Six-channel triangulation
A company's token throughput leaves a footprint in six exhaust streams: hardware, energy, money, users, third-party telemetry, and occasional direct disclosure. No single stream is trustworthy — but where independent streams converge, the convergence is the estimate, and where they diverge, the divergence is the error bar. Bottom-up, that reconciles to 360.4 T/day (308.4–477.1), against a disclosed floor of 300.1. Full method: METHODOLOGY_V2.md.
Reconciled per-company estimates
The reconciled band is the primary channel's, cross-checked by the others — channels are never averaged. The dispersion ratio (high ÷ low) is the honesty metric: a small number means several independent channels agree; a big one means the figure rests on one channel and you should treat it accordingly.
| Entity | Low | Mid | High | Dispersion | Primary channel | Notes |
|---|---|---|---|---|---|---|
| 94.6 | 105.1 | 115.6 | ×1.2 | anchor | disclosed by Pichai at I/O 2026; nearly ground truth | |
| OpenAI | 21.6 | 32.6 | 67 | ×3.1 | demand | band-wise geometric mean of three bracketing routes, low clamped to the disclosed API-only anchor (all-surfaces cannot be below API); COGS mid lands within 15% of Epoch's independent ~19T (2025) after growth |
| Anthropic | 13.7 | 31.9 | 69.4 | ×5.1 | financial | no token disclosure exists; widest per-company band on the board and honestly so |
| Microsoft | 1.4 | 2.2 | 3 | ×2.2 | anchor | Foundry-only by construction; OpenAI serving on Azure excluded (counted under OpenAI); the undisclosed Azure remainder sits in GLOBAL unattributed |
| Meta | 0.1 | 0.6 | 2.6 | ×32.5 | demand | supply signals strong but $0 revenue; demand and capacity pincer it per doc 5 |
| xAI | 0.1 | 0.9 | 7.5 | ×111.9 | demand | no token or revenue signal usable; driver is tokens/message on Grok reasoning traffic |
| DeepSeek | 0.2 | 1.5 | 9 | ×40.9 | demand | EXCLUDED from GLOBAL sum: domestic usage sits inside the China aggregate and third-party-hosted DeepSeek sits under its hosts |
| China aggregate | 140 | 140 | 140 | ×1 | anchor | NDB-measured; STALE (Mar 2026) and growing ~40%/3mo per Yicai - treated as a floor-anchor, not nowcast |
| Fireworks AI | 30 | 30 | 30 | ×1 | anchor | disclosed; serves open models incl DeepSeek/Llama |
| Together AI | 2 | 2 | 2 | ×1 | anchor | disclosed Sep 2025; stale but small |
| GLOBAL | 308.4 | 360.4 | 477.1 | ×1.5 | composition | excluded by construction: DeepSeek (in China agg + hosts), Doubao (in China agg), OpenRouter (router), Harvey (demand ledger), rented fleets (in owners); closure: ExpView-via-Epoch 432T/day sits inside our band -> PASS; realized mid is 24% of global hardware capacity mid |
T tokens/day, all-surfaces scope unless noted. Disclosed-anchor entities (Google, China, Fireworks, Together) carry their as-of dates — China's 140 is March 2026 and growing ~40%/quarter, so the global figure is conservative.
The channel fan, per company
Every channel estimate we computed, including the ones that failed. A channel that cannot reproduce a company's disclosed anchor is excluded from reconciliation but published here — the failures are part of the evidence.
| Entity | Channel | Low | Mid | High | Back-test vs anchor | Conf. | Formula |
|---|---|---|---|---|---|---|---|
| anchor | 94.6 | 105.1 | 115.6 | — | high | disclosed 3,200T/month (all surfaces, incl multimodal) | |
| hardware | 41.1 | 405.2 | 1,250.4 | 3.9x (ceiling/anchor; PASS if > 1) | low | fleet x tok/s x inference_fraction x utilization x 86400 | |
| OpenAI | demand | 22.9 | 26.6 | 44 | — | medium | disclosed API 21.6T/day + messages/day x tokens/message |
| OpenAI | telemetry | 48 | 61.7 | 86.4 | — | low | disclosed API / (API:all-surfaces ratio 0.25-0.45) |
| OpenAI | financial | 4.1 | 21.2 | 79.1 | — | low | COGS: inference spend/day / frontier cost-to-serve ($/Mtok x5-6 frontier premium) |
| Anthropic | financial | 13.7 | 31.9 | 69.4 | — | low | API run-rate revenue/day / realized $/Mtok (1.5-6) |
| Microsoft | anchor | 1.4 | 2.2 | 3 | — | medium | Foundry 500T/FY2025 disclosed, grown to FY26 (+30% QoQ customer trend) |
| Meta | demand | 0.1 | 0.6 | 2.6 | — | low | DAU x msgs/user/day x tokens/message |
| Meta | hardware | 12.8 | 131.7 | 416.8 | — | low | fleet x tok/s x inference_fraction x utilization x 86400 |
| xAI | demand | 0.1 | 0.9 | 7.5 | — | low | DAU x msgs/user/day x reasoning tokens/message |
| xAI | hardware | 1.9 | 23.3 | 78.1 | — | low | fleet x tok/s x inference_fraction x utilization x 86400 |
| DeepSeek | demand | 0.2 | 1.5 | 9 | 1.9x vs 2025-03-01 anchor (PASS: within band) | medium | DAU x msgs/user/day x reasoning tokens/message |
| DeepSeek | financial | 4.9 | 10 | 17.8 | 12.9x vs 2025-03-01 anchor (FAIL: excluded) | very low | API revenue/day / realized $/Mtok |
| DeepSeek | hardware | 0.3 | 1.3 | 2.6 | 1.7x (ceiling/anchor; PASS: disclosed volume implies ~10-25% duty cycle, matching the global utilization prior) | low | fleet x tok/s x inference_fraction x utilization x 86400 |
| China aggregate | anchor | 140 | 140 | 140 | — | high | disclosed (2026-03-24) |
| Fireworks AI | anchor | 30 | 30 | 30 | — | high | disclosed (2026-06-23) |
| Together AI | anchor | 2 | 2 | 2 | — | high | disclosed (2025-09-15) |
| GLOBAL | telemetry | 5 | 15 | 40 | — | very low | unattributed remainder: self-hosted open-weights + Bedrock non-Claude + Groq/Cerebras + sovereign |
Hardware and energy rows are capacity ceilings, never summed into demand. Every input behind these rows lives in channel_inputs.csv with its source URL, band, and confidence.
The estimate across time
The same reconciliation, evaluated quarterly since January 2024. Anchored entities follow piecewise-exponential curves through their dated disclosures (Google's five points, China's three, OpenAI's four, Fireworks' four); modeled entities ride dated proxies (Anthropic's seven-point revenue series, Meta's MAU points). No hand-set correction factors — the historical line is the model, which is what makes the back-tests below meaningful.
| Quarter | Low | Mid | High | QoQ growth |
|---|---|---|---|---|
| 2024-01-15 | 1.5 | 2.2 | 4.3 | — |
| 2024-04-15 | 2.2 | 3.1 | 5.6 | ×1.4 |
| 2024-07-15 | 4 | 5.2 | 8.7 | ×1.69 |
| 2024-10-15 | 7.5 | 9.4 | 14.5 | ×1.81 |
| 2025-01-15 | 14.8 | 18.2 | 26.4 | ×1.92 |
| 2025-04-15 | 30.3 | 36.1 | 50.1 | ×1.99 |
| 2025-07-15 | 68.4 | 79.8 | 105.4 | ×2.21 |
| 2025-10-15 | 121.9 | 139.8 | 182.1 | ×1.75 |
| 2026-01-15 | 205 | 230.3 | 288.9 | ×1.65 |
| 2026-04-15 | 271.3 | 310.1 | 397.1 | ×1.35 |
| 2026-07-15 | 308.3 | 360.1 | 476.6 | ×1.16 |
| 2026-07-16 | 308.4 | 360.4 | 477.1 | ×1 |
T tokens/day. Part of the early-period growth is coverage arriving (China's series begins) on top of real growth — the two are disentangled in the crunch scenarios below by quoting both the full-period and trailing-12-month rates.
Dated closure checks
External global estimates compared against our band at their own dates — not just today.
| Date | External estimate | Their figure | Our band then | Verdict |
|---|---|---|---|---|
| 2025-09-15 | Tomasz Tunguz (The Token Race) Tunguz's sum has no China aggregate; scope-matched comparison is our ex-China figure (~76 T/day mid) -> consistent | 88 | 101 – 152.4 | below our band |
| 2026-06-15 | Exponential View via Epoch | 432 | 295.6 – 448.8 | inside |
When does demand meet capacity?
Global serving capacity is ~1,520 T/day mid-case (167–4,428), growing ~×3.3/yr (Epoch: global compute, 90% CI 2.7–4.1; the oft-quoted ~2.3× is NVIDIA-only). Demand is 360.4 T/day. Whether they cross depends on which demand growth you believe:
demand ×4.49/yr
Demand meets mid-case serving capacity around 2031-03.
demand ×7.69/yr
Demand meets mid-case serving capacity around 2028-03.
Where the honesty lives
Channels must earn trust
Each channel is tested against companies that disclose. DeepSeek's revenue-implied estimate landed ~13× above its disclosed 776B/day — so the financial channel is excluded for DeepSeek, and the failure is published, not hidden.
Top-down reconciliation
The bottom-up sum must contain the independent top-down estimates. Exponential View's ~432 T/day (relayed by Epoch) sits inside our 308.4–477.1 band — agreement we did not manufacture, since no external number feeds our estimate.
The remainder is published
Self-hosted open-weights, Bedrock's non-Claude traffic, Groq/Cerebras, sovereign clouds, and the undisclosed non-Foundry Azure inference are carried as an explicit unattributed band (5–40 T/day), not silently absorbed or guessed per-company.