Key Takeaways

  • SF Compute CEO Evan Conrad argues that financializing compute through cash-settled futures fails because compute requires physical settlement and direct infrastructure control.
  • Before centralized electrical grids existed, factories built private generators sized for peak capacity. Conrad sees AI labs repeating this exact inefficiency with private clusters.
  • Financial desks try to treat compute like agricultural commodities, but supercomputer clusters depend on complex networking topologies and physical delivery that index pricing cannot capture.
  • Venture capital portfolios currently absorb the risk of volatile GPU pricing, creating an artificial bubble that transparent forward contracts can defuse.

The Factory Generator Problem

Before centralized power grids existed, manufacturing plants could not buy steady electricity off a shared line. Every factory owner had to buy an on-site generator and hire staff to maintain it. Because factory power demands fluctuated throughout the week, owners had to size their equipment for peak loads. Most of that capacity sat idle during off-hours, draining capital and slowing expansion.

Evan Conrad sees the AI sector trapped in the exact same bottleneck. Foundation model startups, research labs, and application developers currently lease or purchase private clusters to handle their maximum training spikes. “Right now, compute kind of looks like the early days before there was a power grid,” Conrad explains. “When that happened, every factory ran its own generator and they had to schedule that generator to exist for peak capacity.”

Supercomputers Are Not Soybeans

Wall Street wants to turn raw compute into a standard asset class. Several trading platforms and financial desks have tried to launch derivative markets for GPU hours. Their playbook looks familiar: build a cash-settled futures contract based on an aggregate index price of cloud compute.

Conrad argues this approach collapses the moment engineering teams run real training workloads. “And the way that they're trying to do it is they're typically trying to create some sort of index price and then you can create a cash settled future on top of this. Our impression is that by doing this, you're divorcing it from the actual technology under the hood.”

Compute cannot be treated like a uniform commodity. “These are supercomputers, not soybeans,” Conrad points out. A working AI cluster requires specific networking configurations, interconnect speeds, thermal management, and co-located hardware. If an index-linked financial contract settles purely for cash, the buyer still has no access to physical silicon when models need training. To create a functioning market, the exchange operator must guarantee the delivery of actual cluster capacity. “That market operator also needs to do what's called physical settlement,” Conrad notes. “And in this case, physical settlement we think means running the whole damn thing.”

Deflating the Venture Capital GPU Bubble

Because developers cannot reliably lock in future GPU pricing, cloud providers force multi-year lock-in agreements, and AI startups burn equity cash to secure compute reserves. This dynamic pushes operational infrastructure risk directly onto balance sheets and venture funds.

“So if you're looking for a bubble, it's held currently by the venture capitalists,” Conrad says. “And we're trying to solve that problem by de-risking compute in the future by basically making it possible to know like what the price of GPUs will be off into the future.”

When engineering teams can lock in future delivery dates and standardized cluster capacity through physically backed contracts, compute shifts from an existential venture gamble into a predictable operational expense.

What to Do With This

Audit your infrastructure commitments this week. Calculate the spread between your average GPU capacity and your peak usage over the last ninety days. Stop signing rigid multi-year capacity reservations for speculative workloads, and demand shorter, physically settled delivery terms from your hardware vendors.