We use cookies to improve your experience on our website, analyse traffic, and for marketing purposes. By clicking "Accept All", you consent to our use of cookies. You can also customise your preferences or reject non-essential cookies. Learn more
SemiAnalysis recently modelled how much Anthropic and OpenAI may be subsidising their AI subscription plans. The numbers are worth understanding if your business runs on cloud AI.
A Claude Max 20x subscriber pays $200 a month. Their maximum API-equivalent spend at that tier is $8,000 a month. Assuming compute costs are 25% of API list prices, Anthropic’s breakeven point is around 10% average utilisation. Use it more than that, and the user is costing more to serve than they pay.
ChatGPT Pro 20x is in a tighter spot. Same $200 a month, but $14,000 of potential spend. Breakeven is below 6% utilisation. Any developer who uses it consistently through a normal working week is probably a net cost to OpenAI.
None of this is news to anyone running these businesses. They know the maths. The question is whether it holds at scale, and for how long.
The gym membership logic
The model being applied here is the one every gym uses. Most members pay every month and barely show up. The gym prices for the average member, not for the person there every morning for two hours.
Applying that to AI subscriptions has a problem. The people buying $200 developer plans are, almost by definition, not average users. They’re buying the 20x plan because they use it 20x. The adverse selection is built into the product design.
SemiAnalysis assumes a 75% gross margin on API list prices, consistent with what’s been reported about OpenAI’s economics. At that margin, a Claude Max user running at 25% average utilisation costs Anthropic $500 a month to serve. They’re paying $200.
That $300 monthly gap per heavy user has to close somewhere.
Three realistic paths exist. Compute costs fall fast enough to close the gap, which requires chip efficiency gains to keep pace with usage growth. The majority of subscribers stay low-utilisation and effectively subsidise heavy users, the gym model working as intended. Or prices go up.
Realistically it’s all three, weighted differently over time. But as Anthropic and OpenAI move toward IPO and face pressure to show real margins, raising prices is the most controllable lever available to their boards.
OpenAI has already used it. Prices have increased across several tiers since 2023. Anthropic has moved similarly. Neither company publishes what their subscription economics look like internally, but IPO filings will change that. When the paperwork lands, the numbers will be there.
What this means for your AI cost planning
Your team’s AI subscription costs right now are partly socialised across the broader user base. Power users and casual users are pooled, and the maths only works if enough casual users balance the books.
That’s probably fine today.
The planning risk isn’t just a price increase from $200 to $350 on a given plan. Pricing changes on AI subscriptions can arrive fast, with little notice, and have real effects on teams whose workflows are built around a specific cost assumption. A model replacement, a context window change, a rate limit adjustment – any of these shifts your effective per-seat cost without a single headline price change.
For teams in regulated industries this matters more than most. If you’re a financial services firm, healthcare provider, or legal team that has integrated cloud AI into compliance-relevant workflows, you have a material dependency on pricing that SemiAnalysis’ analysis suggests may be loss-making at the power-user end. At some point that corrects.
What on-premises AI inference actually changes
When you run models on your own hardware or private cloud, the cost structure is different in a fundamental way. You’re paying for compute you control. There’s no repricing event, because there’s no external vendor with a board deciding margins need to improve. Your inference cost doesn’t move because a US-listed company had a bad quarter.
For regulated organisations, there’s a second argument. Your data doesn’t leave your environment. You’re not dependent on a data processing agreement that a US-headquartered company can revise. You’re not downstream of whatever policy decisions come out of Washington next month – and right now that matters.
The managed cloud subscription model for AI is cheap because it’s currently subsidised. When that changes, the case for on-premises AI and private cloud inference strengthens considerably. Organisations that have already built that capability will have a stable cost base and data environment. Organisations that haven’t will be making a rushed infrastructure decision at the moment it’s hardest to make one clearly.
Locai runs LLMs, audio transcription, video processing, and image analysis on your own infrastructure. Predictable costs, no data residency risk, no dependency on a US vendor’s pricing decisions. Book a demo.
FAQ
Why are Anthropic and OpenAI subsidising AI subscriptions? SemiAnalysis estimates that at 20%+ utilisation rates, the cost to serve a power user exceeds their subscription fee. The companies are betting that most users stay below breakeven utilisation, and that compute costs fall fast enough to close the gap over time.
Will AI subscription prices increase? OpenAI has raised prices across multiple tiers since 2023. As both companies approach IPO and face margin scrutiny, further increases are likely, particularly on higher-tier plans where the subsidy is largest.
What is on-premises AI inference? Running large language models and other AI workloads on hardware your organisation controls, rather than sending requests to a cloud provider’s servers. It removes dependency on external pricing, keeps data within your environment, and gives your team predictable compute costs.
Is on-premises AI practical for small regulated organisations? Yes, with the right orchestration layer. Locai runs on standard server hardware and supports models ranging from 1B to 70B+ parameters, depending on your use case and available compute.
