Apr 11, 2025 · 1h 12m · latent-space
SF Compute: Commoditizing Compute
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this interview, SF Compute founder Evan Conrad explores the shifting economics of GPU infrastructure, explaining how the unbundling of hardware from software, bare-metal auditing, and financial instruments like spot markets and futures contracts are turning compute into a standardized commodity.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 2% of the talking time here. How this is scored →
The hosts as informed peer 4.4 Guest teaching 5.9 Guest disagreement 1.9 The hosts pushing back 1.5
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 34:20 Dismissing decentralized crypto compute
Evan flatly rejects the entire premise of decentralized crypto compute networks, calling himself wildly skeptical due to unbeatable physical speed-of-light interconnect limits.
Hardest push from the hosts ▶ 35:54 Swyx pushes back with algorithmic architecturesSwyx refuses Evan's physical networking framing by detailing how 200-expert MoE architectures and block attention can reshape algorithms to bypass collocated hardware constraints.
Biggest teaching moment ▶ 58:50 Financial futures as the antidote to VC valuation bubblesEvan educates the hosts on how unhedged long-term compute contracts force startups to shift risk to VCs at inflated valuations, proving how financial futures stabilize tech ecosystems.
The host holds their own ▶ 19:25 Swyx cites Casado on custom ASIC scaling thresholdsSwyx demonstrates deep domain knowledge by citing Martin Casado's economic calculations showing where run costs ($50M to $5B) mathematically justify designing in-house ASICs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Structural Economics of GPUs vs. CPUs | 4 | 7 | 2 | 1 | Swyx asks Evan to analyze CoreWeave's success; Evan delivers an extensive economic masterclass distinguishing GPU elasticity from traditional CPU clouds. The hosts interject primarily to validate and summarize Evan's points. | |
| The GPU Risk Matrix and Depreciation Curves | 5 | 7 | 2 | 1 | Swyx describes Evan's 2x2 depreciation and risk matrix from an earlier talk. Evan explains how selling short-term compute at high margins fails because price-sensitive AI labs bypass software markups. | |
| Strategic Dynamics: Why Big Tech Relies on Intermediaries | 4 | 6 | 2 | 2 | Swyx challenges why NVIDIA or Microsoft wouldn't capture CoreWeave's margin directly. Evan explains channel conflict and NVIDIA's deliberate effort to prevent monopsony power among hyperscalers. | |
| The Failure of Bundling Software with Hardware | 6 | 6 | 3 | 2 | Evan warns that bundling proprietary software on top of raw hardware leads to bankruptcy, while Swyx brings in Martin Casado's framework regarding extreme compute scale justifying in-house ASICs. | |
| The Genesis of SF Compute: From Survival to Marketplace | 3 | 7 | 1 | 0 | Alessio asks how SF Compute came to be; Evan gives an honest monologue recounting their near-bankruptcy subleasing compute month-to-month before realizing the need for a liquid spot market. | |
| Maximizing Cluster Utilization and Contract Flexibility | 4 | 6 | 1 | 1 | Alessio queries marketplace cluster utilization rates. Evan explains dynamic pricing dropping to clear idle capacity and offering risk-free cancellation mechanisms for suppliers. | |
| Navigating the GPU Glut and the Test-Time Inference Boom | 5 | 6 | 2 | 1 | Swyx brings up the H100 glut, noting Evan was quoted in their newsletter. Evan clarifies that supply chain logistics, not demand contraction, caused the temporary glut, and test-time compute will spike demand again. | |
| Evaluating Decentralized and Crypto Compute Networks | 6 | 5 | 3 | 4 | Alessio inquires about crypto/decentralized compute, which Evan dismisses due to physical speed-of-light constraints. Swyx pushes back, citing fine-grained mixture-of-experts (MoE) architectures and block attention designed to handle latency. | |
| Empowering Non-Traditional Buyers and VC Compute Arbitrage | 5 | 6 | 2 | 3 | Evan describes how SF Compute serves academic grants and startups, then explains how VC compute clusters (like Andromeda) represent credit risk arbitrage. Swyx challenges his claim about where value concentrates. | |
| Mastering SF Compute Pricing Mechanics and Spot Strategies | 5 | 7 | 1 | 1 | Alessio questions why 1-week reservations are more expensive than 1-day or 1-month terms. Evan teaches the mechanics of hourly spot reservations and expiring compute floors, which Swyx equates to Spotinst primitives. | |
| Financialization and Futures Contracts in Compute Markets | 6 | 6 | 3 | 2 | Swyx draws on his derivatives trading background to frame SF Compute as a forward exchange. Evan immediately clarifies that they operate as a spot index market to enable future cash-settled hedging rather than existing derivatives. | |
| Hardware Verification: Cluster Auditing and Bare-Metal Tooling | 4 | 7 | 1 | 1 | Swyx asks about hardware auditing procedures. Evan explains their Linpack stress-testing, active/passive telemetry, remote BMC cluster management, and proprietary UEFI shims. | |
| De-Risking the AI Ecosystem to Deflate Venture Bubbles | 4 | 7 | 3 | 2 | Evan delivers a compelling thesis on how lack of compute futures forces startups to pass risk to VCs, blowing up valuation bubbles. He asserts that financialization is the mechanism to de-risk and calm AI markets. | |
| The Anti-Hype Brand and San Francisco Optimism | 3 | 3 | 1 | 1 | Alessio and Swyx ask about SF Compute's pastoral, anti-hype visual branding and minimal single-page website. Evan explains that lowering expectations upfront creates delightful upside while celebrating San Francisco. | |
| Evan Conrad's Founder Journey and the Intricacies of Email AI | 4 | 5 | 2 | 2 | Swyx asks why building email AI products seems intractable based on Evan's previous startup. Evan clarifies that his pivot was driven by founder burnout rather than product impossibility, acknowledging Intercom's customer service fit. | |
| Hiring Systems and FinTech Engineers at SF Compute | 3 | 4 | 1 | 0 | Evan outlines hiring needs for low-level Linux systems engineers in Rust and fintech ledger engineers. Swyx asks about TigerBeetle accounting databases as Evan wraps up. |
Statements from this episode (25)
Conrad: Incremental GPUs always drive model performance and revenue, unlike CPUs
“Gusto isn't going to make like, you know, five percent more money. They're going to make zero, like literally zero money from every incremental GPU or CPU after a certain point. This is not the case for anyone who is training models. And it's not the case for …”
Conrad: Software margins on GPU clusters drive customers to build in-house
“So if you have a 10% margin increase because you have great software on your billion dollars, the customers are that price sensitive. They will immediately switch off if they can, because why wouldn't you? You would just take that hundred million dollars, you'…”
Conrad: CoreWeave's debt-financed long-term contract model is optimal for GPUs
“So that means that the best way to make money in GPUs was to do basically exactly what CoreWeave did which is go out and sign only long-term contracts, pretty much ignore the bottom end of the market completely, and then maximize your long-term contracts with …”
Conrad: Hyperscalers will probably lose significant money reselling Nvidia GPUs
“My intuition is that the hyperscalers are probably going to lose a lot of money, and they know they're going to lose a lot of money on reselling NVIDIA GPUs at least.”
Swyx: Microsoft and OpenAI account for 77% of CoreWeave revenue
“Which are together, 77% of the revenue of CoreWeave.”
Conrad: Selling short-term GPU contracts is the worst financial model
“Selling short-term contracts for low prices paid over time which is the worst place to be in the worst financial place to be in because it has the highest interest rate which means that your costs go up at the same time, your incoming cash goes down and squeez…”
Conrad: NVIDIA launching a cloud would impair hyperscaler sales
“If they launched their own core weave, then it would make it much harder for them to sell to the hyperscalers.”
Conrad: NVIDIA avoids customer concentration to prevent hyperscaler price setting
“It's really bad for NVIDIA if you have customer concentration, and Microsoft and Google and Amazon, like, Oracle to, like, buy up your entire supply and then you have four or five customers or so who pretty much get to set prices.”
Conrad: GPU businesses succeed as pure real estate or pure software, not both
“The GPU clouds are fantastic real estate businesses. If you treat them like real estate businesses, you will make a lot of money. The, Cloud services you can make on that, all the software you want to make on that, you can do that fantastically. If you don't o…”
Swyx: At $5B+ training runs, designing custom chips makes economic sense
“When you get the five billion dollar runs, when you get the fifty billion dollar runs it is actually makes sense to build your own chips, like to, for OpenAI to get into chip design”
Conrad: Custom chips make sense for inference, not training experimentation
“It only works if you really know which chip you're going to do. If you don't, then it's a little harder. So it makes, in my head, it makes more sense for inference where you've already established it, but for training there's so much, like, experimentation.”
Conrad: SF Compute is the only true liquid bid-ask GPU market
“Turned into what is today SF compute, which is a compute market, which we think we are the functionally the most liquid GPU market of any capacity. Honestly, I think we're the only thing that actually is like a real market that there's like bids and asks and t…”
Conrad: Spot GPU cluster utilization nears 100% through dynamic price clearing
“Assuming there are not, like, hardware problems or software problems, the utilization rate is, like, near a hundred percent, because the price dips until the utilization is a hundred percent.”
Conrad: GPU clouds cannot offer contract cancellation without risking capital costs
“If you're just the GPU cloud, you can never cancel your contract because that introduces so much risk that you would otherwise like not get your cheap cost capital or whatever.”
Conrad: GPU market will likely return to a shortage by winter
“My general prediction is that like by the winter we will be back towards shortage, but then also this very much depends on
The rollout of future chips.”
Conrad: Test-time inference will significantly expand inference compute demand
“The thing I do feel reasonably confident about saying is that the test time inference is probably going to quite significantly expand the amount of compute that was used for inference.”
Conrad: OpenRouter open-source traffic required only around 10 H100 nodes
“The entirety of Open Router that was not Anthropic or Google like, or Gemini or OpenAI or something. It was like, 10 H 100 nodes or something like that. It's just, like, not that much. It's like, not that many GPUs, actually, to service that entire demand.”
Conrad: Decentralized compute networks will never beat co-located InfiniBand clusters
“I just don't really think this is gonna ever be more efficient than a fully interconnected cluster with Infiniband, or, you know, whatever sort of next spec might be. Like, I could be completely wrong, but Speedolite is really hard to beat. And regardless of w…”
Conrad: Grad students are the worst customers for traditional GPU clouds
“And the grad students are like the worst possible customer for the traditional GPU clouds, because they will immediately turn if you sell them a thing, because they're going to graduate and then like, I'm going to go anywhere or they're not going to like, that…”
Conrad: VCs trading compute for equity is credit risk arbitrage
“And so the hack of a VC or some capital partner offering equity for compute is always some arbitrage on the credit risk.”
Conrad: SF Compute plans cash-settled futures to de-risk data centers
“What we're trying to do is create an underlying spot market that gives you an index price that you can use. And then with that index price, you can create a cash settled future. And with a cash settled future, you can go back to the data centers and you can sa…”
Conrad: Compute Spot Markets Fail Without Direct Control and Cluster Auditing
“You really cannot make a smart market work if you don't run the clusters, if you don't have control over them, if you don't know how to audit them, because these are super computers, not soybeans.”
Conrad: HPC Hardware Reliability Issues Are Unlikely to Disappear
“Like the hardware problems aren't going away until the underlying vendors fix things. But honestly, I don't think that's likely because you're always pushing the limits at HPC. This is the case of trying to build a supercomputer”
Conrad: The AI VC bubble will pop and fail to return capital
“So what you've done by not having a future is you've inflated the venture capital market. And that is a bubble that's totally going to pop at some point. Like a lot of the companies are not going to work. And the valuations are not going to work. And what's go…”