Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / How Cloud Contracts Now Include GPU Inventory Reservations
Transcript
- Luna: So Lucas, I was looking at a new AWS enterprise agreement last week, and there was a clause I hadn't seen before — something about 'GPU inventory reservation.' It basically let the customer lock in a specific number of P5 and P4d instances for a six-month window, with a penalty if AWS couldn't deliver. Lucas: Yeah, that's one of the newer additions to the hyperscaler playbook. It's not just AWS — Azure has something similar with their nd series reservation guarantees, and Google Cloud started offering what they call 'accelerator capacity commitments' late last year. The basic idea is the same: you pay a premium — typically 10 to 15 percent above the on-demand rate — and in return, the provider guarantees that those GPU instances will be available when you need them, or they pay a service credit. Luna: Right. And the premium is interesting because traditionally, reserved instances were cheaper than on-demand. This is the opposite. You're paying more to get a guarantee of supply. Lucas: Exactly. And that inversion tells you everything about the current market. We're in mid-2026, and demand for training and inference GPUs — especially NVIDIA's H100 and B200 — is still outstripping supply. The hyperscalers have been building out capacity as fast as they can, but they also have to manage their own internal AI projects and a growing base of external customers. So they've started offering these reservation clauses as a way to prioritize allocation. Luna: And if these conversations are useful for what you're building or running — whether you're negotiating a cloud contract or just trying to understand where the industry is headed — the way we keep this show ad-free and independent is listener support. If today's episode adds something to your thinking, you can buy me a coffee dot com slash fexingo. It's a small way to say this kind of deep-dive is worth keeping around. Lucas: Yeah, that's right. And it really does make a difference. So back to the GPU reservations — I think the key question for procurement teams is whether the premium is worth it. For a startup that needs to train a model on a tight timeline, the cost of delay could be massive. Paying 15 percent more to guarantee access might be a no-brainer. But for a company with more flexible workloads, it might make sense to stick with on-demand and accept the risk of spot-instance interruptions. Luna: Let's talk about the actual contract language. What does a typical GPU inventory reservation clause look like? Lucas: So most of them have four key components. First, the customer specifies a minimum number of GPU instances — usually in increments of eight or sixteen — and a target region. Second, the reservation period is typically three to twelve months. Third, the provider commits to a specific availability SLA, often 99.9 percent uptime for those reserved instances. And fourth, if the provider fails to deliver the instances within a certain window — say, two weeks from the requested start date — the customer gets a service credit, usually 10 to 20 percent of the reservation fee. Luna: And there's usually a catch, right? The customer has to actually use the instances or still pay for them. Lucas: Right. You're committing to a minimum spend. If you reserve 100 H100s for six months, you're on the hook for that capacity whether you use it or not. Some providers allow a one-time reduction of up to 20 percent during the term, but it's not a flexible arrangement. It's designed for predictable, heavy usage. Luna: So who's actually using these clauses? Is it the big AI labs, or are we seeing enterprise adoption too? Lucas: Initially, it was the usual suspects — OpenAI, Anthropic, the large model builders. But over the past six months, I've seen more large financial services firms and pharmaceutical companies signing these agreements. A bank might need a cluster of GPUs for fraud detection model retraining on a quarterly cycle, and they can't afford to have that delayed because of capacity constraints. So they lock in the reservation. Luna: And from the cloud provider's perspective, this is basically a way to smooth demand. They can see exactly how much GPU capacity is committed months in advance, which helps them plan data center builds and negotiate with NVIDIA for chip allocations. Lucas: Exactly. It's a capacity planning tool for both sides. But it also creates a two-tier market: customers with reservation clauses get priority access, while everyone else competes for whatever leftover capacity there is, often at higher spot prices or with longer wait times. That's a significant shift from the early days of cloud, where on-demand was the default and everyone had roughly equal access. Luna: I wonder if this will eventually lead to a secondary market — companies that reserved capacity but don't need it all selling their allocation to others. Lucas: It's already happening informally. I've heard of startups with reservation contracts subleasing GPU time to other companies through brokers. The hyperscalers' terms of service usually prohibit reselling, but enforcement is spotty. I wouldn't be surprised if we see official marketplaces emerge in the next year or two, similar to how AWS has a reserved instance marketplace for compute. Luna: Let's talk about pricing. You mentioned 10 to 15 percent above on-demand. Is that consistent across providers? Lucas: It varies. AWS's premium has been around 12 percent for the P5 instances, but for the newer H100-based instances, it's closer to 18 percent because demand is higher. Azure has a tiered system where the premium decreases the longer your commitment — 15 percent for three months, 10 percent for six, 8 percent for twelve. Google Cloud is the most aggressive; they've been offering a 5 percent premium for six-month commitments, likely to grab market share. Luna: And what about the service credit structure? Is it enough to actually compensate for a delay? Lucas: It's a fraction of the revenue the provider would lose. Typically, the credit is 10 percent of the monthly reservation fee per month of delay. So if you're paying $100,000 a month for reserved GPUs, and AWS is two weeks late delivering, you get a $10,000 credit. But if your business depends on that capacity to launch a product, $10,000 might not cover the lost revenue. It's more of a signal that the provider takes the commitment seriously. Luna: Right. And the credit often doesn't apply to consequential damages — it's capped at the reservation fee or a multiple thereof. Lucas: Exactly. The clause is really about priority placement in the provisioning queue, not a financial guarantee. That said, I've seen some large customers negotiate additional terms, like a requirement that the provider maintain a buffer of at least 10 percent reserved capacity in the same region to cover potential failures. That's a tougher negotiation. Luna: So what's the advice for a company that's considering signing one of these? What should they watch out for? Lucas: First, make sure the reservation period aligns with your actual usage cycle. If you only need the GPUs for two months during a training run, don't lock in for six. Some providers allow a shorter minimum, but you'll pay a higher premium. Second, negotiate the definition of 'availability.' Some contracts count instances as 'available' even if they're being used for maintenance or patching, as long as the total uptime meets the SLA. You want language that specifically covers your ability to launch and run jobs. Luna: And third? Because I know there's always a third. Lucas: Third, check the geographic scope. If your reservation is in us-east-1 and there's a power outage there, you might be out of luck. Some contracts allow you to spill over to another region, but that's not automatic. You may need to negotiate a cross-region fallback clause. And fourth, watch the renewal terms. Some providers automatically renew your reservation at the same premium unless you cancel with 30 days' notice. That can catch you off guard if your demand drops. Luna: Alright, I want to zoom out a bit. How does this change the overall relationship between cloud providers and large customers? Does it give the hyperscalers more leverage, or does it actually help customers by making capacity more predictable? Lucas: I think it's a double-edged sword. On one hand, customers who can commit get a guaranteed supply, which is valuable in a constrained market. On the other hand, it deepens the lock-in. Once you've committed to a specific provider's GPU reservation for six months, you can't easily switch to another cloud for that workload. You're tied to that provider's ecosystem, pricing, and availability. It's a sticky contract. Luna: So it's a trade-off: supply predictability for flexibility. Lucas: Exactly. And I think we'll see more customers adopt a multi-cloud GPU strategy specifically to avoid that lock-in. They might reserve H100s on AWS for one training job, but keep a separate reservation on Azure for another, and use Google Cloud for inference. That way, if one provider has an outage or a capacity crunch, they can shift workloads. But managing multiple reservation contracts is complex and expensive. Luna: I've also heard about some customers using spot instances for a portion of their GPU needs and reserving only the base load. Is that common? Lucas: Yes, that's a smart hybrid approach. A company might reserve 60 percent of their peak GPU demand to guarantee a baseline, then rely on spot instances for the remaining 40 percent during training runs. Spot prices for GPUs can be 60 to 70 percent below on-demand, but they can be interrupted with two minutes' notice. So you need a workload that can handle preemption — like distributed training with checkpointing. Luna: And that's where tools like Kubernetes with node auto-scaling and checkpointing frameworks become crucial. Lucas: Exactly. The technology stack has to support the reservation strategy, not the other way around. I think the companies that are most successful with these GPU inventory clauses are the ones that have already invested in workload portability and resilience. They're not just signing a contract — they're architecting their entire AI infrastructure around the reality of a constrained supply chain. Luna: So where do you see this going in the next year or two? Will GPU inventory reservations become as standard as compute reserved instances? Lucas: I think yes, especially as more enterprises deploy AI in production. The GPU is becoming a strategic resource, just like data center space or fiber connectivity. Companies that need predictable AI performance will demand these clauses. And the hyperscalers will likely expand them to cover not just GPUs but also custom AI accelerators like Google's TPUs and AWS's Trainium. We might even see reservation clauses for networking bandwidth between GPU clusters, because that's becoming a bottleneck too. Luna: So the message for procurement teams: start thinking about GPU reservations now, even if you don't think you need them yet. Lucas: Exactly. The contracts are evolving fast, and the early movers are getting better terms. If you wait until you're desperate for capacity, you'll be paying a higher premium and accepting weaker guarantees. That's the practical takeaway: treat GPU availability as a critical part of your cloud strategy, not an afterthought. Luna: Good advice. And on that note, I think we've covered the landscape pretty well. Lucas, thanks for the deep dive. Lucas: You bet. Thanks, Luna. And to our listeners, if this episode made you think differently about your cloud contracts, we'd love to hear about it. Until next time.