Latent Space

Evan Conrad on why GPU compute needs physical settlement

Evan Conrad· Founder at SF Compute
·~37 min·English·Latent Space
GPUInferenceAI InfrastructureBusiness Strategy
TL;DR

Evan Conrad explains how SF Compute is turning GPU time into a tradable commodity, and why you have to be able to physically deliver the chips before you can safely build a market on top of them.

01Core Thesis

Physical settlement first

SF Compute spent three years learning that you must be able to actually deliver the compute before you can safely trade a price on it.

cash settlement as a future concept is a great idea. Um, we just think it should be derived from physical settlement.

— Evan Conrad, Latent Space
Key Insight
The deeper claim is that a price is only trustworthy if it is anchored to something you can actually be handed. By operating delivery itself, while owning none of the hardware, SF Compute makes its price hard to fake, and that credibility, not the trading engine, is what a lender or hedger is really paying for.
02The Hard Problem

Even oil is not oil

A GPU-hour is not a commodity yet because chips differ by generation, memory, interconnect, storage, network, and even manufacturing batch.

these things are not as fungible as you think. We're going to try and make them fungible. That's the industry's job to do that.

— Evan Conrad, Latent Space
Key Insight
Standardization is usually step one for a commodity market. Conrad inverts it: force a standard too early and you freeze the wrong grade while the hardware is still changing every quarter. The non-fungibility is temporary, but pretending it is already gone is dangerous.
03The Mechanism

Make your own SKU

Instead of waiting for standard contracts, the order book lets a buyer attach constraints to an order and the engine matches whatever supply fits.

the goal is to make it so that somebody else can make their own SKU. So if I want to buy black wells, I can say I'm willing to buy black wells at $6 an hour or less.

— Evan Conrad, Latent Space
Key Insight
By pushing complexity into a query instead of a contract, SF Compute turns a standardization problem into a matching problem. It is the same move exchanges made when they let traders post limit orders instead of negotiating every trade by hand.
04What Trades

The market for breaking leases

Compute splits into a live market and a forward market, and the price of interrupting you is set by how strict your service guarantee is.

the price you would pay, um, in order to interrupt your batch job? It's very different than the price you would pay to interrupt your, um, live job.

— Evan Conrad, Latent Space
Key Insight
Pricing interruption by service level quietly turns reliability into a tradable asset. The real product is not compute, it is the right to interrupt someone else, which is exactly what a spot and forward split in power markets sells.
05What It Reveals

One of these numbers has to change

You can buy a B300 for about five dollars an hour but earn twenty-five to forty running inference on it, a gap that signals the market has not cleared.

right now if you took a B300 and you tried to run an inference on it, um you can make $40 an hour or so um selling that

— Evan Conrad, Latent Space
Key Insight
A five-dollar input that produces twenty-five to forty dollars of output is not a stable price, it is a sign the market is too young to have found the clearing level. Either chip prices rise or inference margins fall, and the exchange exists to let both sides find out which, faster.
06Why It Is Rare

Nobody who sells chips can build it

The exchange has to stay independent, so Nvidia, the hyperscalers, and the inference clouds all benefit from it but cannot run it themselves.

our impression is that Nvidia can't make it but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough for people to trade on it.

— Evan Conrad, Latent Space
Key Insight
Neutrality is a business model, not a virtue. The same independence that makes the exchange trustworthy is what makes it un-buildable by anyone with chips to sell, which is why the incumbents would rather it exist than own it.
07How Delivery Is Made Real

The cloud, without being the cloud

To make delivery real, SF Compute runs the full operational stack of a cloud, from site visits to backplane management, while owning none of the hardware.

we do all the things the cloud does without being the cloud.

— Evan Conrad, Latent Space
Key Insight
The pitch is really a bet that operations, not ownership, is the scarce skill. If that holds, the asset-heavy clouds are carrying a cost SF Compute shed while keeping the hard part, which is why Conrad jokes they look like Palantir for compute, a new rodeo each time.
08Where It Goes

Make the data center programmable

SF Compute is virtualizing the data center so capacity can be redrawn in software, then exporting that capability as sovereign clouds around the world.

we're virtualizing data center and making it programmable. So, all the stuff that you would have gotten if you had owned the hardware, we're trying to virtualize that

— Evan Conrad, Latent Space
Key Insight
Making the data center programmable is the same unlock the public cloud rode: once the hardware controls are virtual, a customer can redraw an InfiniBand fabric or resize a cluster without anyone touching a cable. Exporting that operating capability as sovereign clouds is how a US operations team becomes infrastructure other countries can run themselves.