The “subsidy” in AI coding subscriptions is a compute bill, not a loss bill — Contested · resolves 2027-03-31

The debate: heavy users of monthly AI coding subscriptions burn far more compute than they pay for, and Armin Ronacher, creator of Flask, and others ask whether providers are losing money and will have to slash allowances or raise prices.

Why it matters: if it is a loss problem, companies should brace for price rises, but if it is a compute problem, allowances will loosen and tighten with providers' compute supply, which buyers can anticipate.

Who is on each side: skeptics say subscriptions are a money-losing business; SecondSource reads them as near break-even at average use, with compute, not cash, as the scarce thing.

SecondSource’s reading: the subscription subsidy is mostly a compute bill, not a loss bill. SemiAnalysis estimated on 2026-10-05 that subscriptions are only about 10% of Anthropic's revenue but may take more than 40% of its inference compute.

At a more realistic 20% average utilization, the same estimate puts subscription gross margins at about 6% for Opus 5.5 and 80% for Fable 5.1; only at full use does Opus 5.5 fall to -369%. The losses sit with the few users who max out their limits.

So allowances move with compute and rivals' offers, not with losses: on 2026-09-29 OpenAI halved usage on its $200 plan and added a $500 tier, while on September 22 Anthropic raised subscriber limits alongside Opus 5.5.

These figures come from one analyst firm's estimates, so we mark this call as contested.

SemiAnalysis
“Despite being just 10% of overall revenue…”

At the due date, check: (1) whether a provider discloses, or a second independent meter estimates, clearly negative subscription gross margins at average utilization (below -50%); (2) whether a major frontier lab still tightens mid-tier model subscription allowances within 60 days of announcing the arrival of a large batch of new compute; (3) once Anthropic publishes a breakdown of its company gross margin, whether the gap between overall gross margin and API gross margin comes mainly from subscriptions; (4) whether broad cuts to enterprise AI coding budgets appear.

What would overturn it

If (1) holds, “roughly break-even on average, mainly a compute bill” is withdrawn and the reading returns to a loss bill; if (2) holds, “allowances follow compute” is withdrawn and the driver is reread as pricing or losses; if (3) holds, the profit-and-loss estimate is revised toward larger losses; if (4) holds, “the risk is not on the demand side” is rewritten, while the structure of the call stands.