{ "meta": { "model_version": "0.3.0-public", "as_of": "2026-08-19", "generated_at": "2026-08-19T20:14:55Z", "channel_checks_used": false, "forced_to_balance": false, "assumption_count": 49 }, "principles": [ "No proprietary channel checks or private customer interviews.", "Observed, disclosure-derived, benchmark-derived, and modeled values remain separate.", "Application, API, router, and infrastructure views are not additive unless an explicit attribution boundary says they are.", "Supply capacity is a feasibility envelope, not demand and not revenue.", "The reconciliation reports residuals and never forces supply and demand to balance." ], "components": { "observed_demand": { "title": "Observed and disclosed demand", "formula": "\u03a3 non-overlapping disclosed tokens/day; overlapping platform, geography, and product views remain cross-checks.", "methodology": "Pin the freshest compatible provider or model-family disclosures. ByteDance's June Doubao-family row supersedes the March China aggregate in the additive partition; the China total remains a dated cross-check. Google remains one all-surfaces row, while its model-API and Gemini-app metrics stay contained. OpenRouter is metered but non-additive because host and lab disclosures may contain the same tokens.", "sources": [ { "name": "Google all AI surfaces", "url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "type": "company_disclosure" }, { "name": "OpenAI API platform", "url": "https://openai.com/index/accelerating-the-next-phase-ai/", "as_of": "2026-03-31", "type": "company_disclosure" }, { "name": "China aggregate", "url": "https://www.ecns.cn/cns-wire/2026-03-24/detail-ihfaytev9463369.shtml", "as_of": "2026-03-16", "type": "curated_disclosure" }, { "name": "ByteDance Doubao model family", "url": "https://en.jiemian.com/article/14629930.html", "as_of": "2026-06-23", "type": "company_statement_via_press" }, { "name": "Fireworks AI", "url": "https://fireworks.ai/", "as_of": "2026-06-12", "type": "company_disclosure" }, { "name": "Together AI", "url": "https://tomtunguz.com/trillion-token-race/", "as_of": "2025-09-15", "type": "curated_disclosure" }, { "name": "Google model APIs", "url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "as_of": "2026-07-22", "type": "company_disclosure" }, { "name": "Microsoft Foundry customer floor", "url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3", "as_of": "2026-04-29", "type": "company_disclosure" }, { "name": "OpenRouter data API", "url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "type": "platform_metered" } ], "rows": [ { "id": "google_all_surfaces_tokens_tpd", "channel": "Google all AI surfaces", "scenario_tpd": { "low": 106.6666666667, "base": 106.6666666667, "high": 106.6666666667 }, "additive": true, "confidence": "high", "measurement_class": "company_disclosure", "as_of": "2026-05-19", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "rationale": "Google disclosed more than 3.2 quadrillion monthly tokens across its surfaces; the daily value divides by 30 days. It includes first-party products and APIs, so no separate Google product row is added." }, { "id": "openai_reported_tokens_tpd", "channel": "OpenAI API platform", "scenario_tpd": { "low": 21.6, "base": 21.6, "high": 21.6 }, "additive": true, "confidence": "high", "measurement_class": "company_disclosure", "as_of": "2026-03-31", "source_url": "https://openai.com/index/accelerating-the-next-phase-ai/", "rationale": "OpenAI disclosed more than 15B API tokens per minute; 15B \u00d7 1,440 minutes/day = 21.6T/day if sustained. OpenAI labels the boundary API-platform traffic, so a separately modeled ChatGPT consumer estimate can be added without treating the API row as an all-surfaces total." }, { "id": "china_aggregate_tokens_tpd", "channel": "China aggregate", "scenario_tpd": { "low": 100.0, "base": 140.0, "high": 180.0 }, "additive": false, "confidence": "medium", "measurement_class": "curated_disclosure", "as_of": "2026-03-16", "source_url": "https://www.ecns.cn/cns-wire/2026-03-24/detail-ihfaytev9463369.shtml", "rationale": "March country-level cross-check. It is superseded for the current additive partition by ByteDance's later June Doubao-family disclosure plus an explicitly modeled non-Doubao China residual; adding all three would double count." }, { "id": "bytedance_doubao_tokens_tpd", "channel": "ByteDance Doubao model family", "scenario_tpd": { "low": 180.0, "base": 180.0, "high": 180.0 }, "additive": true, "confidence": "medium", "measurement_class": "company_statement_via_press", "as_of": "2026-06-23", "source_url": "https://en.jiemian.com/article/14629930.html", "rationale": "Volcano Engine president Tan Dai said Doubao-family daily token usage reached 180T at the June FORCE conference. The scope is the model family across consumer and enterprise surfaces and can include multimodal token accounting." }, { "id": "fireworks_tokens_tpd", "channel": "Fireworks AI", "scenario_tpd": { "low": 15.0, "base": 30.0, "high": 35.0 }, "additive": true, "confidence": "medium", "measurement_class": "company_disclosure", "as_of": "2026-06-12", "source_url": "https://fireworks.ai/", "rationale": "Platform disclosure increased from 15T/day to 30T/day. We do not separately add Fireworks customer application estimates." }, { "id": "together_tokens_tpd", "channel": "Together AI", "scenario_tpd": { "low": 1.5, "base": 2.0, "high": 3.0 }, "additive": true, "confidence": "low", "measurement_class": "curated_disclosure", "as_of": "2025-09-15", "source_url": "https://tomtunguz.com/trillion-token-race/", "rationale": "Secondary report of 2T/day. The row is aged and receives a broad range; customers are not separately added." }, { "id": "google_model_api_tokens_tpd", "channel": "Google model APIs", "scenario_tpd": { "low": 31.68, "base": 31.68, "high": 31.68 }, "additive": false, "confidence": "high", "measurement_class": "company_disclosure", "as_of": "2026-07-22", "source_url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "rationale": "Alphabet disclosed approximately 22B model-API tokens per minute; 22B \u00d7 1,440 = 31.68T/day if sustained. This is a contained cross-check within Google's all-surfaces disclosure and is not added again." }, { "id": "openrouter_metered_7d", "channel": "OpenRouter public platform", "scenario_tpd": { "low": 11.148303367825143, "base": 11.148303367825143, "high": 11.148303367825143 }, "additive": false, "confidence": "high", "measurement_class": "platform_metered", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "rationale": "Direct public-platform meter; excluded from the provider total because the same tokens may be represented in model-provider or API-host disclosures." }, { "id": "microsoft_foundry_customer_floor", "channel": "Microsoft Foundry customer floor", "scenario_tpd": { "low": 0.8219178082191781, "base": 0.8219178082191781, "high": 0.8219178082191781 }, "additive": false, "confidence": "high", "measurement_class": "company_disclosure_floor", "as_of": "2026-04-29", "source_url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3", "rationale": "300 customers \u00d7 the disclosed minimum 1T annual tokens \u00f7 365. Cross-check only because Foundry traffic overlaps model-provider totals." } ], "summary": { "additive_total_tpd": { "low": 324.76666666669996, "base": 340.26666666669996, "high": 346.26666666669996 }, "router_crosscheck_tpd": 11.148303367825143 }, "caveats": [ "Disclosure dates and token conventions differ.", "ByteDance and China totals can include multimodal token accounting.", "Some company statements are available only through reputable secondary reporting." ] }, "modeled_app_demand": { "title": "Modeled first-party application and financial demand", "formula": "users \u00d7 active share \u00d7 interactions \u00d7 processed tokens; or annualized revenue \u00f7 realized $/token \u00f7 365; China residual = max(current China total \u2212 Doubao, 0)", "methodology": "Use the freshest direct audience, message, SEC, and revenue numerators. ChatGPT consumer traffic is now additive to OpenAI's explicitly API-only disclosure. Anthropic is counted once at the lab boundary from the latest Q2/July revenue pace and a Claude-specific realized-price range; AWS and Vertex reseller totals are not separately added. Gemini app is a contained cross-check against Google's all-surfaces total. Grok uses the 117M Grok-feature MAU from SpaceX's S-1. The China residual explicitly subtracts Doubao before addition.", "sources": [ { "name": "Gemini app cross-check", "url": "https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/", "as_of": "2026-08-11", "type": "company_disclosure" }, { "name": "Gemini app cross-check", "url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "as_of": "2026-08-11", "type": "modeled_from_company_growth_disclosure" }, { "name": "Gemini app cross-check", "url": "https://inferencex.semianalysis.com/about", "as_of": "2026-08-19", "type": "benchmark_informed_assumption" }, { "name": "Google AI Overviews reference", "url": "https://blog.google/alphabet/investor-presentation-june-2026/", "as_of": "2026-06-01", "type": "company_disclosure_reference_only" }, { "name": "ChatGPT consumer estimate", "url": "https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/", "as_of": "2026-08-06", "type": "modeled_from_company_disclosures" }, { "name": "ChatGPT consumer estimate", "url": "https://openai.com/signals/data/", "as_of": "2026-08-18", "type": "modeled_workload_assumption" }, { "name": "Anthropic financial cross-check", "url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "as_of": "2026-07-31", "type": "reported_investor_update" }, { "name": "Anthropic Q2 recognized-revenue anchor", "url": "https://finance.yahoo.com/technology/ai/articles/anthropic-revenue-surges-over-11-210857853.html", "as_of": "2026-06-30", "type": "reported_company_documents" }, { "name": "Anthropic financial estimate", "url": "https://docs.anthropic.com/en/docs/about-claude/pricing", "as_of": "2026-08-19", "type": "modeled_from_public_pricing" }, { "name": "Microsoft 365 Copilot seat cross-check", "url": "https://www.sec.gov/Archives/edgar/data/789019/000119312526323632/msft-ex99_1.htm", "as_of": "2026-07-29", "type": "sec_filing" }, { "name": "Microsoft 365 Copilot seat cross-check", "url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4", "as_of": "2026-08-18", "type": "modeled_workload_assumption" }, { "name": "Microsoft 365 Copilot seat cross-check", "url": "https://inferencex.semianalysis.com/about", "as_of": "2026-08-18", "type": "benchmark_informed_assumption" }, { "name": "Meta AI", "url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "as_of": "2025-10-01", "type": "company_disclosure" }, { "name": "Meta AI", "url": "https://about.fb.com/news/2024/09/metas-ai-product-news-connect/", "as_of": "2026-08-18", "type": "modeled_from_company_disclosure" }, { "name": "Meta AI", "url": "https://about.fb.com/news/2025/04/introducing-meta-ai-app-new-way-access-ai-assistant/", "as_of": "2026-08-18", "type": "modeled_workload_assumption" }, { "name": "Grok consumer estimate", "url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "as_of": "2026-03-31", "type": "sec_filing" }, { "name": "Grok consumer estimate", "url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "as_of": "2026-08-19", "type": "modeled_workload_assumption" }, { "name": "China current total estimate", "url": "https://news.cgtn.com/news/2026-06-18/China-builds-computing-network-as-AI-token-usage-skyrockets-1O4tgBLAnxS/index.html", "as_of": "2026-06-23", "type": "modeled_from_national_and_company_disclosures" } ], "rows": [ { "id": "chatgpt_consumer_estimate", "channel": "ChatGPT consumer estimate", "scenario_tpd": { "low": 1.4, "base": 5.1428571432, "high": 15.0 }, "additive": true, "confidence": "low", "source_url": "https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/", "formula_inputs": [ "openai_chatgpt_messages_per_day", "openai_tokens_per_message" ], "rationale": "Current weekly users \u00d7 the prior disclosed messages/user rate \u00d7 processed tokens/message. Additive to the API disclosure because OpenAI labels that disclosure API-only." }, { "id": "google_gemini_app_crosscheck", "channel": "Gemini app cross-check", "scenario_tpd": { "low": 0.02, "base": 0.54, "high": 7.2 }, "additive": false, "confidence": "low", "source_url": "https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/", "formula_inputs": [ "google_gemini_monthly_users", "google_gemini_daily_active_share", "google_gemini_interactions_per_active_day", "google_gemini_tokens_per_interaction" ], "rationale": "Latest 1B MAU converted through explicit engagement assumptions. Cross-check only because it is contained in Google's all-surfaces 106.7T/day disclosure." }, { "id": "meta_ai_modeled", "channel": "Meta AI", "scenario_tpd": { "low": 0.03, "base": 0.405, "high": 3.0 }, "additive": true, "confidence": "low", "source_url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "formula_inputs": [ "meta_ai_monthly_users", "meta_ai_daily_active_share", "meta_ai_interactions_per_active_day", "meta_ai_tokens_per_interaction" ], "rationale": "User-driven application estimate; excludes ranking, ads, image/video, and non-Meta Llama hosting." }, { "id": "anthropic_financial_estimate", "channel": "Anthropic / Claude financial estimate", "scenario_tpd": { "low": 15.753424657534246, "base": 50.68493150684931, "high": 178.08219178082192 }, "additive": true, "confidence": "low", "source_url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "formula_inputs": [ "anthropic_run_rate_revenue_usd", "anthropic_realized_price_per_million_tokens" ], "rationale": "Fresh Q2/July revenue pace \u00f7 a Claude-specific realized-price range \u00f7 365. Counted once at the model-lab boundary; AWS and Vertex reseller totals are not otherwise added." }, { "id": "xai_grok_consumer_estimate", "channel": "xAI / Grok consumer estimate", "scenario_tpd": { "low": 0.005616, "base": 0.06318, "high": 0.5265 }, "additive": true, "confidence": "low", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "formula_inputs": [ "xai_grok_monthly_users", "xai_grok_daily_active_share", "xai_grok_interactions_per_active_day", "xai_grok_tokens_per_interaction" ], "rationale": "Uses the Grok-specific 117M MAU disclosed in SpaceX's S-1 rather than combined X/Grok reach. Multimodal generation is excluded." }, { "id": "china_non_doubao_residual", "channel": "China ex-Doubao residual", "scenario_tpd": { "low": 20.0, "base": 40.0, "high": 70.0 }, "additive": true, "confidence": "low", "source_url": "https://news.cgtn.com/news/2026-06-18/China-builds-computing-network-as-AI-token-usage-skyrockets-1O4tgBLAnxS/index.html", "formula_inputs": [ "china_current_total_tokens_tpd", "bytedance_doubao_tokens_tpd" ], "rationale": "Current China total envelope minus Doubao, floored at zero. This prevents national and ByteDance disclosures from being added twice." }, { "id": "microsoft_copilot_seat_crosscheck", "channel": "Microsoft 365 Copilot seat cross-check", "scenario_tpd": { "low": 0.0075, "base": 0.12, "high": 1.08 }, "additive": false, "confidence": "low", "source_url": "https://www.sec.gov/Archives/edgar/data/789019/000119312526323632/msft-ex99_1.htm", "formula_inputs": [ "microsoft_copilot_paid_seats", "microsoft_copilot_interactions_per_seat_day", "microsoft_copilot_tokens_per_interaction" ], "rationale": "Paid seats \u00d7 interactions/seat/day \u00d7 processed tokens/interaction. Excluded because OpenAI/Anthropic lab totals may contain the same inference." } ], "summary": { "additive_total_tpd": { "low": 37.18904065753425, "base": 96.2959686500493, "high": 266.6086917808219 }, "chatgpt_consumer_tpd": { "low": 1.4, "base": 5.1428571432, "high": 15.0 }, "gemini_app_crosscheck_tpd": { "low": 0.02, "base": 0.54, "high": 7.2 }, "anthropic_financial_tpd": { "low": 15.753424657534246, "base": 50.68493150684931, "high": 178.08219178082192 }, "xai_grok_consumer_tpd": { "low": 0.005616, "base": 0.06318, "high": 0.5265 }, "china_non_doubao_residual_tpd": { "low": 20.0, "base": 40.0, "high": 70.0 }, "microsoft_copilot_crosscheck_tpd": { "low": 0.0075, "base": 0.12, "high": 1.08 } }, "caveats": [ "Per-interaction token distributions are not public.", "ARR is a run-rate metric, not audited annual revenue or token volume.", "Image, video, voice, ads, ranking, and recommender inference remain mostly excluded." ] }, "relative_strength": { "title": "Modeled aggregate relative strength by lab and model family", "formula": "global additive denominator partitioned once: direct lab totals + lab-specific application/financial estimates + normalized allocation of the hosted/open-model pool", "methodology": "Start with the same low/base/high global demand denominator used in reconciliation. Pin ByteDance and Google direct lab-family totals; combine OpenAI's API disclosure with its separately modeled ChatGPT consumer surface; backsolve Anthropic from fresh revenue using a lab-specific price range; model Meta and xAI from first-party users. Allocate only the residual pool\u2014China ex-Doubao plus Fireworks and Together\u2014across named open-model labs using normalized app/router priors. Base shares sum to exactly 100%. Low/high shares are uncertainty envelopes and are not designed to sum to 100%. Exact model-version splits remain unassigned because labs do not disclose production routing.", "sources": [ { "name": "ByteDance Doubao model family", "url": "https://en.jiemian.com/article/14629930.html", "as_of": "2026-06-23", "type": "company_statement_via_press" }, { "name": "Google all AI surfaces", "url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "type": "company_disclosure" }, { "name": "OpenAI API platform", "url": "https://openai.com/index/accelerating-the-next-phase-ai/", "as_of": "2026-03-31", "type": "company_disclosure" }, { "name": "ChatGPT consumer estimate", "url": "https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/", "as_of": "2026-08-06", "type": "modeled_from_company_disclosures" }, { "name": "Anthropic financial cross-check", "url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "as_of": "2026-07-31", "type": "reported_investor_update" }, { "name": "Anthropic financial estimate", "url": "https://docs.anthropic.com/en/docs/about-claude/pricing", "as_of": "2026-08-19", "type": "modeled_from_public_pricing" }, { "name": "Meta AI", "url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "as_of": "2025-10-01", "type": "company_disclosure" }, { "name": "Grok consumer estimate", "url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "as_of": "2026-03-31", "type": "sec_filing" }, { "name": "China current total estimate", "url": "https://news.cgtn.com/news/2026-06-18/China-builds-computing-network-as-AI-token-usage-skyrockets-1O4tgBLAnxS/index.html", "as_of": "2026-06-23", "type": "modeled_from_national_and_company_disclosures" }, { "name": "Fireworks AI", "url": "https://fireworks.ai/", "as_of": "2026-06-12", "type": "company_disclosure" }, { "name": "Together AI", "url": "https://tomtunguz.com/trillion-token-race/", "as_of": "2025-09-15", "type": "curated_disclosure" }, { "name": "DeepSeek", "url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "as_of": "2026-08-17", "type": "modeled_blend_of_app_and_router_proxies" }, { "name": "Tencent", "url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "type": "modeled_blend_of_app_and_router_proxies" } ], "rows": [ { "lab": "ByteDance", "model_family": "Doubao / Seed family", "scenario_tpd": { "low": 180.0, "base": 180.0, "high": 180.0 }, "base_share_pct": 41.231196955140355, "share_range_pct": { "low": 29.369756430729904, "base": 41.231196955140355, "high": 49.729841623621326 }, "evidence_class": "company_disclosure", "confidence": "medium", "source_url": "https://en.jiemian.com/article/14629930.html", "as_of": "2026-06-23", "momentum_pct": 50.0, "momentum_period": "Mar\u2013Jun token growth", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "Volcano Engine president Tan Dai said Doubao-family daily token usage reached 180T at the June FORCE conference. The scope is the model family across consumer and enterprise surfaces and can include multimodal token accounting." }, { "lab": "Google", "model_family": "Google model family", "scenario_tpd": { "low": 106.6666666667, "base": 106.6666666667, "high": 106.6666666667 }, "base_share_pct": 24.433301899350067, "share_range_pct": { "low": 17.40430010710464, "base": 24.433301899350067, "high": 29.469535776969995 }, "evidence_class": "company_disclosure", "confidence": "high", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "momentum_pct": 37.5, "momentum_period": "Q1\u2013Q2 API-token growth proxy", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "Google disclosed more than 3.2 quadrillion monthly tokens across its surfaces; the daily value divides by 30 days. It includes first-party products and APIs, so no separate Google product row is added." }, { "lab": "Anthropic", "model_family": "Claude family", "scenario_tpd": { "low": 15.753424657534246, "base": 50.68493150684931, "high": 178.08219178082192 }, "base_share_pct": 11.61000218675946, "share_range_pct": { "low": 2.570412473009086, "base": 11.61000218675946, "high": 49.19999551804241 }, "evidence_class": "financial_backsolve", "confidence": "low", "source_url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "as_of": "2026-07-31", "momentum_pct": 38.3, "momentum_period": "May\u2013Jul ARR growth proxy", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "CNBC reported that Anthropic told investors its July run rate exceeded $65B; Bloomberg separately reported preliminary Q2 revenue above $11.5B, equivalent to a $46B quarterly annualization. The scenario uses annualized Q2 as low, the midpoint between Q2 pace and July run rate as base, and the reported July run rate as high. Neither figure is audited and Anthropic declined comment." }, { "lab": "OpenAI", "model_family": "GPT / Codex family", "scenario_tpd": { "low": 23.0, "base": 26.742857143200002, "high": 36.6 }, "base_share_pct": 6.125777833413673, "share_range_pct": { "low": 3.7528022105932655, "base": 6.125777833413673, "high": 10.11173446346967 }, "evidence_class": "mixed_direct_and_modeled", "confidence": "medium", "source_url": "https://openai.com/index/accelerating-the-next-phase-ai/", "as_of": "2026-03-31", "momentum_pct": 20.0, "momentum_period": "July ARR month-over-month proxy", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "OpenAI disclosed more than 15B API tokens per minute; 15B \u00d7 1,440 minutes/day = 21.6T/day if sustained. OpenAI labels the boundary API-platform traffic, so a separately modeled ChatGPT consumer estimate can be added without treating the API row as an all-surfaces total." }, { "lab": "DeepSeek", "model_family": "DeepSeek family", "scenario_tpd": { "low": 12.166666666666664, "base": 21.599999999999998, "high": 29.589041095890412 }, "base_share_pct": 4.947743634616842, "share_range_pct": { "low": 1.9851779809660024, "base": 4.947743634616842, "high": 8.17476848607474 }, "evidence_class": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior for non-Doubao China and unattributed hosted open-model tokens. Blends QuestMobile app reach with current OpenRouter model-lab share; neither is a complete token census." }, { "lab": "Alibaba", "model_family": "Qwen family", "scenario_tpd": { "low": 11.06060606060606, "base": 19.44, "high": 26.630136986301366 }, "base_share_pct": 4.452969271155158, "share_range_pct": { "low": 1.8047072554236387, "base": 4.452969271155158, "high": 7.357291637467264 }, "evidence_class": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior informed by Qwen's 167M China app MAU, Alibaba cloud distribution, and router traffic." }, { "lab": "Other / unattributed", "model_family": "Other / unallocated open models", "scenario_tpd": { "low": 5.53030303030303, "base": 10.799999999999999, "high": 18.493150684931507 }, "base_share_pct": 2.473871817308421, "share_range_pct": { "low": 0.9023536277118194, "base": 2.473871817308421, "high": 5.1092303037967115 }, "evidence_class": "modeled_residual_allocation", "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Residual allocation for Baidu, MiniMax, Xiaomi, NVIDIA, Cohere, AI21, self-hosted and other labs not independently bounded." }, { "lab": "Tencent", "model_family": "Hunyuan family", "scenario_tpd": { "low": 3.318181818181818, "base": 7.2, "high": 11.095890410958903 }, "base_share_pct": 1.649247878205614, "share_range_pct": { "low": 0.5414121766270916, "base": 1.649247878205614, "high": 3.0655381822780265 }, "evidence_class": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior for Hunyuan and Tencent-hosted demand; public absolute token volume is unavailable." }, { "lab": "Mistral", "model_family": "Mistral family", "scenario_tpd": { "low": 1.659090909090909, "base": 5.040000000000001, "high": 8.876712328767123 }, "base_share_pct": 1.15447351474393, "share_range_pct": { "low": 0.2707060883135458, "base": 1.15447351474393, "high": 2.4524305458224216 }, "evidence_class": "modeled_router_informed_allocation", "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior for hosted Mistral-family tokens. First-party Le Chat and private enterprise usage are not directly metered." }, { "lab": "Moonshot AI", "model_family": "Kimi family", "scenario_tpd": { "low": 1.659090909090909, "base": 4.32, "high": 7.397260273972603 }, "base_share_pct": 0.9895487269233685, "share_range_pct": { "low": 0.2707060883135458, "base": 0.9895487269233685, "high": 2.043692121518685 }, "evidence_class": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior for Kimi consumer, API, and open-weight use. Model releases can move router share much faster than domestic production use." }, { "lab": "Z.ai", "model_family": "GLM family", "scenario_tpd": { "low": 1.106060606060606, "base": 3.6, "high": 5.917808219178083 }, "base_share_pct": 0.824623939102807, "share_range_pct": { "low": 0.18047072554236387, "base": 0.824623939102807, "high": 1.6349536972149477 }, "evidence_class": "modeled_router_informed_allocation", "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "momentum_pct": null, "momentum_period": "See OpenRouter platform signal", "momentum_class": "not_available_globally", "version_split_available": false, "rationale": "Allocation prior informed primarily by current GLM traffic on OpenRouter; domestic and private deployments remain unobserved." }, { "lab": "Meta", "model_family": "Meta AI / Muse family", "scenario_tpd": { "low": 0.03, "base": 0.405, "high": 3.0 }, "base_share_pct": 0.0927701931490658, "share_range_pct": { "low": 0.004894959405121651, "base": 0.0927701931490658, "high": 0.828830693727022 }, "evidence_class": "application_model", "confidence": "low", "source_url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "as_of": "2025-10-01", "momentum_pct": 60.0, "momentum_period": "Q2 daily-interaction growth proxy", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "Meta disclosed more than 1B monthly Meta AI users. User count is direct; daily activity and intensity remain modeled." }, { "lab": "xAI", "model_family": "Grok family", "scenario_tpd": { "low": 0.005616, "base": 0.06318, "high": 0.5265 }, "base_share_pct": 0.014472150131254264, "share_range_pct": { "low": 0.000916336400638773, "base": 0.014472150131254264, "high": 0.14545978674909238 }, "evidence_class": "application_model", "confidence": "low", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "as_of": "2026-03-31", "momentum_pct": 31.5, "momentum_period": "Dec\u2013Mar Grok-feature MAU growth", "momentum_class": "comparable_public_signal_not_token_share", "version_split_available": false, "rationale": "SpaceX's S-1 disclosed approximately 117M MAUs that used Grok AI features. This is broader than standalone grok.com/app usage because it includes embedded X features, but it is now a clean Grok-specific numerator rather than the prior combined X/Grok reach." } ], "summary": { "global_demand_tpd": { "low": 361.9557073242342, "base": 436.56263531674927, "high": 612.8753584475219 }, "leading_lab": "ByteDance", "leading_lab_base_share_pct": 41.231196955140355, "top3_base_share_pct": 77.27450104124988, "direct_token_anchor_pct": 77.94223305890998, "lab_count": 13, "base_share_sum_pct": 100.00000000000001, "hosted_open_pool_tpd": { "low": 36.5, "base": 72.0, "high": 108.0 }, "model_family_rows": [ { "model_family": "Doubao / Seed family", "lab": "ByteDance", "scenario_tpd": { "low": 180.0, "base": 180.0, "high": 180.0 }, "base_share_pct": 41.231196955140355, "share_range_pct": { "low": 29.369756430729904, "base": 41.231196955140355, "high": 49.729841623621326 }, "confidence": "medium", "source_url": "https://en.jiemian.com/article/14629930.html", "as_of": "2026-06-23", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Google model family", "lab": "Google", "scenario_tpd": { "low": 106.6666666667, "base": 106.6666666667, "high": 106.6666666667 }, "base_share_pct": 24.433301899350067, "share_range_pct": { "low": 17.40430010710464, "base": 24.433301899350067, "high": 29.469535776969995 }, "confidence": "high", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Claude family", "lab": "Anthropic", "scenario_tpd": { "low": 15.753424657534246, "base": 50.68493150684931, "high": 178.08219178082192 }, "base_share_pct": 11.61000218675946, "share_range_pct": { "low": 2.570412473009086, "base": 11.61000218675946, "high": 49.19999551804241 }, "confidence": "low", "source_url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "as_of": "2026-07-31", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "GPT / Codex family", "lab": "OpenAI", "scenario_tpd": { "low": 23.0, "base": 26.742857143200002, "high": 36.6 }, "base_share_pct": 6.125777833413673, "share_range_pct": { "low": 3.7528022105932655, "base": 6.125777833413673, "high": 10.11173446346967 }, "confidence": "medium", "source_url": "https://openai.com/index/accelerating-the-next-phase-ai/", "as_of": "2026-03-31", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "DeepSeek family", "lab": "DeepSeek", "scenario_tpd": { "low": 12.166666666666664, "base": 21.599999999999998, "high": 29.589041095890412 }, "base_share_pct": 4.947743634616842, "share_range_pct": { "low": 1.9851779809660024, "base": 4.947743634616842, "high": 8.17476848607474 }, "confidence": "low", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Qwen family", "lab": "Alibaba", "scenario_tpd": { "low": 11.06060606060606, "base": 19.44, "high": 26.630136986301366 }, "base_share_pct": 4.452969271155158, "share_range_pct": { "low": 1.8047072554236387, "base": 4.452969271155158, "high": 7.357291637467264 }, "confidence": "low", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Other / unallocated open models", "lab": "Other / unattributed", "scenario_tpd": { "low": 5.53030303030303, "base": 10.799999999999999, "high": 18.493150684931507 }, "base_share_pct": 2.473871817308421, "share_range_pct": { "low": 0.9023536277118194, "base": 2.473871817308421, "high": 5.1092303037967115 }, "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Hunyuan family", "lab": "Tencent", "scenario_tpd": { "low": 3.318181818181818, "base": 7.2, "high": 11.095890410958903 }, "base_share_pct": 1.649247878205614, "share_range_pct": { "low": 0.5414121766270916, "base": 1.649247878205614, "high": 3.0655381822780265 }, "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Mistral family", "lab": "Mistral", "scenario_tpd": { "low": 1.659090909090909, "base": 5.040000000000001, "high": 8.876712328767123 }, "base_share_pct": 1.15447351474393, "share_range_pct": { "low": 0.2707060883135458, "base": 1.15447351474393, "high": 2.4524305458224216 }, "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Kimi family", "lab": "Moonshot AI", "scenario_tpd": { "low": 1.659090909090909, "base": 4.32, "high": 7.397260273972603 }, "base_share_pct": 0.9895487269233685, "share_range_pct": { "low": 0.2707060883135458, "base": 0.9895487269233685, "high": 2.043692121518685 }, "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "GLM family", "lab": "Z.ai", "scenario_tpd": { "low": 1.106060606060606, "base": 3.6, "high": 5.917808219178083 }, "base_share_pct": 0.824623939102807, "share_range_pct": { "low": 0.18047072554236387, "base": 0.824623939102807, "high": 1.6349536972149477 }, "confidence": "low", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Meta AI / Muse family", "lab": "Meta", "scenario_tpd": { "low": 0.03, "base": 0.405, "high": 3.0 }, "base_share_pct": 0.0927701931490658, "share_range_pct": { "low": 0.004894959405121651, "base": 0.0927701931490658, "high": 0.828830693727022 }, "confidence": "low", "source_url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "as_of": "2025-10-01", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." }, { "model_family": "Grok family", "lab": "xAI", "scenario_tpd": { "low": 0.005616, "base": 0.06318, "high": 0.5265 }, "base_share_pct": 0.014472150131254264, "share_range_pct": { "low": 0.000916336400638773, "base": 0.014472150131254264, "high": 0.14545978674909238 }, "confidence": "low", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "as_of": "2026-03-31", "version_split_available": false, "rationale": "Family-level attribution only. Providers do not publish production routing by exact model version; forcing GPT, Claude, Gemini, Doubao, or Qwen version shares would create false precision." } ], "exact_model_version_share_available": false }, "caveats": [ "The 180T Doubao disclosure can include multimodal/internal accounting unlike some text-only API totals.", "Anthropic's wide range dominates high-case uncertainty because ARR does not uniquely determine tokens.", "App reach and OpenRouter are allocation priors only for the residual pool, never global-share meters.", "Momentum fields use each lab's latest comparable public signal; they are not directly comparable token-growth series." ] }, "supply_capacity": { "title": "Public operating supply capacity", "formula": "accelerators \u00d7 inference share \u00d7 effective utilization \u00d7 benchmark processed tokens/sec/accelerator \u00d7 86,400", "methodology": "Count only dated operational hardware when calculating token capacity. Map only compatible fleets to a named public InferenceX benchmark with an explicit equivalence range. Preserve other chip, MW, capability-envelope, and national-FLOP disclosures in native units because their boundaries and throughput are not comparable. Keep planned contracts separate and never add capacity to demand.", "sources": [ { "name": "xAI Colossus 1", "url": "https://x.ai/news/anthropic-compute-partnership", "as_of": "2026-05-06", "type": "company_disclosure" }, { "name": "xAI Colossus 1", "url": "https://x.ai/news/anthropic-compute-partnership", "as_of": "2026-08-18", "type": "modeled_allocation_assumption" }, { "name": "xAI Colossus 1", "url": "https://inferencex.semianalysis.com/about", "as_of": "2026-08-18", "type": "modeled_operations_assumption" }, { "name": "xAI Colossus 1", "url": "https://inferencex.semianalysis.com/api/v1/tco-feed?format=json", "as_of": "2026-08-18", "type": "benchmark_informed_assumption" }, { "name": "Amazon\u2013Anthropic Project Rainier", "url": "https://aws.amazon.com/blogs/aws/aws-weekly-roundup-project-rainier-online-amazon-nova-amazon-bedrock-and-more-november-3-2025/", "as_of": "2025-11-03", "type": "company_operational_disclosure" }, { "name": "Microsoft Azure capacity added", "url": "https://azure.microsoft.com/en-us/blog/accelerating-open-source-infrastructure-development-for-frontier-ai-at-scale/", "as_of": "2025-10-13", "type": "company_operational_lower_bound" }, { "name": "CoreWeave active power", "url": "https://investors.coreweave.com/news/news-details/2025/CoreWeave-Reports-Strong-Second-Quarter-2025-Results/default.aspx", "as_of": "2025-06-30", "type": "company_operational_disclosure" }, { "name": "Google global TPU capability", "url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "type": "company_capability_envelope" }, { "name": "China national intelligent compute", "url": "https://english.www.gov.cn/news/202508/28/content_WS68b00fe2c6d0868f4e8f522a.html", "as_of": "2025-08-28", "type": "government_operational_stock" }, { "name": "Meta\u2013AMD planned deployment", "url": "https://about.fb.com/news/2026/02/meta-amd-partner-longterm-ai-infrastructure-agreement/", "as_of": "2026-02-24", "type": "company_contract_disclosure" }, { "name": "InferenceX TCO feed", "url": "https://inferencex.semianalysis.com/api/v1/tco-feed?format=json", "as_of": "2026-08-12", "type": "public_hardware_benchmark" } ], "rows": [ { "id": "xai_colossus_operating_capacity", "channel": "xAI Colossus 1", "scenario_tpd": { "low": 8.866568887344, "base": 58.386657707136, "high": 199.0454240016 }, "operating": true, "confidence": "low", "unit": "trillion_tokens_per_day", "operating_status": "operational_modeled", "convertible_to_tokens": true, "source_url": "https://x.ai/news/anthropic-compute-partnership", "as_of": "2026-05-06", "benchmark": { "model": "dsv4", "hardware": "b200", "workload": "8192x1024", "interactivity_tps": 50.0, "output_tps_per_gpu": 3447.462, "boundary": "interpolated", "is_interpolated": true, "frontier_points": 13, "latest_date": "2026-08-12", "evidence_date": { "from": "2026-08-07", "to": "2026-08-07" } }, "formula_inputs": [ "xai_operational_gpus", "xai_inference_share", "xai_capacity_utilization", "xai_b200_equivalence" ], "rationale": "Feasibility envelope from a mixed operational fleet and a public open-model benchmark; not an assertion of actual Grok production." }, { "id": "meta_amd_planned_capacity", "channel": "Meta\u2013AMD agreement", "scenario_mw": { "low": 0.0, "base": 0.0, "high": 6000.0 }, "operating": false, "unit": "megawatts", "operating_status": "contracted_planned", "convertible_to_tokens": false, "source_url": "https://about.fb.com/news/2026/02/meta-amd-partner-longterm-ai-infrastructure-agreement/", "as_of": "2026-02-24", "confidence": "medium", "formula_inputs": [ "meta_amd_planned_mw" ], "rationale": "Contracted multi-year capacity. Excluded from current token-production capacity." }, { "id": "amazon_trainium2_operational_chips", "channel": "Amazon\u2013Anthropic Project Rainier", "scenario_native": { "low": 500000.0, "base": 500000.0, "high": 500000.0 }, "unit": "trainium2_chips", "operating_status": "operational", "operating": true, "convertible_to_tokens": false, "confidence": "high", "source_url": "https://aws.amazon.com/blogs/aws/aws-weekly-roundup-project-rainier-online-amazon-nova-amazon-bedrock-and-more-november-3-2025/", "as_of": "2025-11-03", "formula_inputs": [ "amazon_trainium2_operational_chips" ], "rationale": "AWS disclosed nearly 500K Trainium2 chips online and already training and serving Claude. No comparable public InferenceX Trainium2 throughput row exists, so the count is not converted to tokens." }, { "id": "microsoft_added_capacity_mw", "channel": "Microsoft Azure capacity added", "scenario_native": { "low": 2000.0, "base": 2000.0, "high": 2000.0 }, "unit": "megawatts", "operating_status": "operational_lower_bound", "operating": true, "convertible_to_tokens": false, "confidence": "medium", "source_url": "https://azure.microsoft.com/en-us/blog/accelerating-open-source-infrastructure-development-for-frontier-ai-at-scale/", "as_of": "2025-10-13", "formula_inputs": [ "microsoft_added_capacity_mw" ], "rationale": "Microsoft disclosed more than 2GW of new capacity added in the prior year. The source does not define utility, facility, or IT-load boundaries or the share allocated to inference." }, { "id": "coreweave_active_power_mw", "channel": "CoreWeave active power", "scenario_native": { "low": 470.0, "base": 470.0, "high": 470.0 }, "unit": "megawatts", "operating_status": "operational_power", "operating": true, "convertible_to_tokens": false, "confidence": "medium", "source_url": "https://investors.coreweave.com/news/news-details/2025/CoreWeave-Reports-Strong-Second-Quarter-2025-Results/default.aspx", "as_of": "2025-06-30", "formula_inputs": [ "coreweave_active_power_mw" ], "rationale": "Company-wide active power, not necessarily energized accelerator IT load or average draw. No model/workload/SKU allocation is disclosed, so it is not converted to tokens." }, { "id": "google_tpu_global_capability", "channel": "Google global TPU capability", "scenario_native": { "low": 1000000.0, "base": 1000000.0, "high": 1000000.0 }, "unit": "tpu_chips", "operating_status": "capability_envelope", "operating": false, "convertible_to_tokens": false, "confidence": "medium", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "formula_inputs": [ "google_tpu_global_capability" ], "rationale": "Google said JAX/Pathways can scale training across more than one million TPUs globally. This is a distributed capability envelope, not a dedicated inference inventory or simultaneous active cluster." }, { "id": "china_intelligent_compute_pflops", "channel": "China national intelligent compute", "scenario_native": { "low": 780000.0, "base": 780000.0, "high": 780000.0 }, "unit": "reported_pflops", "operating_status": "reported_national_stock", "operating": false, "convertible_to_tokens": false, "confidence": "medium", "source_url": "https://english.www.gov.cn/news/202508/28/content_WS68b00fe2c6d0868f4e8f522a.html", "as_of": "2025-08-28", "formula_inputs": [ "china_intelligent_compute_pflops" ], "rationale": "National reported stock. Precision and benchmark convention are not specified, so it cannot be directly compared with vendor low-precision AI FLOPS or converted to tokens." } ], "summary": { "operating_capacity_tpd": { "low": 8.866568887344, "base": 58.386657707136, "high": 199.0454240016 }, "publicly_modeled_operator_count": 1, "native_unit_evidence_rows": 6, "planned_capacity_mw": { "low": 0.0, "base": 0.0, "high": 6000.0 } }, "caveats": [ "This is only the publicly token-convertible xAI subset, not global capacity.", "Amazon Trainium2, Microsoft/CoreWeave MW, Google TPU capability, Meta planned MW, and China national FLOPS remain in native units.", "Closed-model throughput, workload allocation, utilization, and power boundaries are not publicly comparable." ] }, "economics": { "title": "Realized token economics", "formula": "additive trillion tokens/day \u00d7 realized dollars per million tokens \u00d7 1,000,000", "methodology": "Apply a broad realized-price range to the additive demand estimate. The range is intentionally below many frontier list prices because the volume includes free, cached, batch, discounted, subscription, internal, and self-hosted usage. Treat the result as token-equivalent economic value, not recognized revenue.", "sources": [ { "name": "Market-wide paid inference", "url": "https://openrouter.ai/api/v1/models", "as_of": "2026-08-18", "type": "modeled_from_public_list_prices" } ], "rows": [ { "id": "implied_daily_token_value", "channel": "Modeled additive demand", "scenario_usd_per_day": { "low": 126684497.56348197, "base": 654843952.9751239, "high": 3064376792.2376094 }, "confidence": "low", "formula_inputs": [ "realized_price_per_million_tokens" ], "rationale": "Gross token-equivalent value, not recognized revenue. Free, self-hosted, subscription, and internal traffic do not necessarily monetize per token." } ], "summary": { "realized_price_per_million": { "low": 0.35, "base": 1.5, "high": 5.0 }, "implied_value_usd_per_day": { "low": 126684497.56348197, "base": 654843952.9751239, "high": 3064376792.2376094 } }, "caveats": [ "List price is not realized price.", "Subscription and internal use cannot be cleanly expressed as per-token revenue.", "Input, cached-input, output, and reasoning token prices differ." ] }, "reconciliation": { "title": "Demand, disclosure, and capacity reconciliation", "formula": "modeled additive demand \u2212 publicly quantified operating capacity; residual is reported, not allocated or forced to zero", "methodology": "Compare the additive demand ledger with the small subset of supply capacity that can be modeled from public operational disclosures. A large residual primarily means the public supply ledger is incomplete; it is not evidence that demand is wrong or that the named operator serves the whole market.", "sources": [ { "name": "Google all AI surfaces", "url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "as_of": "2026-05-19", "type": "company_disclosure" }, { "name": "OpenAI API platform", "url": "https://openai.com/index/accelerating-the-next-phase-ai/", "as_of": "2026-03-31", "type": "company_disclosure" }, { "name": "China aggregate", "url": "https://www.ecns.cn/cns-wire/2026-03-24/detail-ihfaytev9463369.shtml", "as_of": "2026-03-16", "type": "curated_disclosure" }, { "name": "ByteDance Doubao model family", "url": "https://en.jiemian.com/article/14629930.html", "as_of": "2026-06-23", "type": "company_statement_via_press" }, { "name": "Fireworks AI", "url": "https://fireworks.ai/", "as_of": "2026-06-12", "type": "company_disclosure" }, { "name": "Together AI", "url": "https://tomtunguz.com/trillion-token-race/", "as_of": "2025-09-15", "type": "curated_disclosure" }, { "name": "Google model APIs", "url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "as_of": "2026-07-22", "type": "company_disclosure" }, { "name": "Microsoft Foundry customer floor", "url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3", "as_of": "2026-04-29", "type": "company_disclosure" }, { "name": "OpenRouter data API", "url": "https://openrouter.ai/docs/cookbook/administration/data-api", "as_of": "2026-08-17", "type": "platform_metered" }, { "name": "InferenceX TCO feed", "url": "https://inferencex.semianalysis.com/api/v1/tco-feed?format=json", "as_of": "2026-08-12", "type": "public_hardware_benchmark" } ], "rows": [], "summary": { "additive_demand_tpd": { "low": 361.9557073242342, "base": 436.56263531674927, "high": 612.8753584475219 }, "public_supply_tpd": { "low": 8.866568887344, "base": 58.386657707136, "high": 199.0454240016 }, "unresolved_residual_tpd": { "low": 162.9102833226342, "base": 378.17597760961326, "high": 604.0087895601779 }, "public_supply_coverage_pct": { "low": 1.446716492208784, "base": 13.374176574862712, "high": 54.99165228614513 }, "forced_to_balance": false }, "caveats": [ "Supply and demand boundaries are intentionally asymmetric in v0.3.", "Amazon, Microsoft, Google, Meta, CoreWeave, and China have public native-unit evidence but remain unconverted because SKU, workload, utilization, or power boundaries are missing.", "Oracle, private enterprise, and self-hosted capacity remain unresolved; the residual is not idle capacity or unmet demand." ] } }, "assumptions": [ { "id": "google_all_surfaces_tokens_tpd", "component": "observed_demand", "channel": "Google all AI surfaces", "metric": "tokens_per_day", "low": 106.6666666667, "base": 106.6666666667, "high": 106.6666666667, "unit": "trillion_tokens_per_day", "as_of": "2026-05-19", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "source_type": "company_disclosure", "confidence": "high", "additive": true, "rationale": "Google disclosed more than 3.2 quadrillion monthly tokens across its surfaces; the daily value divides by 30 days. It includes first-party products and APIs, so no separate Google product row is added." }, { "id": "openai_reported_tokens_tpd", "component": "observed_demand", "channel": "OpenAI API platform", "metric": "tokens_per_day", "low": 21.6, "base": 21.6, "high": 21.6, "unit": "trillion_tokens_per_day", "as_of": "2026-03-31", "source_url": "https://openai.com/index/accelerating-the-next-phase-ai/", "source_type": "company_disclosure", "confidence": "high", "additive": true, "rationale": "OpenAI disclosed more than 15B API tokens per minute; 15B \u00d7 1,440 minutes/day = 21.6T/day if sustained. OpenAI labels the boundary API-platform traffic, so a separately modeled ChatGPT consumer estimate can be added without treating the API row as an all-surfaces total." }, { "id": "china_aggregate_tokens_tpd", "component": "observed_demand", "channel": "China aggregate", "metric": "tokens_per_day", "low": 100.0, "base": 140.0, "high": 180.0, "unit": "trillion_tokens_per_day", "as_of": "2026-03-16", "source_url": "https://www.ecns.cn/cns-wire/2026-03-24/detail-ihfaytev9463369.shtml", "source_type": "curated_disclosure", "confidence": "medium", "additive": false, "rationale": "March country-level cross-check. It is superseded for the current additive partition by ByteDance's later June Doubao-family disclosure plus an explicitly modeled non-Doubao China residual; adding all three would double count." }, { "id": "bytedance_doubao_tokens_tpd", "component": "observed_demand", "channel": "ByteDance Doubao model family", "metric": "tokens_per_day", "low": 180.0, "base": 180.0, "high": 180.0, "unit": "trillion_tokens_per_day", "as_of": "2026-06-23", "source_url": "https://en.jiemian.com/article/14629930.html", "source_type": "company_statement_via_press", "confidence": "medium", "additive": true, "rationale": "Volcano Engine president Tan Dai said Doubao-family daily token usage reached 180T at the June FORCE conference. The scope is the model family across consumer and enterprise surfaces and can include multimodal token accounting." }, { "id": "fireworks_tokens_tpd", "component": "observed_demand", "channel": "Fireworks AI", "metric": "tokens_per_day", "low": 15.0, "base": 30.0, "high": 35.0, "unit": "trillion_tokens_per_day", "as_of": "2026-06-12", "source_url": "https://fireworks.ai/", "source_type": "company_disclosure", "confidence": "medium", "additive": true, "rationale": "Platform disclosure increased from 15T/day to 30T/day. We do not separately add Fireworks customer application estimates." }, { "id": "together_tokens_tpd", "component": "observed_demand", "channel": "Together AI", "metric": "tokens_per_day", "low": 1.5, "base": 2.0, "high": 3.0, "unit": "trillion_tokens_per_day", "as_of": "2025-09-15", "source_url": "https://tomtunguz.com/trillion-token-race/", "source_type": "curated_disclosure", "confidence": "low", "additive": true, "rationale": "Secondary report of 2T/day. The row is aged and receives a broad range; customers are not separately added." }, { "id": "google_model_api_tokens_tpd", "component": "observed_demand", "channel": "Google model APIs", "metric": "tokens_per_day", "low": 31.68, "base": 31.68, "high": 31.68, "unit": "trillion_tokens_per_day", "as_of": "2026-07-22", "source_url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "source_type": "company_disclosure", "confidence": "high", "additive": false, "rationale": "Alphabet disclosed approximately 22B model-API tokens per minute; 22B \u00d7 1,440 = 31.68T/day if sustained. This is a contained cross-check within Google's all-surfaces disclosure and is not added again." }, { "id": "google_gemini_monthly_users", "component": "modeled_app_demand", "channel": "Gemini app cross-check", "metric": "monthly_active_users", "low": 1000000000, "base": 1000000000, "high": 1000000000, "unit": "users", "as_of": "2026-08-11", "source_url": "https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/", "source_type": "company_disclosure", "confidence": "high", "additive": false, "rationale": "Google's latest app-specific disclosure says more than 1B monthly Gemini app users. This supersedes the 950M Q2 figure and is modeled only as a cross-check because Gemini app tokens are contained in Google's all-surfaces total." }, { "id": "google_gemini_daily_active_share", "component": "modeled_app_demand", "channel": "Gemini app cross-check", "metric": "daily_active_share_of_mau", "low": 0.05, "base": 0.15, "high": 0.3, "unit": "share", "as_of": "2026-08-11", "source_url": "https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/", "source_type": "modeled_from_company_growth_disclosure", "confidence": "low", "additive": false, "rationale": "Google disclosed that Gemini daily active users tripled year over year but not the DAU count. The range converts the latest 1B MAU into an active-day envelope." }, { "id": "google_gemini_interactions_per_active_day", "component": "modeled_app_demand", "channel": "Gemini app cross-check", "metric": "interactions_per_active_day", "low": 1.0, "base": 3.0, "high": 6.0, "unit": "interactions_per_day", "as_of": "2026-08-11", "source_url": "https://blog.google/innovation-and-ai/products/gemini-app/one-billion-monthly-users/", "source_type": "modeled_workload_assumption", "confidence": "low", "additive": false, "rationale": "Google publishes feature mix and users but not prompts per active user. The range covers one-shot embedded use through a multi-turn session." }, { "id": "google_gemini_tokens_per_interaction", "component": "modeled_app_demand", "channel": "Gemini app cross-check", "metric": "tokens_per_interaction", "low": 400, "base": 1200, "high": 4000, "unit": "tokens_per_interaction", "as_of": "2026-08-19", "source_url": "https://inferencex.semianalysis.com/about", "source_type": "benchmark_informed_assumption", "confidence": "low", "additive": false, "rationale": "Processed tokens include prompt, retained context, output, and reasoning/tool steps. Voice, image, and video units are not converted." }, { "id": "google_ai_overviews_monthly_users", "component": "modeled_app_demand", "channel": "Google AI Overviews reference", "metric": "monthly_active_users", "low": 2500000000, "base": 2500000000, "high": 2500000000, "unit": "users", "as_of": "2026-06-01", "source_url": "https://blog.google/alphabet/investor-presentation-june-2026/", "source_type": "company_disclosure_reference_only", "confidence": "high", "additive": false, "rationale": "AI Overviews exceeds 2.5B monthly users, but Google does not disclose attributable queries or processed tokens. The row is reach context only and remains contained in Google's all-surfaces total." }, { "id": "openai_chatgpt_messages_per_day", "component": "modeled_app_demand", "channel": "ChatGPT consumer estimate", "metric": "messages_per_day", "low": 3500000000, "base": 4285714286, "high": 5000000000, "unit": "messages_per_day", "as_of": "2026-08-06", "source_url": "https://openai.com/index/how-the-world-is-putting-chatgpt-to-work/", "source_type": "modeled_from_company_disclosures", "confidence": "medium", "additive": true, "rationale": "OpenAI disclosed 3B messages/day at 700M users in August 2025 and more than 1B weekly users in August 2026. The base scales the old message rate by current users; the range allows per-user intensity to change. The API disclosure explicitly covers API tokens, so this consumer estimate is additive in the lab partition." }, { "id": "openai_tokens_per_message", "component": "modeled_app_demand", "channel": "ChatGPT consumer estimate", "metric": "tokens_per_message", "low": 400, "base": 1200, "high": 3000, "unit": "tokens_per_message", "as_of": "2026-08-18", "source_url": "https://openai.com/signals/data/", "source_type": "modeled_workload_assumption", "confidence": "low", "additive": false, "rationale": "Total processed tokens per user message include the current prompt, retained context, output, and sometimes reasoning/tool tokens. No public representative token distribution exists, so the range is intentionally wide." }, { "id": "anthropic_run_rate_revenue_usd", "component": "modeled_app_demand", "channel": "Anthropic financial cross-check", "metric": "annualized_revenue_usd", "low": 46000000000, "base": 55500000000, "high": 65000000000, "unit": "usd_per_year", "as_of": "2026-07-31", "source_url": "https://www.cnbc.com/2026/08/17/anthropic-says-annualized-revenue-climbed-to-65-billion-in-july.html", "source_type": "reported_investor_update", "confidence": "medium", "additive": true, "rationale": "CNBC reported that Anthropic told investors its July run rate exceeded $65B; Bloomberg separately reported preliminary Q2 revenue above $11.5B, equivalent to a $46B quarterly annualization. The scenario uses annualized Q2 as low, the midpoint between Q2 pace and July run rate as base, and the reported July run rate as high. Neither figure is audited and Anthropic declined comment." }, { "id": "anthropic_q2_preliminary_revenue_usd", "component": "modeled_app_demand", "channel": "Anthropic Q2 recognized-revenue anchor", "metric": "quarterly_revenue_usd", "low": 11500000000, "base": 11500000000, "high": 11500000000, "unit": "usd_per_quarter", "as_of": "2026-06-30", "source_url": "https://finance.yahoo.com/technology/ai/articles/anthropic-revenue-surges-over-11-210857853.html", "source_type": "reported_company_documents", "confidence": "medium", "additive": false, "rationale": "Bloomberg reported preliminary Q2 revenue above $11.5B from company documents. Four-times this quarterly pace gives the $46B low annualized input. It is preliminary, not audited, and not an additional demand row." }, { "id": "anthropic_realized_price_per_million_tokens", "component": "modeled_app_demand", "channel": "Anthropic financial estimate", "metric": "realized_usd_per_million_tokens", "low": 1.0, "base": 3.0, "high": 8.0, "unit": "usd_per_million_tokens", "as_of": "2026-08-19", "source_url": "https://docs.anthropic.com/en/docs/about-claude/pricing", "source_type": "modeled_from_public_pricing", "confidence": "low", "additive": true, "rationale": "Realized revenue per processed token can be well below list output pricing because Anthropic mixes input, cached input, batch, subscriptions, commitments, API, Claude Code, and cloud resale. The lab-specific range replaces the overly broad market-wide price used in v0.2." }, { "id": "microsoft_foundry_large_customers", "component": "observed_demand", "channel": "Microsoft Foundry customer floor", "metric": "customers_over_one_trillion_annual_tokens", "low": 300, "base": 300, "high": 300, "unit": "customers", "as_of": "2026-04-29", "source_url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3", "source_type": "company_disclosure", "confidence": "high", "additive": false, "rationale": "Microsoft said more than 300 customers were on track to process over one trillion Foundry tokens in the year. Multiplying only the disclosed floor gives a minimum, not total Foundry volume. It overlaps underlying model providers." }, { "id": "microsoft_foundry_tokens_per_large_customer_year", "component": "observed_demand", "channel": "Microsoft Foundry customer floor", "metric": "minimum_annual_tokens_per_customer", "low": 1000000000000, "base": 1000000000000, "high": 1000000000000, "unit": "tokens_per_year", "as_of": "2026-04-29", "source_url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3", "source_type": "company_disclosure", "confidence": "high", "additive": false, "rationale": "The wording is 'over one trillion'; using exactly one trillion produces a conservative disclosed floor." }, { "id": "microsoft_copilot_paid_seats", "component": "modeled_app_demand", "channel": "Microsoft 365 Copilot seat cross-check", "metric": "paid_seats", "low": 30000000, "base": 30000000, "high": 30000000, "unit": "seats", "as_of": "2026-07-29", "source_url": "https://www.sec.gov/Archives/edgar/data/789019/000119312526323632/msft-ex99_1.htm", "source_type": "sec_filing", "confidence": "high", "additive": false, "rationale": "Microsoft disclosed more than 30M paid Microsoft 365 Copilot seats. The row is a workload cross-check because its model-provider tokens may already appear in OpenAI/Anthropic aggregates." }, { "id": "microsoft_copilot_interactions_per_seat_day", "component": "modeled_app_demand", "channel": "Microsoft 365 Copilot seat cross-check", "metric": "interactions_per_paid_seat_day", "low": 0.5, "base": 2.0, "high": 6.0, "unit": "interactions_per_day", "as_of": "2026-08-18", "source_url": "https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4", "source_type": "modeled_workload_assumption", "confidence": "low", "additive": false, "rationale": "Paid seat count is not active use. The range spans light licensed-seat penetration through repeated daily use." }, { "id": "microsoft_copilot_tokens_per_interaction", "component": "modeled_app_demand", "channel": "Microsoft 365 Copilot seat cross-check", "metric": "tokens_per_interaction", "low": 500, "base": 2000, "high": 6000, "unit": "tokens_per_interaction", "as_of": "2026-08-18", "source_url": "https://inferencex.semianalysis.com/about", "source_type": "benchmark_informed_assumption", "confidence": "low", "additive": false, "rationale": "Office grounding, retrieved documents, and agent/tool steps can make Copilot interactions substantially larger than simple chat turns." }, { "id": "meta_ai_monthly_users", "component": "modeled_app_demand", "channel": "Meta AI", "metric": "monthly_active_users", "low": 1000000000, "base": 1000000000, "high": 1000000000, "unit": "users", "as_of": "2025-10-01", "source_url": "https://about.fb.com/news/2025/10/improving-your-recommendations-apps-ai-meta/", "source_type": "company_disclosure", "confidence": "high", "additive": true, "rationale": "Meta disclosed more than 1B monthly Meta AI users. User count is direct; daily activity and intensity remain modeled." }, { "id": "meta_ai_daily_active_share", "component": "modeled_app_demand", "channel": "Meta AI", "metric": "daily_active_share_of_mau", "low": 0.1, "base": 0.18, "high": 0.3, "unit": "share", "as_of": "2026-08-18", "source_url": "https://about.fb.com/news/2024/09/metas-ai-product-news-connect/", "source_type": "modeled_from_company_disclosure", "confidence": "low", "additive": true, "rationale": "Meta disclosed 185M weekly and 400M monthly users in September 2024. The daily share is not disclosed and is bounded below the weekly/monthly ratio." }, { "id": "meta_ai_interactions_per_active_day", "component": "modeled_app_demand", "channel": "Meta AI", "metric": "interactions_per_active_day", "low": 1.0, "base": 2.5, "high": 5.0, "unit": "interactions_per_day", "as_of": "2026-08-18", "source_url": "https://about.fb.com/news/2025/04/introducing-meta-ai-app-new-way-access-ai-assistant/", "source_type": "modeled_workload_assumption", "confidence": "low", "additive": true, "rationale": "Meta describes chat and voice use but does not disclose prompt frequency. Range covers light embedded use through short multi-turn sessions." }, { "id": "meta_ai_tokens_per_interaction", "component": "modeled_app_demand", "channel": "Meta AI", "metric": "tokens_per_interaction", "low": 300, "base": 900, "high": 2000, "unit": "tokens_per_interaction", "as_of": "2026-08-18", "source_url": "https://inferencex.semianalysis.com/about", "source_type": "benchmark_informed_assumption", "confidence": "low", "additive": true, "rationale": "Counts processed input plus generated output for a short assistant turn. Image/voice generation is excluded because it is not consistently expressed as text tokens." }, { "id": "xai_grok_monthly_users", "component": "modeled_app_demand", "channel": "Grok consumer estimate", "metric": "monthly_active_users", "low": 117000000, "base": 117000000, "high": 117000000, "unit": "users", "as_of": "2026-03-31", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "source_type": "sec_filing", "confidence": "high", "additive": true, "rationale": "SpaceX's S-1 disclosed approximately 117M MAUs that used Grok AI features. This is broader than standalone grok.com/app usage because it includes embedded X features, but it is now a clean Grok-specific numerator rather than the prior combined X/Grok reach." }, { "id": "xai_grok_daily_active_share", "component": "modeled_app_demand", "channel": "Grok consumer estimate", "metric": "daily_active_share_of_mau", "low": 0.08, "base": 0.15, "high": 0.25, "unit": "share", "as_of": "2026-03-31", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "source_type": "modeled_from_sec_disclosure", "confidence": "low", "additive": true, "rationale": "The filing gives monthly users but no Grok DAU. Embedded explain/search features likely have lower daily intensity than deliberate standalone chatbot use." }, { "id": "xai_grok_interactions_per_active_day", "component": "modeled_app_demand", "channel": "Grok consumer estimate", "metric": "interactions_per_active_day", "low": 1.5, "base": 3.0, "high": 6.0, "unit": "interactions_per_day", "as_of": "2026-08-19", "source_url": "https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm", "source_type": "modeled_workload_assumption", "confidence": "low", "additive": true, "rationale": "Range covers light embedded feature use through a deliberate multi-turn Grok session. Image/video generation remains outside text-token conversion." }, { "id": "xai_grok_tokens_per_interaction", "component": "modeled_app_demand", "channel": "Grok consumer estimate", "metric": "tokens_per_interaction", "low": 400, "base": 1200, "high": 3000, "unit": "tokens_per_interaction", "as_of": "2026-08-19", "source_url": "https://inferencex.semianalysis.com/about", "source_type": "benchmark_informed_assumption", "confidence": "low", "additive": true, "rationale": "Processed text-token envelope includes prompt, retained context, output, reasoning, and tools; multimodal outputs are excluded." }, { "id": "china_current_total_tokens_tpd", "component": "modeled_app_demand", "channel": "China current total estimate", "metric": "tokens_per_day", "low": 200.0, "base": 220.0, "high": 250.0, "unit": "trillion_tokens_per_day", "as_of": "2026-06-23", "source_url": "https://news.cgtn.com/news/2026-06-18/China-builds-computing-network-as-AI-token-usage-skyrockets-1O4tgBLAnxS/index.html", "source_type": "modeled_from_national_and_company_disclosures", "confidence": "low", "additive": false, "rationale": "Nowcasts China's March 140T/day national total to June using the observed Doubao step from 120T to 180T. The model adds only max(current-China-total minus Doubao, zero), preventing ByteDance double counting. Different token and multimodal conventions make this a broad envelope." }, { "id": "rs_weight_deepseek", "component": "relative_strength", "channel": "DeepSeek", "metric": "open_pool_allocation_weight", "low": 0.22, "base": 0.3, "high": 0.4, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "source_type": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "additive": false, "rationale": "Allocation prior for non-Doubao China and unattributed hosted open-model tokens. Blends QuestMobile app reach with current OpenRouter model-lab share; neither is a complete token census." }, { "id": "rs_weight_qwen", "component": "relative_strength", "channel": "Alibaba / Qwen", "metric": "open_pool_allocation_weight", "low": 0.2, "base": 0.27, "high": 0.36, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://technode.com/2026/07/14/questmobile-chinas-ai-native-apps-reach-499-million-monthly-active-users/", "source_type": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "additive": false, "rationale": "Allocation prior informed by Qwen's 167M China app MAU, Alibaba cloud distribution, and router traffic." }, { "id": "rs_weight_tencent", "component": "relative_strength", "channel": "Tencent", "metric": "open_pool_allocation_weight", "low": 0.06, "base": 0.1, "high": 0.15, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "source_type": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "additive": false, "rationale": "Allocation prior for Hunyuan and Tencent-hosted demand; public absolute token volume is unavailable." }, { "id": "rs_weight_moonshot", "component": "relative_strength", "channel": "Moonshot / Kimi", "metric": "open_pool_allocation_weight", "low": 0.03, "base": 0.06, "high": 0.1, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "source_type": "modeled_blend_of_app_and_router_proxies", "confidence": "low", "additive": false, "rationale": "Allocation prior for Kimi consumer, API, and open-weight use. Model releases can move router share much faster than domestic production use." }, { "id": "rs_weight_zai", "component": "relative_strength", "channel": "Z.ai / GLM", "metric": "open_pool_allocation_weight", "low": 0.02, "base": 0.05, "high": 0.08, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "source_type": "modeled_router_informed_allocation", "confidence": "low", "additive": false, "rationale": "Allocation prior informed primarily by current GLM traffic on OpenRouter; domestic and private deployments remain unobserved." }, { "id": "rs_weight_mistral", "component": "relative_strength", "channel": "Mistral", "metric": "open_pool_allocation_weight", "low": 0.03, "base": 0.07, "high": 0.12, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "source_type": "modeled_router_informed_allocation", "confidence": "low", "additive": false, "rationale": "Allocation prior for hosted Mistral-family tokens. First-party Le Chat and private enterprise usage are not directly metered." }, { "id": "rs_weight_other_open", "component": "relative_strength", "channel": "Other / unallocated open models", "metric": "open_pool_allocation_weight", "low": 0.1, "base": 0.15, "high": 0.25, "unit": "weight", "as_of": "2026-08-17", "source_url": "https://openrouter.ai/docs/cookbook/administration/data-api", "source_type": "modeled_residual_allocation", "confidence": "low", "additive": false, "rationale": "Residual allocation for Baidu, MiniMax, Xiaomi, NVIDIA, Cohere, AI21, self-hosted and other labs not independently bounded." }, { "id": "xai_operational_gpus", "component": "supply_capacity", "channel": "xAI Colossus 1", "metric": "operational_accelerators", "low": 180000, "base": 220000, "high": 220000, "unit": "gpu_equivalents", "as_of": "2026-05-06", "source_url": "https://x.ai/news/anthropic-compute-partnership", "source_type": "company_disclosure", "confidence": "medium", "additive": true, "rationale": "xAI says Colossus 1 has over 220K mixed H100, H200, and GB200 GPUs. The model excludes the separate one-million H100-equivalent statement until site/SKU attribution is clearer." }, { "id": "xai_inference_share", "component": "supply_capacity", "channel": "xAI Colossus 1", "metric": "inference_share", "low": 0.15, "base": 0.3, "high": 0.5, "unit": "share", "as_of": "2026-08-18", "source_url": "https://x.ai/news/anthropic-compute-partnership", "source_type": "modeled_allocation_assumption", "confidence": "low", "additive": true, "rationale": "The cluster supports training, fine-tuning, inference, and HPC. No public workload allocation exists." }, { "id": "xai_capacity_utilization", "component": "supply_capacity", "channel": "xAI Colossus 1", "metric": "effective_utilization", "low": 0.35, "base": 0.55, "high": 0.75, "unit": "share", "as_of": "2026-08-18", "source_url": "https://inferencex.semianalysis.com/about", "source_type": "modeled_operations_assumption", "confidence": "low", "additive": true, "rationale": "Effective utilization includes availability, batching/interactivity, maintenance, and scheduling losses; it is deliberately below benchmark saturation." }, { "id": "xai_b200_equivalence", "component": "supply_capacity", "channel": "xAI Colossus 1", "metric": "b200_throughput_equivalence", "low": 0.35, "base": 0.6, "high": 0.9, "unit": "share_of_b200_throughput", "as_of": "2026-08-18", "source_url": "https://inferencex.semianalysis.com/api/v1/tco-feed?format=json", "source_type": "benchmark_informed_assumption", "confidence": "low", "additive": true, "rationale": "Colossus contains mixed H100/H200/GB200 hardware; equivalence maps that mix to the public B200 DeepSeek V4 benchmark rather than calling every device a B200." }, { "id": "amazon_trainium2_operational_chips", "component": "supply_capacity", "channel": "Amazon\u2013Anthropic Project Rainier", "metric": "operational_accelerators", "low": 500000, "base": 500000, "high": 500000, "unit": "trainium2_chips", "as_of": "2025-11-03", "source_url": "https://aws.amazon.com/blogs/aws/aws-weekly-roundup-project-rainier-online-amazon-nova-amazon-bedrock-and-more-november-3-2025/", "source_type": "company_operational_disclosure", "confidence": "high", "additive": false, "rationale": "AWS disclosed nearly 500K Trainium2 chips online and already training and serving Claude. No comparable public InferenceX Trainium2 throughput row exists, so the count is not converted to tokens." }, { "id": "microsoft_added_capacity_mw", "component": "supply_capacity", "channel": "Microsoft Azure capacity added", "metric": "added_capacity_mw", "low": 2000, "base": 2000, "high": 2000, "unit": "megawatts", "as_of": "2025-10-13", "source_url": "https://azure.microsoft.com/en-us/blog/accelerating-open-source-infrastructure-development-for-frontier-ai-at-scale/", "source_type": "company_operational_lower_bound", "confidence": "medium", "additive": false, "rationale": "Microsoft disclosed more than 2GW of new capacity added in the prior year. The source does not define utility, facility, or IT-load boundaries or the share allocated to inference." }, { "id": "coreweave_active_power_mw", "component": "supply_capacity", "channel": "CoreWeave active power", "metric": "active_power_mw", "low": 470, "base": 470, "high": 470, "unit": "megawatts", "as_of": "2025-06-30", "source_url": "https://investors.coreweave.com/news/news-details/2025/CoreWeave-Reports-Strong-Second-Quarter-2025-Results/default.aspx", "source_type": "company_operational_disclosure", "confidence": "medium", "additive": false, "rationale": "Company-wide active power, not necessarily energized accelerator IT load or average draw. No model/workload/SKU allocation is disclosed, so it is not converted to tokens." }, { "id": "google_tpu_global_capability", "component": "supply_capacity", "channel": "Google global TPU capability", "metric": "tpu_capability_envelope", "low": 1000000, "base": 1000000, "high": 1000000, "unit": "tpu_chips", "as_of": "2026-05-19", "source_url": "https://blog.google/innovation-and-ai/sundar-pichai-io-2026/", "source_type": "company_capability_envelope", "confidence": "medium", "additive": false, "rationale": "Google said JAX/Pathways can scale training across more than one million TPUs globally. This is a distributed capability envelope, not a dedicated inference inventory or simultaneous active cluster." }, { "id": "china_intelligent_compute_pflops", "component": "supply_capacity", "channel": "China national intelligent compute", "metric": "reported_intelligent_compute_capacity", "low": 780000, "base": 780000, "high": 780000, "unit": "reported_pflops", "as_of": "2025-08-28", "source_url": "https://english.www.gov.cn/news/202508/28/content_WS68b00fe2c6d0868f4e8f522a.html", "source_type": "government_operational_stock", "confidence": "medium", "additive": false, "rationale": "National reported stock. Precision and benchmark convention are not specified, so it cannot be directly compared with vendor low-precision AI FLOPS or converted to tokens." }, { "id": "realized_price_per_million_tokens", "component": "economics", "channel": "Market-wide paid inference", "metric": "realized_usd_per_million_tokens", "low": 0.35, "base": 1.5, "high": 5.0, "unit": "usd_per_million_tokens", "as_of": "2026-08-18", "source_url": "https://openrouter.ai/api/v1/models", "source_type": "modeled_from_public_list_prices", "confidence": "low", "additive": false, "rationale": "Realized price is below many frontier list prices because global volume includes cached input, batch, free/subscription use, small models, private discounts, and self-hosting." }, { "id": "meta_amd_planned_mw", "component": "supply_capacity", "channel": "Meta\u2013AMD planned deployment", "metric": "planned_capacity_mw", "low": 0, "base": 0, "high": 6000, "unit": "megawatts", "as_of": "2026-02-24", "source_url": "https://about.fb.com/news/2026/02/meta-amd-partner-longterm-ai-infrastructure-agreement/", "source_type": "company_contract_disclosure", "confidence": "medium", "additive": false, "rationale": "Up to 6GW is a multi-year contract with first deployments beginning in 2H26. It is planned, not current operating capacity, and is excluded from the operating supply total." } ], "unresolved_channels": [ "Exact production routing by model version inside GPT, Claude, Gemini, Doubao, Qwen, and Grok families", "Anthropic first-party versus AWS Bedrock and Google Vertex reseller attribution", "Microsoft Copilot and Azure model-routing mix", "AWS Bedrock aggregate token volume", "Meta ads/recommendation and MTIA token conversion", "Google TPU product allocation and token conversion", "Oracle operating fleet", "CoreWeave active-power token conversion", "private enterprise gateways", "self-hosted and on-device inference", "image/video/audio token equivalents" ] }