● GAMBIT DATA · THE AI SUPPLY CHAIN
Follow the AI trade from dollars committed to tokens sold
The AI economy moves through four measurable layers. Capital expenditure shows what companies intend to build. Datacenter filings show how much of that ambition is becoming powered capacity. GPU rental prices show whether compute stays scarce once that capacity reaches the market. Model prices show what customers finally pay for the intelligence compute produces.
AS OF 10 AUGUST 2026 · SOURCE: GAMBIT · CURRENT VALUES IN THE TERMINAL ↗
The thesis
No layer answers the question on its own
Rising capex is not the same as operational capacity. Announced megawatts are not available compute. Available compute does not guarantee firm rental prices. Lower GPU costs do not necessarily reach customers as lower model prices. The information is in the transitions, and Gambit tracks each one, dates every observation and keeps the source attached.
The lag between the layers is measurable rather than rhetorical. Capital gets committed on an earnings call. Capacity energises a median 8.7 months after it is announced, across the 534 SEC-filed sites in the tracker. That supply then competes in the rental market, and the cost of a GPU-hour sets the floor under what a million tokens can profitably sell for.
| Layer | Core question | Gambit measure | Desk |
|---|---|---|---|
| 01 Capital | Is investment accelerating? | Quarterly capex and finance leases | Hyperscaler capex |
| 02 Physical | Is spending becoming operational capacity? | Megawatts by pipeline stage | Datacenter buildout |
| 03 Compute | Is usable compute scarce or abundant? | Dollars per GPU-hour | GPU rental prices |
| 04 Output | What are customers paying for intelligence? | Dollars per million tokens | AI model prices |
Reading the chain
Five patterns worth recognising
These are interpretive lenses, not proven causality. The four datasets differ in frequency, coverage and lag, and a filing quarter is not a quote snapshot. What they do reliably is tell you which question to ask next.
01 Capital rises, capacity does not
- Capex accelerating
- Announced projects increasing
- Operational megawatts flat
- Grid queues or delivery lags widening
Possible reading. Demand for power, construction and equipment is strong, but compute supply is being held up by physical constraints rather than by willingness to spend.
02 Capacity rises and compute prices fall
- Energized megawatts increasing
- Announced to operational conversion improving
- Comparable GPU medians declining
- Spot discounts widening
Possible reading. New supply is reaching the market faster than demand absorbs it. Scarcity is easing and the economics of owning compute are getting harder.
03 Capacity rises and compute prices hold
- Operational megawatts increasing
- Quote depth improving
- Rental medians stable
- Committed terms still rich against spot
Possible reading. Demand is absorbing the new supply, and buyers are still paying for guaranteed access rather than taking what is available.
04 Compute gets cheaper, output does not
- Rental prices declining
- API prices stable
- Cost per capability improving
Possible reading. Lower input costs are staying with the provider: as margin, as more inference spent per request, or as a better model at an unchanged price.
05 Output prices fall faster than compute
- GPU prices stable
- Model API prices falling
- Capability gaps narrowing
Possible reading. Competition at the model and application layer is passing value to customers faster than infrastructure costs are coming down. Undifferentiated inference is losing its premium.
Questions
What is the AI economy stack?
Four measurable layers that the AI trade moves through in sequence: the capital committed to building capacity, the physical capacity that capital becomes, the price compute clears at once that capacity is live, and the price customers pay for the intelligence that compute produces. Gambit publishes one dataset per layer, each dated and traceable to the filing or quote it came from.
Why read the four layers together?
Because no layer answers the question on its own. Rising capex is not delivered capacity. Announced megawatts are not available compute. Cheap compute does not mean cheap tokens. The transitions between layers are where the information is, and reading one in isolation is how the AI trade gets misread in both directions.
How long does capital take to become operational capacity?
On Gambit's own tracking of SEC-disclosed sites, the median lag from announcement to operational is 8.7 months. That figure is why capex and buildout cannot be read as the same signal in the same quarter.
Is any of this a forecast?
No. Every figure is an observation with a date on it and a source behind it: filed cash-flow statements, disclosed site records with their EDGAR links, provider quotes with the sample size attached, and listed API prices. The patterns on this page are interpretive lenses for reading those observations, not predictions.
Start anywhere