(9) The Subsidy Ends at the Bell

August 3, 2026/5 min read

The Disclosure Test

Anthropic filed a confidential draft S-1 with the SEC on 1 June 2026 and OpenAI followed on 8 June, which means the argument about subsidised AI tokens will be settled by an accounting policy note. Gross margin is the number that matters, and it depends on a classification choice: whether inference compute sits in cost of revenue or gets pushed into research and development. Global Capital Network's reading guide to the two filings makes the point plainly, that investors will use the resulting margin to decide whether these are software businesses deserving software multiples or something structurally closer to capital-intensive infrastructure. A person familiar with OpenAI told Reuters that the company does not expect to be profitable until 2030, and it filed anyway.

What Uber Did After the Bell

Uber is the closest precedent for a business that priced below cost at scale and then had to answer to public shareholders. Hubert Horan documented the subsidy phase in detail, with GAAP net losses of $2.6bn in 2015, $3.8bn in 2016, $4.5bn in 2017 and $3.9bn in 2018 while gross fares grew from $9bn to $45bn, and passengers not covering the full cost of the service they received. Uber listed on the New York Stock Exchange on 10 May 2019. The repricing that followed took years rather than quarters, and Uber's 10-K shows an operating loss of $1.8bn in 2022 turning into operating income of $1.1bn in 2023 and $2.8bn in 2024.

The mechanism matters more than the timeline. Rakuten data cited by Slate put the rise in average Uber prices at 92% between 2018 and 2021, but the heavier lifting happened in the split rather than the fare. Len Sherman of Columbia Business School found Uber's take rate rose from 15 to 20% of each fare a decade ago to above 50% in some cities by 2026, with rider fares and driver earnings staying coupled until the 2019 listing and diverging afterwards. Uber disputes the estimate and says it kept 21% of each fare on average in the third quarter of 2025.

The Same Play, Running Faster

AI is doing this before the listing rather than after it. Between February and June 2026, OpenAI, Anthropic and GitHub moved customers from flat-rate billing toward charging on token consumption, and Business Insider attributes the shift to agentic workloads that providers could no longer afford to subsidise for heavy users. GitHub's usage-based monthly quotas took effect on 1 June 2026, with a single long-running agentic prompt estimated at over $100. Corporate buyers responded with rationing, and Coinbase introduced weekly AI spend caps of $500 to $5,000 by employee seniority.

The list price tells you almost nothing about this. The Silicon Data Token Expenditure Index shows single-token prices down more than 90% since 2023 while spending on large language models doubled after late 2025, and Bain measured token costs halving between December 2024 and December 2025 while consumption rose 450%. Uber had to raise the fare because it had no other lever. AI labs can hold the fare down and still take more, because consumption per task is climbing faster than price is falling.

Where the Analogy Breaks

Uber never had a cost curve underneath it, because a driver-mile does not get 90% cheaper. SemiAnalysis estimates Anthropic's inference gross margins moved from minus 94% in 2024 to 38% in 2025 and into the 60 to 70% range during 2026, driven by cost reduction rather than price increases, through Blackwell replacing Hopper and higher throughput per GPU. A listed lab with that curve can serve a public margin target and cut prices in the same quarter, which is an option Uber never held. Reuters reported in June that OpenAI was weighing significant price cuts including on token usage, with Synovus Securities' Christopher Brown describing a price-war dynamic as the two labs compete to reach public markets first, and Citi notes Chinese models available at $0.18 per million tokens against a $4 average for frontier models.

What Buyers Should Underwrite

The transferable lesson from Uber concerns which variable moves. Fares rose, but the take rate re-rated the business, and it moved quietly over four years in a line item most riders never examined. The equivalent variable in AI is consumption per task, and it is already moving. Any team that has built agent workflows on today's economics holds an unhedged short position on someone else's cost curve, and the available hedges are architectural: routing, caching, model right-sizing, and a credible self-hosting option whose crossover point shifts by orders of magnitude depending on which API you are escaping.

Uber listed in May 2019 and did not turn an operating profit until 2023. Read the cost-of-revenue definition in the first AI prospectus that becomes public, then watch tokens per task over the twelve quarters that follow.

Sources