Anthropic's IPO Filing Puts a $518 Billion Number on AI Inference Costs

The number that matters in Anthropic's filing is not the valuation, which Reuters says could top $2 trillion. It is the $518 billion the company plans to spend on cloud, computing and infrastructure obligations in coming years. Reuters reported the figure on 29 September, citing the prospectus, which it says it has seen. CNBC's write-up of that reporting summarises it.

The same document records a $42 billion net loss for 2025. Revenue grew 12-fold to nearly $4.6 billion. The operating loss, excluding writedowns tied to earlier fundraising, exceeded $8 billion. Anthropic spent $7.33 billion on compute and infrastructure last year, a threefold rise from 2024 and more than half of its $12.65 billion in total operating expenses.

The price of a token is not the price of the GPU

Anyone reading those numbers as a straightforward GPU bill should look at what the pricing trackers measured the same week. Meetrix.io checked on-demand list prices on 28 September. It found an H100 at $6.88 an hour on AWS against $10.98 on GCP, roughly 37% cheaper, with the AWS figure sourced from Vantage's EC2 tracker. The same analysis puts the cost per million tokens at $0.76 on a fully busy H100 and $3.06 at 25% utilisation, using a labelled placeholder throughput of 2,500 tokens per second. The ratio, a fourfold swing, does not depend on that assumption. No cloud switch delivers that.

GPUAdvisor says it tracks 14 providers and refreshed its table on 1 October. It lists per-GPU rates across AWS, GCP, Azure, Lambda, CoreWeave, RunPod, Nebius and others, and notes that GPU-specialist clouds offer two to four times lower per-GPU pricing than hyperscalers. Its headline finding is narrower than it sounds: pricing has stopped behaving like a lookup table. The comparison only holds if you check the provider page before you commit budget.

Anthropic is not shopping on price alone. Its prospectus says nearly a quarter of revenue came from two customers last year. It also warns that many of its largest clients are not locked into long-term contracts.

Routing, not procurement, is where the savings sit

Unblocked published its own numbers on 29 September. It described a router that moves traffic between Baseten, Fireworks and CoreWeave, all serving the same open-weight model. Under round-robin, Fireworks served 51% of tasks at prices 25% higher than Baseten's, and slower. A fixed order pushed Baseten to 98.5% of tasks, but every change needed a deploy. The adaptive version scores providers 0.7 on cost and 0.3 on speed, and publishes a summary to Redis every five minutes so no network call is made per request. CoreWeave's list prices were about 45% below Baseten's, and the router could sample it without a human.

Artificial Analysis, in a 29 September note, adds a model-side variable. Claude Sonnet 5.5, priced at $2 per million input tokens and $10 per million output, reaches 56 on its Intelligence Index. It uses roughly 193,000 output tokens per task at max effort, the heaviest the firm has measured and about seven times GPT-6 Astra (max). Cost per task lands at $7.60, about 50% above Sonnet 5.

The two figures do not disagree so much as measure different things: list price per GPU-hour and cost per completed task. Meetrix.io says plainly that it did not benchmark throughput and that its tokens-per-second figure is a placeholder. Anthropic's $518 billion commitment is a contract, not a benchmark.