the field. So that's another reason why I think that this year is going to set the course for the future.”
Margins:
Meta’s margins and income are impressive at levels not seen since the company was the defacto Wall Street darling in 2017.
- Q4 gross margin was 81.7%, falling slightly from Q3 gross margin of 81.8%.
- Operating margin steadily climbed to one of its highest levels (ever) at 48.3%. Looking back, it was Q4 of 2017 when Meta last reported a higher operating margin. This is up from an OM of 41% last year.
- Net margin of 43.1% is similar – one of Meta’s highest on record, up from a margin of 35% last year.
Cash and Debt Close 2024 at Their Highest Levels for the Year
Q4 operating cash flow reached its highest level for 2024 at $27.99 billion with a margin of 58%. This is up from $19.4 billion for a margin of 48.4%.
Free cash flow fell sequentially to $13.15 billion compared to last year at $11.5 billion. The roughly $14B difference in OCF and FCF is from high capex spend.
Q4 cash and cash equivalents rose 9.75% QoQ for $77.81 billion in cash. Debt was $28.82 billion.
Conclusion:
Clearly, AI is extremely nascent in terms of what it can do with non-stop development and progress coming from tech’s largest players as well as startups. However, also consider that very few companies have a 3B+ global user base to convert to an AI app.
Meta is positioning itself to become a dominant player in the AI landscape led by its Llama 4 models and standalone Meta AI assistant app. AI has enhanced its core advertising business with ad pricing reaching an important inflection point. It’s also interesting to consider its AI tool Advantage+ has surpassed Azure in annual run rate at $20B compared to $13B.
The market is rife with noise, it is hard to know what to pay attention to right now. You can expect us to build a strong pipeline of opportunities to seize when the timing is right.
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: