Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / Why Enterprise Cloud Contracts Now Include Spot Instance Guarantees
Transcript
- Lucas: I've been digging into enterprise cloud contracts lately — specifically a clause that I'm seeing more often in recent renewals. It's about spot instance guarantees. Luna: Spot instances have always been the 'use at your own risk' option. Are hyperscalers actually guaranteeing something now? Lucas: They're starting to. I looked at contract language from a mid-size fintech that renegotiated its Azure deal in April. The new clause caps the reclamation rate at 1% per week — meaning Azure can't reclaim more than 1% of their spot VMs in any given week. And it limits spot price increases to no more than 20% above the baseline. Luna: That is a big shift. Historically, spot pricing could spike 3x or more during demand surges. I remember one AWS re:Invent where spot prices for GPU instances jumped 400% in a single day. Lucas: Exactly. And that volatility made spot essentially unusable for any workload that needed predictable cost or uptime. Enterprises would dabble — run batch jobs, some dev/test — but they'd never bet critical training pipelines on it. The risk was just too high. Luna: So what changed? Why are AWS, Azure, and GCP suddenly willing to put guardrails on spot? Lucas: Competition for AI workloads is the main driver. Enterprises training large models need massive compute — think thousands of GPUs running for weeks. On-demand pricing for that is astronomical. Spot could cut costs by 60-80%, but only if the enterprise can trust it won't get yanked mid-training. Hyperscalers are now bending contract terms to win those deals. Luna: It's like they're saying, 'We'll give you near-guaranteed spot, but you have to commit to a certain volume.' Lucas: Yes — typically the guarantee comes with a minimum spend commitment. The fintech I mentioned committed to $5 million in Azure compute over two years. In exchange, they got the 1% reclamation cap and a price ceiling at 80% of on-demand. Without that commitment, the hyperscaler has no incentive to reserve capacity for spot. Luna: Makes sense. The hyperscaler takes on more risk, so they want a guaranteed revenue stream. But how enforceable are these clauses? If AWS reclaims 2% in a week, what happens — a credit? Lucas: In the contracts I've seen, it's a service credit — usually 10% of the monthly spot spend for each percentage point above the cap. So if the cap is 1% and they reclaim 3%, the enterprise gets a 20% credit on that month's spot bill. It's not huge, but it creates a financial incentive for the hyperscaler to manage reclaims. Luna: That's interesting. It turns spot from a 'maybe' into something you can actually budget around. I wonder if smaller companies can negotiate these terms, though. Lucas: That's the catch — these clauses are showing up in enterprise agreements with seven-figure commitments. If you're a startup spending $50k a month, you probably won't get a reclamation cap. But the trend might trickle down. As hyperscalers build the operational muscle to offer guarantees, they might offer lighter versions for smaller customers. Luna: It also makes me think about the flip side: if spot becomes more reliable, does on-demand pricing come under pressure? Because the value proposition of on-demand is 'always available, no interruptions.' If spot is almost as reliable at half the price, why pay full freight? Lucas: Right — and that's exactly why hyperscalers are cautious. They don't want to cannibalize their on-demand revenue. So the guarantees are carefully calibrated: the price ceiling is still above what spot normally costs, and the reclamation cap still allows some interruptions. It's not a 'five nines' SLA. It's more like 'best effort with teeth.' Luna: You know, speaking of making things work with small contributions — this podcast is a good example. A couple of dollars a month is genuinely what keeps these going. If you've gotten something out of our episodes, buy me a coffee dot com slash fexingo really does help us stay ad-free and focused on the real details. Lucas: Yeah, it's true. We put a lot of time into sourcing these contract clauses and talking to procurement folks. Listener support makes that possible without sponsors dictating the angle. So if you can chip in, we appreciate it. Luna: All right — back to the clauses. Lucas, you mentioned the price ceiling. How exactly is that structured? Is it a fixed number or a formula? Lucas: It varies. In the Azure contract I saw, the price ceiling is set at 80% of the equivalent on-demand instance price. So if the on-demand price for a GPU instance is $3 per hour, spot can never exceed $2.40. That's a hard cap. For AWS, I've seen a different approach: a 'spot price volatility guard' that limits price changes to no more than 15% per hour. It doesn't cap the absolute price, but it smooths out spikes. Luna: Which approach do enterprises prefer? Lucas: From the conversations I've had, the hard cap is more popular for budgeting. You can model your maximum compute cost. The volatility guard still leaves some uncertainty — you could have a gradual rise over a day that ends up 50% higher. But it's better than nothing, and it's easier for the hyperscaler to implement without reserving capacity. Luna: So it sounds like this is still early stage. Are there any standard terms emerging, or is every contract bespoke? Lucas: Very bespoke. I've seen reclamation caps range from 0.5% to 5% per week. Price ceilings from 60% to 90% of on-demand. And the commitment periods vary — some are annual, some are three-year. The common thread is that enterprises with leverage — large spend, multiple cloud providers, willingness to walk — are getting better terms. Luna: What about GCP? They have preemptible VMs with a 24-hour max lifetime. Are they offering guarantees too? Lucas: GCP is actually the most aggressive. They recently introduced 'Spot VMs with extended lifetime' — up to 48 hours without forced termination. And in enterprise contracts, they're offering 'committed use spot discounts' where you get a guaranteed spot price for a one-year term, with a reclamation rate under 2%. They're using this to win AI training workloads from AWS and Azure. Luna: That makes sense — Google has a lot of TPU capacity and wants to fill it. If they can offer near-guaranteed spot on TPUs, that's a differentiator. Lucas: Exactly. And it's working. I talked to a computer vision startup that runs all its training on GCP spot TPUs with a 1% reclamation cap. They're saving 70% compared to on-demand and have had zero interruptions in six months. That kind of reliability was unthinkable two years ago. Luna: So what's the downside? If hyperscalers are giving away spot guarantees, something must give. Higher on-demand prices? Less capacity for smaller customers? Lucas: Both, probably. To offer guarantees, hyperscalers need to reserve capacity — which means less spot capacity available for everyone else. So the 'free for all' spot market might shrink. And on-demand prices might rise to compensate for the discounts on spot. It's a redistribution of cost and risk. Luna: Interesting. For enterprise FinOps teams, this changes the calculus. Instead of just 'use spot when you can,' it's 'negotiate spot guarantees as part of your contract.' Lucas: Exactly. And that requires a different skill set — procurement expertise, not just cost monitoring. The teams that succeed will be the ones that treat cloud contracts like any other strategic vendor negotiation, with specific terms around pricing, capacity, and reliability. Luna: So for our listeners who manage cloud spend — what's the one thing they should ask for in their next renewal? Lucas: A reclamation rate cap. That's the most impactful term. Even a 5% weekly cap makes spot viable for workloads that can handle occasional interruptions. Pair it with a price ceiling at 80% of on-demand, and you've got a predictable, cost-effective compute option. Start the conversation early — these clauses take time to negotiate. Luna: Good advice. And if you want more details on how to structure that ask, we'll put some resources in the show notes. Lucas: I think this trend is only going to accelerate. As AI workloads grow, the demand for cheap, reliable compute will push hyperscalers to offer more guarantees. The spot instance of tomorrow might look a lot like today's on-demand — just with a discount. Luna: Let's hope so. Thanks for breaking this down, Lucas. Lucas: Always a pleasure.