Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / Why Cloud Contracts Now Include GPU Decommissioning Fees
Transcript
- Lucas: There's a clause showing up in more hyperscaler contracts that I think most infrastructure buyers haven't fully priced in yet: GPU decommissioning fees. Luna: Wait — you mean a fee for returning hardware you reserved? Like, a penalty for leaving a GPU commitment early? Lucas: Exactly. And it's not just an early termination penalty in the traditional sense. It's a specific charge tied to the physical and logistical cost of taking a GPU out of your dedicated environment, wiping it, testing it, and reallocating it to another customer. Think of it like a hotel charging a deep-cleaning fee after a long-term stay, except the bill can run into the millions. Luna: How often are we seeing this? Is it standard in AWS, Azure, GCP now? Lucas: Not universal yet, but it's become common in reserved-instance contracts for NVIDIA H100s and the upcoming B200 clusters — especially for commitments of twelve months or longer. I've talked to three cloud procurement leads in the past month who all had decommissioning language added in their latest renewals. The fee typically ranges from 10 to 25 percent of the remaining contract value. Luna: Ouch. So if you have a $5 million GPU reservation with a year left, the hyperscaler could charge you over a million just to walk away. Lucas: Right. And the real-world impact hit me when I looked at a case from late last year. A generative ai startup called VoxelGen — about 80 employees, raised a Series B in early 2025 — reserved 512 H100s on a twelve-month commitment through Azure. Nine months in, they got acquired by a larger enterprise platform. The acquirer had its own cloud strategy and didn't want to keep the GPU cluster. The decommissioning fee came out to roughly $1.2 million. Luna: That's a big number for a mid-stage startup. Did it become a deal-breaker? Lucas: It became a major negotiating point in the acquisition terms. The acquirer argued the fee was an operational liability that should reduce the purchase price. VoxelGen's team pushed back, saying the reservation was signed before acquisition talks even started. Ultimately, the hyperscaler agreed to reduce the fee to 15 percent — about $720,000 — in exchange for the acquirer signing a new reserved-instance deal for a different region. So it got resolved, but it added weeks of legal back and forth. Luna: That's the kind of hidden cost that could mess up an exit timeline. Why are hyperscalers adding these fees now? Feels like a shift from a few years ago. Lucas: It really is a shift. Two main drivers. First, GPU scarcity. For the last two years, demand for H100s has far outstripped supply. Hyperscalers have been allocating capacity months in advance. When a customer returns a GPU early, the cloud provider can't immediately resell that exact unit — it needs to be decommissioned, wiped, tested, and sometimes physically moved between data centers. During that downtime, they're losing revenue on a scarce asset. The fee covers that gap. Luna: And the second driver? Lucas: Oversubscription risk. Cloud providers overcommit their GPU fleets by a certain margin, banking on some customers not using their full reservation. But if a large customer suddenly drops out, the oversubscription math can break. Decommissioning fees act as a deterrent against strategic gaming — like reserving a ton of capacity just to block competitors, then walking away. Luna: So it's partly about protecting their own utilization rates. Makes sense. But for a legitimate startup that just gets acquired or pivots, it feels punitive. Lucas: That's the tension. And I think negotiators need to be more proactive about this. There are at least three tactics I've seen work. Luna: Let's hear them. Lucas: First, tie the decommissioning fee to a force majeure clause or a change of control provision. If the company is acquired, or if there's a material shift in business direction — like a pivot away from AI — the fee gets waived or capped at, say, 5 percent. Second, negotiate a reallocation credit. If the hyperscaler can place your GPUs with another customer within 30 days, the fee should be reduced by the revenue they earn from that new customer. Luna: That seems fair — they shouldn't double-dip. And the third tactic? Lucas: Structure the commitment as a rolling month to month after a minimum term. For example, commit to six months upfront, then automatically renew monthly with a 30-day cancellation notice. The per-unit price might be slightly higher — maybe 3 to 5 percent — but you avoid the lump-sum decommissioning fee entirely. For startups with uncertain timelines, that premium is often worth it. Luna: That's a good trade-off. I'm curious — do hyperscalers publish these fees openly, or is it buried in contract appendices? Lucas: Almost always in appendices or custom pricing addendums. You won't see it in the standard service terms. It's negotiated. I've looked at Azure's enterprise agreement templates and AWS's reserved-instance terms — neither mention decommissioning in their public docs. It surfaces when you start talking about multi-year GPU reservations, especially for large clusters. Luna: So the onus is on the buyer to ask. And many probably don't until it's too late. Lucas: Exactly. And that's why we're talking about it. The GPU market is only getting more capital-intensive. As the next generation of chips comes online — Blackwell Ultra, AMD's MI400 — the reservation values will climb. A decommissioning fee on a $10 million commitment is a real risk. Luna: You know, this is exactly the kind of practical insight that a lot of our listeners tell me they don't get from the vendor webinars. It's the stuff that actually matters when you're signing paper. Lucas: And that's why we keep doing this show ad-free — because these conversations are useful for people building and running infrastructure. Quick honest thing: a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that literally funds making this many episodes. No ads, no sponsors, just listener support. If you find value in episodes like this, that's the whole engine. Luna: Yeah, it's a small group that keeps it going. And it means we can talk about stuff like decommissioning fees without anyone telling us to soften the critique. Lucas: Totally. So back to the tactics — I want to add one more angle I've seen work. Some buyers are now including a 'GPU buyback' clause, where the hyperscaler agrees to purchase the remaining commitment at a discount — say 10 percent below market — if the customer needs to exit. That turns a fee into a sale. Luna: Interesting. So instead of paying to leave, you get some money back. Does that actually happen in practice? Lucas: I've seen it once, with a large financial services firm that reserved a cluster for a risk-modeling project that got deprioritized. The hyperscaler, in that case, was GCP, and they agreed to a buyback because they had a waiting list for GPUs in the same region. It's not common, but it shows that everything is negotiable if you understand the provider's incentives. Luna: That's the key — understanding their incentives. If they're oversubscribed, they might be more flexible because they can resell quickly. If they're not, you're stuck. Lucas: Right. And the other thing to watch is the definition of 'decommissioning.' Some contracts include data wiping as a separate line item. If you're dealing with sensitive data — healthcare, finance — that could add another 2 to 3 percent. Always ask for a breakdown. Luna: So the takeaway: read the appendices, negotiate the exit terms upfront, and consider rolling commitments if your timeline is fuzzy. Anything else? Lucas: One more thing: if you're an enterprise with multiple business units, try to centralize GPU reservations under a single procurement team. Decentralized teams often sign separate deals with different decommissioning terms. Consolidating gives you leverage to negotiate a master decommissioning cap for the whole company. Luna: That's a smart structural move. I wonder how many companies actually do that versus just letting each team run its own cloud spend. Lucas: From what I've seen, very few. But the ones that do save a lot of money — and headache. Luna: Great episode. Thanks, Lucas. Lucas: Yeah, thanks Luna. And thanks to everyone who makes this possible.