Anthropic Has Committed $900 Billion to Compute. Revenue Is Not Locked In.
Anthropic has committed roughly $900 billion to compute and data centers, with $71 billion due within twelve months, against Q3 revenue of $17 billion and token prices down nearly 60% since May, per Nvidia.

Why it matters
- Anthropic has committed roughly $900 billion to future purchases and leases, per Bryan Rich's analysis, with about $71 billion due within twelve months.
- Of $18.8 billion spent on compute in the first nine months, only $4.8 billion served customers; the rest funded internal research and model development.
- Third-quarter revenue was $17 billion against $9.9 billion in compute costs.
- Nvidia reports token prices have dropped nearly 60% since May, while rental prices for even older chips are rising.
Anthropic has committed roughly $900 billion to future purchases and leases, much of it tied to computing and data centers, according to analysis by macro investor Bryan Rich — and about $71 billion of that falls due over the next twelve months. The figure anchors a structural tension at the heart of the AI infrastructure boom: the computing bill is locked in well into the future, while the revenue that must service it is not.
Rich, founder of Logic Fund Management and author of the Pro Perspectives daily briefing read by more than 25,000 investors, laid out the arithmetic in a note published on Forbes. His central observation is blunt: the commitments are contractual, and the demand that justifies them is not.
Where will Anthropic's IPO money actually go?
Not to customers, and not to shareholders. According to Rich, much of the capital Anthropic raised in its IPO will flow directly to chipmakers — to secure more compute.
That reversal carries market consequences. Rich notes that investors may have sold the big chipmakers to buy Anthropic in the IPO. But the money Anthropic raised then moves back upstream, to the very chip suppliers investors rotated out of. As Rich puts it: "That's the boom."
The scale of Anthropic's internal consumption, relative to customer workloads, is the more striking datapoint. Per the figures Rich cites:
- Of $18.8 billion Anthropic spent on compute in the first nine months, only $4.8 billion served customers.
- The remainder — roughly $14 billion — supported internal research and model development.
- Anthropic's own compute, in Rich's words, "is working around the clock, training and testing the next generation of models in the race to achieve superintelligence."
In other words, customers are not the biggest buyers of Anthropic's computing capacity. Anthropic is. Nearly three-quarters of the company's compute spend in the period covered internal model training and testing rather than revenue-bearing inference for clients — a ratio that reframes what kind of company Anthropic is at this stage: a research engine that also sells access, rather than a services business that also does research.
How big is the gap between commitments and revenue?
The core numbers, all attributed to Rich's analysis of Anthropic's filings:
- Committed future purchases and leases: roughly $900 billion, much of it compute- and data-center-related.
- Due within twelve months: about $71 billion.
- Third-quarter revenue: $17 billion.
- Third-quarter compute costs: $9.9 billion.
The Q3 figures mean Anthropic currently earns roughly $1.7 of revenue for every dollar of compute cost — a positive gross spread on compute, but one that must persist and grow at scale to absorb a $71 billion near-term obligation. And that obligation sits within a total commitment roughly fifty times the company's quarterly revenue run rate.
Rich returns to the language of Anthropic's own filings to underline the point. As discussed in his October analysis, Anthropic's first filing stated that it operates in a "compute-constrained environment where customer demand exceeds available compute." The latest filing, Rich indicates, continues in the same vein. Demand exceeding supply is, on its face, favorable. But it also means the company's growth is gated by infrastructure it has promised to buy at fixed prices, on fixed schedules, against revenue that arrives only if usage materializes.
The asymmetry is the story. Compute commitments are contractual and locked in well into the future. Revenue is usage-based, price-sensitive, and contingent.
What must Anthropic overcome to sustain this spending?
Rich's answer is a single condition: usage must grow faster than AI prices fall.
The price side of that equation is moving fast against incumbent economics. Citing Nvidia, Rich reports that token prices (the per-unit pricing of AI model usage) have dropped nearly 60% since May. Falling per-token prices are a deliberate feature of the AI ecosystem — cheaper inference drives adoption — but they compress revenue per unit of usage, which means volume has to outrun deflation for total revenue to climb.
The compute cost side is moving the other way. Rich notes that rental prices for even Nvidia's older chips are rising. That combination — falling output prices, rising input prices — is the squeeze. Anthropic has locked in its input costs through its $900 billion commitment structure while its output prices deflate at roughly 60% in under a year, at least on Nvidia's reported token-price trajectory.
If usage growth outruns price decline, the fixed compute base becomes an advantage: Anthropic owns capacity competitors must buy later at higher spot rates. If usage growth lags, the commitments become a fixed burden against shrinking unit revenue, with $71 billion due inside a year.
Why does this matter beyond one company?
Because the same dynamic now runs through the entire AI supply chain. Rich's framing connects three layers: investors rotate out of chipmakers into model developers like Anthropic; the model developers then spend the proceeds back into chips and data centers; and the whole cycle depends on end-demand growing fast enough to service the accumulated commitments.
The October filing language — demand exceeding available compute — is the bull case. The $900 billion commitment against unlocked future revenue is the bear case. Both are drawn from the same company's disclosures, and both are true simultaneously: Anthropic cannot buy compute fast enough today, and has promised to buy compute it cannot yet pay for from committed revenue tomorrow.
The $4.8 billion customer figure is the one to watch. When the portion of compute spend serving customers starts approaching, and then exceeding, the portion serving internal research, the business inverts from a research program subsidized by investors into a self-funding services company. Until then, every additional dollar of IPO capital that flows to chipmakers is a bet that superintelligence-adjacent model breakthroughs arrive before the deflation in token prices catches up with the fixed-cost base.
The near-term test is the twelve-month window: $71 billion in obligations against a quarterly revenue base of $17 billion. Whether usage outruns a 60%-since-May decline in token prices will determine whether Anthropic's compute commitments read as foresight or as leverage.
Source: seekingalpha.com
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
791 articles
Related articles
- Anthropic IPO Buyers Face a $900 Billion Compute Bill, Cramer Warns
- Anthropic's IPO: $26 Billion From Tech Giants and $900 Billion in Bills
- Anthropic’s IPO Filing Shows $16B Revenue and $80.7M Loss
- Anthropic IPO Valuation Slides From $2.5 Trillion Talk to $1 Trillion
- Anthropic Warns of 'Catastrophic' AI Risks in Its Own IPO Filing