The AI subsidy is the actual business model

David Rosenthal’s post today, “AI’s Affordability Crisis”, pulls together the most useful numbers I’ve seen on what AI inference actually costs the labs versus what the labs charge for it. The two findings worth being explicit about, both originally reported by Ed Zitron at Where’s Your Ed At.

The first is the SemiAnalysis subscription-burn benchmark. On a $200-a-month Anthropic plan, a user can burn $8,000 in tokens. On a $200 OpenAI plan, $14,000. Assuming the platforms are not (also) subsidizing the per-token price, that’s a 40x subsidy at Anthropic and a 70x subsidy at OpenAI on the top-tier subscriptions. The same analysis estimates gross margin at “negative 25% if you use as little as 25% of your rate limit.” This is the drug-dealer pricing pattern Rosenthal flagged a year ago, now with receipts.

The second is OpenAI’s just-leaked 2025 financials, reported by Zitron last week: $13.07B in revenue, $34B in costs and expenses, $20.92B in operating loss. The numbers will be argued over for months; the order of magnitude will not.

This matters for anyone designing an agent program right now in a way most decks have not absorbed. The unit economics you’re planning against today are a sticker price held up by venture capital. The unit economics you’ll plan against in 2027 will be closer to the actual cost of generating tokens.

Three practical reads worth being explicit about.

• Treat today’s per-token pricing as a temporary subsidy, not a baseline. If your agent program’s TCO model assumes the current ratio of frontier-model price to value holds through 2027, you are quietly building on a forecast nobody at the lab has signed up to. The honest sensitivity analysis is “what does this workload cost at 2x, 5x, and 10x current token prices, and where does the value still hold.” Run it now.
• Cost-aware architecture is the durable hedge. Routing the right work to the right model, caching aggressively, using small or open-weights models for triage, and reserving frontier-tier calls for genuinely hard turns are no longer “advanced” practices. They are the design choices that make a 2027 budget survive a 2026 repricing. Microsoft’s Agent Optimizer in Foundry Agent Service explicitly optimizes the quality-cost trade-off across model deployments, with per-task token costs in the output. That is the right shape of tool for the moment.
• For SI and ISV partners, packaged subsidy-risk analysis is the new procurement deliverable. Customers in regulated industries are about to start asking, on paper, what their AI program’s cost looks like under each vendor’s plausible re-pricing scenarios. A partner that ships that analysis as a one-page deliverable will look like an adult in the room. A partner that doesn’t will quietly lose the conversation.

The headline AI race in 2026 has been about capability. The 2027 race is going to be about repricing without losing customers. The vendors and customers that build for that earlier are the ones who will be calm when the subsidy ends, instead of scrambling. Worth running the math on your own program this week.