What happened
In a November 3, 2026 post, Hadley Wickham asks whether frontier-model token prices are being held artificially low by venture-backed subsidies. He focuses on Kimi K3, an open model he says he has used heavily for a couple of months, and describes it as a reasonable proxy for a capable frontier model, though he finds Opus 5.5 somewhat smarter. The analysis compares buying inference from providers listed on OpenRouter with operating infrastructure to serve the model directly.
The provider comparison uses an assumed agentic-coding mix: input tokens exceed output tokens by 20 to 1, and 95% of input tokens are cached. Under this blended calculation, listed prices range from about $0.59 to $1.81 per million tokens. Wickham notes multiple competing providers, including six at the higher end around $1.13 per million blended tokens. He argues that competition among suppliers of downloadable open models puts pressure on prices. The figures are a snapshot from the post, not a lasting price schedule.
For self-hosting, Wickham cites Moonshot AI’s vLLM recipe recommending at least eight GB300 GPUs. He prices one possible Supermicro 8U GPU server at $550,000. Financing it over three years at an assumed 8% interest rate would cost about $17,000 monthly; renting eight GPUs at roughly $4 per hour each would be similar once hosting is included. His estimated monthly colocation, bandwidth and power costs are $3,750, $600 and about $700. Adding smaller expenses brings his estimated total to roughly $23,000 a month.
To estimate capacity, he uses a SemiAnalysis figure of about 6,000 tokens per second per GPU, or 48,000 across eight GPUs. At 75% utilization, he estimates 124 billion tokens monthly and a serving cost of approximately $0.25 per million. Under the same coding mix, he calculates an equivalent API cost of $1.13 per million. He concludes that token prices may be reasonably close to serving costs, with room for healthy margins, while noting that real operators may gain scale and negotiate better terms but must also pay staff.
Wickham distinguishes those per-token economics from other subsidies. He cites a SemiAnalysis post estimating that a $200 monthly subscription provides between $2,000 and $10,000 worth of tokens; he says such plans are for individuals and small businesses, while enterprises pay per token. He also says his calculation excludes model-training costs, which may have reached several hundred million dollars, and identifies investor funding—including Chinese government investment in Moonshot AI—as a possible subsidy.
What the archive adds
This is a new analysis within a known pricing thread. A dive into why OpenAI, Anthropic cut LLM prices by up to 80% examined price cuts and cache-read costs; Wickham instead estimates serving costs for an open model. The spending has outrun the math. If your team’s AI coding bill keeps growing, you’re not alone. documented AI coding spend and the difficulty of quantifying returns. This post adds a supply-side cost estimate, not a new price change.
Trend signal
This adds evidence to the broader pricing discussion in our prior analysis of LLM price cuts: cache mix and provider competition can change the economics hidden behind a per-token headline. It does not establish actual provider margins; the self-hosting model depends on uncertain cost and utilization assumptions. No supplied SITE_PATTERNS entry directly covers inference pricing or cost estimation, so this is not a fit for a named pattern.
Why now
Wickham frames the analysis around suspicion that low token prices are a strategy to attract users before prices rise. The post’s timing is also relevant to his own ongoing use of Kimi K3 alongside the release of Opus 5.5, which he describes as smarter but less cost-effective for his needs. It does not report a pricing change or a new subsidy announcement.
Why it matters
Teams often compare model prices per million tokens as though that number reveals either a provider’s margin or the cost of completing a task. This analysis shows how much the answer can depend on cache rates, input/output mix, GPU utilization and hardware financing. A team considering self-hosting can use the structure of the calculation to identify which assumptions it needs to measure; it should not treat Wickham’s estimate as a universal break-even point.
The distinction between API pricing, subscriptions and training costs is also useful for procurement conversations. A low subscription price does not prove that per-token API inference is loss-making, and a plausible serving-cost estimate does not show that a provider has recovered model-training investment. The post offers a framework for asking those separate questions, not audited accounts from any provider.
Limits and uncertainty
Wickham calls the analysis crude and says he used Claude to help determine self-hosting costs, while checking individual facts. Colocation, bandwidth, power, utilization, hardware financing and throughput are estimates or assumptions, not provider disclosures. The throughput figure comes from SemiAnalysis, and the post does not demonstrate that a real deployment would sustain it at the assumed utilization or token mix. Provider rates can change, and the comparison cannot establish actual margins. It excludes staff, model training and other business costs; the subscription value estimate is attributed to a separate SemiAnalysis post.
Reader takeaway
The useful lesson is not that every AI token costs exactly $0.25 to serve. It is that any claim about token economics depends on workload mix, caching, utilization and what costs are included. Use the post’s calculation as a checklist for your own scenarios, not as proof that APIs are subsidized—or that they are not.