AivexaNewsSearch
AI news for builders and product teamsChecked every hour

The Programmable Data Center — Evan Conrad, SF Compute

Collected Oct 23, 2026

SF Compute CEO Evan Conrad used a session on the new Compute Trading track at AI Engineer New York to describe how his company has grown from a cluster subleasing startup into something closer to a full cloud operator, and how that work underpins the GPU-hours market it runs. The company, which began life as Junark trying to sublease a cluster, has spent the past several years building what Conrad calls the operating capacity of a cloud in order to actually deliver the compute traded on its market. That is what he means by physical settlement: a buyer receives usable compute, not simply a cash payment against an index.

His analogy is a hotel. A customer asks for a room, then for cleaning, room service and security. Eventually the seller discovers that delivering the room means running the hotel. In SF Compute's case, the supplier side of its market now looks less like a random cluster owner and more like a fund or a cloud with access to the data center.

Settlement, in this framing, is an engineering problem. Conrad said the company co-designed its physical settlement layer and its order book at the same time, so that the order book can capture the real constraints of a cluster. Those include whether a running workload can be interrupted, how quickly it can recover from storage, which GPU generation, networking and software configuration is involved, and whether the storage underneath is something like a Weco or Vast system. Connectivity between clusters matters too: many GPU clusters are sited next to power rather than next to customers, so a fast wide-area transit path between sites cannot be assumed, which affects how easily a preempted workload can pop back up elsewhere.

On liquidity, Conrad was blunt: the compute market today is not very mature, deep or liquid. SF Compute first had to bring supply online, then build the whole platform, which he said has just started working. About 5% of total capacity is now on the order book, with more being moved over and new capacity arriving already connected to it. Volume is small but liquidity is concentrated, and the company tracks whether that liquidity helps commercial participants. As a rough measure, for every $100 a customer spends buying capacity, they recover $25 on average by selling it back to someone else.

Contract standardization, which was a theme of the last conversation with Latent Space two years ago, remains far off in Conrad's view. He argued that you cannot standardize a compute contract until the underlying operations are made the same, because two virtual machines from a single operator can differ sharply. He cited GPU kernel launch times as one source of divergence, with one VM type showing roughly 40% degradation versus another, which shows up in the end as worse model flops utilization. That is the compute equivalent of crude grades: without established equivalents, there is nothing to standardize around.

SF Compute's alternative was to build an order book with multiple constraints attached to each order — a time window, performance ranges and other properties — and match buy and sell orders in a large matching engine. This lets the market generate liquidity before the technology settles down, but the trade-off, which Conrad acknowledged, is less concentrated liquidity and wider spreads between bids and offers. He framed the flip side for customers: a product that can use many types of capacity gets more matching options and a chance to capture that spread.

The dimensions that matter in matching include whether capacity is running a workload right now or is future capacity. Selling current capacity means interrupting a live workload, so an interruption-tolerant buyer and one that cannot be preempted are not interchangeable. Capacity beginning a month out can be bought and sold much more freely. Hardware cohorts also matter: early H100s and early GB200s broke, Conrad said, so sophisticated buyers may pay more for units deployed after a certain date rather than the first units off the rack.

On why Nvidia has not built such a market itself, Conrad's stated view is that the chipmaker cannot, because it would not be independent enough for people to trade on, but benefits tremendously from one existing. That independence argument is central to his case for a neutral market operator: companies with large hardware or inference businesses have incentives that make running an exchange harder. The unglamorous side of the business, he said, involves visiting sites, designing clusters, running the backplane and making infrastructure programmable; one example he gave is the ability to subdivide InfiniBand fabric on the fly.

The discussion also covered spot versus forward pricing, batch versus live inference leases, and why an index can drift away from the real cost of buying compute. The episode was recorded on the sidelines of the Compute Trading track at AIE NYC, and Latent Space pointed listeners to its previous SF Compute conversation, Conrad's AI Engineer talk on how compute markets work, coverage of SF Compute's $245M compute sale, and Abel, the company's settlement ledger.

Why it matters: teams that rent GPU capacity increasingly face a choice between opaque contracts and a market whose prices depend on hardware cohort, interruptibility and storage behind the cluster. If SF Compute's matching approach works, buyers could resell unused hours more reliably, though thin liquidity and wide spreads, by Conrad's own account, mean early savings may be limited. Developers building batch or flexibly scheduled workloads are the most likely near-term beneficiaries; inference teams needing guaranteed, non-interruptible capacity are likely to stay on conventional cloud contracts for now.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

From GPU trading to physical settlement, independent markets, and programmable compute