Layer 2.5 · head to head
SF Compute vs Lambda
Buying a market window against buying a reservation.
| SF Compute | Lambda | |
|---|---|---|
| Can you leave | Portable. The workload is standard enough to move to another vendor without a rewrite. | Movable with effort. Expect to redo images, storage wiring and networking. |
| Counterparty | Market mechanics rather than a vendor relationship. | Private, so you are diligencing a company that does not have to tell you anything. Reported among the group carrying GPU-collateralised debt maturing 2026-2028. |
| What it is | Brokers capacity somebody else owns. Cheapest headline rates, and the machine you get is not the machine you chose. | Purpose-built GPU cloud. Owns or leases its own datacentre capacity and sells it directly. |
| How you reach it | You get the physical machine. Maximum control, and every layer above it is yours to run. | You get virtual machines. The familiar cloud model. |
| Accelerators | H100 clusters | H100, H200, B200, GH200, A100 |
| Regions | US | US |
| Pricing model | A genuine spot market for cluster time, with prices that clear and are published. | On-demand by the hour, plus reserved clusters. Pricing is published on the website, which is rarer at this layer than it should be. |
| Getting started | Market participation. | Credit card, single GPU, minutes. |
| Capacity | Cluster-scale blocks. | Frequently sold out on the newest parts. Availability is the constraint, not price. |
| Ownership | Private. | Private, widely reported as IPO-track. |
SF Compute
For: Training runs that can start when the price is right rather than when the calendar says.
The catch: You are buying a window, not a service. If your run overruns, you are back in the market at whatever it now costs.
Economics: The closest thing to a true commodity market for training compute. Its clearing price is a genuinely useful benchmark for what a reserved contract should cost.
Lambda
For: Researchers and small teams who want a real GPU in the next ten minutes without a procurement conversation.
The catch: The thing you want is often unavailable. Reserved capacity solves it and turns the credit-card product into a contract.
Economics: Among the lowest published rates on the newest silicon — reported lowest on B200 in an August 2026 comparison. Cheap when you can get it.
Neither table row is a price
Deliberately. Published on-demand rates at this layer move weekly, and essentially nobody signing a real contract pays them — every serious buyer pays less than every list figure either of these companies publishes. Quoting one here would date this page within a month.
The GPU rental price index carries dated, sourced figures instead, and the durable finding there is the spread: the identical H100 rents from roughly $1.38 to $12.29 an hour depending only on who you rent it from.
The layers underneath both
Whichever you pick is renting you chips in a building that needs power. In 2026 that is the constraint that binds: Microsoft has disclosed an Azure backlog it cannot fill for want of megawatts rather than accelerators, and the US interconnection queue exceeds 2,600 GW with roughly 80% of projects withdrawing before they energise.
Layer 2 — Silicon · Layer 1 — Energy · The interconnection queue