06 Evidence & Methods

Who Runs Open Models?

Where each chart’s numbers come from, in one card per chart. The full calculations are in the files named on each card.

The charts

  1. Who serves

    What it shows
    The share of OpenRouter’s monthly tokens served by model makers, independent hosts and the big clouds, and the largest hosts.
    Source
    OpenRouter providers page, as displayed on 9 November 2026; JetBrains Developer Ecosystem Survey 2025.
    How we calculated it
    Monthly tokens as displayed, open and closed models together, grouped by us; providers per open model from the endpoints API.
    Main limit
    The traffic one router sends; not market share; a host’s headquarters is not where it runs the model.
    Files
    analysis/infra/p29/or_provider_volumes_2026_11_09.csv · analysis/infra/p29/or_models_open_2026_11_09.csv
  2. API or capacity

    What it shows
    Three ways to run an open model, with example prices; how many hosts charge the developer’s own price.
    Source
    Providers’ price pages and OpenRouter endpoints, 9 November 2026; GPU list prices, November 2026.
    How we calculated it
    Standard list prices on each host’s own page, the developer’s price as reference; median on-demand GPU price.
    Main limit
    List prices only; discounts and committed-use prices are not public; cheaper OpenRouter endpoints often declare 4-bit precision.
    Files
    analysis/infra/p29/primary_prices_focus_summary_2026_11_09.csv · analysis/infra/p29/or_models_open_2026_11_09.csv
  3. When it pays

    What it shows
    The cost of a rented node per million output tokens against the share of the month it runs, compared with the developer’s API; the same break-even for other models.
    Source
    Providers’ price pages; SemiAnalysis InferenceX throughput measurements.
    How we calculated it
    A modelled scenario: node cost at the median on-demand price (730 hours) against the developer’s API price with three input tokens per output token; break-even utilisation = throughput needed at full use / measured throughput.
    Main limit
    Not the total cost of production: staff, storage, network, redundancy, set-up and idle peaks are not counted; measured throughput comes from agentic coding traces; no Helmcode figure is used.
    Files
    analysis/infra/p29/breakeven_models_headline.csv · analysis/infra/p29/breakeven_models.csv · analysis/infra/p29/model_serving_configs.csv
  4. Hardware and place

    What it shows
    GPU prices and memory, documented configurations, declared precision of endpoints, formats of downloaded copies, and Europe.
    Source
    Providers’ price pages; model cards and vLLM and SGLang recipes; OpenRouter endpoints (as displayed); Hugging Face hub-stats; EU providers’ price pages.
    How we calculated it
    Medians and ranges of public on-demand prices; price per GB of memory; shares of endpoints by declared precision; downloads of quantised copies by format; EU-headquartered providers’ share of the page’s tokens.
    Main limit
    Price per GB is not an inference cost; declared precision is not verified; GGUF downloads are inflated by how the Hub counts files; Vast.ai is left out; the EU share is of one router’s traffic.
    Files
    analysis/infra/p29/gpu_price_summary_2026_11.csv · analysis/infra/p29/or_precision_open_2026_11_09.csv · analysis/infra/p29/quant_formats_2026_10_22.csv · analysis/infra/europe_p29/or_eu_share_summary_2026_11_09.csv

More data

Figures from the research that the chapter’s charts do not show.

  • Other GPUs, median public on-demand list price per GPU and hour in November 2026: AMD MI300X $4.73, L40S $1.57. Spot capacity, which can be withdrawn at short notice, was cheaper: a median $4.90 for a B200 and $2.68 for an H100.
  • The throughput behind the break-even scenario was measured on agentic coding sessions with about 127,000 input tokens and 986 output tokens per request, around 129 to one, with 84% of the input served from cache; the API price assumes three input tokens per output token. The two sides do not describe the same workload.

Sources and notes

Show the 20 notes

  1. 1
    OpenRouter, providers page (https://openrouter.ai/providers), captured 2026-11-09 (research/editorial/extreme-edit-pass29/ch06-infra/raw/openrouter/2026-11-09/pages/providers.html, SHA-256 f79bf51e…7a83), and OpenRouter endpoints API, same date. Figures used as displayed; the page gives no dates or window, and the monthly column is read as a rolling month (our inference). OpenRouter's terms reserve the materials (ch06-infra/sources.md, P1 and P2).
  2. 2
    analysis/infra/p29/or_provider_volumes_2026_11_09.csv and or_provider_volumes_summary_2026_11_09.csv: sum of monthly tokens 797.0T over 97 providers; shares 0.52 (first-party lab), 0.39 (third-party host), 0.09 (hyperscaler: Google Vertex, Claude Platform on AWS, Amazon Bedrock, Azure); hosts with ≥90% of endpoints on open or downloadable weights 0.35; top 10 third-party hosts 0.73 of third-party-host tokens. Classification (kind_our_classification) is ours. Fifteen largest by monthly_tokens; TypeSafe gives no headquarters. Traffic routed by OpenRouter, not market share. Method: ch06-infra/scripts/parse_or_providers_page.py.
  3. 3
    analysis/infra/p29/or_models_open_2026_11_09.csv and key_numbers.csv (or_verified_open_models_with_endpoints 99, or_median_providers_per_verified_open_model 2, or_verified_open_models_with_1_provider 39, or_top10_open_by_tokens_median_providers 17.5). "Verified open": classes OW-P/OW-N in the effective openness registry with DEC-099 suspensions. Token ranking from OpenRouter rankings-daily, 2026-10-01 to 14 (CC BY 4.0); provider counts from endpoints on 2026-11-09. Method: ch06-infra/scripts/openrouter_providers.py.
  4. 4
    JetBrains Developer Ecosystem Survey 2025, analysis/adoption/jetbrains_t2b_route_by_family.csv, family "DeepSeek (R1, V3)": website or app 70.5% (95% CI 67.3–73.8), developer's API 27.8% (24.6–31.0), weighted, base 1,267 respondents; options are not exclusive. Survey method as in the previous version of this chapter (research/editorial/editorial-rewrite/before/06-infrastructure.md).
  5. 5
    analysis/infra/p29/primary_prices_focus_summary_2026_11_09.csv, column share_at_developer_price (Kimi K3 0.89 of 9; GLM-5.3 0.7 of 10; DeepSeek-V4.1-Flash 0.8 of 10; Qwen3.8-27B 0.14 of 7). Standard list price on each host's own page or API, one price per host, 2026-11-09, from ch06-infra/raw/web/model_prices_primary.csv (169 rows). Method: ch06-infra/scripts/primary_prices.py.
  6. 6
    analysis/infra/p29/or_models_open_2026_11_09.csv (GLM-5.3: 34 providers, output 1.40–8.80; Kimi K3: 21, 9.50–22.50) and key_numbers.csv (spread_models_with_5plus_priced_providers 42, spread_median_max_over_min_output 2.62). OpenRouter endpoints, 2026-11-09, as displayed. Not a like-for-like comparison: endpoints differ in precision, speed and tier.
  7. 7
    analysis/infra/p29/focus_prices_summary_2026_11_09.csv, cheapest_provider and cheapest_declared_precision (Kimi K3: Wafer, mxfp4; GLM-5.3: Wafer, fp4). DeepInfra promotion, DeepSeek off-peak tariff and cloud regional variants: research/editorial/extreme-edit-pass29/ch06-infra/ch06-infra.md §1e and §2a, from the providers' pages on 2026-11-09.
  8. 8
    analysis/infra/p29/gpu_price_summary_2026_11.csv, commitment = on-demand: columns n_list_prices, min, median, max. Combines the 2026-10-26 observations used in Chapter 04 with 149 rows collected on 2026-11-09 (ch06-infra/raw/web/gpu_prices_2026_11.csv). Per GPU and hour, USD, list prices; the big clouds sell eight-GPU instances with CPU, memory and disk included. Method: ch06-infra/scripts/gpu_price_summary.py.
  9. 9
    Spot medians: gpu_price_summary_2026_11.csv, commitment = spot. Announced increases: Lambda (https://lambda.com/pricing) and Crusoe (https://www.crusoe.ai/cloud/pricing), both read on 2026-11-09 (ch06-infra/ch06-infra.md §3a); a fact about two price lists, not a market trend. Vast.ai: terms of 2026-09-01; its rows are kept out of every median (ch06-infra/sources.md, P3; column vast_internal_only_median).
  10. 10
    analysis/infra/p29/model_serving_configs.csv (gpus, nodes, precision, hardware_basis, with the source row of each configuration in ch06-infra/raw/web/models_serving_observations.csv and models_serving_notes.md). On B200, B300 and RTX PRO 6000 the configurations use NVFP4, a third-party conversion for every model: memory from the NVFP4 variants of the vLLM recipes (Kimi K3 1,650 GB; GLM-5.3 558 GB, within four B200s by our arithmetic; GLM-5.3-Flash 229 GB; Qwen3.5-397B-A17B 238 GB; Qwen3.8-27B 24.6 GiB) and profiles from the Nvidia NIM support matrix, a vendor source and the only one for DeepSeek-V4-Pro and DeepSeek-V4.1-Flash at NVFP4. H200 and H100 rows keep the developer's release. Parameters from the model cards. Monthly cost = GPUs × median on-demand price in gpu_price_summary_2026_11.csv × 730 hours, rounded to the nearest $100, as in the chapter's visual report (pipelines/web/ch06_2026.py, gpu_widget). B300 median from two list prices.
  11. 11
    analysis/infra/p29/breakeven_models_headline.csv (rows with api_price = developer_api) and the full grid breakeven_models.csv. Break-even V* = monthly node cost ÷ API price per output token including R input tokens (R = 1, 3 or 10; 3 in the headline). Node cost = GPUs × price × 730 h; B200 at the 2026-10-26 median of $8.35 (node of 8: $48,735 a month; Kimi K3's 16: $97,470). Sensitivities from the same file (gpu_price min, max, spot; input_per_output_R; api_price = openrouter_cheapest). Throughput never enters the calculation. Column excluded_costs: labour; storage; egress; load balancing; redundancy; setup engineering; model updates. Label (DEC-114): "Modelled infrastructure break-even — not observed production cost". Published under the owner's lifting of DEC-110 rule 7 (external review), 2026-11-09. Method: ch06-infra/scripts/breakeven_models.py. Break-even utilisation = throughput needed per GPU at full use / measured throughput per GPU (pipelines/web/ch06_2026.py, be_rows()): DeepSeek-V4-Pro on 8 × B200, 292.7 / 523.0 = 0.56; Qwen3.5-397B-A17B on 2 × B200, 429.3 / 627.4 = 0.68; DeepSeek-V4.1-Flash on 8 × B200, 1,103.8 / 2,799.5 = 0.39; Kimi K3 on 16 × B200, 96.6 / 56.3 (not reached); GLM-5.3-Flash on 8 × B200 FP4, 2,440.1 / 264.7 (not reached). Cost per million output tokens at full use: $48,735 / (523 × 8 × 3,600 × 730 / 10⁶) = $4.43. DeepSeek-V4-Pro developer price: input $1.32, output $3.96 per million (primary_prices_focus_summary_2026_11_09.csv), $3.96 + 3 × $1.32 = $7.92. Against the cheapest OpenRouter endpoint (3.96 for the same mix), the throughput needed is 585.4, above the 523 measured. Sensitivity for DeepSeek-V4-Pro on 8 × B200 (breakeven_models.csv, throughput needed at full use / 523): spot median 172.0 (0.33); highest on-demand price 499.5 (0.96); R = 10, 135.1 (0.26); R = 1, 439.0 (0.84). Single-turn measurement: 381.3 output tokens per second per GPU, 8 × B200, TensorRT-LLM, 8,192 input / 1,024 output tokens, at least 50 tokens per second per user (ch06-infra/raw/web/models_serving_observations.csv); 292.7 / 381.3 = 0.77 (ch06_2026.sensitivity()).
  12. 12
    Needed throughput: breakeven_models_headline.csv, needed_output_tok_s_per_gpu_at_100pct and _at_50pct. Published throughput: reported_tok_s_per_gpu, reported_conditions and reported_source in model_serving_configs.csv, each an observation from SemiAnalysis InferenceX (https://inferencex.semianalysis.com, accessed 2026-11-09; rows of ch06-infra/raw/web/models_serving_observations.csv). GLM-5.3: the only independent row is 8 × B300 FP4 with synthetic MTP acceptance; GLM-5.3-Flash at 265 is on 8 × B200 FP4. InferenceX labels its Blackwell runs FP4 without naming the variant (NVFP4 or MXFP4), so a measured throughput is not tied to the NVFP4 conversion. Kimi K3 retired from InferenceX on 2026-11-03, DeepSeek-V4-Pro on 2026-10-29. InferenceX code is Apache-2.0; the dashboard reserves its rights; observations cited individually (sources.md, P5). No MLPerf figure is used in this comparison. Agentic traces of 85,000 to 240,000 input tokens per request (ch06-infra/ch06-infra.md §3); the 8 × B200 vLLM run behind the 523: mean input 127,081 and output 986 tokens per request, server GPU cache hit 0.84 (ch06-infra/raw/web/models_serving_observations.csv).
  13. 13
    David Abitbol (Revolut), talk at AI Engineer, New York, published on YouTube by AI Engineer on 2026-10-29; automatic captions in ch06-infra/raw/web/alt/talk_transcript_Ok6QdZHCcI8_asr.txt; notes in ch06-infra/raw/web/revolut_glm_notes.md; key_numbers.csv (revolut_engineers 2000, revolut_b200_gpus 56, revolut_blended_usd_per_m 0.3). Single primary source; press coverage (The Register, 2026-11-03) repeats the talk. Self-reported; no token volume is derived from it.
  14. 14
    analysis/infra/p29/or_precision_open_2026_11_09.csv: 595 endpoints; fp8 0.43 (token-weighted 0.49); fp4 0.18 + mxfp4 0.02 + nvfp4 0.02 + int4 0.02 = 0.24; bf16 0.08 + fp16 0.01 + fp32 <0.01 = 0.09; unknown 0.25. OpenRouter endpoints, 2026-11-09, as displayed; declared by each provider, not verified.
  15. 15
    analysis/infra/p29/quant_formats_2026_10_22.csv (share_of_downloads: GGUF 0.55, MLX 0.11, AWQ 0.10, NVFP4/FP4 0.08, FP8 0.07) and key_numbers.csv (hub_quant_bits_*: 4 bit 0.28, 8 bit 0.10, multiple in one repo 0.52). Population: 103,023 repositories declaring a "quantized" relation to a model on the ATOM list, 313.1 million 30-day downloads, Hugging Face hub-stats read on 2026-10-22. Format classified from the repository name; sensitivity without the largest repositories in quant_formats_sensitivity_2026_10_22.csv. Counting difference: ch06-infra/ch06-infra.md §4a. Method: ch06-infra/scripts/quant_formats.py.
  16. 16
    Moonshot AI, K2 Vendor Verifier, 2025-11-15, as summarised in ch06-infra/ch06-infra.md §4d.
  17. 17
    research/editorial/extreme-edit-pass29/europe/data/eu_focus_price_summary_2026_11_09.csv (also analysis/infra/europe_p29/), column eu_median_over_first_party: GLM-5.3 1.06×, DeepSeek V4.1 Flash 1.06×, Kimi K3 1.0×, Qwen3.8-27B 0.98×, GLM-5.3-Flash 1.41×. Providers headquartered in the EU/EEA with a declared European location for the model (eu_providers_catalog_2026_11_09.csv, 68 rows); output price plus three times input, USD per million, VAT excluded, euros at the ECB rate of 2026-11-06 (1 € = $1.12, our conversion). Helmcode not included. Method: europe/scripts/eu_price_comparison.py.
  18. 18
    analysis/infra/europe_p29/eu_gpu_compare_2026_11_09.csv: H100 6 European list prices, median $3.76, against $4.46 (18); H200 4, median $5.93, against $6.15 (10). Per GPU and hour, on demand; euro prices converted at the ECB rate of 2026-11-06. Method: europe/scripts/eu_gpu_compare.py.
  19. 19
    analysis/infra/europe_p29/or_eu_share_summary_2026_11_09.csv: EU/EEA-headquartered providers (Mistral, Nebius, Inceptron, NextBit) 0.01 of all displayed tokens; EU hosts 0.01 of third-party-host tokens (Inceptron reclassified as a host). Headquarters as displayed by OpenRouter (note 1). europe/data/mistral_shares.csv: hf_tracked_downloads_share_mistralai 0.02 (30-day counter to 2026-10-01, rank 8; Mistral is the only developer with region "eu" on the tracked list); OpenRouter rankings-daily top-50 share 0.04 (2025-01) and <0.01 (2026-09), CC BY 4.0.
  20. 20
    Mistral model pages ("served without Mistral modifications"), and provider pages and regions, 2026-11-09: europe/europe.md §0 and §2a–2b; europe/data/nebius_catalog_regions_2026_11_09.csv (31 models: 10 eu-north1, 1 eu-west2, 3 uk-south1, 17 in US regions). Residency and sovereignty statements quoted as the providers' own. Revolut's headquarters and location statement: ch06-infra/raw/web/revolut_glm_notes.md.