Key Takeaways

  • Trading compute looks like software until you have to physically deliver it, forcing SF Compute to run physical data center operations instead of staying a clean marketplace broker.
  • Early manufacturing runs of chips like Nvidia H100s and GB200s ship with OEM firmware bugs and silicon flaws that break under initial load.
  • Reliable compute providers run stress tests to deliberately kill weak GPUs and secure hardware replacements (RMAs) before handing clusters to customers.
  • Physical settlement requires controlling the backplane: managing power distribution units (PDUs), baseboard management controllers (BMCs), data processing units (DPUs), and InfiniBand switches.
  • Workload migration between clusters fails without dedicated private wide area networks (WANs) and matched storage backends, which standard data centers lack.

Software Brokerages Die in the Server Rack

Evan Conrad started SF Compute to build an open financial exchange for compute power. The theory sounded simple: build an order book, match buyers who need compute with sellers who have spare capacity, and collect a fee on the trade.

Then physical reality hit. Compute is not a fungible digital token that clears through an API. It is hot silicon sitting inside metal racks. If a seller delivers an unreliable cluster to a buyer, the exchange takes the blame. Conrad found that running an exchange meant taking direct operational control over the facility. “And at a certain point, you're just running the whole hotel,” Conrad said. “And the moment you're running the whole hotel, you say, 'Oh, wait a second. Instead, my supplier on this market is somebody that looks closer to a fund or a cloud that has access in our case, a cloud that has access to the data center, and we're going to level up.'”

Settling a trade means managing the backplane directly. Conrad's team had to hook into power distribution units (PDUs), baseboard management controllers (BMCs), data processing units (DPUs), unified fabric managers (UFM), and top-of-rack switches. You cannot guarantee uptime or enforce physical delivery without direct hardware access.

The First Manufacturing Runs Always Break

Hardware quality is not uniform across a chip's lifespan. When Nvidia releases a new architecture, early deployments suffer from unpatched OEM firmware defects and silicon infant mortality.

Conrad saw this pattern repeat across multiple chip generations. “The early GB200s broke. The early H100s broke,” he explained. “If you were buying H100s or you were buying GB200s, you might really want the ones that were deployed after a certain time because you're not getting the first GD1s off the rack.”

Because early hardware breaks, physical settlement requires a brutal burn-in procedure. SF Compute cannot simply accept a cluster and hand the keys to an AI team. Instead, they run intense stress tests designed to destroy fragile components upfront. “When you get a cluster, a good cloud, what it does is it tries to kill all the GPUs before they ever get in the hands of the customer,” Conrad said. “And then the GPUs that are broken or the nodes that are broken, you get RMAd. You basically go and replace them.” If you do not break the bad cards in week one, your customer will break them mid-training run in week three.

Workload Migration Demands Dedicated Private Fiber

Even when the GPUs survive stress testing, trading spot or short-term compute creates another physical bottleneck: preemption and relocation. When an expensive long-term buyer books a cluster, an exchange needs to evict lower-tier batch workloads and migrate them elsewhere.

That migration fails without specific networking. “To make that work you also need like a fast transit path between one cluster and another,” Conrad noted. “Typically this is implemented with a WAN but most of these clusters don't necessarily have a wide area network.”

Without private wide area network links and compatible storage backends between facilities, moving terabytes of model weights and training checkpoints takes hours or days. The exchange grinds to a halt. If you want to trade compute like a commodity, you must own the physical pipes that connect the server rooms.

What to Do With This

Before you sign a contract for cheap GPU capacity, inspect the rack delivery date and ask the provider for their burn-in failure log. If they cannot prove they pushed those H100 or GB200 nodes through full-load thermal stress testing and replaced the dead cards through RMAs, schedule 72 hours of NCCL and GEMM stress tests on day one before you launch your real training job.