Seth DeGroot

The most important party in your GPU contract is one you'll never meet. The provider's lender. Buyers think they're negotiating with a neocloud or a marketplace. They're really negotiating with whoever financed the cluster. That disconnect is why basic terms keep getting harder to win. Start with how capacity gets built. Most providers don't speculatively self-finance clusters. They line up colo space, lean on an OEM relationship (sometimes adding units to an order already in flight), and take a deposit from the customer. The rest comes from an asset-backed lender. Those lenders are essential. They're also pricing risk. Money runs from roughly 6% for a deal backed by an investment-grade customer to the high teens for a smaller operator. Here's the part most buyers miss. GPUs depreciate fast, so lenders discount them heavily as collateral. What they're really underwriting is your contract. Your commitment to pay is the asset. Now look at what customers commonly ask for: A walkaway right if the POC (proof of concept) underperforms A fast refund if delivery slips An exit for chronic failures Every one of those is reasonable from the buyer's seat. From the lender's seat, every one turns contracted revenue into optional revenue. Lenders typically size loans so contracted cash flow covers debt payments with a cushion, often around 1.3x. Optional revenue doesn't count toward the cushion. So each exit right has to be approved by the lender, or the provider absorbs it. And most providers can't absorb much. They aren't sitting on piles of cash. They put what they have into more capacity, and your deposit is already committed to hardware worth tens or hundreds of millions. So the "no" you got on a walkaway clause probably didn't come from the provider. It came from a credit committee. As the market has sped up, this has gotten worse. Less time per deal, less room to redline, more capacity lost while lawyers trade drafts. The obvious pushbacks: "I'll find a provider who gives me the terms." Some will. Ask who's carrying that risk, and whether they can pay out if it actually triggers. "Walkaway rights are standard in enterprise software." Enterprise software doesn't have a hardware loan sitting behind every seat. "So buyers should just accept whatever's offered?" No. Good terms are still out there. The skill is knowing which ones a lender will bend on and which ones are dead on arrival. That's how we run it at Shadeform. When new capacity comes to us, our first question is: send the MSA, the SLA, and every rider. We start that negotiation before a customer needs the capacity, so a legal gap doesn't cost them a great cluster. We know the game, and we know where lenders move and where they don't. The lenders aren't the villain. Neither are the providers. They're the reason the capacity exists. In a fast market, the GPUs aren't the bottleneck. The paper is.

If nodes are <5 weeks from RFS and priced reasonably, you have 24–72 hours to buy them. Max. That’s true even for 16–128 node deals worth tens or hundreds of millions of dollars. If your process requires exhaustive DD, legal review, committees, and multiple layers of approval before you can commit, you’re simply not going to win near-term capacity. Sub-4-week RFS supply is in absurd demand. The companies actually getting it have adapted their buying process to the market. Everyone else either pays 60% more for the same hardware so they can run their normal process, or gets nothing. It really is that crazy out there.