GAIB

New episode of Where Capital Meets Compute: Evan Conrad, founder and CEO of SF Compute, on why the company had to stop brokering compute and start running the whole hotel. SF Compute began as a small AI lab about $5 million in the hole on a GPU cluster it had to sublease every month to survive. Today it operates clusters it doesn't own and sells the capacity on a physically settled forwards market where every buyer writes their own spec, with several hundred million dollars of volume so far ("not huge, but we're scaling"). We get into why he thinks credit is the real bottleneck for GPU clouds, and how SF Compute wants to make the GPU hour usable as collateral. This episode is sponsored by @computedesk, which builds financial infrastructure for compute traders. Their Compute Indexes publish daily reference prices for GPUs including the H100, H200, B200 and B300, and their platform helps teams price, procure, hedge, finance and settle compute from first offer to final agreement. Learn more at computedesk.com.

New episode of Where Capital Meets Compute: Thomas Galbraith, founder and CEO of Barkr, on what it takes to price hard assets accurately enough to stand behind the number. Barkr has run about $40B in valuations this year, backs roughly $500M of them with a Munich Re-backed guarantee, and works only from resale transactions. We get into how its GPU depreciation view has changed, the difference between a residual value guarantee and residual value insurance, and why he expects the coming correction to land unevenly. This episode is sponsored by @computedesk, which builds financial infrastructure for compute traders. Their Compute Indexes publish daily reference prices for GPUs including the H100, H200, B200 and B300, and their platform helps teams price, procure, hedge, finance and settle compute from first offer to final agreement. Learn more at computedesk.com.

How do you tell whether a customer is a person or an agent? Travis Cannell @traviscannell, COO of @vast_ai, says the clearest signal they have is API usage patterns over time, and those are shifting. Beyond the data, his point is simpler: the way people approach problems has changed, and the systems on the other end are getting longer memories and running for longer stretches. On building the procurement model for it: “we built this system for what was called the singularity, and now we’re calling it this agent thing.” Full episode available now.

“Humans love intelligence, and we can manufacture it now.” Travis Cannell @traviscannell, COO of @vast_ai, on what’s actually driving compute demand: not hype, but how much intelligence people want to produce. As the big labs absorb more capacity, the mid-tier buyers scramble for what’s left, a waterfall effect running down the market. He thinks it may be a little frothy, with cycles that let some error out, but sees it as underpinned by real breakthroughs rather than a bubble. Full episode available now.

Travis Cannell @traviscannell, COO of @vast_ai, on why “marketplace” is a confusing word for what Vast does. Some GPU marketplaces really do sell physical hardware. His preferred description is an infrastructure layer, because Vast does more than match buyers and sellers, starting with virtualization. Full episode available now.

“Google is for information. Vast is for computation.” Travis Cannell @traviscannell, COO of @vast_ai, on how he describes the company: Google doesn’t make websites, it organizes information. Vast doesn’t own compute, it organizes, tests and standardizes it. His five-to-ten-year view: training may start to look more like inference, which means fewer long contracts and more on-demand compute from different places and GPU types. Full episode available now.

Even identical servers, in the same data center, can vary by about 5% in GPU performance. Travis Cannell @traviscannell, COO of @vast_ai, citing a study, on why: the silicon lottery, heat dissipation, and where a GPU physically sits relative to the others. If compute varies even when you control for everything, pricing it all the same means you either underprice or overprice it. Full episode available now.