Invest1 publisher2 min readPublished
SemiAnalysis estimates Anthropic's subscriptions take over 40% of inference compute for 10% of revenue
Subscriptions cut Anthropic's blended revenue per megawatt by about $36 million while bringing in a tenth of its revenue, SemiAnalysis estimates. Plan limits, set through credit costs the labs can change without notice, decide how much of its compute earns API rates.
The Investor · Invest desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- SemiAnalysis says Anthropic's plans give about five times the monthly API-equivalent value of OpenAI's for agentic work, comparing Claude Opus 5.5 with GPT-6.1 Sol.
- The firm credits OpenAI's generous usage resets with part of a recent surge in Codex adoption and says they forced Anthropic to walk back planned plan cuts more than once.
- OpenAI once advertised 20 times Plus usage on its $200 Pro plan, then halved that plan's usage and removed relative-usage claims from its pricing page.
- SemiAnalysis sizes each limit by running prompts that isolate one token type at a time and watching how far the provider's usage meter moves.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Anthropic pays for plan generosity in yield: compute serving subscribers earns roughly a sixth of what its other compute earns, so each point of compute share the plans hold lowers revenue per MW.
- decision Anthropic can act on limits without an announcement, since credit-cost changes let it shift compute toward API buyers while leaving list prices untouched.
- constraint Outside analysts cannot read a lab's cost per subscriber from its pricing page, because plans expose only a usage meter, so margin estimates depend on meter tests like SemiAnalysis's.
- exposure OpenAI's margins are more exposed to the same limits than Anthropic's, because subscriptions make up a larger share of OpenAI's revenue, according to SemiAnalysis.
Per unit of compute, subscription plans earn Anthropic about a quarter of its average rate, since 10% of revenue is spread across at least 40% of inference compute [1][2]. The other 90% of revenue runs on at most 60% of the compute and earns one and a half times the average [12]. Compute sold outside the plans therefore earns about six times what plan compute earns. The multiple is higher if the subscription share is above 40% [13].
The $36 million figure can be checked against that split [3]. If it measures the gap between Anthropic's blended rate and what its non-subscription compute earns, then half the blended rate equals $36 million. Solved that way, blended revenue comes out near $72 million per MW, with plan compute near $18 million and the rest near $108 million [17]. SemiAnalysis calls its figures rough, and the published excerpt does not state the period they cover [1].
Limits are set in credits. A subscription buys a pool of them, and each model and token type draws it down at its own rate. Those ratios can differ dramatically from API price ratios, so SemiAnalysis values a plan only as a combination of plan, model and workload [4]. Labs can change limits silently by adjusting those credit costs, the firm says [5].
Anthropic can raise credit costs on its heaviest workloads and win back part of the six-to-one gap [13][5]. It can keep its limits where they are, as it did each time it walked back a planned cut under pressure from OpenAI's resets [15]. Or OpenAI tightens first, as it did when it halved usage on its $200 plan, and Anthropic follows with cover [6]. I'd expect Anthropic to hold for now, given how often it has already retreated [15].
SemiAnalysis's own text makes the case against. It says the plans are subsidized on purpose and can make sense as customer acquisition and marketing [7]. API revenue is only lost if a paying API customer wanted that compute. If Anthropic raised credit costs and its API revenue per MW stayed flat, the freed compute had no API buyer waiting, and the $36 million overstates the cost.
SemiAnalysis also sells the tool it recommends. Its Subscriptions Dashboard tracks plans at nine providers, OpenAI and Anthropic among them, and is open only to subscribers of its Tokenomics Model [8]. The firm wrote that "if you want to accurately model AI lab financials, you need to understand their subscription limits" [9].
What to watch
- A rise in Anthropic's credit costs for Opus 5.5 on the $200 plan, visible as the usage meter moving faster for the same agentic workload.
- Whether OpenAI extends its usage resets or cuts plan usage again after halving its $200 Pro tier.
- Any Anthropic disclosure of subscription revenue share or compute allocation that tests SemiAnalysis's 10% and 40% estimates.