On 10 August 2026, Octave Klaba, the founder and chairman of OVHcloud, one of Europe’s largest independent cloud providers, posted the company’s new prices in a thread on X. Almost everything is going up. The steepest increases fall on the newest dedicated servers, which will cost as much as 87 percent more, and the rises run down the whole catalogue from there, through older machines and into public cloud instances. New orders reprice on 1 September. Renewals follow a month later.
Behind the increases is the cost of memory. Since 1 July, adding RAM to a new OVHcloud order costs 127 percent more, and adding storage 89 percent more. By Klaba’s account, DRAM and NAND prices have risen roughly sixfold year on year through June, he projects ninefold by September and twelvefold by 2027, and the reason is that the manufacture of GPUs and high-bandwidth memory for AI has pulled fabrication capacity away from the ordinary components every server is built around.
Which means OVHcloud does not set its own prices. Its cost base moves with the purchasing decisions of the American companies it competes against.
For comparison, Amazon buys DRAM and NAND in the same overheated market. Yet so far its prices have barely moved. It has raised EC2 Capacity Blocks for ML, one reserved GPU product, by about 20 percent from July on top of roughly 15 percent in January, and left everything else untouched, including instances running Trainium, the accelerator it designs itself. This was also done without an announcement. The new numbers appeared on the pricing page, and after Business Insider reported them AWS said only that Capacity Blocks prices “are updated periodically based on supply and demand.”
Amazon can hold its prices because it designs its own accelerators and because it buys memory years in advance, in volumes big enough that suppliers serve it first. None of the European providers has those options. Klaba described OVHcloud’s own position in the same thread, placing orders “month by month, over 12 months, with no guarantee of the purchase price and without knowing what will be the real demand from our customers.” There is no reason to think StackIT, IONOS, Scaleway or Hetzner, the other European Cloud providers, buy on better terms.
And yet, even after all of this, OVHcloud is still the cheap option. Klaba says it remains “the cheapest on the market for bare metal and public cloud,” which Hetzner might quarrel with, but against the American clouds his numbers hold, “where before we could be 3x cheaper, we will be 2x cheaper (if our competitors don’t increase their prices).” This raises an awkward question.... If OVHcloud is two or three times cheaper, why do the American clouds hold roughly 70 percent of the European market while European providers, all of them combined, hold about 15 percent? Europe has had a cheaper European cloud for years but it seems that european companies and governments prefer to keep buying the expensive American cloud. So what gives? The answer is that most customers are not buying raw compute. They are buying the layer on top of it, managed databases, identity systems, AI services, hundreds of products that come bundled with the platform. Little of that layer exists at the European providers. A company that moves to OVHcloud pays half as much for its servers and then pays its own engineers to build and run what AWS was running for it, and for most companies coding their own infra is a task they could not manage.
In the first chapter of the book I sort Europe’s technological dependency into four forms, and two of them show up in this story. The first is lock-in, switching costs and installed ecosystems that hold customers in place even where alternatives exist. That is what I was referring to before when talking about the market-share situation. While alternative exists to some degree, it still remains cheaper to stay or extend the use of products within the existing vendors catalogue, because switching or signing up an additional service provider costs more in either time or capital than the savings are worth. The second is economic capture, value and market power flowing to the incumbents while the alternatives stay too small to fight back. This is what is currently leading to the differentiated price rise. The AI buildout is what drove memory prices up. The companies doing that buying are also the ones insulated from the result, because they bought their memory before the price moved. OVHcloud buys month to month, so it pays what the market asks today, and passes it on in September.
Public buyers across the continent are meanwhile being told, in procurement frameworks, in security certifications and in ministerial speeches, that moving workloads to sovereign providers is the prudent thing to do. That case rests on jurisdiction rather than on price. But every migration still has to be paid for, and cheap servers were a big part of what made the arithmetic work. Raw infrastructure is now getting more expensive, which makes each move harder to justify on none sovereignty logic exactly when Europe needs to transition. But the worste part of this is that the price of the transformation Europe keeps calling necessary is being set by the companies it is supposed to migrate away from.
Memory prices are cyclical, so the obvious objection is that this shock will pass, and it probably will. Fabs add capacity, the AI buildout eventually slows, and DRAM has crashed as violently as it has spiked more than once. Klaba expects elevated prices to run through 2028, with a return to normal hoped for in 2029. But 2029 sits beyond the horizon of most of the European migration programmes now underway, so the budgets written for them will carry the higher prices for as long as they run.
In the book I call this last piece an operational choke point. Europe makes almost no memory of its own, and no European buyer is big enough to be served first when supply runs short, so when DRAM and NAND prices move there is nothing in Europe that can absorb them or redirect them. OVHcloud’s invoices go out from Roubaix in euros, but the numbers on them are decided in fabs and purchasing departments Europe has no part in.
A single price change should not normally be this laden. But follow the cost back, through the memory market and the contracts that protect the companies doing most of the buying, and you arrive at the dependencies I describe in the first chapter of Digital Sovereignty, which I am serialising here, and at how little room they leave a European business that wants to move. If you negotiate cloud contracts for a European organisation, I would like to hear how these numbers land in your spreadsheet, and if you can fault a figure above, let me know.
Sources:
https://www.businessinsider.com/amazon-raises-ai-cloud-prices-memory-chip-costs-soar-2026-6
True sovereignty cannot be negotiated in software if your memory bitlines are forged in someone else's fab. 🌐
When European providers get hit with an 87% price shock on bare metal and 127% on RAM, it isn't just an inflationary bump. It's a structural eviction. American hyperscalers didn't just outbuild the competition. They absorbed the global supply of silicon wafer starts to feed high-bandwidth AI memory banks. When you buy DRAM month-to-month on the spot market, you aren't running an independent infrastructure. You're paying the raw marginal cost of someone else's training cluster buildout. 📉
We must confront a fundamental reality of compute topology: software autonomy built on unhedged physical hardware is an illusion. When the underlying physical memory capacity gets redirected toward high-margin AI accelerators, commodity server economics collapse overnight. The hyperscalers didn't keep prices flat through goodwill. They did it through massive multi-year advance wafer commitments and proprietary accelerator designs that insulate their core margins from the spot floor. 🛡️
Moving workloads to regional clouds based purely on jurisdiction ignores the physical layer of the stack. If a sovereign provider is forced to pass through a 12x memory price surge by 2027, the economic incentive to migrate evaporates. The real bottleneck isn't just the software ecosystem or managed services moat. It's the complete lack of domestic semiconductor fabrication capable of shielding local infrastructure from global AI capital expenditure cycles. ⚡
Can an infrastructure ever truly be sovereign if its operational unit economics are dictated by the hardware procurement roadmap of its direct competitors?
(¬_¬)