Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / Why Cloud Contracts Now Include Reserved Compute Swaps
Transcript
- Lucas: So you sign a three-year commitment for a specific compute instance family — say, AWS's M5 or Azure's D-series — and six months in, your team realizes they actually need GPU instances for inference workloads. Historically, you were stuck. Luna: Stuck paying for something you don't use, while buying full-price compute for what you actually need. Lucas: Right. But a growing number of enterprise cloud contracts now include what's called a reserved-compute swap right. It lets you convert your existing reserved instances to a different instance family — sometimes even across regions — without penalty. Luna: Is this standard now, or is it still something you have to negotiate? Lucas: Definitely still negotiated. The hyperscalers — AWS, Azure, GCP — all have some flavor of instance flexibility baked into their standard reserved-instance products. AWS, for example, has had 'Convertible Reserved Instances' for years. But the new thing is that enterprises are pushing for broader swap rights that cover GPU instances, high-memory instances, even bare metal. Luna: And the hyperscalers are actually agreeing to that? Seems like it could hurt their utilization planning. Lucas: It does hurt their planning, but they're doing it because the alternative is losing the deal. I spoke with a procurement lead at a large European retailer — they signed a three-year deal with AWS last year worth about forty million dollars. They insisted on a clause that let them swap up to forty percent of their reserved compute between instance families once per quarter. Luna: Forty percent is huge. What triggered the swap? Lucas: They shifted from general-purpose compute to memory-optimized instances for their SAP HANA workloads. Without that clause, they'd have been paying for M5 instances they didn't need while buying R5 instances at on-demand rates. Luna: So the swap right directly saved them money, but also let them adapt to a changing workload mix. Lucas: Exactly. And that's the crux of it. Cloud demand is shifting faster than ever — AI inference, real-time analytics, edge computing — and enterprises don't want to bet on a single instance family for three years. Luna: I've seen a similar trend on the Azure side. A financial services client of mine negotiated a swap right that covers not just instance families but also regions. They wanted the ability to move workloads from US East to US West if data-residency rules changed. Lucas: That's a smart addition. The geographic swap is still rare — most contracts limit it to within the same region. But as data sovereignty becomes more complex, I expect that to become a standard ask. Luna: What about GPUs? Those are the hardest to get committed capacity for right now. Lucas: GPU swap rights are the frontier. Right now, if you reserve an A100 or H100 instance, you're typically locked into that specific GPU type. But some large AI labs are negotiating clauses that let them swap between, say, NVIDIA and AMD GPU instances if one vendor's supply dries up. Luna: So it's not just flexibility on compute — it's hedging against supply-chain risk. Lucas: Precisely. And the hyperscalers are willing to entertain it because they want to lock in the commitment. A three-year reservation gives them predictable revenue, which helps them plan data-center builds. They'll give up some flexibility on the back end to get that front-end commitment. Luna: But isn't there a cost to that flexibility? I assume you pay a premium for the swap right. Lucas: You do. Typically, the discount on a convertible reserved instance is lower — maybe 30 percent off on-demand versus 40 percent for a standard reserved instance. So you're trading some savings for optionality. Luna: So the key is to model whether the flexibility is worth the lower discount. For the retailer, it clearly was. Lucas: Right. And that modeling is exactly what FinOps teams should be doing now. The old approach was 'buy reservations for what you know you'll use.' The new approach is 'buy reservations with swap rights for what you might need.' Luna: Let me ask you this — are there any gotchas in the fine print? I've heard some contracts have a 'swap fee' or limit the number of swaps per year. Lucas: Yes, both. AWS's Convertible RIs, for example, allow unlimited swaps, but only to instances of equal or greater value. If you swap to a cheaper instance, you lose the remaining value. And some providers charge a nominal fee — like one percent of the reservation value — per swap. Luna: That one percent can add up if you're swapping tens of millions in capacity. Lucas: Exactly. So the smart move is to cap that fee or negotiate it away. I know of a gaming company that got the swap fee waived by agreeing to a five-year term instead of three. Luna: So leverage is key. And right now, with GPU demand so high, enterprises have more leverage than they think. Lucas: They really do. Hyperscalers are desperate for long-term commitments to justify their capacity expansions. If you're signing a seven-figure deal, you should absolutely push for swap rights across instance families, across regions, and ideally across GPU vendors. Luna: And that's the takeaway — treat your cloud reservation like an option, not a fixed bet. Lucas: If today was actually useful to you, the way these stay ad-free is listener support — buy me a coffee dot com slash fexingo. Luna: Yeah, I mean, this is exactly the kind of detail that saves real money, and that's why we do it. Lucas: So back to the retailer — they also included a clause that requires AWS to notify them ninety days before any instance family is deprecated. That way they can plan a swap instead of getting stuck. Luna: That's smart. Deprecation is a real issue. AWS has deprecated several older instance types over the years, and if you have reservations on them, you're forced into a migration. Lucas: Right. And without a swap right, you either pay to break the reservation or you run on deprecated hardware until the term ends. The notice clause gives you a window to move. Luna: I'm also seeing some enterprises demand 'forward swap' rights — where they can swap into instance families that don't even exist yet, like future-generation CPUs. Lucas: That's ambitious. I haven't seen many hyperscalers agree to that, but it's a logical extension. If you're committing for three years, you want to be able to take advantage of new hardware. Luna: One last thing — what about convertible reservations on Azure? How do they compare? Lucas: Azure calls them 'reserved instances with flexibility' — they allow swaps within the same VM series but not across series. So you can go from D2s v3 to D4s v3, but not from D-series to E-series. AWS is more flexible there. Luna: Good to know. So if you're multi-cloud, you need to track these differences. Lucas: Absolutely. And the contracts are getting longer and more complex. I think within two years, swap rights will be table stakes for any enterprise deal above ten million dollars. Luna: Which means if you're not asking for them now, you're leaving money on the table. Lucas: That's the bottom line. Next episode, we'll look at how some enterprises are bundling swap rights with early-termination options — a new hybrid clause that's starting to appear. Luna: Looking forward to it.