Who powers the Token Economy?
Explore the companies behind AI tokens: their chips, data centers, contracts, prices and inference traffic. Every figure links to its source.
How the companies connect
Facility owners lease powered space to cloud operators. Clouds supply AI labs and other clouds. Follow the disclosed links to see who depends on whom.
Swipe across the diagram to follow facilities → clouds → buyers, or choose a company above.
Equal-width links show selected public agreements, including unpriced ones. Motion indicates direction, not volume. Small numbers count counterparties. A connected path does not allocate particular racks or megawatts through a reseller. Undisclosed relationships stay absent.
Most numbers in this business are announced rather than measured. Contract terms can include floors or ceilings, a GPU count is often worked out from megawatts, and a hyperscaler's fleet is somebody else's model. So every figure here carries its source and a tag distinguishing reporting, calculation, and estimation.
- R
- Reported by a company, filing, or named press source; this tag is not independent verification.
- D
- Derived from reported numbers by a stated formula, such as value ÷ term ÷ GPUs.
- E
- Estimated by an analyst or inferred from indirect evidence. Source ranges are retained where available.
Hover over any mark for its terms and source; click to pin the box and open its links. Close it with Escape, the close button, or a click outside. Click a buyer or supplier in the contract flow to see only its deals. Tables sort when you click a column. A source-note marker opens the two dated claims: “source conflict” means unresolved disagreement, “different bases” means the definitions differ, and “later evidence” identifies dated updates. Source-date warnings remain separate from verified revisions.
Chips and power
No hyperscaler says how many accelerators it owns. The best public estimate is Epoch AI's model of chip deliveries, which converts every chip into H100-equivalents so that a TPU and a GPU can share an axis. Neoclouds and hosts rarely disclose GPU counts either, but they do disclose power, so the second view compares them in megawatts, including clearly labelled partial AI-site estimates for the hyperscalers.
Accelerator fleets by chip designer, and power by provider
Native accelerator counts: NVIDIA, AMD, TPU and Trainium
Estimated delivered chips, stacked by owner. These are physical chip counts, not equal units of performance. Renting capacity does not transfer chip ownership.
Individual disclosures and source records
The table also includes reported deliveries, customer-use subsets and planned deployments. These records can overlap the chart and one another. H100/H200 is one combined source category; it is not split into invented counts.
Compute deals
A compute contract is a promise to pay for capacity over several years. The flow below runs from buyer to supplier, and the default view shows relationships even when prices are undisclosed. Switch to disclosed dollars to size the bands by value. Framework deals such as Stargate's $500B plan are left out, because they commit no one to a price.
Building leases get their own view. A host that rents a powered hall to Fluidstack, which then sells the same capacity to Anthropic, would otherwise be counted twice.
Who buys capacity from whom
Relationships include unpriced TPU access and AMD supply agreements. Link widths and circle areas scale with disclosed nominal value in this view; unpriced links stay thin and unpriced nodes hollow. Minimum node sizes preserve visibility. Filter smaller agreements with the slider, or select Disclosed dollars for a stacked value view. Amounts retain their conditions and can overlap.
How implied prices are worked out
$/GPU-hour is shown only when a deal gives a GPU count. Otherwise the page gives an all-in $/kW-month from the megawatts, which stays comparable across chip generations. For GPU deals it adds an estimated $/GPU-hour range at pre-Rubin power draw, which is a floor for 2027 and later deliveries. Leases of powered buildings are priced per kW-month and include no chips.
Every deal in the current view
Supplier growth
All selected companies share one set of axes. Absolute values show scale; Indexed compares growth from a Q2 2025 baseline. Revenue uses a cloud segment where available and otherwise the reported group; exact scope and dates appear below.
Growth on shared axes
Dates, values and reporting scope
Compute prices
Compare market benchmarks first, then published provider rates by commitment term. A hyperscaler’s one-year retail reservation is not a negotiated neocloud contract. TPU and Trainium rates use physical chip-hours; price alone does not compare model throughput.
Market price movements
Same-series changes from public benchmark data, calculated over calendar windows. Hover a change for its exact start and end dates. A rental index is not a GPU resale price.
H100 one-year rental eased about 3.1% over the last 30 days while B200/B300 one-year indices rose. SemiAnalysis’s B200 composite fell about 6.3% in the last week but remains higher over 30 and 90 days. These observations do not show a broad collapse across GPU markets.
Published rates per physical accelerator-hour
One dot per provider: its lowest published region/configuration quote for the selected terms. “From” rates stay separate. The table retains every region, configuration, billing formula and source.
Rate cards, original billing units and conversions
Rental benchmarks, with different methodologies
Silicon Data standardizes rental observations; Ornn describes executed trades. These are different benchmarks, not interchangeable one-year contract prices. Current values and source availability are shown above.
What the big deals imply per kW-month
Spot prices over time
Nebius hourly posted spot quotes, October 26–November 2. These can be interrupted and are not executed trades or reserved contracts.
First and last daily medians use 24 hourly observations each; the final endpoint is outside the last daily window. The line includes all 169 endpoints. Prices have no volume weighting or delivery guarantee.
Inference and token use
Public disclosures describe inference through shares of spending or usage, and through tokens served. These views answer different questions: revenue, compute, spending and token counts are distinct measures. Each chart retains its sources, dates and scope.
Reported inference shares
Shares of revenue, usage, spending or compute retain their original basis.
Reported token use
Gold bars and D mark conversions. OpenRouter’s daily rate is its weekly total ÷ 7; OpenAI’s is its minute rate × 1,440.
Dates and scopes differ. Bars show one latest dated figure per subject in the selected period; the table retains every collected series. API-only, chatbot, single-cluster and company-consumption disclosures have different coverage. A ≥ marker retains a lower bound. Unavailable totals appear without a bar. Router and provider traffic can overlap, so these figures do not form a market total.
Coverage and undisclosed totals
If this traffic ran on DeepSeek
Compare V4 Pro and V4.1 Flash on 72-GPU GB200 NVL72 racks. There is no verified, non-overlapping total for the entire inference market here. Choose a disclosed traffic boundary or enter an explicit market assumption.
“Usable capacity” reserves room for downtime, peaks and other losses relative to the benchmark. At 70%, 30% is reserved; it is not a measured GPU utilization rate. Per-user speeds below are reciprocals of P90 response inter-token latency.
Enter a positive traffic volume and nonnegative finite cost and power assumptions.
Tokens are not all the same work
Cache hits apply only to input tokens. This control changes the token composition above; it does not change the rack estimate. The runs do not isolate uncached prefill capacity, so they cannot calibrate a cache-hit-to-rack curve. Rack estimates retain the chosen benchmark’s context and cache behavior.
Assumptions, packing and source checks
Defaults: SemiAnalysis’s public GB200 specifications and ownership model, retrieved November 2, 2026. Annual cost = all physical GPUs × hourly ownership cost × 8,760. Power includes the source’s host, networking and cooling allowance; no extra PUE multiplier is applied. Neither figure is a purchase quote.
Method: group equivalents = max(input tokens/day ÷ 86,400 ÷ group input TPS, output tokens/day ÷ 86,400 ÷ group output TPS) ÷ usable capacity. Round up to groups, then to whole racks. This is a fixed-rate extrapolation from coding-trace replay. A different input/output mix, context length, cache distribution or serving stack can change the attainable rates. Full-rack packing was not measured.
Benchmark scope: AgentX variable-length coding traces, prefix caching and speculative decoding. The Flash configurations use different engines and cache setups; this is a configuration comparison, not a controlled model-only comparison. The vLLM Flash run also uses synthetic speculative acceptance. Input throughput counts reused context as well as fresh input.
Why not use the headline? The run pages’ 100-token/s figures interpolate total logical throughput. They are not output throughput. Pro’s overview also applies a decode-share factor after the raw producer has already divided by all allocated GPUs. This calculator uses the raw absolute input and output rates and explicit physical groups instead. Producer · Overview transformation.
Unresolved source metadata: Pro’s API labels precision “fp4”; inspected worker logs name deepseek_v4_fp8 with bfloat16 dtype. The API’s October 14 date also differs from the September run timestamps. These fields do not establish uniform arithmetic precision or an exact measurement date. The selected raw values and source URLs are preserved in the benchmark snapshot.
The earlier 40T/day example can be reproduced with 2% output, 100% usable capacity and Pro result 444911: 244 racks. Its 60 allocated GPUs per rack include both prefill and decode; only 36 are decode GPUs. This configuration reaches 95.8, rather than 100, tokens/s/user at P90 latency.
The historical Fireworks Pro 0813 serverless tariff is not used as current revenue: Fireworks says that endpoint was retired from serverless. Current GPU ownership cost and hypothetical API billing are different quantities.
Compute in China
Chinese clouds disclose spending and revenue but almost never fleet sizes. Epoch AI tracks China as a single bucket, so the chart shows its estimate of accelerators delivered to China by the end of 2025, by chip rather than by company.
Accelerators delivered to China, end-2025
What the Chinese clouds disclose
Company profiles
Empty fields mean not disclosed, not zero. Hosts show megawatts only, because the GPUs in their buildings belong to their tenants.
Sources and methods
Selected primary-source evidence, with dates and scope
Source conflicts, later evidence and differences in definition
Each badge opens both sources, their dates, the exact issue and our treatment. A difference in date or unit is not automatically a contradiction.
Related surveys: where they are stronger and what this page adds
Established surveys already cover this market. This comparison describes inspected public features; it is not a ranking of accuracy or a claim to be first.
Download the presentation data: providers CSV · contracts CSV · prices CSV. Source URLs, units and dates are retained.