Meta spent $72 billion and the stock fell

The premise is half right, and the half that is wrong matters

Meta closed at $673.31 on 16 September 2026. That is a 28.1% rise from its 2026 low of $525.72 on 27 March, which is a real surge and a datable one. It is also 14.8% below the $790.00 close of 12 August 2025, the highest in the price record kept here ^[2].

So the stock has not been surging. It has been recovering. The distinction is the whole story, because the period it spent falling is the period Meta spent most heavily on the thing now credited for the rebound.

What the money bought

Meta's annual report puts the number plainly: "Capital expenditures, including principal payments on finance leases, were $72.22 billion for the year ended December 31, 2025" ^[1]. Against that, net income was $60.46 billion ^[1] — the company spent more building than it earned.

The spending shows up in the balance sheet rather than the income statement. Total assets rose from $276.05 billion to $366.02 billion across 2025, and total liabilities from $93.42 billion to $148.78 billion ^[1]. Revenue grew 22% to $200.97 billion and operating income 20% to $83.28 billion, yet net income slipped from $62.36 billion to $60.46 billion ^[1]. A business growing operating profit by a fifth while its bottom line goes nowhere is a business paying for something.

What it is paying for is legible in the supply chain. Meta and Broadcom announced an expanded partnership "to co-develop multiple generations of our next-generation MTIA (Meta Training and Inference Accelerator) chips — the custom silicon that helps power AI across all of Meta's apps and services" ^[3]. Separately, AMD and Meta agreed "a definitive multi-year, multi-generation partnership to deploy up to 6 gigawatts of AMD Instinct™ GPUs", with the first gigawatt shipping from the second half of 2026 ^[4].

Two facts describing one posture: own silicon for inference at the bottom, a second merchant GPU vendor at the top, and less of the stack rented from anyone.

The model line, and the year it changed its name

Meta's public model programme has a clean arc, and it bends sharply in 2026.

The open phase peaked on 23 July 2024, when Meta released Llama 3.1 405B and called it "the world's largest and most capable openly available foundation model", making it available "for download on llama.meta.com and Hugging Face" ^[5]. That is as strong a claim to open leadership as the period produced, and Meta made it about a model anyone could pull down.

Llama 4 arrived on 5 April 2025 with two models shipped — Scout and Maverick, both downloadable — and one held back. Of the largest, Meta wrote: "While we're not yet releasing Llama 4 Behemoth as it is still training, we're excited to share more technical details about our approach" ^[6]. The frontier model was described but not handed over.

Then the line changed its name. On 8 April 2026 Meta announced "Muse Spark, the first in a new series of large language models built by Meta Superintelligence Labs", explaining that "over the last nine months, Meta Superintelligence Labs rebuilt our AI stack from the ground up" ^[7]. On distribution, the same post is precise in a way the Llama posts never had to be: the technology "will be available in private preview via API to select partners, and we hope to open-source future versions of the model" ^[7]. Hope, and future versions.

Three months later the preview widened rather than the weights: on 9 July 2026 Meta put Muse Spark 1.1 into "the new Meta Model API, now in public preview" ^[8].

The filing had already telegraphed it. Meta's 10-K keeps the open-source language — "We have a history of open-sourcing in AI, including releasing our Llama foundation models" — and then adds the qualifier that matters: "We have not released everything we have developed historically and expect to continue training a combination of open and closed models going forward" ^[1].

Read the arc without adjectives: a company whose distinguishing practice was giving the frontier away now describes that practice in the past tense and its frontier model as an API.

Four rivals, and the one thing none of them shares with Meta

Over roughly the same window, the other four published something Meta did not.

OpenAI shipped GPT-6 Astra, "the world's most intelligent and aligned model", and priced it: "OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens." It distributes through "the OpenAI API, Microsoft Azure, and AWS Bedrock" ^[9].

Anthropic shipped Claude Fable 5.1 and Mythos 5.1, "the world's most advanced models for coding and knowledge work", at the same headline rate — "$10 per million input tokens and $50 per million output tokens" — and "available today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure" ^[10].

Google put Gemini 3, "our most intelligent AI model yet", into a filing that is candid about what selling AI does to a business built on advertising: cloud revenues carry margins that "vary significantly and are generally lower than the margins on our advertising revenues" ^[11].

xAI stopped being a standalone company. On 2 February 2026 it announced that "SpaceX announced today that it has acquired xAI" ^[12], and Grok now reports inside a Nasdaq-listed registrant whose CEO "organizes the Company, manages resource allocations, and measures performance among three operating and reportable segments: (i) Space, (ii) Connectivity, and (iii) AI" ^[13].

Four companies, four ways of selling inference: an API, a cloud marketplace, a cloud segment, a reporting segment. Meta has none of them. Its filing says instead: "We generate substantially all of our revenue from advertising" ^[1].

That is the comparison that actually holds. Not which model is best — which we will come to — but that four of the five convert model quality into a per-token price, and the fifth converts it into a better ad auction.

What "catching up" would have to mean

The question assumes one race. The record describes two.

If the race is selling inference, Meta is barely entered: a public-preview API ^[8] against three rivals on every major cloud ^[9][10][11] and a fourth consolidated into a listed filer ^[13]. On that board Meta is behind, and nothing in the cited record suggests it is trying very hard to be ahead.

If the race is converting compute into profit, the scoreboard reverses. Meta's ad engine already does what the others are still building a business model for — and it does it at a scale their disclosures cannot match, because Alphabet concedes that cloud margins run below advertising margins ^[11] and Meta's advertising is the margin.

There is a third possibility the record permits and does not settle: that these converge, and whoever holds the best model eventually holds the best ad auction too. Meta is buying the option on that outcome with $72.22 billion a year ^[1]. The others are buying it with a price list.

Why he is betting, in the company's own words

The filing does not leave the motive to inference. Meta writes that it is "working to develop the next generation of AI models and advance our vision to build superintelligence, which we define as AI that surpasses human intelligence", and then, in the same breath, concedes the timing: "Although it is inherently difficult to predict when superintelligence may be achieved, we are investing now because we believe this has the potential to begin a new era of individual empowerment" ^[1].

Read that as a capital-allocation sentence rather than a mission statement. It says the payoff is undated, the company knows it is undated, and it is spending anyway. That is not a forecast. It is an option being bought, and the premium is the $72.22 billion.

What the company will not claim

The risk factors are where a filing stops selling. Meta's carries a heading most peers would soften — "We may not be successful in our artificial intelligence initiatives, which could adversely affect our business, reputation, or financial results" — and the text beneath withholds the usual reassurance: "there can be no assurance that the usage of AI will enhance our products or services or be beneficial to our business, including our efficiency or profitability" ^[1]. On cost it offers only that "we expect our AI initiatives will require increased investment in infrastructure and headcount" ^[1].

An investor asking whether Meta can catch up should notice that Meta has not claimed it will.

Where the AI already pays

The part of the bet already earning is not the frontier model. It is the ad auction.

Meta reports that "Ad impressions delivered across our Family of Apps increased by 12% year-over-year in 2025", while "In 2025, the average price per ad increased by 9%, as compared with an increase of 10% in 2024, year-over-year" ^[1]. Price and volume rising together is the signature of demand, not of a squeeze, and the filing attributes the price rise to "an increase in advertising demand" ^[1]. Underneath sits an audience still growing: "Family daily active people (DAP) was 3.58 billion on average for December 2025, an increase of 7% year-over-year" ^[1].

This is the asymmetry the phrase "catch up" obscures. Meta does not need to win the model race to fund it. It needs its recommendation and ad-ranking systems to keep converting compute into price per ad, and on the 2025 record they did.

So is it misunderstood?

The record supports a narrower claim than the question invites.

What is verifiable is that the market repriced Meta downward through the heaviest spending year — from $790.00 in August 2025 to a low of $525.72 in March 2026 ^[2] — while revenue grew 22% and daily users 7% ^[1]. However that is read, "the market rewarded the AI build" is not what happened. The rebound to $673.31 came later, and this record does not say what caused it.

Also verifiable: the risks are not all technological. Meta settled teen-safety litigation for $18 billion in August 2026 ^[14], a figure of the same order as a quarter of its annual capital budget.

Will Meta win, and can this record tell you?

No, and the reason is worth more than a guess would be.

There is no neutral scoreboard. Every head-to-head number in the sources above is a vendor scoring its rivals on its own harness, and each says so in its own footnotes — OpenAI publishes Claude columns ^[9], Anthropic publishes a GPT column ^[10]. Both caution that evaluation setups differ between models. A ranking assembled from those tables would be a collage of marketing, and this page does not run one.

Nor is there a like-for-like revenue comparison. Alphabet and SpaceX file audited segment figures ^[11][13]; OpenAI and Anthropic publish none. Four numbers in one column would need two of them invented.

And on Meta's own side the record is thinner than its rivals'. Its frontier results are published mostly as charts rather than prose, and it has published no head-to-head figure against any of the four.

So: nothing cited here dates the payoff of the AI spending, and Meta explicitly declines to ^[1]. Nothing here says whether the recovery from March continues — the price record ends at 16 September 2026 ^[2], and a past close is not a forecast.

What the record does support is a sharper question than who wins. Every rival has to sell inference at a price that covers what the inference costs. Meta has to keep an ad auction compounding faster than a capital budget that has already outrun its net income ^[1]. Those are different failure modes, and the next filing tests Meta's — not the leaderboard.