The currency of Silicon Valley right now might look like dollars – billions of them – but talk to the founders of AI labs, and they’ll point to something else: compute.
Compute – and specifically, an ongoing shortage of it – is what leads some startups to raise billions of dollars without a product, and others to join Big Tech. And it’s a double-edged sword if you over-commit to costly GPU clusters that you end up not needing. Then, you need to find someone else to take the compute off your hands – or face bankruptcy.
That was the situation that Evan Conrad found himself in years ago, when the small voice-focused AI lab he ran realized it was on the hook for $400,000 in monthly payments, with only $500,000 in the bank. Becoming a compute marketplace wasn’t some lofty philosophical decision – it was a necessity, a founder trying to keep his team together, and his reputation intact.
That lab became SF Compute, a startup that’s built what Conrad claims is the first physically-settled compute market of its kind.
SF Compute helps companies set up and maintain their GPU clusters, flagging broken hardware and cleaning up data centers dropped into, in at least one instance, a literal chicken coop. Then SF Compute manages a marketplace where businesses can secure such GPU access, and resell excess capacity when demand reaches a favorable price.
SF Compute recently signed a $245 million agreement with a leading AI lab. But the numbers involved in compute are so big, that Conrad says that’s a drop in the bucket. He almost didn’t bother to announce it.
On The Upstarts Podcast, Conrad talks about why credit, not land or power, is AI’s true bottleneck; why upfront compute costs are driving eye-popping valuations from VCs; and why data centers could become geopolitical bargaining chips.
Plus, he shares his Upstart Moment: pitching early investors on a new “power grid” for AI they just couldn’t understand.
Below, we’ll break down highlights of the episode, and Alex’s thoughts.
A MESSAGE FROM OUR SPONSOR
Move from intent to impact
Jira makes it easy to capture tasks from Slack, turn your ideas into structured plans, and hand off work to agents — in Jira or from where you work. Plus, it taps your team's context so you never start from ground zero. Run your work from your favorite AI tools with the Atlassian MCP or CLI.
The $30 billion problem
If you’ve wondered why AI labs raise billions pre-launch at unprecedented valuations – Jeff Dean’s startup Discovery Loop reportedly raising at a $50 billion valuation, Yann LeCun’s raising $1 billion, or Thinking Machines raising $2 billion last year in its early days – a big reason for it, Conrad says, is the deposit game.
“It’s not going to be, ‘click here to get a button, or click here to get a supercomputer.’ You’re going to pay a massive deposit. You have to pretty much put your entire bank balance online, and you may need to even raise more money. So this is why folks raise big, giant rounds at pretty high valuations relative to where they are.”
The cash demands, and valuations, are only going up.
“A year or two ago, it was a $100 million initial valuation. Now, it’s like a $1 billion initial valuation. It’s gonna be $5 billion, and it’s gonna be $15 billion, and it’s gonna be $30 billion. At some point, it’s gonna become unreasonable that you cannot actually do that.”
The $30 billion number stands out to me today, and maybe you, too, because Tech Twitter has recently spread the rumor that AI assistant Instinct is raising new funds at that valuation now, just a couple of months after its public launch.
Investor Sarah Guo seemed to pour some cold water on that by implying “investors are fishing.” And there are arguments to be made that big VC firms, now with tens of billions to deploy, are increasingly comfortable making bets at such skyhigh prices on the chance that the commensurate upside – a $100 billion, or $1 trillion company if it worked – can actually move the needle. But compute costs, and the crunches involved with scale, can also help explain how we get to startups like Instinct being discussed at such numbers.
Coming home to roost
When talking about data centers and massive cloud computing platforms, it’s easy to picture them in the abstract. One refresher I welcomed in the conversation with Conrad: the hardware underpinning all of this lofty AI talk is tangible, physical, and often a bit chaotic, too.
Conrad talks on the episode about partners installing the wrong GPUs into the wrong data centers, causing brownouts; turning on a new cluster will pretty much always fry some percentage of the chips, too, a “burn-in” issue that then requires hands-on work to replace.
“ We’ve seen basically everything that can go wrong. We’ve seen data centers that were literally chicken coops. We’ve gotten hit by a monsoon, at one point, and got flooded during a site visit.”
Companies looking to build their own compute capacity have to be prepared to rouse those chickens, and swap in spare GPUs when nodes die.
A wonky commodity
That’s also what makes a potential futures market around compute so complicated, Conrad argues. We’ve written about Andromeda, a company that’s creating more of a spot market around compute, and is interested in helping compute to be traded as an asset like oil or wheat.
One challenge with a compute futures market is that it’s only partially fungible, Conrad says.
”Depending on how your VM implementation works, you can have a 40% performance difference on GPU kernel launches from one operator to another. Another example is if you had the early H100s, they were shit, and if you had the later H100s, they were great, because there were problems in the initial ones and they got fixed over time. If you’re talking about just an H100 price, that’s not the same thing. It matters which one you get. It depends on which one you get.”
A challenge with a cash-settled futures market for compute, versus the physically-settled one that SF Compute wants to run with its own direct access to the hardware, is that even if you bet right on the index price, you’re getting paid in cash, not more compute. If compute is what you need, you then need to turn those dollars into it – paying the same prices.
“You have de-risked yourself on an index price that is divorced from the actual cost of compute,” Conrad says.
Speaking of de-risking: at one point I ask Conrad why, if compute is so scarce, he doesn’t simply orient his entire business around selling to the top few frontier labs. Can’t they pay top dollar?
Conrad doesn’t rule out selling to the big labs in the future, but he insists the mission is to make sure there is liquid capacity available to all. Above all, he wants the next little AI lab like the one he initially launched to be able to get going without raising a ton of money to commit.
“ Right now, we’re living in the good times. We’re living in [a time] when super high valuations are possible; markets are hot. That’s not gonna be sustainable for a long period of time. We don’t think that’s possible. When it is no longer sustainable, what that means is all of the startups that could have existed are impossible to start because you can’t go get a cluster.”
The GPU credit problem is the quiet bottleneck nobody at the top of the YC feed wants to talk about. Every AI founder chasing multi-billion-dollar revenue assumes compute is just another cost centre, and the first thing that breaks in a downturn is the balance sheet on the other side of those leases. How do you think compute suppliers actually underwrite these deals today, do they model the AI company’s own revenue like a traditional bank or is it still mostly reputation and backing