The "subsidy" in AI coding subscriptions looks mostly like a compute bill, not a loss. Our call: allowances mostly move with compute.

Key takeaways

In early August, in the August 8 issue, this publication made a call. Everyone assessing the financial risk of AI coding tools had been asking whether enterprises would cut budgets. SemiAnalysis, a research firm covering semiconductors and AI economics, interviewed more than fifty enterprises in June and matched the interviews against corporate card data. The answer was no. But the real risk might sit at the other end: a heavy user paying US$200 a month consumes far more compute than the price covers, and the vendor absorbs the difference. If that held, the correction would not look like enterprises cutting budgets. The correction would look like vendors tightening allowances, redrawing plan tiers, and quietly serving a cheaper model behind the same monthly fee. On September 29, OpenAI halved the allowance for new subscribers on the US$200 plan and opened a US$500 tier. That looked like the prediction coming true.

This piece takes the call apart and re-tests it. The result comes in three layers.

First, the structure holds: the customers who make money and the customers being subsidized are two different groups. TechCrunch, the tech news outlet, cited sources in April saying that Cursor, the AI coding tool, is already gross-margin positive on sales to large enterprises while individual developer accounts still lose money. GitHub announced in April that Copilot would move to usage-based billing, and the announcement said in plain words that GitHub had "absorbed much of the escalating inference cost" and that the existing model was "no longer sustainable". Enterprise budget surveys ask the profitable group, so they cannot answer the subsidy question. In August this step was only a logical distinction. Now two sources with no connection to SemiAnalysis support it. Cursor splits enterprise from individual accounts directly, and GitHub shows that heavy usage under a flat fee is absorbed by the vendor.

Second, half of "the prediction came true" has to be withdrawn. Tightening allowances is not new. In July 2025 Anthropic announced weekly caps, effective at the end of August, citing a user who had consumed tens of thousands of dollars of usage on a US$200 plan. Cursor, Replit and GitHub all changed their pricing around June of that year. When this publication made the call in August, rationing had already been running for a year. So the "there will be rationing" half of the call described something that had already happened. Most of the public reasons vendors gave for tightening in 2026 were compute shortages, not losses. Before the September 29 halving, OpenAI had paused sales of the US$200 plan on September 10, citing demand at a level it had not seen before.

Third, the "subsidy" has three different numbers behind it (API list-price equivalent, cost at full utilization, cost at average utilization), and they differ by multiples. Once separated, the bigger line is compute. SemiAnalysis estimates that subscriptions bring in about a tenth of Anthropic's revenue but may consume more than forty percent of its inference compute. API here means the pay-per-use channel where customers buy tokens directly from the model vendor. SemiAnalysis computed a gross margin of minus 369 percent for Anthropic's US$200 plan on the mid-tier Opus 5.5, with the allowance fully used. The same article assumes a 92 percent API gross margin. On that assumption, its next sentence puts the same plan at a small profit of about 6 percent at a more realistic 20 percent average utilization, and the top-tier Fable 5.1 at about 80 percent. The loss sits with the users who max out the cap. Taken together, by dividing those two rough estimates (our calculation), one unit of compute spent on subscribers earns about a sixth of what the same unit earns from API and enterprise customers. That money is a real cost only when compute is scarce. So allowances appear to follow compute: they loosen when capacity arrives and tighten when demand surges.

Why this is worth twenty minutes: the usual "subsidy" figures mix up what they are measured against, and this piece separates them. "A US$200 plan can consume US$8,000 of tokens" measures the API list-price equivalent, meaning what the same usage would cost at list price through the API. The list price includes the vendor's margin, so it is not a cost. "Minus 369 percent gross margin" measures the extreme case of a maxed-out cap. Neither is the cost of an average user. Once the three numbers are separated, the loss estimate shrinks and compute becomes the larger item. The August line that "plan tightening is the prediction coming true" was also tested here, and only half of it held up. What would count as this piece being wrong is stated in advance: if a major vendor keeps tightening mid-tier model allowances after a large block of compute arrives, "allowances follow compute" fails.

The piece runs in five steps: ① why enterprise budget surveys cannot answer this question; ② rationing began long ago, and the reasons were mostly compute; ③ the "subsidy" split into three ledgers; ④ forty percent of compute for a tenth of revenue; ⑤ which readings to watch.

If only this part gets read: for a company running its daily development on US$200 subscriptions, the main exposure is allowance cuts, more than a price rise. When a vendor's compute runs tight, subscribers' allowances get cut first. In OpenAI's September change, new subscribers on the US$200 plan got half the allowance at the same price, which roughly doubles the price per unit of allowance. The ceiling for a budget is what the same usage would cost through the API, not the subscription price. That ceiling can be several times the subscription price. SemiAnalysis estimates a US$200 plan can consume up to forty times its price in API list-price equivalent. That is the worst case, not the expected spend.

One: enterprise budget surveys ask the profitable group

SemiAnalysis's June study is the most solid answer available to the question of whether enterprises will cut AI coding budgets. The study interviewed more than fifty enterprises and matched the interviews against actual card data from Ramp, a corporate spend-management platform. The conclusion: spending is highly concentrated in a long tail. The Ramp customer at the 99th percentile spends nearly US$90,000 per employee per year. The median Ramp customer spends about US$136. The original does not say whether that median figure is also per employee per year. The big customers are highly sticky. SemiAnalysis judged that AI budgets for the second half of 2026 carry no material risk.

That method measures how much customers paid. The question this publication raised in August is a different one: whether the money a vendor receives covers the compute it burns for that customer. Both can be true at once. Most customers spend little and budgets are safe, while a few heavy users are subsidized by the vendor.

In August this was only a logical distinction. Now two sources with no connection to SemiAnalysis support the distinction from two angles:

• Cursor: TechCrunch, citing sources on April 17, reported that Cursor had been gross-margin negative until recently and turned slightly positive through in-house models and switching to cheaper models. Split out, Cursor has "positive gross margins on its sales to large enterprises, but continues to lose money on individual developer accounts".
• GitHub: on April 27 GitHub announced that Copilot would move fully to usage-based billing from June. The announcement said GitHub had "absorbed much of the escalating inference cost" behind that usage and that the existing per-request model was "no longer sustainable". The reason given: under the same monthly fee, a short question and a multi-hour autonomous coding session were billed the same.

Cursor's reading is that enterprise makes money and individual accounts lose it, though the sources did not say whether the losses concentrate in heavy users. GitHub's announcement supplies the other half: under a flat fee, the cost of heavy usage is absorbed by the vendor. Enterprise budget surveys reach the profitable enterprise side. So "budgets are not at risk" can be right and still not reach the subsidy question.

Two: rationing began long ago, and the reason was mostly compute

The August 8 issue wrote: "if the subsidy is real, the correction will not show up as budget cuts; it will show up as product-side rationing: rate limits tightening, plan tiers redrawn, the same subscription quietly served by a cheaper model." That issue also stated plainly that vendors' term changes were not yet being tracked. The cheapest test of the call was to check whether allowances had always been limited, for reasons unrelated to cost. If so, the leading indicator fails.

This time the check was done. The table below lists allowance and pricing changes at the main coding tools from 2025 to October 2026, in date order:

Three things can be read off the table.

First, rationing was under way a year before this publication made the prediction. The "there will be rationing" half of the August call describes something that had already happened. That half was not a prediction. The October 6 issue called the September 29 cut the first case pointing in the predicted direction. That is withdrawn. SemiAnalysis itself writes that labs publicly change their allowances all the time with promotions and new model releases, and in its own testing it caught one vendor running an allowance A/B test on a small share of accounts. Against a base rate that high, a single tightening has very little discriminating power.

Second, the vendors that say "cost" out loud are the three middle-layer tool makers; the frontier labs mostly say "compute". Cursor, Replit and GitHub all buy tokens in volume from frontier labs (Cursor also has in-house models). The losses of those three show up more directly on their own books, which may be why the three say cost. Of the 2026 tightenings by Anthropic and OpenAI that came with a public reason, most cited surging demand and short compute. OpenAI's September 29 halving itself came with no reason; this piece reads it alongside the September 10 sales pause. Anthropic's 2025 move was officially framed around users running 24/7 in the background, which is closer to abuse prevention. That weakens half of the August call. But the evidence Anthropic cited was a usage amount, someone consuming tens of thousands of dollars of usage on a US$200 plan, so the move is not entirely unrelated to cost either.

Third, allowances have bounced back at least once. Anthropic tightened at peak hours in March, citing demand growth. In May, when the SpaceX capacity arrived, Anthropic removed the peak-hour reduction and doubled the allowance. OpenAI's September 29 halving was preceded by the September 10 sales pause. A loss does not disappear when compute arrives, so this looks less like a loss-driven correction and more like compute allocation. The sample is one pair: Anthropic's March tightening and May loosening. OpenAI's pause followed by a halving adds one more case. And "short of compute" is each vendor's own account, not measured capacity. Anthropic's net cut at the end of August was also attributed to compute management, so it is not counted here as a counterexample. With a sample this small, that choice cannot be checked either way.

Three: the "subsidy" split into three ledgers

The magnitude in the August call rested on an example in a post by Armin Ronacher, the author of the Flask web framework: "spent 200 USD for a codex sub, but you used 2000 USD worth of tokens" (his post, August 3, 2026). Measured figures now exist. But the figures come in three forms, and mixing them produces conclusions that differ by multiples.

The first ledger: API list-price equivalent. SemiAnalysis published an article in June whose title says a US$200 Claude plan can consume US$8,000 worth of tokens. This converts the subscription allowance into what the same usage would cost through the API: forty times. The forty-times figure measures the value the customer receives, not the vendor's cost, because the API list price itself includes gross margin.

The second ledger: cost to serve at a maxed-out cap. On October 5 SemiAnalysis assumed a 92 percent API gross margin and a user consuming 100 percent of the allowance. On those assumptions, Anthropic's US$200 plan runs a gross margin of minus 369 percent when the user picks the mid-tier Opus 5.5 (the model Anthropic positions as the daily driver) and 1 percent with the top-tier Fable 5.1. One reason the two differ so much is that the plan lets Fable consume only half of the allowance, so "maxed out" on Fable is half the volume to begin with.

The third ledger: cost to serve at average utilization. The same article's next sentence: with a more realistic average utilization of 20 percent, the gross margins rise to 6 percent and 80 percent respectively. This condition is the one most often dropped when minus 369 percent is quoted on its own. These two numbers are sensitive to the API gross-margin assumption. The plan's cost is derived backwards from API list prices. If the API gross margin were only the 70 percent mentioned below, the same average utilization would put the plan at a clear loss.

With the three ledgers side by side, the shape of the August call becomes clearer. Under this set of assumptions, the average subscriber carries no structural loss. The loss sits in the tail of users who max out the cap. That is the "healthy median, subsidized tail" shape. But the magnitude changes. If only the tail loses money, the loss across the whole subscription pool is far smaller than a figure like minus 369 percent. This shape is derived from utilization scenarios. No utilization distribution is available. Anthropic says its weekly cap affects under 5 percent of subscribers, which points the same way but is not a distribution.

How small can it get? Put SemiAnalysis's own figures together (the following is this publication's calculation; every input is one analyst firm's rough estimate). The inputs are coarse. "Forty percent" is SemiAnalysis's "more than 40%" phrasing, and "a tenth" is also its estimate. Both rest on that single source. If the actual shares differ, every multiple below moves with them.

• Subscriptions bring in about a tenth of revenue and use forty percent of compute; the remaining ninety percent of revenue uses sixty percent of compute.
• So per unit of compute, subscriptions earn about a quarter of the average (0.10 ÷ 0.40), while API and enterprise earn about 1.5 times the average (0.90 ÷ 0.60). The two differ by about six times.
• If the API gross margin is 92 percent as SemiAnalysis assumes, the cost per unit of compute is only 8 percent of API revenue, because a 92 percent gross margin means cost is 8 percent of revenue. Subscriptions still earn more than cost. Summed across all subscribers, the pool earns a small profit under this assumption.
• If the API gross margin is only 70 percent (SemiAnalysis's May figure for inference infrastructure gross margin was "over 70%"), the subscription pool loses money, about 8 percent of company revenue.

Under either assumption, the profit-and-loss effect of subscriptions lands between "a single-digit percentage of company revenue" and "a small profit". Which assumption is closer to reality depends on one number that does not reconcile, set out in the next paragraph. The test is this: if the heavy tail's loss is a single-digit share of revenue, the subsidy is an affordable marketing cost. SemiAnalysis says the same: subscriptions are highly subsidized but can still make economic sense as customer acquisition and marketing tools.

But one number does not reconcile, and it has to be laid out. PitchBook and The Information have reported Anthropic's company-wide gross margin at around forty percent, far below 92 percent. Where the gap falls (free users, subscriptions, internal usage) cannot be separated from public data. If the gap comes mainly from subscriptions, the third ledger is too optimistic.

Four: forty percent of compute for a tenth of revenue

The profit-and-loss effect is small, yet vendors keep tightening. The main reason looks like the opportunity cost of compute.

The same unit of compute, spent on subscribers, earns about a sixth of what it earns from API and enterprise customers. When compute is plentiful, the marginal cost of that is low: idle machines serve subscribers, every extra dollar helps, and the plans pull in customers. When compute is short, the same unit carries a large opportunity cost. If those forty percent of compute were sold instead to API customers waiting in line, at the API and enterprise revenue per unit (about 1.5 times the average), the compute would earn about sixty percent of total revenue. Subtract the tenth that subscriptions earned, and company revenue could rise by about half (0.40 × 1.5 − 0.10). That is revenue by this publication's calculation, not profit, and it holds only on two premises: API demand is really queued up, and compute can be shifted between customer groups. SemiAnalysis frames it as a compute ledger too: by its estimate, subscriptions lower the blended revenue per MW, across subscriptions and API combined, by about US$36 million. MW is megawatts, the data center's electrical capacity, which the industry uses to measure compute scale. The original does not state the period or baseline for that figure, and it has not been recomputed here.

This explains the pattern in the section two table:

• Tighten when compute is tight: Anthropic's peak-hour reduction in March and its net weekly-cap cut at the end of August; OpenAI's demand-driven sales pause on September 10, then the halving three weeks later (the halving itself came with no reason).
• Loosen when compute arrives: Anthropic doubled allowances and raised API limits in May as soon as the SpaceX capacity arrived.
• Competitors' offers matter too: SemiAnalysis writes that OpenAI's earlier generous allowance resets forced Anthropic to repeatedly walk back planned subscription cuts.

The two vendors also tighten in different ways. In SemiAnalysis's testing, OpenAI halved across every model tier, while Anthropic gives the newer and more expensive model the smaller allowance. On the same-priced plan, Anthropic's mid-tier Opus 5.5 offers about five times the API-equivalent value of OpenAI's mid-tier model. That five times is value converted at API list prices. The two vendors' list prices differ, so it does not mean five times the work gets done. A new top-tier model launches with a low allowance. Our read: capping the allowance first on the new model, where demand concentrates, avoids going back to cut older plans. The vendors have not stated their motive.

This reading connects to a trend line this publication has been tracking: once model vendors go from losing money on tokens to making it, whether the new margin stays.

Where this piece plugs in: the three paths above argue over API gross margin. This piece adds a reading from the subscription side. Its margin is not mainly being pushed down by competition. It is being used as a valve when compute is short. That gives the "temporary rent during the shortage" path one more observable reading: when the shortage lifts, subscription allowances should loosen first.

Five: which readings to watch next

Our call

The subsidy in AI coding subscriptions is real, but it looks mostly like a compute bill, not a loss. By SemiAnalysis's estimate, subscriptions bring in about a tenth of Anthropic's revenue and may consume more than forty percent of its inference compute. From that, by this publication's calculation, one unit of compute spent on subscribers earns about a sixth of what the same unit earns from API and enterprise customers. Under SemiAnalysis's assumed scenarios, the mid-tier plan at average utilization is a small profit, and the loss sits with users who max out the cap. Those figures rest on a single source, and Anthropic's company-wide gross margin of about forty percent does not reconcile with the 92 percent API gross margin SemiAnalysis assumes; the gap cannot be separated. So for now, when vendors tighten or loosen allowances, they appear to follow compute first and competitors' offers second, not losses. The sample behind this reading is still small.

To anticipate the next tightening, two things to watch: the allowance per model tier on each vendor's plan page, and announcements of large compute deliveries (for example the data center Anthropic contracted from SpaceX in May). If allowances do not loosen within sixty days of a compute delivery, or if tightening continues while compute is loose, the reading should change to pricing- or loss-driven. Enterprise budget surveys measure only the paying customers and cannot see this.

What this changes, and for whom

• Enterprise engineering leads and finance: a team's US$200 subscriptions are a discount that holds while the vendor's compute is loose, not a fixed price. The budget ceiling is what the same usage would cost through the API. In OpenAI's September change, new subscribers got half the allowance for the same money. That halving came three weeks after OpenAI paused sales on September 10 citing demand at a level not seen before, and OpenAI gave no reason for the halving.
• Procurement and contracts: for teams whose daily capacity rides on subscriptions, allowances can be changed unilaterally and even silently (SemiAnalysis caught one A/B test). Workflows that need stable capacity are better placed on API or enterprise plans with contractual limits; subscriptions suit work that can wait.
• Investors evaluating AI coding startups: the question to ask is the gross margin of enterprise and individual accounts separately, more than the count of paying users. Cursor's reading is enterprise positive, individual negative. A middle layer that buys tokens from frontier labs has no compute to reallocate, so its losses go straight to the books.
• Operators of model vendors and clouds: launching a new top-tier model with a low allowance is less trouble than going back to cut older plans. When OpenAI halved across the board, Anthropic's approach became the obvious point of comparison. This is read from the allowance structure by this publication, not stated by the vendors.
• People tracking the compute market: subscription allowances are a public reading of compute tightness, faster than earnings reports. Within a month or two of a large compute delivery announcement, allowances should loosen. If allowances do not loosen, the new compute was consumed by other demand.

What would prove this wrong

1. A vendor discloses, or a second independent measurement firm estimates, that subscriptions at average utilization run a clearly negative gross margin (for example below minus fifty percent). Then "no structural loss at average usage, mostly a compute bill" fails and the loss ledger returns. Verdict date, set by this publication: March 31, 2027.

2. A major frontier lab keeps tightening mid-tier model subscription allowances within sixty days of announcing a large block of new compute. Then "allowances follow compute" fails, and the driver should be read as pricing or losses. Same verdict date.

3. Anthropic's gross-margin breakdown becomes public (for example in a listing filing) and shows the gap between company-wide and API gross margin comes mainly from subscriptions. Then section three's profit-and-loss calculation has to be redone with the loss revised upward. Same date.

4. Widespread AI coding budget cuts appear on the enterprise side. This does not overturn the structure of this piece, but "the risk is not on the demand side" would have to change. So far there are only individual cases (SemiAnalysis's June study mentions Uber exhausting its annual budget in four months) and a McKinsey survey, relayed by The Register, in which most enterprises overspend; the first-hand survey hasn't been obtained. Overspending is not the same as cutting budgets. Same date.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.

Sources this deep dive draws on

Compiled mechanically from the records of the facts this piece draws on; each line is the source of one of them. Whether we read the original or only reporting is stated in the body.

• SemiAnalysis · 2026-10-05: SemiAnalysis 2026-10-05:Anthropic Subscriptions Offer 5x+ More Value Than OpenAI
• WinBuzzer · 2026-09-30: winbuzzer.com/2026/09/30/openai-adds-500-chatgpt-pro-cuts-allowance-new-200-subs
• X / @mikeyk · 2026-09-22: x.com/mikeyk/status/2102441797482152313
• SemiAnalysis: TokenBudgeting: Our Conversations with Enterprises on Token Spend
• X / @mitsuhiko · 2026-08-04: x.com/mitsuhiko/status/2084563924435845350
• X / @teortaxesTex: x.com/teortaxesTex/status/2084492223962362331