Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / Why Enterprise Cloud Contracts Now Have GPU Resale Clauses
Transcript
- Lucas: If your enterprise is training large language models or running heavy inference workloads, there is a new clause appearing in cloud contracts that you need to know about. It's the GPU resale right. Luna: Resale right — meaning you can sell unused GPU compute back to someone else? Lucas: Exactly. Cloud providers are now offering customers the ability to resell unused GPU capacity on secondary markets. It's a response to the fact that GPU supply is still constrained and spot prices are volatile. For buyers, it's a way to offset some of the cost of committing to large reservations. Luna: I've heard about this in the context of AWS's capacity auctions, but it sounds like this is more formalized now. Lucas: It is. Let me give you a concrete example. A multinational pharmaceutical company — let's call them PharmaCo — signed a three-year deal with a major cloud provider in early 2025. They committed to a certain number of NVIDIA H100 instances for drug discovery workloads. But their actual usage fluctuated. Sometimes they needed all of them, sometimes only sixty percent. Luna: So they were paying for idle capacity. Lucas: Right. Under the old model, that idle time was pure waste. But their contract included a GPU resale clause. They were allowed to resell unused capacity on a secondary market that the provider operates. In Q1 2026, PharmaCo resold about fifteen percent of their reserved GPU hours and reduced their net compute cost by eighteen percent. Luna: Eighteen percent is significant. How does the pricing work on that secondary market? Lucas: Good question. The resale price isn't fixed — it's determined by a bidding mechanism similar to spot instances. But the key is that the enterprise sets a floor price. If the market clears above that floor, the GPU hours get sold and the enterprise gets a credit on their bill. If it doesn't clear, they keep the capacity. Luna: So it's like a put option for compute. Lucas: Exactly. It's a financial instrument more than a technical one. And it changes the negotiation dynamics. Enterprises that can forecast their GPU demand reasonably well — but not perfectly — can now afford to commit to larger reservations because they have an exit valve. Luna: Which cloud providers are offering this today? Is it all of them? Lucas: As of June 2026, the three major providers — AWS, Azure, and GCP — all have some form of GPU resale program. But the terms vary significantly. AWS has the most mature market, called Compute Resale. Azure has a similar program tied to their reservation model. GCP is still in pilot but has the most flexible terms — they allow resale across different instance families, not just the same GPU type. Luna: That flexibility could be a big differentiator. If you've got H100s but you need L40s for inference, being able to swap would be powerful. Lucas: Exactly. And that leads to a broader point: these clauses are forcing cloud providers to compete on contract terms, not just on raw pricing or performance. The resale right becomes a negotiation lever. An enterprise can say, 'We'll commit to $50 million in GPU reservations if you include a resale clause with no penalty and a floor price that tracks market benchmarks.' Luna: And the provider gets the commitment they need to plan their own capacity investments. It's a hedge for both sides. Lucas: Precisely. But there are risks. The biggest legal risk is that the resale market isn't deeply liquid. If too many enterprises try to sell at the same time — say, because demand for AI training drops — prices could collapse and the floor might not protect you. Luna: So you could still end up paying for capacity you can't sell. Lucas: Yes. And there's also the complexity of managing the resale process. If you're a large enterprise with hundreds of GPU instances across multiple regions, you need automated tools to decide when to sell and at what price. Some providers offer APIs for this, but it's not turnkey yet. Luna: This is where the cloud broker role that we talked about in Episode 26 comes back in. A good broker can help enterprises optimize these decisions. Lucas: That's a really good connection. The resale clause basically creates a secondary market that needs active management. For enterprises that don't have the in-house expertise, third-party cloud brokers are stepping in to handle the bidding and allocation. Luna: So what should our listeners do if they're negotiating a cloud contract right now? Lucas: First, ask for the resale clause explicitly. It's not automatically included in standard contracts. Second, negotiate the floor price. You want it tied to a transparent benchmark, like the average spot price over the trailing thirty days. Third, understand the minimum resale period — some providers require you to list capacity for at least an hour, others allow shorter windows. Luna: And what about the legal terms on liability? If you sell capacity and the buyer's workload causes a security issue, who's responsible? Lucas: That's a great point. In most contracts, the provider indemnifies the seller against actions by the buyer. But you need to check. Also, some contracts include a non-compete: you can't sell to a direct competitor of the provider. That's rare but worth watching for. Luna: This is really shifting the enterprise cloud procurement landscape. It used to be about locking in discounts. Now it's about flexibility and financial engineering. Lucas: Exactly. And it's a sign that the GPU market is maturing. We're seeing the same evolution that happened with general-purpose compute: reserved instances, spot instances, and now resale markets. But it's happening faster because the stakes are higher. A single large GPU cluster can cost tens of millions per year. Luna: Speaking of the economics of this show — we're able to dig into these contract details because of a small group of listeners who chip in monthly through buy me a coffee dot com slash fexingo. It keeps the podcast ad-free and lets us spend time reading actual contract clauses so you don't have to. Lucas: Yeah, and it's not a big ask — it's just that the listener-supported model means we don't have to chase advertisers or dumb down the content. If these deep dives are useful for what you're building or running, that's the whole point. Luna: Alright, back to the GPU resale clause. One more thing I want to ask: how do the cloud providers themselves benefit from this? It seems like they're giving up potential revenue. Lucas: They benefit in two ways. First, they get the commitment. A multi-year reservation locks in revenue that helps them order chips from NVIDIA or AMD. Second, they capture a spread on the resale — they charge a small transaction fee, usually around five percent of the resale price. So they make money on the volume. Luna: And they also avoid the reputational damage of having customers wastefully idle capacity. It's a better story for sustainability reporting too. Lucas: Good point. Some providers are marketing these programs as 'sustainability initiatives' because they reduce the number of GPUs that need to be manufactured. Whether that's greenwashing or genuine depends on the accounting, but it's part of the narrative. Luna: So what's the future of this? Do you see the secondary market becoming independent of the cloud providers? Lucas: That's the big question. There are startups trying to create neutral marketplaces for GPU capacity, but they face a chicken and egg problem: without liquidity, enterprises won't list; without listings, buyers won't come. The cloud providers have the advantage of existing customer relationships and billing infrastructure. Luna: So for now, it's a provider-controlled market. But if the terms are favorable, enterprises can still benefit. Lucas: Exactly. And the key takeaway for today is: if you're signing a cloud contract for GPU capacity in the second half of 2026, make sure the resale clause is on the table. It's not just a nice to have — it could be worth millions in cost savings. Luna: And if you need help navigating those terms, there are consultants who focus specifically on cloud contract negotiation. Might be worth the investment. Lucas: Absolutely. Alright, that's it for today. Next episode, we'll look at how cloud regions are starting to offer dedicated fiber connections for GPU clusters — and why that changes the latency math for distributed training. Luna: Looking forward to it.