The biggest tech companies are now fighting over a very small group of people. Only a small number of researchers have ever trained the most capable AI models. The firms that want them are offering huge pay deals, buying startups and building data centers on a scale that would have sounded absurd a few years ago.
If you own shares in one of those firms, the news is easy to follow and the cost is hard to find. Headlines name the hires and the new models, while the money they eat up is spread across several lines of the accounts, mixed in with everything else the company spends.
The way through is to follow what the company itself tells its owners. Each quarter it sets a range for its total costs and for what it will spend on buildings and machines. On the earnings call it explains what is pushing those numbers up, and read in order, those notes show where a new project lands in the books.
Meta Platforms set up exactly such a project in 2025. On BloFin, the METAX/USDT Spot pair lets you follow its effect on the company, since each METAX token is backed by Meta stock. You can also use the METAUSDT Perpetual, a contract for betting on a rise or a fall in Meta's price.
What is Meta Superintelligence Labs?
Meta Superintelligence Labs is the group that runs all of Meta's AI work. Meta set it up in June 2025 under Alexandr Wang, who ran Scale AI before he became Meta's chief AI officer (source: TechCrunch). The lab builds the Muse family of AI models, starting with Muse Spark in April 2026, and the products that use them (source: Fortune).
Mark Zuckerberg described the new group on Meta's earnings call of July 30, 2025. It took in Meta's existing AI teams, including FAIR, its long-running research lab, and added a new team to build the next generation of models. He named Alexandr Wang to lead the whole group, Nat Friedman to run AI products and applied research, and Shengjia Zhao as its chief scientist (source: Meta Q2 2025 earnings call).
In August 2025 Meta split the lab into four groups. TBD Lab builds Meta's largest new models, while the other three cover research, putting AI into Meta's apps, and the computers the models run on (source: TechCrunch on the reorganization).
Muse Spark now runs Meta AI, the assistant inside Facebook, Instagram, WhatsApp and Messenger (source: Meta Q2 2026 earnings call). So the lab shapes both the apps billions of people use and a large share of Meta's spending, and Meta's AI capex and cash flow show the full size of that spending.
Why Meta set up a new AI lab in 2025
Meta set up the lab to chase what Zuckerberg calls personal superintelligence: an AI that each person can direct towards what they value in their own lives. He set out the idea on July 30, 2025, telling investors that Meta's vision is to bring personal superintelligence to everyone (source: Meta Q2 2025 earnings call).
Before the lab, Meta's best-known AI work was Llama, a family of models it gave away so anyone could download and run them. The lab moved Meta toward models it keeps to itself. Muse Spark became a closed model, which outside developers reach through Meta's own paid access point, known as an API (source: Meta Q2 2026 earnings call).
Meta has since gone back to open models in part. On August 10, 2026 it released Muse Glimmer, a smaller model that anyone can download and run on a home computer. Meta trained it on data made by its closed Muse Spark models (source: SiliconANGLE).
Zuckerberg explained the mix on the July 29, 2026 call. He said Meta has always planned to give some models away and keep others closed, and that he wanted the lab free to build the smartest models it could (source: Meta Q2 2026 earnings call). For an investor, the shift matters because a closed model can be sold, while an open one mostly helps the apps that already earn Meta its money. The products built on the closed models, from Meta AI to the Muse agent and its paid plans, are where that selling starts.
The size of Meta's AI bet in October 2026
Fortune published a profile of Alexandr Wang on October 17, 2026 that put a rough size on the bet. Recalling talks late in 2025, Zuckerberg said on a podcast that Meta was going to "write this check for like, I don't know, like $100 billion or $200 billion". Fortune called the figure off the cuff (source: Fortune).
The same profile carries a warning, as the lab's chief remembers it. He recalled that Susan Li, Meta's chief financial officer, took him aside at a meeting of top staff in Hawaii. She told him, "if this doesn't work, this is going to be really, really bad." That is his account of a private talk, and Fortune presents it that way (source: Fortune).
The profile also gives a number that explains why the cost runs so far ahead of the head count. By the lab chief's count, its core research team has fewer than 200 people, nearer 100. Much larger teams build the computers and the products around it, inside a company of about 75,000 people (source: Fortune).
The profile came out days after the lab released Muse Wonder, the first model Meta calls frontier-class, a label for models at the leading edge of the field (source: Fortune). A small team building costly models is the pattern behind every cost line that follows.
What Meta Superintelligence Labs costs in talent
Meta does not report a separate budget for the lab, so its talent cost shows up in two places: the Scale AI deal and pay given in stock. The biggest single item is Scale AI. Meta booked $13.8 billion of shares with no voting rights in Scale AI in June 2025 (source: Meta 10-Q for the second quarter of 2025).
Fortune puts the deal at $14.3 billion for 49% of Scale AI (source: Fortune). Meta's filing counts $13.8 billion of that as a stake in a private company it does not steer. Stakes like this sit in a line Meta calls non-marketable equity investments. It grew from $6.1 billion at the end of 2024 to $22.0 billion on June 30, 2025, the month of the deal (source: Meta 10-Q for the second quarter of 2025).
Pay is the second place. Meta's accounts show it through share-based pay, the part of pay given in stock. It rose 22.4% in 2025 to $20.4 billion (source: Meta Q4 2025 results). The rise sped up in 2026, as the June quarter shows.
Share-based pay per employee equals the quarter's share-based pay divided by head count at the end of the quarter.
Quarter | Share-based pay ($ million) | Head count at quarter end | Share-based pay per employee |
|---|---|---|---|
Q2 2025 | 4,834 | 75,945 | about $63,700 |
Q2 2026 | 7,658 | 75,472 | about $101,500 |
Run that sum on the two June quarters and the shift is plain. Meta's staff was almost the same size, down 0.6%, yet share-based pay per person rose 59.4% (source: Meta Q2 2026 results; Meta 10-Q for the second quarter of 2025). The June 2026 head count still includes about 8,000 people affected by the May 2026 job cuts. Meta said in July that most of them would drop out of its head count by the end of September (source: Meta Q2 2026 results).
Susan Li named the cause on the July 29, 2026 call. Leaving out severance, she said, the growth in pay came from technical hires made over the past year, especially in AI (source: Meta Q2 2026 earnings call).
Where the lab shows in Meta's expense and capex guidance
Meta has named the lab, or the AI talent hired for it, in most of its outlooks since July 2025. The plan it set on July 29, 2026 put 2026 costs at $165 billion to $169 billion. It also planned $130 billion to $145 billion of capital spending, or capex, which pays for data centers and servers (source: Meta Q2 2026 results).
The trail starts in July 2025, a month after the lab was formed. Li told analysts that Meta was putting special weight on hiring the industry's top researchers for the lab. She said the pay for those hires would push 2026 costs up (source: Meta Q2 2025 earnings call).
Meta narrowed its 2025 cost range on the July 2025 call (source: Meta Q2 2025 earnings call) and raised the bottom of it on the October call (source: Meta Q3 2025 earnings call). It set the 2026 range with its January results (source: Meta Q4 2025 results) and raised the bottom of that range with its July results (source: Meta Q2 2026 results).
Release date | What Meta said about the lab or AI talent | Cost range it set |
|---|---|---|
July 30, 2025 | Pay for technical hires becomes the second-largest cause of 2026 cost growth | 2025 costs $114 billion to $118 billion (source: Meta Q2 2025 earnings call) |
October 29, 2025 | A full year of pay for 2025 hires, especially AI talent; the lab's computing needs grow fastest | 2025 costs $116 billion to $118 billion (source: Meta Q3 2025 earnings call) |
January 28, 2026 | 2026 capex growth driven by "our Meta Superintelligence Labs efforts and core business" | 2026 costs $162 billion to $169 billion (source: Meta Q4 2025 results) |
July 29, 2026 | Pay growth from technical hires, especially in AI; fees for outside AI models added | 2026 costs $165 billion to $169 billion (source: Meta Q2 2026 results) |
The October 2025 call matters most for capex. Li said 2026 capex would grow across the lab, Meta's other AI work and its non-AI spending, but that "the MSL AI needs are growing the most" (source: Meta Q3 2025 earnings call). MSL is Meta's short name for the lab. In January, the written outlook named the lab itself (source: Meta Q4 2025 results).
To see the size of the step, take the middle of the July 2026 cost range. $167 billion is $49.3 billion, or 41.9%, more than the $117.7 billion Meta spent in 2025 (source: Meta Q4 2025 results). In 2025, total costs had grown 23.7%.
On the July 2026 call, Li named a new cost: third-party AI token costs, which are the fees Meta pays outside firms each time it uses their AI models. They sit beside pay, data centers and legal costs (source: Meta Q2 2026 earnings call). Much of the computing power behind these numbers comes from hardware Meta buys or designs, including its own AI chips. The rest comes from capacity it rents from cloud providers.
How Meta stock moved after each guidance update
Meta's share price moved between 8.0% and 11.3% on the next trading day after each of its five outlooks from July 2025 to July 2026. It rose after two and fell after three. Each release also carried sales, profit and many other figures, so these moves sit beside the lab's mentions without showing that the lab caused them (source: Yahoo Finance META history).
Release date | Next trading day | Change in Meta's closing price from the day before |
|---|---|---|
July 30, 2025 | July 31, 2025 | +11.3% |
October 29, 2025 | October 30, 2025 | −11.3% |
January 28, 2026 | January 29, 2026 | +10.4% |
April 29, 2026 | April 30, 2026 | −8.6% |
July 29, 2026 | July 30, 2026 | −8.0% |
Meta publishes after the Nasdaq close at 4PM ET (8PM UTC in summer, 9PM UTC in winter), so the next day's session is the first full session to price each update.
To watch how the market takes Meta's next outlook, the METAX/USDT Spot page carries the token's live price. The METAUSDT Perpetual page gives the perpetual's current price along with its funding rate.
Meta's AI acquisitions and the Manus deal
Meta has also bought AI startups, which is faster than hiring one by one. One of them, Manus, a startup that builds AI agents, cost roughly $2 billion and was meant to bring its agent technology into Meta AI. China ordered the deal undone in April 2026 (source: TechCrunch on the Manus order).
Date | Company | What it does | Where the team went, or what happened |
|---|---|---|---|
December 2025 | Manus | AI agents | Ordered undone by China on April 27, 2026 |
March 2026 | Moltbook | A social network for AI agents | Meta Superintelligence Labs |
May 2026 | Assured Robot Intelligence | AI for human-shaped robots | Meta Superintelligence Labs |
A Meta spokesperson confirmed that Moltbook was joining the lab, and TechCrunch reported that its founders were coming with it (source: TechCrunch on Moltbook). The founders of Assured Robot Intelligence moved into the lab as well (source: TechCrunch on Assured Robot Intelligence).
Manus shows the risk in buying across borders. The company had moved its head office from China to Singapore before the sale. Even so, China's National Development and Reform Commission, its top economic planner, required both sides to withdraw the deal. Meta told TechCrunch that the deal "complied fully with applicable law" and that it expects the matter to be settled properly (source: TechCrunch on the Manus order).
Acqui-hire scrutiny of the Scale AI deal
An acqui-hire is a deal in which a big company hires a startup's leaders and takes a stake or a license, stopping short of buying the whole firm. In early 2026 the US competition regulator said it was studying such deals, and three senators asked regulators to look into Meta's Scale AI deal by name.
In January 2026, Andrew Ferguson, chair of the Federal Trade Commission (FTC), announced that the agency would look into acqui-hires by big tech firms (source: Senator Warren's office).
On February 4, 2026, Senators Elizabeth Warren, Ron Wyden and Richard Blumenthal asked the FTC and the Justice Department to look into three such deals, starting with Meta and Scale AI. They argued that the deals work like mergers while skipping merger review, and asked the agencies to block or reverse them if they break the law on competition (source: Senator Warren's office).
A letter from senators asks the agencies to act, and only the agencies or a court could force a change. The risk for a shareholder is a later order to undo part of the deal, which is what Meta has already faced in China over Manus. Meta's separate court fight with the FTC over Instagram and WhatsApp is a different case.
AI rules in Europe and the US
Rules on AI models also shape how freely the lab can release its work. In July 2025, Meta chose not to sign the European Union (EU) code of practice for general-purpose AI models, a voluntary guide to the EU's AI Act. Joel Kaplan, then Meta's global affairs chief, wrote that Europe was heading down the wrong path on AI (source: CNBC via NBC San Diego).
In the US, the push has been toward fewer state rules. A December 11, 2025 executive order told the Attorney General to set up an AI Litigation Task Force to fight state AI laws in court. It left child-safety laws outside the national plan it called for (source: The White House).
How Meta says its AI work pays off
Meta says its AI spending pays off in three ways: better ads in its apps, new products that people and businesses pay for, and selling access to its models or its computing power. Ads still bring in almost all of Meta's sales (source: Meta Q2 2026 earnings call).
The first route works through ranking. Better AI models pick which posts and ads each person sees more accurately. That keeps people in the apps longer and makes each ad worth more to the business buying it. Advertising made up $59.4 billion of the $60.8 billion Meta earned from April to June 2026 (source: Meta Q2 2026 results), and Meta's advertising business shows how better ranking reaches that line.
The second and third routes are newer and smaller. On the July 2026 call, Li listed AI agent tools, Meta's API and selling computing power directly among the things Meta could sell to businesses. Zuckerberg called personal AI agents the base for Meta's next products and sources of revenue (source: Meta Q2 2026 earnings call).
Li also drew a line between the two kinds of payoff. She said Meta's reach should let it serve useful AI products to its users and businesses "regardless of whether our models are on the frontier". Being at the front of the field, she added, would open new markets that may need even more computing power (source: Meta Q2 2026 earnings call).
Meta's AI lab and the October 21 report
If you hold METAX or trade the METAUSDT Perpetual, the lab matters most on report nights, when Meta updates the cost and capex ranges the lab helps drive. Meta set its results for the July to September quarter for after the US stock market shuts on Wednesday, October 21, 2026, and each release's outlook is the update to watch (source: Meta Investor Relations).
Four items carry most of the lab's news. The first is whether the $165 billion to $169 billion cost range moves, and the second is whether the capex range moves. The third is what Li says about pay and AI talent, and the fourth is any first word on 2027 spending. Meta's earnings explained shows where each one sits in the release.
Tokenized Meta (METAX) follows the share price and trades through the night on weekdays under its written schedule, which lets it react on the evening the report lands. METAX trading hours against Nasdaq sets out when each market is open. The METAUSDT Perpetual adds leverage to a bet on a rise or a fall, which also means a big move after the report can close a leveraged position fast. Outside the Nasdaq session the perpetual may accept only orders that reduce or close a position (source: BloFin stock futures notice).
Looking to trade Meta around its next AI spending update? To get started, you'll need to first create a BloFin account, fund your account with cryptocurrency, and go to the METAX/USDT Spot trading page or the METAUSDT Perpetual trading page.
Frequently asked questions
How much does Meta Superintelligence Labs pay?
Meta keeps the pay of each person in the lab private. The best-known figure is the $100 million that Meta reportedly offered some OpenAI researchers, a figure Fortune repeated in October 2026 as a report (source: Fortune). Meta's closest public figure is its total share-based pay, which came to $13.7 billion in the first half of 2026, against $9.0 billion a year earlier (source: Meta Q2 2026 results).
Who leads Meta Superintelligence Labs?
Alexandr Wang leads the lab as Meta's chief AI officer, and he co-founded Scale AI. Mark Zuckerberg started asking him for advice in April 2025. One of their calls ran about 44 minutes, by Zuckerberg's recollection, and two days later Wang sent him a long memo on what he would do in Zuckerberg's place. During the recruiting push that followed, Zuckerberg cleared his calendar for two months (source: Fortune). He told investors in July 2025 that he had spent a lot of his own time that quarter building the team (source: Meta Q2 2025 earnings call).
What is TBD Lab at Meta?
TBD Lab is the team Meta set up inside Meta Superintelligence Labs in August 2025 to build its largest new models (source: TechCrunch on the reorganization). When Meta cut roughly 600 jobs across the lab's research, product and computer groups in October 2025, TBD Lab was left out of the cuts and kept hiring (source: AP via Yahoo Finance).
What happened to Scale AI after Meta's investment?
Scale AI kept running as a separate company after Meta took its stake and hired its chief executive. Weeks after the June 2025 deal, Scale AI cut 200 full-time staff and ended work with 500 contractors, according to the three senators who later asked regulators to look at the deal (source: Senator Warren's office). Meta's stake is in shares with no voting rights, and its June 2025 filing says Meta has no significant influence over how Scale AI is run (source: Meta 10-Q for the second quarter of 2025).
What is the difference between Meta AI and Meta Superintelligence Labs?
Meta AI is the assistant people chat with inside Facebook, Instagram, WhatsApp and Messenger, while Meta Superintelligence Labs is the group that builds the models behind it. Since the lab's Muse Spark model took over Meta AI, Zuckerberg says the number of people using the assistant each day has risen 60% (source: Meta Q2 2026 earnings call).
Researched and written by the BloFin Academy editorial team with AI-assisted drafting. All facts independently verified. Primary sources include Meta's quarterly results and earnings call transcripts from July 2025 to July 2026, its 10-Q filing for the second quarter of 2025 and its notice of the third-quarter 2026 report, with share prices from Yahoo Finance. News and policy sources include Fortune, TechCrunch, SiliconANGLE, CNBC, the Associated Press, Senator Warren's office, the White House and BloFin's stock futures notice, current as of October 21, 2026.
Nothing in this article constitutes financial advice, and nothing in it is a recommendation to buy, sell or hold Meta in any form. Meta's AI lab has raised its costs well before reporting any revenue of its own, and the share price has moved roughly 8% to 11% in a day after each recent outlook. A holder of the stock or of METAX can lose money if spending rises again or the lab's models disappoint. A leveraged position on the METAUSDT Perpetual held through a results report can be liquidated by one sharp move, losing the margin behind it. Past performance does not indicate future results. Do your own research and consider your risk tolerance before you trade on BloFin.