Data assembled 17 September 2026. Every figure is from a fetched source (Sources list at the end); derived cost figures show their formulas. Machine-readable data and per-provider tables live in research/.
Thesis. The commodity tier of open-weight LLM inference — the DeepSeek-V3-class models priced at $0.18–0.38 per million output tokens — is being sold at or below the hardware cost of a provider renting GPUs at market rates. This is sustainable only while investor capital and GPU-collateralized debt absorb the gap between price and cost. Roughly $43B of private equity in US inference clouds, another $50–85B in open-model lab valuations, and a $45B+ public debt wall secured against the very GPUs the cheapest tokens feed are exposed to that gap. The first repricing has already printed — Groq's valuation was marked down 50% in August 2026. The question is not whether the commodity tier earns its cost of capital; it is whose balance sheet this year's discounts are running on.
1. What's actually cheap: the price ladder
The cheapest posted open-weight output price in a scrape of 332 live models: DeepSeek V4 Flash on DeepInfra at $0.06 input / $0.18 output per million tokens[159][13]. At the top, Claude Fable 5.1 and gpt-6-class list at $50 per million output[175][163]. The spread between commodity open weights and the closed frontier is far wider than the "10–20× cheaper" the market usually quotes:
| Comparison (output $/1M) | Ratio |
|---|---|
| Claude Fable 5.1 ($50.00) ÷ DeepInfra DeepSeek Flash ($0.18) | 278× |
| GPT-5.5 ($30.00) ÷ $0.18 | 167× |
| Claude Opus 5 ($25.00) ÷ $0.18 | 139× |
| GPT-5 / Claude Sonnet 5 ($10.00) ÷ $0.18 | 56× |
| Even Gemini 3.8 Flash promo ($3.75) ÷ $0.18 | 21× |
Sources: OpenAI[163], Anthropic[175], Google[165], DeepInfra[159] pricing pages, fetched 2026-09-17.
Two nuances before the economics:
- The open frontier is no longer uniformly cheap. Kimi K3 (2.8T params, the largest open-weight model ever released) lists at $3.00/$15.00 — more than Claude Sonnet 5's output tier[173][199][175]. The 50–280× spread is a property of the commodity tier, not of "open weights" as a category.
- Intra-open dispersion is wild. The identical DeepSeek V3.2 weights sell for $0.31 output on GMI Cloud and $4.50 on SambaNova — a 14.5× spread on byte-identical inference[178]. Llama 3.3 70B: $0.32→$1.04 (3.3×)[190]. This is what a commodity market looks like: nobody can price on quality, so prices spread across hardware cost, utilization luck, and burn tolerance[93].
2. Who's buying: volume without revenue
The cheap tier isn't a niche. On OpenRouter's live leaderboard (2026-09-16), open-weight models fill nearly the entire token-volume table: DeepSeek V4 Flash variants alone ran ~25.8T tokens/day, GLM-5.3-Flash 11.27T, Xiaomi MiMo-V2.5 7.83T[14]. Chinese-origin models went from under 2% to ~46% of identified OpenRouter tokens in twelve months[116]. And the quality gap is nearly closed: GLM-5.3 and Kimi K3 sit 6–9 points below the closed frontier on Artificial Analysis's index (45 vs 53) and ~20–25 Elo on LMArena — within a few confidence intervals — at ~1/15 the cost per task ($0.25–0.27 vs $3.91)[117][118].
Vercel's production gateway data pins the revenue problem: 29% of June-2026 tokens were open-weight, carrying under 4% of spend[137]. High volume, low dollars. The market is consuming enormous quantities of tokens at prices that raise the central question of this piece: do those dollars even cover the GPUs?
3. The cost side: what a token actually costs
No provider publishes its cost of goods. It can be derived from two public inputs — GPU rental rates and measured throughput — with the formula:
cost_per_1M_output = 1e6 × $/GPU-hr ÷ (tok/s/GPU × 3600 × utilization)
GPU rates (Sep 2026, $/GPU-hr): a barbell. Silicon Data's neocloud index: H100 $2.66, H200 $3.31, B200 $5.73, MI300X $2.64; hyperscalers charge $7.18+ for the same H100[28]. On-demand medians run $3.25 / $4.40 / $7.00[10][6]. At the floor, Vast.ai lists H100s from $1.49, and an H100 card buys outright for ~$31K — meaning three-year straight-line depreciation alone is $1.17/hr; renting at $3.99 costs ~$35K/year, more than buying[8]. SemiAnalysis's owned-at-scale TCO (depreciation-embedded): H100 $1.17, H200 $1.22, B200 $1.73, MI300X $0.95[1].
Throughput (measured, not assumed): SemiAnalysis InferenceX publishes DeepSeek R1 0528 at 739.5 tok/s/GPU on H100 ($0.44/M on owned TCO) vs 5,215 on B200 ($0.09/M)[1]. SGLang's EP72 max-batch run gives 1,400 tok/s/GPU on H100 → $0.40/M at $2/hr rental (TensorEconomics' published derivation)[68]. DeepSeek's own fleet disclosure implies ~1,850 output tok/s/GPU on H800 at $2/hr → ~$0.30/M hardware cost[131][132]. GLM-4.6-FP8 lands ~872 tok/s/GPU on H200 and ~1,005 on B200 in independent node benchmarks[90][91].
Derived cost per 1M output tokens, and how it compares to market:
| Model class | Derived cost (full util) | Realistic cost (40–60% util) | Market price | Verdict |
|---|---|---|---|---|
| DeepSeek V3-class commodity MoE | $0.23–0.65 | $0.9–1.6 | $0.28–0.38 | At/below cost for renters |
| DeepSeek R1 / Kimi K2 reasoning | $0.44–1.22 | $0.8–2.4 | $2.00–2.15 | Healthy, 2–5× markup |
| GLM / Qwen-235B-class | $1.05–1.94 | $1.5–4.0 | $2.00–3.48 | Thin; negative below ~60% util |
| 8–32B small models | cents | cents | $0.02–0.30 | Loss leader, not physics |
The commodity row is the thesis's load-bearing finding: if you rent GPUs at even neocloud-index rates, selling DeepSeek-V3-class output at $0.28–0.38/M is selling below hardware cost at any realistic utilization. The only paths out: own depreciated silicon (TCO math: $0.23/M), run B200-class FP4 at max batch ($0.09–0.31/M)[1], or cross-subsidize. DeepSeek itself proves the point from the other side — its claimed 545% "cost profit margin" comes from owning its H800 fleet and pricing to roughly a 10-month payback, a rule that structurally disciplines every reseller who rents[131][135]. SemiAnalysis's summary of the dynamic: "plenty of room for DeepSeek to make money while crushing everyone else's inference economics"[2].
One more nail in the "sell at list price" idea: DeepSeek V4 Pro's list price ($1.74/$3.48) is identical across Fireworks, Together, and Baseten, while DeepInfra ($1.30/$2.60) and Novita ($1.60/$3.20) quietly undercut it[133]. A reference tariff plus spread competition — the classic shape of a market where the product is a commodity and the price is set by whoever's marginal cost (or burn rate) is lowest.
To be precise about scope: the thesis is tier-specific. The reasoning tier ($2+/M output) carries a real 2–5× hardware markup and needs no rescue capital. If the whole market looked like the reasoning tier, there would be no story. The story lives entirely in the gap between the two rows of the table above — a gap currently filled by someone else's money.
4. Whose capital rides on the gap?
Below-cost selling is rational only while another balance sheet absorbs the difference. Here is that balance sheet, September 2026.
US inference neoclouds — ~$3.8B of fresh 2026 equity onto ~$43.5B of post-money valuations:
| Company | Latest round | Valuation | Revenue (est.) | Multiple |
|---|---|---|---|---|
| Fireworks AI | $1.505B Series D, Jul 2026[19] | $17.5B[19] | >$1B ARR[21][22] | ~17× |
| Baseten | $1.5B Series F, Jun 2026[41] | $13B[41] | ~$600M run-rate[43] | ~22× |
| Together AI | $800M Series C, Jul 2026[15] | $8.3B[15] | ~$1B ARR (Sacra est.)[17] | ~8× |
| Modal | $355M Series C, May 2026[53] | $4.65B[54] | >$300M ARR[53] | >15× |
| Groq | $350M Aug 2026 at a markdown to $3.5B[36] | $3.5B (was $6.9B[39]) | ~$172M ARR (est.)[104] | ~20× |
| DeepInfra | $107M Series B, May 2026[32] | undisclosed | n/d | n/d |
| Parasail | $32M Series A, Apr 2026[47] | undisclosed | n/d | n/d |
Groq is the cautionary case study: peak valuation $6.9B (Sep 2025)[39], then Nvidia's $20B licensing "not-acqui-hire", a $7.6B shareholder payout[103] — and an August 2026 round at $3.5B, a 50% markdown[36]. The first marked-down print in the category.
Chinese labs monetize the same weights from cheaper positions (state funds and IPO markets, not US VC): DeepSeek closed a maiden ~$7.4B round at up to ~$50B — structured through an LP controlled by founder Liang Wenfeng with a five-year lockup[66][67]. Moonshot raised $3.5B at $35B anchored by China's national AI fund[82]. Zhipu (GLM) listed in Hong Kong in January 2026 and peaked at a ~$125B market cap[81][101]; MiniMax listed at ~$10.8B[84][85]; StepFun closed $2.5B at ~$10B[86].
And the GPU layer beneath is debt, not equity: CoreWeave carries $35.6B of GPU-collateralized debt against a $1.37B quarterly net loss[144], with ~$104B backlog[62] and $20B+ raised in 2026 capital alone[88]. Nebius raised ~$11B in under a year, mostly convertibles — $5B of it layered on a ~$1.4B TTM revenue base[106][145]. Private neocloud losses (where they exist) hit shareholders; these hit creditors.
5. Bubble, bargain, or both?
Three honest readings of the same ledger:
Not (yet) a revenue bubble. Fireworks, Together, Modal and Baseten all show $300M–$1B+ real revenue at 8–22× multiples — rich, but customers are paying. Aggregate open-weight token demand is real and exploding (29% of Vercel gateway tokens, ~46% of OpenRouter identified volume)[137][116].
A unit-economics bubble at the commodity tier. The $0.18–0.40/M price band is below derived rental cost by 3–5×. That band survives on owned depreciated fleets, credits, and the hope of moving up-market — which is exactly what Kimi K3's $15 price and DeepSeek's own August price increases look like: the market discovering its cost floor from below[156][173]. Commodity token processing prices have already collapsed >25× since early 2025[70]; the providers whose valuations were set assuming a durable margin on those tokens — at 17–22× ARR — are exposed twice: to the price floor, and to the depreciation base (H100 rental) sinking under their fleets[6].
The exposure, quantified. At least $43.5B of private US valuation across the four big inference clouds[15][19][41][54], ~$50–83B in open-model lab valuations/market caps (DeepSeek ~$50B, Moonshot $35B, Zhipu $83B post-IPO)[67][82][81], and a public debt wall (CoreWeave $35.6B, Nebius ~$11B in converts)[144][106] whose collateral is the same GPUs the $0.18/M token is supposed to feed. Groq's $6.9B→$3.5B markdown[36] is the first observable repricing; the second-order risk is that a DeepSeek price cut (not a hike) at the next release wave resets the commodity floor again — and everything rented, not owned, goes underwater with it.
Where this analysis could be wrong: the cost floors are decode-only lower bounds (prefill excluded, partly offset by KV-cache hits[68]); utilization is unobservable and banded at 40–60%; revenue figures are third-party estimates (Sacra/GetLatka), not audited; SemiAnalysis bodies are partly paywalled (excerpts labeled where used). Novita AI has no verifiable funding data at all — a genuine gap, flagged rather than filled. Even so, the directional claim survives the uncertainty band: for the commodity tier to break even on rented silicon at these prices, utilization would have to sit near theoretical maximum — which is not what the latency numbers show.
Data provenance: prices scraped from 10 provider pricing pages + OpenRouter API on 2026-09-17 (research/prices.json, 137 verified entries); cost model with explicit formulas in research/cost_model.json; funding table in research/funding.json. Underlying research memos: research/01_models.md, 02_prices.md, 03_costs.md, 04_funding.md.
Sources
[1] https://inferencex.semianalysis.com/compare/deepseek-r1-b200-vs-h100 — InferenceX (SemiAnalysis): DeepSeek R1 B200 vs H100 benchmark
[2] https://newsletter.semianalysis.com/p/deepseek-debates — SemiAnalysis: DeepSeek Debates - costs, margins
[6] https://www.thundercompute.com/blog/ai-gpu-rental-market-trends — Thunder Compute: AI GPU Rental Market Trends Sep 2026
[8] https://www.cloudzero.com/blog/h100-gpu-cost — CloudZero: Nvidia H100 price in 2026
[10] https://aimultiple.com/gpu-index — aimultiple: Cloud GPU Rental Price Index
[13] https://openrouter.ai/api/v1/models — OpenRouter models API (live pricing JSON, fetched 2026-09-17)
[14] https://openrouter.ai/rankings — OpenRouter model rankings / token share
[15] https://www.reuters.com/legal/transactional/together-ai-raises-800-million-83-billion-valuation-2026-07-01 — Reuters: Together AI raises $800 million at $8.3 billion valuation (Jul 1 2026)
[17] https://sacra.com/c/together-ai — Sacra: Together AI revenue, valuation & funding
[19] https://fireworks.ai/blog/series-d-announcement — Fireworks AI: $1.505B Series D at $17.5B valuation
[21] https://www.cnbc.com/2026-07-16/fireworks-nvidia-cloud-ai-startup-value.html — CNBC: Fireworks hits $17.5B valuation as companies pursue cheaper AI models
[22] https://sacra.com/c/fireworks-ai — Sacra: Fireworks AI revenue, valuation & funding (ARR past B)
[28] https://www.silicondata.com/products/silicon-index — Silicon Data: GPU Rental Price Indices
[32] https://deepinfra.com/series-b — DeepInfra: $107M Series B (May 4 2026)
[36] https://www.bloomberg.com/news/articles/2026-08-17/groq-valued-at-3-5-billion-in-funding-round-after-nvidia-deal — Bloomberg: Groq valued at $3.5B in funding round after Nvidia deal (Aug 17 2026)
[39] https://finance.yahoo.com/news/chip-startup-groq-raises-750-113709471.html — Reuters/Yahoo: Groq raises $750M at $6.9B valuation (Sep 2025)
[41] https://www.baseten.co/blog/announcing-our-series-f — Baseten: Series F, $1.5B at $13B valuation (Jun 22 2026)
[43] https://sacra.com/c/baseten — Sacra: Baseten revenue, valuation & funding ($2.085B total)
[47] https://siliconangle.com/2026-04-15/parasail-raises-32m-pay-per-token-inference-cloud — SiliconANGLE: Parasail raises $32M Series A for pay-per-token inference cloud
[53] https://modal.com/blog/modal-series-c — Modal blog: Series C $355M at $4.65B valuation
[54] https://www.reuters.com/business/modal-labs-valued-465-billion-ai-coding-takes-off-2026-05-21 — Reuters: Modal Labs valued at $4.65B as AI coding takes off (May 21 2026)
[62] https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx — CoreWeave IR: Q2 2026 results; revenue backlog ~$104B (Jun 30 2026)
[66] https://www.forbes.com/sites/anishasircar/2026-06-17/deepseek-just-raised-74-billion-heres-the-catch — Forbes: DeepSeek just raised $7.4B (June 2026)
[67] https://www.reuters.com/business/retail-consumer/deepseek-slated-draw-7-billion-maiden-fundraising-sources-say-2026-06-03 — Reuters: DeepSeek slated to raise ~B maiden round from Tencent, CATL (Jun 3 2026)
[68] https://www.tensoreconomics.com/p/moe-inference-economics-from-first — TensorEconomics: MoE Inference Economics from First Principles
[70] https://longyield.substack.com/p/inference-economics-the-hidden-war — LongYield: Inference Economics - The Hidden War
[81] https://www.reuters.com/world/asia-pacific/chinas-zhipu-ai-plans-apply-shanghais-sci-tech-board-issue-shares-2026-06-02 — Reuters: Zhipu plans Shanghai STAR listing; raised HK$4.35B in Jan HK IPO; mkt cap $83B (Jun 2 2026)
[82] https://www.bloomberg.com/news/articles/2026-07-29/china-s-moonshot-ai-passes-funding-goal-to-hit-35-billion-value — Bloomberg: Moonshot AI passes funding goal, $35B valuation after $3.5B round (Jul 29 2026)
[84] https://www.reuters.com/world/asia-pacific/china-ai-firm-minimax-set-surge-hong-kong-debut-2026-01-09 — Reuters: MiniMax doubles in HK debut; raised HK$4.8B ($620M) IPO (Jan 9 2026)
[85] https://www.caproasia.com/2026-07-11/china-11-billion-ai-startup-hong-kong-listed-minimax-to-raise-2-billion-in-new-shares-sale-bonds-issue-plans-china-secondary-listing-ipo-founded-in-2022-by-yan-junjie-yang-bin-zhou-yucong-cu — Caproasia: MiniMax to raise $2B in new shares/bonds; mkt cap $10.8B (Jul 2026)
[86] https://www.caproasia.com/2026-05-09/china-ai-startup-stepfun-raised-2-5-billion-in-new-funding-round-plans-hong-kong-ipo-in-2026-june-at-10-billion-valuation-founded-in-2023-by-ex-microsoft-vice-president-jiang-daxin-investors-inc — Caproasia: StepFun raised $2.5B, plans HK IPO Jun 2026 at $10B valuation
[88] https://finance.yahoo.com/markets/stocks/articles/coreweave-raised-20-billion-capital-144406639.html — Yahoo Finance: CoreWeave raised >$20B capital in 2026 (term loan, NVIDIA equity, GPU-backed facility)
[90] https://docs.gpustack.ai/2.0/performance-lab/glm-4.x/h200 — GPUStack: GLM-4.x throughput on 8x H200 (SGLang/vLLM)
[91] https://medium.com/data-science-collective/benchmarking-llm-inference-on-nvidia-b200-h200-h100-and-rtx-pro-6000-66d08c5f0162 — Medium: Benchmarking LLM inference B200/H200/H100 (GLM-4.6, Qwen3-Coder)
[93] https://www.mindstudio.ai/blog/open-source-vs-closed-source-ai-pricing-war — MindStudio: Open-weight models turn inference into a commodity market
[101] https://www.ventureatlas.org/company/zhipu-ai — Venture Atlas: Zhipu +1,700% post-IPO, mkt cap US$125-137B late Jun/Jul 2026
[103] https://sacra.com/c/groq — Sacra: Groq - $7.6B shareholder distribution Feb 2026 after Nvidia deal
[104] https://getlatka.com/companies/groq.com — GetLatka: Groq ARR $172.5M (2025)
[106] https://techfundingnews.com/nebius-4-5b-convertible-notes-share-price-drop — TFN: Nebius raised $11B in under a year; $5B converts Aug 2026
[116] https://presenc.ai/research/openrouter-model-usage-rankings-2026 — Presenc AI: OpenRouter Model Usage Rankings 2026 - Real Token Share
[117] https://artificialanalysis.ai/leaderboards/models — Artificial Analysis LLM Leaderboard - Intelligence Index (Sept 2026)
[118] https://lmarena.ai/leaderboard/text — LMArena Text Arena Leaderboard (Sep 13 2026, 8.1M votes)
[131] https://www.globaltimes.cn/page/202503/1329353.shtml — Global Times: DeepSeek 545% theoretical cost-profit margin (H800 /hr)
[132] https://officechai.com/ai/deepseek-claims-545-cost-profit-margin-understand-what-this-means — OfficeChai: DeepSeek claims 545% cost profit margin
[133] https://www.morphllm.com/llm-api-pricing — Morph: LLM API pricing 2026 across providers
[135] https://chinaresearchcollective.substack.com/p/what-if-deepseek-never-wanted-the — China Research Collective: DeepSeek pricing strategy phases
[137] https://vercel.com/blog/ai-gateway-production-index-july-2026 — Vercel AI Gateway Production Index July 2026: open-weight 29% of volume
[144] https://www.stocktitan.net/sec-filings/CRWV/10-q-core-weave-inc-quarterly-earnings-report-d3b596a1f167.html — StockTitan (10-Q): CoreWeave Q2'26 rev $2.6B, net loss $1.37B, debt $35.6B
[145] https://www.fool.com/investing/2026-08-22/nebius-just-raised-usd5-billion-in-convertible-notes-against-a-usd1-4-billion-revenue-base-are-we-in-a-bubble — Motley Fool: Nebius $5B convertible notes vs $1.4B revenue base (Aug 2026)
[156] https://api-docs.deepseek.com/quick_start/pricing — DeepSeek API Pricing (official docs)
[159] https://deepinfra.com/pricing — DeepInfra Pricing
[163] https://platform.openai.com/docs/pricing — OpenAI API Pricing
[165] https://ai.google.dev/gemini-api/docs/pricing — Google Gemini API Pricing
[173] https://www.together.ai/pricing — Together AI Serverless Pricing (per 1M tokens)
[175] https://claude.com/pricing — Anthropic / Claude Pricing (per MTok)
[178] https://openrouter.ai/api/v1/models/deepseek/deepseek-v3.2/endpoints — OpenRouter endpoints: DeepSeek V3.2
[190] https://openrouter.ai/api/v1/models/meta-llama/llama-3.3-70b-instruct/endpoints — OpenRouter endpoints: Llama 3.3 70B
[199] https://docs.fireworks.ai/serverless/pricing — Fireworks AI Serverless Pricing (docs, per 1M tokens)