Meta Platforms has internally projected that it could spend as much as $10 billion a year on Anthropic’s AI models and services, according to a report from The New York Times published August 27-28, 2026. The figure is described as an internal planning number rather than a signed contract, but it lands at a pointed moment: Anthropic is preparing an IPO that Reuters has reported could rank among the largest public listings on record, and any sign that a top customer is trimming its AI budget carries weight with the bankers drafting that prospectus.
Neither company has confirmed the $10 billion figure on the record. Meta has not issued a public statement quantifying its Anthropic spend on its newsroom, and Anthropic has not addressed the number directly on its own news page. What is confirmed, across multiple outlets, is that Meta has become one of Anthropic’s largest customers, reportedly paying hundreds of millions of dollars a month for Claude access, and that the two companies are separately negotiating a deal for Anthropic to lease AI data center capacity from Meta. That compute-leasing talk, first reported by The New York Times and covered by Reuters, Bloomberg, and CNN in mid-July 2026, was pegged at up to $10 billion over two years, a number CNN’s own reporting flagged as speculative and sourced to people familiar with early-stage discussions.
Two different $10 billion figures, two different deals, two different directions of cash flow, all wrapped into headlines that read like a single mega-deal. That confusion is itself part of the story, and it matters because Anthropic’s IPO math depends heavily on how durable its largest customer relationships look to public investors.
What Meta actually projected, and what it didn’t confirm
The core of the story, per The New York Times, is an internal Meta projection made earlier in 2026 estimating that the company could spend up to $10 billion annually on Anthropic’s models. Internal projections are not budgets, and they are not contracts. Companies model best-case, worst-case, and everything-in-between scenarios for major vendor relationships constantly, and a chunk of that modeling never survives contact with an actual finance committee. Meta’s own public disclosures don’t break out AI-vendor spend by counterparty, so there is no SEC filing or earnings-call remark to check the number against.
What is independently reported, and treated as more solid, is that Meta already ranks among Anthropic’s largest customers today, paying on the order of hundreds of millions of dollars per month for access to Claude models across internal tools and products. That’s a run-rate in the low billions annually just from what’s already flowing, before any projection about where it could go. The gap between “what Meta pays today” and “what Meta modeled it might pay at the high end” is exactly where the cost-cutting angle comes in: if Meta pulls back toward the lower end of its own internal range, or flattens spend rather than scaling it toward $10 billion, that reads publicly as a pullback even though nothing was ever formally committed.
Meta has its own reasons to want flexibility here. The company has spent 2025 and 2026 pouring capital into Llama development, in-house inference infrastructure, and its Superintelligence Labs unit, all while facing investor pressure over AI capex generally. Leaning on a third-party frontier model vendor for a meaningful share of internal AI workloads is useful for speed, but it is also a line item that finance teams revisit every time internal alternatives get better. Any sign of tightening at Meta, even directional, feeds a narrative that AI-services demand from hyperscalers is less locked-in than markets have priced.
The Meta-Anthropic compute lease: a separate, still-unsettled negotiation
Running in parallel to the spend-projection story is a second, distinct negotiation: Anthropic reportedly proposed in June 2026 that it lease AI data center computing capacity directly from Meta. Reuters broke that story on July 17, 2026, reporting the arrangement could be worth up to $10 billion over two years, with monthly payments and an early-exit clause under discussion. Bloomberg and CNN followed within hours, both describing the talks as early-stage. CNN was explicit that the $10 billion figure, first reported by The New York Times, was speculative and not a locked number. Chosunbiz and other regional outlets picked up the Reuters framing on July 18-19.
This is, in effect, the mirror image of the spend story. In the compute-lease scenario, money would flow from Anthropic to Meta, not the other way around, as Anthropic looks to diversify its infrastructure supply beyond Amazon, Google, and its other cloud partners. Anthropic has publicly leaned on a multi-cloud strategy for training and inference capacity, and leasing capacity from Meta, which has built out enormous data center infrastructure for its own Llama and recommendation-system workloads, would give it another source of GPU-hours without waiting on new hyperscaler contracts.
Reuters reported on July 28, 2026, in a separate piece about Meta’s broader infrastructure strategy, that the compute-lease talks were still being referenced in the press as ongoing, without a signed agreement. As of August 30, 2026, no outlet has reported that the lease deal has closed, and both the spend-projection number and the lease-value number remain, in the language multiple outlets have used, under discussion rather than finalized.
| Reported figure | Direction of payment | Status as of Aug 30, 2026 | First reported by |
|---|---|---|---|
| Up to $10B/year internal projection | Meta pays Anthropic | Internal estimate, not a contract | The New York Times (Aug 27-28, 2026) |
| Hundreds of millions/month, current spend | Meta pays Anthropic | Reported as already occurring | The New York Times, via multiple wire reports |
| Up to $10B over two years, compute lease | Anthropic pays Meta | Early-stage talks, no signed deal | Reuters (July 17, 2026) |
| $65B annual revenue run rate (end of July 2026) | N/A – Anthropic revenue | Reported by a source familiar with the matter | Reuters (Aug 17, 2026) |
| $190B-$200B 2028 revenue projection | N/A – Anthropic forecast | Pre-IPO investor materials, per sourcing | Reuters (Aug 14, 2026) |
Why the timing lands hard on Anthropic’s IPO
Anthropic has been moving toward a public listing for months. Reuters reported on August 17, 2026, that the company’s annual revenue run rate had crossed $65 billion by the end of July, up from a fraction of that a year earlier, and that Anthropic was preparing for what could be one of the largest IPOs on record. A separate Reuters report on August 14 cited a 2028 revenue projection in the $190 billion to $200 billion range as part of the pitch to prospective investors. CNBC reported on August 21 that Anthropic’s IPO filing would list AI backlash as a risk factor, with some investors reportedly floating a valuation near $2 trillion. Reuters also reported, citing the Wall Street Journal, that Anthropic plans to tell investors it sees more than $30 trillion in total addressable market across its target segments.
Reuters reported the prospectus itself is expected to become public after Labor Day, which falls on September 7, 2026, putting the roadshow squarely in the weeks immediately following this Meta spend story. That sequencing is the reason the story matters beyond a single vendor relationship. When a company the size of Anthropic is walking investors through customer concentration and growth durability, any signal that a top-tier customer might flatten or cut spend becomes a data point analysts will ask about directly, regardless of whether the $10 billion figure was ever more than an internal model.
Anthropic has not disclosed customer-level revenue concentration publicly, so it’s not possible to say what percentage of Anthropic’s roughly $65 billion run rate Meta represents. If Meta’s current spend is genuinely in the hundreds of millions per month, that alone could put Meta among Anthropic’s largest handful of customers, given that Anthropic’s overall enterprise and API business, separate from consumer Claude subscriptions, is the segment growing fastest according to the company’s own investor materials as reported by Reuters.
Historical context: how we got here
Meta and Anthropic’s relationship has evolved quickly. Meta built its own foundation models, the Llama family, and has invested heavily in home-grown infrastructure through 2025 and 2026, including the buildout tied to its Superintelligence Labs group. At the same time, Meta has continued to license and integrate outside frontier models, including Anthropic’s Claude family, for specific internal workloads such as coding assistance and select product features where Llama has lagged competitively on benchmarks.
That dual-track approach, build in-house while also paying a rival lab for access, is not unique to Meta. Amazon has poured tens of billions into Anthropic while also developing its own Trainium chips and Nova model family. Google is both an Anthropic investor and cloud supplier while running its own Gemini line. The pattern reflects how expensive and uncertain frontier-model development has become: even companies with enormous balance sheets are hedging by keeping a foot in multiple camps rather than betting everything on one internal roadmap.
Anthropic, for its part, has spent 2026 diversifying its own compute supply chain in the opposite direction, adding infrastructure partners beyond its original backers to avoid being locked into any single cloud provider’s pricing or capacity limits. The reported talks to lease capacity from Meta fit that broader pattern: Anthropic needs GPU-hours wherever it can secure them at workable terms, and Meta has data center capacity built for its own workloads that could, in theory, be sold or leased out when not fully utilized.
Competitive comparison: how Meta’s AI vendor bets stack up
Meta is not the only hyperscaler juggling an in-house model strategy against outside AI vendor spend, and the contrast is useful for sizing up how unusual, or how routine, this story actually is.
| Company | In-house model line | Reported outside AI vendor relationship | Reported annual scale |
|---|---|---|---|
| Meta | Llama family | Anthropic (Claude, for select workloads) | Up to $10B/year, per internal projection |
| Amazon | Nova family | Anthropic (primary strategic partner, AWS-hosted) | Multi-billion dollar equity and compute commitments, disclosed in prior funding rounds |
| Gemini family | Anthropic (investor and cloud supplier) | Multi-billion dollar investment and TPU/cloud compute deals, disclosed in prior rounds | |
| Microsoft | MAI model family | OpenAI (primary partner, Azure-hosted) | Multi-year, multi-billion dollar commercial agreement, publicly disclosed |
The Meta-Anthropic figure stands out mainly because it has not been confirmed by either party, whereas Microsoft’s OpenAI relationship and Amazon’s Anthropic investment have both been the subject of formal disclosures. That distinction matters for anyone trying to size the real financial exposure here: an internal projection surfaced through anonymous sourcing is a different category of fact than a disclosed commercial agreement, even when the headline dollar figures look similar.
Market impact: what moved, and what didn’t
Because Anthropic is not yet public, there is no stock ticker to watch move on this news the way there would be for a listed company. The more immediate market impact runs through Anthropic’s IPO preparation. CNBC’s August 21 report on investor valuation chatter, floating figures near $2 trillion, came before this specific Meta spend story broke, and it remains to be seen whether the NYT report shifts that conversation. Analysts covering the AI infrastructure space will likely treat any Meta pullback signal as a data point on customer concentration risk, one of the standard line items in IPO risk-factor sections.
For Meta, the market impact is more about capital allocation optics than stock price. Meta has faced repeated investor questions through 2025 and 2026 about the scale of its AI capital expenditure, split between its own data center buildout and payments to outside model providers. A projection showing Meta modeling flexibility, or cost discipline, around a major external AI vendor could actually read as a positive to investors focused on Meta’s overall AI spend trajectory, even as it reads as a negative signal for Anthropic’s growth story. The same fact can cut two ways depending on which company’s shareholders are reading the headline.
What a cost-cutting pivot at Meta would actually look like
None of the reporting so far specifies concrete cuts, only that the $10 billion figure represents an internal ceiling rather than a floor or a locked commitment. If Meta does pull back, the mechanics would likely show up gradually: renewing Claude access at a lower usage tier for internal tools, shifting more coding-assistant or content-moderation workloads to Llama variants as they close the capability gap, or simply not scaling Anthropic usage at the pace the original internal projection assumed. None of that requires a public announcement, which is part of why this story is built entirely on anonymous sourcing rather than a company statement from either side.
It’s also worth separating “cost-cutting” from “walking away.” Even a scaled-back Meta relationship with Anthropic at, say, half of the internally modeled $10 billion ceiling would still represent a substantial commercial relationship, and would still count toward whatever concentration figure eventually shows up in Anthropic’s IPO disclosures. The framing of a pivot matters more for sentiment than for the underlying dollar amounts, at least based on what’s been confirmed so far.
Anthropic’s broader IPO risk profile
CNBC’s August 21 report that Anthropic’s IPO filing will flag AI backlash as a risk factor points to a wider set of concerns beyond any single customer relationship. Regulatory scrutiny of AI companies, copyright litigation, and public skepticism about AI’s economic returns have all become standard risk-factor language across the sector, and Anthropic is not immune. Layered on top of that, customer concentration, the exposure to any one large buyer scaling back, is a classic pre-IPO worry for any company where revenue is dominated by enterprise and API contracts rather than a broad base of small customers.
Reuters’ report that Anthropic sees more than $30 trillion in total addressable market, per the Wall Street Journal’s sourcing, is the kind of expansive framing companies use to justify aggressive valuations ahead of a listing. Whether investors buy that framing will depend partly on whether stories like the Meta spend projection make Anthropic’s near-term growth look durable or wobbly. A $10 billion internal ceiling that never gets reached reads very differently from a $10 billion contract that gets signed and scaled.
What outlets agree on, and where the reporting diverges
Stripping away the headline dollar figures, there’s a consistent thread across Reuters, Bloomberg, CNN, and The New York Times: both the spend-projection story and the compute-lease story rely on unnamed sources describing internal or early-stage material, not on statements from named Meta or Anthropic executives. CNN was the most explicit in flagging the $10 billion compute-lease number as speculative. Chosunbiz’s July 19 coverage of the lease talks noted the negotiations were early enough that they might not produce a contract at all.
Where the reporting diverges is mostly in scope and emphasis. Some outlets folded the spend-projection story and the compute-lease story together in headlines, which is understandable given both involve the same two companies and the same round number, but it obscures that they describe cash flowing in opposite directions for entirely different services (AI model access versus physical data center capacity). Readers scanning headlines alone could easily come away thinking Meta and Anthropic struck one $20 billion mega-deal, when what’s actually been reported is two separate, unconfirmed, and differently-sourced negotiations.
Predictions: what to watch next
Based on the current state of reporting, here’s how this is likely to play out over the next several weeks:
- Anthropic’s IPO prospectus, expected after Labor Day per Reuters, will likely include customer concentration language that indirectly addresses relationships like the one with Meta, even without naming Meta specifically.
- Neither Meta nor Anthropic is likely to confirm or deny the exact $10 billion figure on the record before the IPO filing becomes public, since neither company has an incentive to quantify the relationship prematurely.
- The separate compute-lease negotiation is more likely to either quietly close at a smaller scope than the reported $10 billion ceiling, or stall entirely, given how early-stage multiple outlets described it in July.
- Expect continued scrutiny of Meta’s overall AI capex mix, in-house Llama and infrastructure spend versus external model licensing, on Meta’s next earnings call.
- If Anthropic’s IPO roadshow proceeds as expected, investor questions about customer concentration and the durability of enterprise AI spend commitments will likely reference this exact story as a case study, regardless of how the underlying numbers eventually resolve.
Why internal projections keep leaking as if they were deals
This isn’t the first time a tech company’s internal planning number has been reported as though it were a finalized commitment, and it won’t be the last. Large technology companies run scenario models constantly for vendor relationships, and those models get shared internally across finance, product, and infrastructure teams, any one of which can become a source for a reporter. The gap between a modeled ceiling and an actual signed contract is often lost by the time a number reaches a headline, especially when the number is round and large enough to be attention-grabbing on its own.
For readers trying to track the real state of the Meta-Anthropic relationship, the more reliable signal is the confirmed, already-occurring spend, reportedly in the hundreds of millions of dollars per month, rather than the projected ceiling. Ongoing spend that’s already happening is a fact pattern that can be checked against future earnings commentary or eventual IPO disclosures. A projection is, by definition, a guess about the future that may never materialize.
What this means for enterprise AI buyers watching from the outside
For engineering and procurement teams at other companies evaluating multi-vendor AI strategies, the Meta-Anthropic story is a useful case study regardless of the exact dollar figures. It shows a hyperscaler with massive internal model development resources still choosing to pay for outside frontier-model access on specific workloads, rather than assuming in-house models can cover every use case immediately. It also shows that even the largest AI buyers treat vendor spend as adjustable rather than fixed, revisiting allocations as internal alternatives mature.
That flexibility cuts both ways for AI labs like Anthropic, OpenAI, and others whose revenue increasingly depends on a small number of very large enterprise customers rather than a long tail of smaller accounts. The more concentrated that customer base, the more a single company’s internal budget review can move the growth narrative, which is precisely the dynamic playing out here in the run-up to Anthropic’s IPO.
For more on the AI infrastructure spending race, see how Nvidia’s reported $12.9B Hugging Face bid reflects the same land-grab dynamics, how Anthropic’s Claude has adapted to EU AI Act rules, why Nvidia has compressed its AI model release cycle to 4-6 weeks, how the Jalapeño chip project threatens Nvidia’s margins, and how HPE’s stock rally ties back to Nvidia server demand. For broader coverage of the AI sector, visit the AI & Machine Learning section.
Frequently asked questions
Has Meta confirmed it will spend $10 billion a year on Anthropic?
No. The $10 billion figure comes from an internal Meta projection reported by The New York Times on August 27-28, 2026. Neither Meta nor Anthropic has confirmed it as a signed commitment.
Is this the same as the Meta-Anthropic compute lease deal?
No. That’s a separate negotiation, first reported by Reuters on July 17, 2026, in which Anthropic would lease AI data center capacity from Meta for up to $10 billion over two years. In that deal, money would flow from Anthropic to Meta, the opposite direction from the AI-spend story.
How much is Meta currently paying Anthropic?
Multiple reports describe Meta as already paying hundreds of millions of dollars per month for Claude access, though neither company has disclosed an exact figure.
What is Anthropic’s current revenue?
Reuters reported on August 17, 2026, that Anthropic’s annual revenue run rate topped $65 billion by the end of July 2026, citing a person familiar with the matter.
When will Anthropic go public?
Reuters has reported the IPO prospectus is expected to become public after Labor Day (September 7, 2026), though no exact listing date has been confirmed.
Why would Meta cut spending on Anthropic if it’s already paying for Claude?
No confirmed cut has been reported. The story centers on Meta possibly not scaling spend toward the high end of its own internal projection, not on Meta reducing existing usage. Meta has also been investing heavily in its own Llama models and infrastructure, which gives it flexibility to shift workloads in-house over time.
Does this affect Anthropic’s valuation?
It could influence investor sentiment ahead of the IPO. CNBC reported on August 21, 2026, that some investors were discussing a valuation near $2 trillion, and customer concentration risk is a standard factor IPO analysts weigh when a company’s revenue depends heavily on a small number of large customers.
Is Meta an investor in Anthropic?
The reporting reviewed here describes Meta primarily as a customer of Anthropic’s AI models, not as a disclosed equity investor. Amazon and Google are Anthropic’s most prominently disclosed strategic investors.