Garry Tan's $200,000 Token Bill Prices an AI Employee Like a Utility — and It Breaks Seat Pricing

On Walter Isaacson's podcast The Builders, Y Combinator CEO Garry Tan said he spends about $100,000 to $200,000 a year on tokens. The spend covers Claude Code, for which he has published 23 reusable \"gstack\" skills, Y Combinator's internal agent Stray Light, which holds context on every founder the firm has funded, and a personal agent setup he adopted early. It is tokens only: no model training, no headcount, no hardware named. For the first time, the operating cost of an \"AI employee\" has been put on the record by someone who signs the invoices, and it makes agent economics a consumption question rather than a licensing question.

Anthropic's published prices convert the bill into units without guessing at token mixes. Using an 80/20 input-output split, Claude Sonnet 5.5 works out to $3.60 per million tokens and Claude Opus 5.5 to $7.20, so $200,000 a year buys roughly 55.6 billion Sonnet tokens — about 152 million a day — or 27.8 billion Opus tokens. These are list-price estimates: prompt caching cuts input reads to $0.10–$0.20 per million on Sonnet and Opus, and batch inference runs at half price, so real agent fleets consume more tokens per dollar than list prices imply.

The sharpest evidence comes from a firm that sells AI transformation for a living and still had to ration the tokens. From October 1, Accenture centrally funds $29 a month per person toward Claude and charges anything above that to a project; internal messages showed consultants spending $500 to $1,000 a month, one had access revoked outright, and Fortune's reading of the reporting puts the most intensive user at $12,000 a year. A flat per-seat fee cannot price that spread, so any flat per-seat fee mis-prices the distribution by more than an order of magnitude.

Two margins matter here. Software gross margin, the profit on a seat, is high: 79.5% at Datadog, 77.3% at Salesforce, 74.7% at NVIDIA and 67.6% at Microsoft on trailing twelve months. But when a seat vendor meters usage, it carries the model bill inside cost of revenue, while the layers that sell raw capacity or measure it do not. The market's split is already visible in multiples: Datadog, the meter that watches agent workloads, trades at 20.1 times trailing revenue against 4.9 times for Salesforce, and the margin migrates from the vendor that owns the seat to the vendor that owns the meter and the memory.

The force working hardest against seat pricing is that the price of a token is falling quickly. Chinese open-weight models took 57% to 67% of all tokens used on OpenRouter in the week of September 14, 2026, up from 6% to 13% in February, and 55% of Vercel's tokens in August against 11% in January, with CNBC's reporting attributing the shift to price once quality meets the job. Chinese cloud demand is real: Alibaba's AI Cloud and Compute Services revenue reached $7.1 billion in the June 2026 quarter, up 45%, with AI-related product revenue of $1.8 billion growing at a triple-digit rate for a twelfth straight quarter. That deflates the cost of every agent, and it compresses seat revenue per unit of finished work.

Two clocks matter for positions. Over the next two quarters, the December 1 Microsoft billing change and its 4,000-credit default cap will show up in renewal conversations and attach rates, and Anthropic's public listing process will put a market price on token revenue. Over one to three years the question is whether seat licences survive as an access fee while credits, conversations and work units become the revenue line — which is what Salesforce's 7.0 billion delivered agentic work units already look like, a consumption ledger sitting next to a seat base.