Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / The Price Cap That Failed AWS Azure and Google Cloud
Transcript
- Lucas: Let me start with a number that I think explains more about the cloud market right now than any press release: thirty-four percent. That is the premium, on average, that AWS, Azure, and Google Cloud charge for reserved instances versus on-demand compute in Q2 2026. Luna: Wait — reserved instances are supposed to be the discount option. You commit for one or three years, you pay upfront, and you get a lower per-hour rate. How is that a premium? Lucas: That's exactly the confusion. Reserved instances are cheaper per compute-hour than on-demand, but the thirty-four percent number is the premium you pay compared to what you'd pay if you bought compute on the spot market or used a truly flexible commitment model. The catch is that the big three have quietly engineered the pricing so that the 'discount' reserved tier is actually the default high-margin product, and the truly flexible tiers — spot instances, convertible reservations — carry hidden costs that eat the savings. Luna: So the tier that most enterprises actually use is the one the cloud vendors want them on. And they call it a discount. Lucas: Exactly right. And the interesting part is how we got here. In late 2024, there was a lot of noise about a 'cloud pricing revolution.' Google Cloud had introduced committed-use discounts with no upfront payment. Azure matched with reservation terms that seemed friendlier. AWS introduced savings plans that were supposed to be more flexible. But by early 2026, all three had walked back the most aggressive discounts by about eight to twelve percent. Luna: Because enterprises actually took them up on it. They shifted workloads to cheaper tiers, and the hyperscalers saw revenue per customer drop. Lucas: Bingo. The cloud providers learned that when you make pricing too flexible, customers optimize their spend downward — and that hurts the quarterly numbers. So they reintroduced complexity. The reserved-instance tier now comes with a 'capacity guarantee' that sounds great but actually limits your ability to migrate to newer instance families. If you want to move to a newer generation CPU, you have to convert your reservation, which triggers a fee — typically around twelve percent of the remaining commitment value. Luna: I've heard sales engineers call that the 'modernization tax.' And it's not small. For a mid-size fintech I was talking to last month, they had about 4 million dollars in annual reserved-instance commitments, and converting to newer instances would have cost them almost half a million in fees. Lucas: That's a real example. And it gets worse. The same fintech tried to arbitrage by buying reserved instances on the secondary market — there are brokers that resell unused AWS reservations at a discount. They found a deal that saved them about fourteen percent on paper. But when they moved the workloads, they hit data egress fees that wiped out the savings entirely. Luna: Because egress is still the hidden lever. The cloud providers make money on data leaving their network, and they've never really competed on that price. So any cost optimization that involves shifting data between regions or clouds gets eaten by egress. Lucas: Right. And that's the structural advantage. The big three know that once you're in their ecosystem, the cost of leaving — in egress, in re-architecture, in retraining your ops team — is higher than the premium they charge on reserved instances. So they can maintain that thirty-four percent markup and still have customers feel like they're getting a deal. Luna: But is there any movement toward real price transparency? I know the European cloud federation has been pushing for standardized pricing tables. Lucas: There's been some progress. In March 2026, the three providers jointly published a 'cloud cost comparison framework' in response to the EU's Digital Markets Act pressure. But when you actually read the fine print, the framework excludes egress costs, and it uses a 'typical enterprise workload' definition that assumes eighty percent utilization — which almost no one actually achieves. So the comparison is basically meaningless. Luna: It's a compliance checkbox, not a real transparency tool. And enterprises are starting to notice. I've seen more RFPs that include clauses requiring the cloud provider to disclose the effective hourly rate after all discounts and fees, not just the list price. Lucas: That's a smart procurement move. But it's hard to enforce because the cloud providers treat those pricing details as proprietary. The sales teams have discretion to offer 'custom discounts' for large commitments, but they won't put them in writing in the RFP response. They want you to sign the multi-year deal first, then negotiate the discount after you're locked in. Luna: That sounds like a classic bait and switch. You commit to three years of spend, and only then do you find out what you're actually paying per compute-hour. Lucas: It is. And the worst part is that the procurement teams at most enterprises don't have the leverage to push back. If you're a $500 million a year AWS customer, you have some negotiating power. But the vast majority of cloud spend comes from accounts that are under $50 million annually — and those customers get the standard pricing sheet with maybe a five percent courtesy discount. Luna: So the pricing cap that everyone expected — the one that would force margins down to single digits — never really materialized. Instead, the market settled on a stable premium that keeps the hyperscalers' margins fat. Lucas: Exactly. AWS still reports operating margins north of twenty-eight percent. Azure is around twenty-two. Google Cloud finally turned profitable in 2025 and is running at about fifteen percent. Those margins are supported by the reserved-instance pricing structure. If they had to compete purely on spot and on-demand pricing, those margins would compress to maybe ten percent. Luna: And that's why we don't see a real price war. The hyperscalers have effectively colluded — without colluding — on a pricing model that benefits all of them. It's a stable oligopoly. Lucas: Right. And the one wild card that could disrupt it is the rise of specialized cloud providers — companies like CoreWeave that focus on GPU compute for AI workloads. They don't have the legacy reserved-instance baggage. They price purely on supply and demand. In Q1 2026, CoreWeave's on-demand GPU pricing was about forty percent cheaper than AWS's equivalent reserved-instance tier for the same NVIDIA H100 capacity. Luna: But that's a niche. General enterprise workloads — databases, web servers, container orchestration — aren't going to move to a GPU cloud. They need the full ecosystem. Lucas: True. But if AI workloads continue to grow as a share of total cloud spend — and they're already about twenty-five percent of new workloads — then the specialized providers could force the big three to compete on price in that segment. And once you start competing on price for one workload type, it's hard to keep the walled garden intact. Luna: So the real pricing revolution might come from the edge, not from the core. From ai specific clouds that don't play by the reserved-instance rules. Lucas: That's the thesis I'm watching. But for now, the thirty-four percent premium is the reality. And if you're an enterprise cloud buyer, your best strategy is to treat reserved instances as a liability, not a discount. Negotiate shorter terms, demand transparency on effective rates, and always model the egress cost of leaving. Luna: That's a lot of work for a procurement team that's already stretched. But it's the only way to avoid the hidden tax. Lucas: Yeah. And on that note — and I want to be completely honest here — this kind of conversation, where we can dig into the real numbers without an agenda, is exactly why a handful of listeners chip in at buy me a coffee dot com slash fexingo. No ads, no sponsors, just this. Luna: It makes a real difference. We get to spend the time on the details that matter, and that's because of that small group. Lucas: So if these deep dives are useful for what you're building or running, that's the place. And we'll keep doing them either way. Luna: Alright. Back to that fintech case — you mentioned the egress fee wiped out their savings. How exactly did that happen? Lucas: They had a multi-region deployment. The secondary-market reservations were for us-east-1, but their primary compute was in us-west-2. Moving the workloads required cross-region data transfer at nine cents per gigabyte. After moving about 200 terabytes, the egress bill hit $18,000. The reservation savings were about $15,000. Net loss of $3,000. Luna: And that's just one region move. Imagine an enterprise with dozens of workloads across three providers. The complexity alone makes it almost impossible to optimize perfectly. Lucas: Right. And that's the cloud providers' real moat: not technology, but pricing complexity. They don't need to be the best — they just need to be the hardest to leave.