OVHcloud raises prices because the AI race made memory more expensive for everyone

OVHcloud will raise prices across much of its catalog this fall (Northern Hemisphere), and the cause isn't in the products that will get more expensive. According to an InfoQ report, founder Octave Klaba attributes the move to the accelerated buildout of AI infrastructure, which is pulling manufacturing capacity away from the components any ordinary server is built with.

The mechanism: factories shifting to GPU memory

Klaba laid out the reasoning in a post on X (he writes in French; the quotes here come from machine translation). The world's three major RAM suppliers reconfigured their factories toward high-bandwidth memory (HBM), which serves GPUs and carries better margins, at the expense of standard DDR4 and DDR5, the memory used in ordinary servers.

The numbers come from OVHcloud's own purchasing. Indexing June 2025 as base 100, memory reached 604 by June 2026, SSDs hit 323, and HDDs hit 148. In other words: the company now pays roughly 6x the RAM price it paid a year ago. And the forecast gets worse:

"This month, June 2026, we're paying 6x the RAM price we paid in June 2025. We already know it will be 9x in September 2026, while forecasts for early 2027 stand at 12x!?"

Klaba adds that NVMe drives are at 7x, HDDs at 3.5x, and that he's heard of price increases of 15% to 20% on CPUs, motherboards, and network cards.

Who pays more, and how much

The increase isn't uniform. Customers renting the 2026-edition gaming servers face the steepest hike, 87%, while other recent servers rise between 40% and 59%, all effective from September. Customers on 2024-generation hardware see smaller increases upon renewal, which Klaba estimates to be three to six times lower than on a new order. Older lines, including Kimsufi, Rise, and earlier generations of Advance and Scale, are left out, as they already were in an April price increase.

Some charges are being restructured, not just raised. Starting October 1st, storage and IP addresses become separate line items on Gen3 instances, at €0.000146 per GB per hour and €0.0027 per hour. OVHcloud is also dropping the 1-, 6-, and 24-month savings plans, keeping only the 12- and 36-month options, which lock in the price for the contract's duration, a clear sign that the company is betting on cost predictability as a selling point.

Why AWS barely moved

What elevates this case above a simple pricing notice is the asymmetry it exposes. Klaba describes OVHcloud's buying position as orders placed "month by month, over 12 months, with no guarantee on the purchase price and no idea what our customers' actual demand will be."

Amazon buys in the same market and has barely moved. It raised prices on EC2 Capacity Blocks for ML (a reserved-GPU product) by about 20% in July, following ~15% in January, and left the rest of its catalog untouched, including instances with Trainium, the accelerator the company designs itself.

The explanation is purchasing scale and vertical integration. Hyperscalers contract memory years in advance, in volumes that guarantee priority allocation, and several design their own chips. A provider that buys market components on rolling monthly orders has none of those cushions.

What changes for those building software in Brazil

This asymmetry matters more than it seems at first glance. If Klaba's reading is correct, the pressure is market-wide, not specific to OVHcloud: The Register spoke with the founder of a mid-sized managed provider who expects similar announcements from Azure and AWS soon. Klaba himself qualifies his projections with "if our competitors don't raise their prices."

In practice, this means developers and platform teams running ordinary workloads (databases, caches, queues, web applications, nothing AI-related) may see infrastructure costs rise for a reason that has nothing to do with what they run. Bare metal and high-RAM VMs tend to be the most affected, since memory is precisely the component that spiked the most.

Some concrete points for anyone planning capacity from here on:

• Renewal vs. new order: keeping previous-generation hardware can cost three to six times less than provisioning new equipment. Checking whether you can stretch the life of what's already contracted is a real cost decision.
• Long plans lock in price: with short plans discontinued at OVHcloud, 12- and 36-month contracts become the tool for protecting against cascading price increases.
• Exchange rate stacks on top: OVHcloud invoices are issued in euros. For those paying from here, the component price hike compounds with currency fluctuation, doubling budget uncertainty.

What remains open

OVHcloud itself says the situation is exceptional and should last until 2028: "AI demand is insane at every level: data centers, GPUs, token-as-a-service, agentic AI." Klaba maintains that the company is still the cheapest option for bare metal and public cloud, but puts a number on the erosion: "where we used to be 3x cheaper, we'll be 2x cheaper (if our competitors don't raise their prices)."

There's also a structural angle. Andrei Nutas, who is serializing a book on digital sovereignty, points out that the companies driving memory demand are the same ones shielded from the resulting price, because they bought ahead of time. He sums it up: "OVHcloud's invoices come out of Roubaix in euros, but the numbers on them are decided in factories and purchasing departments that Europe isn't part of."

One detail is both comforting and unsettling: memory markets are cyclical and have crashed just as fast as they spiked before, so the shock should pass. The problem is that budgets and contracts written during the peak don't pass. Whoever locks in capacity now is fixing costs at a high point on the curve, and that's the math infrastructure teams in Brazil will have to work out in the coming months.