GPU capacity is tight, scattered and expensive to reach. The next constraint is not silicon. It is selling, routing and recovering compute by the hour.
Moody's forecasts that six companies, Microsoft, Amazon, Meta, Alphabet, Oracle and CoreWeave, will spend about $785 billion on capital equipment in 2026 and close to $1 trillion in 2027.
Then look at what that spending runs into. The International Energy Agency puts data center electricity use at 415 terawatt-hours in 2024, about 1.5 percent of the world's electricity, and projects about 945 terawatt-hours by 2030. Gartner predicted in November 2024 that 40 percent of existing AI data centers would be operationally constrained by power availability by 2027.
Getting power to a site is slow too. One analysis of grid data found that the median wait to connect a power plant to the grid rose from under 20 months in 2005 to 55 months by 2023. That figure is for power plants, not data centers, but it shows how long connections can take.
Chips are not the sole constraint. Power and time are scarce too.
That is the backdrop for the model that defines this cycle. CoreWeave, the sector's best-known player, reported these results for the second quarter of 2026:
| Metric | Q2 2026 |
|---|---|
| Revenue | about $2.58 billion |
| Net loss | $626 million |
| Interest expense | about $640 million, up from $267 million a year earlier |
| Revenue backlog | about $104 billion |
Its 2026 capital spending guidance is $35 billion to $39 billion, according to press coverage of the earnings call.
That is what is happening now, not a verdict. The open question for 2026 into 2027 is whether financing and building every megawatt is the most viable, or the most feasible, way to serve a market where capacity already exists elsewhere. Growth tied to debt and construction schedules carries a different risk profile than demand routed to capacity that is already built.
The capacity that does exist is spread across hyperscalers, neoclouds, regional operators and hosting companies. Much of it is committed under long contracts, and operators often have racks sitting idle between them.
To sell a GPU by the hour, an operator needs self-service provisioning, tenant isolation, metering, automatic payment settlement and failover when a machine dies. That is a software company bolted onto a facilities company, and many operators are not set up to build it.
Gartner's forecast, as reported, puts spending on AI-optimized infrastructure services at $42 billion in 2026 and $66 billion in 2027, with inference overtaking training in 2026. Inference demand is made of many small, automated purchases, and we expect software agents to make a growing share of them. An agent does not sign a multi-year reservation or fill out a sales form. It needs to find capacity, pay and run.
Payments are forming quickly. In July 2026 the Linux Foundation launched the x402 Foundation with 40 founding organizations, including Google, AWS, Visa, Mastercard, Stripe and Coinbase. x402 uses the HTTP 402 "Payment Required" status code so that software can pay for a web resource without an account or an API key.
Kilawatt Cloud owns no data centers and buys no GPUs. What it owns is the control plane.
request -> quote across six connected suppliers
-> pay per job with x402
-> launch the machine
-> watch it
-> if the supplier fails, move the job to another one
Nothing is financed, built or racked in between. Capacity moves from point A, the buyer's request, to point B, a running machine, through whichever supplier is available. Because the platform holds no hardware, it has no reason to favor one source over another, and adding a supplier is an integration rather than a construction project.
We destroyed a running machine and measured what happened next.
Every drill is logged, and the drill log closed on October 21, 2026.
Any platform can add x402. What is hard to copy is a track record and a posture: published proof, neutrality, because we own no hardware, and safety by default, with the full cost of a job authorized before launch.
In 2027 and after, capacity stays tight, buying shifts to small automated purchases, and reliability data becomes the scarce asset. That is the market Kilawatt Cloud is built for. The category is funded: SF Compute, a marketplace that owns no GPUs, raised a $40 million Series A in December 2025.
Next for us: onboarding paying customers, widening the supplier base and bringing regional operators onto the network.
If you build agents that need compute, or run GPU capacity and want to talk, reach us at hello@kilawattcloud.dev or kilawattcloud.dev.
Estimates, opinions and plans are labeled as such. Kilawatt Cloud is a trade name of Right Recruit LLC.