Meta

Meta Platforms (NASDAQ: META) sells advertising against the daily attention of more than three billion people who use Facebook, Instagram, WhatsApp, Messenger, and Threads. That business now funds the largest infrastructure bet in the company’s history: Meta is trying to build a superintelligence lab and its own frontier AI models in-house, rather than resell someone else’s.

TEXXR’s archive shows the coverage pace has broken from its own pattern. Quarterly volume ran between 125 and 171 articles for two full years, 2024Q1 through 2025Q4, an average of about 151. It then rose to 255 in the first quarter of 2026, 275 in the second and 334 in the third.

This is DeadRisk’s mag7 tier, and the point isn’t to call the stock. Benjamin Graham’s test was whether a price is justified by facts already on the table, and Meta’s own spending guidance — set in January, its floor raised three times since — is the sharpest test case the AI-capex record has produced yet.

The company changed its name from Facebook to Meta in October 2021, betting that the metaverse was the next platform worth owning outright rather than renting from Apple or Google. The bet has been expensive: by March 2026 Reality Labs had lost more than $80 billion, though much of that went on hardware rather than virtual worlds. The pivot started with cost cuts. In February 2023 Mark Zuckerberg named 2023 the “Year of Efficiency”; asked on the same call about generative AI, he sounded cautious, wanting “to be careful not to get too ahead of the development of it.”

What followed was an open-model strategy Meta later reversed. Zuckerberg argued in July 2024 that “open source AI is the path forward,” warning that closed models were open to vendor lock-in and state-backed espionage — a framing TEXXR’s graph shows commentators disputing at the time, given Llama’s licensing limits on large commercial users. The strategy then frayed from both ends. Inside FAIR, Meta’s research lab, a former employee told Fortune the unit was “dying a slow death” as Joelle Pineau, who led it, announced her exit in April 2025. The same month a non-public Llama 4 Maverick version ranked second on LMArena’s leaderboard, and LMArena said it would update its rules. Two months later, Zuckerberg announced Meta Superintelligence Labs, and by July 2025 he was walking back the open-source pledge, saying superintelligence would raise “novel safety concerns” and Meta must be “careful about what we choose to open source.” The Information’s sources said Meta planned its fourth AI restructuring in six months; it later cut roughly 600 lab jobs, some in FAIR, and in December CNBC’s sources said Llama’s successor, codenamed Avocado, could be a proprietary model. By July 2026 Meta was selling access: it launched a Meta Model API, which Zuckerberg said would cost about 25% of OpenAI’s and Anthropic’s models — a metered product, not a free download.

The spending behind all of it is the story DeadRisk’s Graham page uses as its central example. On 28 January 2026 Meta guided 2026 capex to $115–135 billion, above the Street’s $110.6 billion estimate and up from $72.2 billion in 2025; the stock rose 10.4% the next day. On 29 April it reported first-quarter revenue up 33% to $56.31 billion and net income up 61%, helped by an $8.03 billion one-time tax benefit, and raised the range to $125–145 billion. The stock fell 8.6% the next day. The quarter before had shown how far the earnings line can swing under strong revenue: Meta’s Q3 2025 net income fell 83% on a one-time $15.93 billion tax charge, even as revenue grew 26%. In May 2026 the cost side reached headcount: Meta began laying off 8,000 people, or 10% of staff, and moved 7,000 more to AI work.

Update, 27 October 2026. The floor kept rising. Meta narrowed the range to $130–145 billion on 29 July and to $135–145 billion on 21 October; the stock fell 8.0% and 7.0% the day after each. The third quarter carried a $9.89 billion charge for a settlement with 52 state attorneys general over teen mental health. The same release created a non-voting Class C stock, to be paid as a one-for-one dividend once it takes effect in January 2027, which Meta said it could use to raise capital, pay staff and fund acquisitions. Five days later Meta launched a $25 billion share sale at $625 and $12.5 billion of mandatory convertible preferred, and said it plans a $60 billion at-the-market program in Class C shares. Nvidia, AMD and Broadcom indicated an interest in buying up to $5 billion of the shares, which the prospectus says is not binding.

TEXXR’s edge-arc tool measures how the kind of thing said about a company changes, not just how much. Between 2024Q1 and 2025Q4, regulatory edges fell from 27% to 17% of Meta’s total and personnel edges rose from 6% to 15%: a company once defending its privacy record, now hiring and reorganizing for AI. Read 2026Q1 with care. The graph logged 1,046 Meta edges that quarter against 192 the quarter before, a fivefold-plus jump Nvidia, Microsoft and Amazon show too, and a “competitor” label that had barely existed tagged 18 of the 130 it labelled, or 14%. That quarter’s shift, drift spike included, mostly records a change in how TEXXR labels coverage.

The record also shows deals running both ways. Meta agreed to buy Manus, a Singapore-based agent startup, in December 2025. By July the FT’s sources said Manus’s investors were discussing unwinding the $2 billion buyout, and in August Manus said it would soon operate independently again, after an order from China in April. On compute, the record holds two different deals. On 2 July SemiAnalysis said Meta was “in final talks with Anthropic” for private instances of Claude, with Meta as buyer. On 17 July the New York Times’ sources said Meta was in talks to rent compute to Anthropic, a deal that could be worth about $10 billion over two years, with Meta as seller. TEXXR still logs that as rumored, with nothing new since 18 July.

The financing underneath is the same shape Nvidia and CoreWeave show from the supply side, run here from the demand side. Meta formed a joint venture with Blue Owl to fund its $27 billion, 2-gigawatt Hyperion data center in Louisiana, its largest private capital deal, keeping about 20% of the equity — a structure that keeps the debt off its balance sheet, one of several such vehicles tech companies now use, and the kind of leverage Ray Dalio’s debt-cycle framework treats as a normal, recurring stage. In July 2026 Meta said it would spend another $40 billion on that campus, pushing its total there past $250 billion. Meta also buys from the neoclouds: it signed a deal worth up to $14.2 billion with CoreWeave in September 2025, then committed another $21 billion in April 2026, three weeks before it raised its capex range.

Positioning: how much of META is sold short, on the record brokers file twice a month. It lags by design — read it as where the crowd stood, not where it stands.

Signals in the coverage record worth tracking, not trade ideas:

• The first 2027 number. In October Meta said it was still setting next year’s budget and expects “to continue investing aggressively.” It gave its first 2026 range with fourth-quarter results in January. The stock fell the day after four of the last five reports; whether the 2027 range breaks that pattern is the cleanest read on how much patience is left.
• The Class C program. The non-voting shares take effect in January 2027, and the $60 billion at-the-market plan waits on them. Filings will show how much of it Meta sells, and how fast.
• Whether the new labels hold. Competitor edges went from almost nothing to 14% of Meta’s labelled edges in 2026Q1, the quarter TEXXR’s labelling changed. Later quarters on the same labels will show whether that rivalry is in the coverage or only in the ledger.
• The Anthropic compute-rental talks. Still rumored. A confirmation, at whatever price, would show Meta renting capacity to a rival lab rather than building only for itself.
• Personnel churn inside Superintelligence Labs. Four AI restructurings in six months and about 600 lab job cuts are already logged, and headcount fell 15% in the year to September. More reorganization edges, or a stretch without one, says something about whether the lab has found its shape.

Meta lives or dies by the same constraint as these names — small equal-weighted groups, one constraint each. Not investment products: see all categories.