Evan Conrad, chief executive of SF Compute, describes his company's growth through a hotel. You start out trying to rent one room. Then the guest asks whether you will clean it, so you need the cleaning closet. Then they want room service, so you need the kitchen, and then security. "And at a certain point," Conrad said, "you're just running the whole hotel."
He was speaking on the Latent Space podcast, in an episode recorded on the sidelines of the AI Engineer conference in New York and published on October 20, 2026. SF Compute runs a market where companies buy and sell time on GPUs, the processors used to train and run AI models, either for now or for future dates. The hotel describes how a trading business ended up doing much of the physical work of a cloud provider. The rest of the conversation explained why Conrad thinks only a company that owns no chips can run that kind of market.
From one room to the whole hotel
The company started as an AI lab with a cluster, a group of GPU servers wired together, that it was trying to sublease. In his talk at the same conference on October 14, Conrad traced the business back to an AI music startup that had raised $500,000 and needed to rent out its cluster to survive. A company announcement names that earlier lab Junelark and says the market grew out of a need to get out of long computing commitments.
Once SF Compute was effectively running the hotel, Conrad said, its suppliers changed. They became "somebody that looks closer to a fund or a cloud" with access to a data center. The company then spent several years "building the entire operating capacity of a cloud" so it could deliver computing reliably. It calls this physical settlement, borrowing the commodity-trading term for a contract that ends with real goods changing hands rather than a cash payment. Someone who buys GPU hours through SF Compute gets machines they can actually run work on.
Why the market operator owns no hardware
"We do all the things a cloud does without being the cloud," Conrad said. SF Compute doesn't own the hardware it sells, and he was blunt about why: "if we did, we would be able to manipulate the forwards market." A forward market is one for capacity that will be delivered later, such as GPU time starting next month or next year. A company that owned GPUs and also ran the exchange could steer those future prices toward its own interests. "So we have to be Switzerland," he said. "We have to be independent, which creates a lot of sort of strange things about the company."
The episode opened with the question of why NVIDIA, the "800 pound gorilla" of the chip business, didn't build this itself. Conrad's answer was that NVIDIA "can't make it, but benefits tremendously from it existing," because a market it ran "would then not be independent enough for people to trade on it." Without trading, he went on, "you don't have enough liquidity for lenders to finance against the forwards market." Liquidity here means enough willing buyers and sellers that a contract can be priced and resold, which is what would let lenders treat future computing contracts as something to lend against.
He applied the same reasoning to GPU cloud companies such as Crusoe, CoreWeave and Lambda, to the hyperscalers (the largest cloud providers), and to inference providers, which run AI models for customers, such as Baseten, Fireworks and Modal. All of them, he said, have "too much demand on that side." In his view, a neutral market would help the whole industry by bringing down financing costs, but nobody can build one "unless they give up the ability to do all the big inference money and all the big cloud deals. You have to kill your business in order to make this. That's why nobody does it."
That financing question is a live one. On August 10, 2026, NVIDIA announced memorandums of understanding with six financial institutions aimed at mobilizing more than $500 billion in outside capital for AI infrastructure, according to a joint announcement. Conrad did not mention that plan, but his argument goes to a question it raises: how lenders can judge what computing capacity will be worth over the life of a loan.
Conrad also acknowledges the pull of becoming an ordinary cloud. In his conference talk, he said that once a market builder has the operating skills needed to deliver capacity reliably, signing large direct contracts with customers becomes commercially attractive. He presented staying a transferable market as a deliberate choice. Carmen Li, presenting Compute Exchange, another GPU marketplace, at the same conference, said it deliberately owns no GPUs so that it has no incentive to favor its own inventory.
What running the whole hotel involves
Asked to describe the company's technology, Conrad listed the physical work first. SF Compute staff visit data centers. They design the bill of materials, the list of every component a cluster needs, and do what he called supply chain consulting, using relationships with equipment manufacturers to keep deliveries on time. They design the cluster's architecture, then roll out the racks of servers themselves or supervise partners who do.
Then comes what Conrad called backplane management: taking over the power distribution units, network switches, DPUs (data processing units, chips that handle networking and security for a server), the remote-management controllers known as BMCs, and NVIDIA's UFM. NVIDIA's documentation describes UFM, or Unified Fabric Manager, as a control system with visibility across the whole network linking the servers. "When you buy a node from a cloud, there's the bare metal," Conrad said, meaning the physical server, "but there's all this shit around the bare metal. And in order to make that bare metal node work, you have to operate everything else."
SF Compute has written down what that means for the companies that own the hardware. Its access requirements, in effect since September 23, 2026, ask providers for the ability to recover and reimage machines remotely, administer DPUs, standardize firmware, partition the network and allocate shared storage. The document lists safeguards in return. Access is limited to the contracted cluster and kept separate from customer traffic, changes follow agreed change-control rules, firmware flashing needs the provider's approval, and original BIOS settings are kept so they can be restored. Some items can be arranged in other ways: shared power units, for example, can be handled through a provider's ticket process.
Turning data centers into clouds
Conrad said SF Compute isn't really in the data center business. It offers what he called "supercomputing as a service." A typical partner has some megawatts of power, or a whole data center, and is weighing whether to sell the GPU cluster, sell the site, or become a cloud itself. SF Compute's answer, he said, is "great, we will turn you into the cloud. We're the supercomputing part. You're the data center." When a partner has only powered land, SF Compute project-manages the work with partners until there is a data center, then takes over the computing layer. That requires some data center expertise in-house, Conrad said, "but we don't self-perform builds."
Someone once described the company as "Palantir for compute," he said, "because it's like these, each time it's this new rodeo." SF Compute works mainly in the US, but Conrad said it is deploying clusters around the world, often with partners who are themselves trying to become clouds and who "look like sovereign compute players": organizations building computing capacity for their own countries. His pitch to them is to "help you set up a cloud in your own country," with SF Compute bringing "the San Francisco offtake," meaning buyers committed to using the capacity, along with its expertise and cluster setup. Asked which partners he could name, he said there was nobody he could name yet and that announcements were expected in a week or two.
A data center you can program
Asked about milestones, Conrad said the company had closed several hundred million dollars of volume and had a significant number of new clusters on the way. One documented deal is a $245 million sale announced on September 10, 2026: two dedicated clusters of NVIDIA's B300 chips for an unnamed AI company, each worth $122.5 million over 36 months, with capacity arriving through the fourth quarter and into 2027.
The feature Conrad said he was most excited about was a networking one, part of a set of changes he said make SF Compute viable for platform-as-a-service companies. GPU clusters are linked by InfiniBand, a high-speed network that lets machines in a training run exchange data quickly. Many clouds give a customer one InfiniBand fabric, he said, and anyone who wants to redraw its boundaries has to contact the provider and wait for them to do it. SF Compute now lets customers subdivide the fabric "on the fly" without asking. He said this fits the market, where customers need to redraw those boundaries as they scale clusters up and down.
The company announced the feature, Elastic InfiniBand Partitions, in public preview on September 25, 2026. Customers create named, isolated networks through a command-line tool or API, and the company configures them through NVIDIA's UFM, with switches and network adapters enforcing the isolation in hardware. Its example starts with two machines for debugging, then reserves 64 for a five-day training run and attaches them to the same network when the reservation begins. If plans change, the unused part of the reservation can be resold. The documentation lists the limits. All machines in a partition must sit on the same physical fabric, and a machine's partition is set when it is created, so moving it means shutting it down and launching it again. A machine launched without a partition has no InfiniBand connection at all.
The conversation turned to whether this amounted to virtualizing the data center and making it programmable. Conrad agreed: "all the stuff that you would have gotten if you had owned the hardware, we're trying to virtualize that." The episode's show notes compare this to software-defined networking, which in the early cloud era separated control of a network from the hardware that carries its traffic.
The pitch to cluster owners
Conrad ended with an offer to anyone who wants to own a GPU cluster for its residual value or as part of a portfolio. SF Compute will help build it, run the supercomputing side and take responsibility for the service-level agreement, the contract's promises about uptime and performance. And if owning the cluster turns out to be a bad idea, he said, "we can immediately sell that cluster to somebody else. So it doesn't hit your balance sheet."