What Nvidia announced on 22 August
Contract manufacturers that build servers on Nvidia's behalf for major data center operators notified their customers - including Microsoft, Google, and Oracle - of price increases above 15 percent on AI servers. Affected are systems with the upcoming Vera Rubin and Grace Blackwell chip generations; the exact amount varies by chip generation and memory configuration. The new prices apply to systems shipping in early 2027. Gaming-oriented graphics cards are also seeing price increases - a sign the cause isn't specific to AI servers, but to a component that sits in both product categories: memory chips.
The cause: not a shipping delay, but a permanent capacity shift
Market data from TrendForce shows the scale: conventional DRAM prices rose 50 to 90 percent quarter over quarter, with many DDR5 modules up 3.5 to 4x. Samsung and SK Hynix have also raised prices for HBM3E memory - High Bandwidth Memory, the memory technology AI accelerators like Nvidia's H200 need for their compute performance - by roughly 20 percent for 2026.
The decisive point is the cause of this price increase: it's structural, not the result of a temporary bottleneck. Memory chip makers have permanently shifted their manufacturing capacity away from conventional DRAM and NAND flash - the memory types used in laptops, smartphones, and ordinary corporate servers - toward HBM. The reason is purely economic: HBM production consumes three times more wafer capacity than standard DRAM, but commands three to five times higher margins. For a memory chip maker, reallocating capacity toward HBM is an obvious economic decision - with the consequence that HBM production capacity for 2026 is already completely sold out at the major manufacturers (Micron, SK Hynix, Samsung).
Why this affects more than AI servers
Data centers now consume an estimated 70 percent of global memory chip production. That explains why the shortage isn't confined to AI infrastructure: any product that needs conventional DRAM or NAND memory - laptops, smartphones, tablets, ordinary corporate servers with no AI connection - competes for the same, increasingly scarce manufacturing capacity. Other major tech companies, including Apple and Qualcomm, have reportedly already had to raise prices on their own products for the same reason.
For a company in our audience, that means: your own IT procurement is affected by this development, even if you run no AI servers at all in-house. Regular laptop refresh cycles, your next server purchase, equipping new workstations - all of it runs through the same memory chip supply chains currently being reallocated toward AI data centers.
The connection to cost trends already observed
This series has already written several times about AI-related cost increases, without examining the underlying hardware economics in detail until now. Microsoft 365's blanket price increase, officially justified by Copilot integration, and the cases documented by Bitkom, where a third of surveyed German companies reported AI costs higher than expected, can be read more plausibly against this backdrop: a structurally rising compute and memory price inevitably affects the pricing of every service running on that infrastructure - regardless of how any individual vendor frames its own price increase.
That doesn't mean every AI vendor's price increase is explained solely by memory chip costs - margin, competitive pressure, and individual business decisions play a role too. But it does mean part of the current cost pressure on AI and cloud services has a real, structural basis that can't be resolved through better contract negotiations with a single vendor.
What this means in practice
Since the underlying capacity shift is structural and not reversible in the short term, per market observers, there are concrete consequences for your own IT budget planning.
- Consider bringing forward planned hardware procurement (laptop fleets, servers, storage) for the next one to two years rather than delaying it, if a price increase is more likely than a normalization - reassessing your own replacement cycles can be worthwhile.
- Scrutinize AI and cloud vendors' price-increase justifications critically, but don't dismiss them as implausible on principle: part of the industry's current cost increases has a real, externally verifiable basis in the memory chip supply chain.
- For larger hardware investments, compare multiple suppliers and timing windows rather than committing to a single vendor or a single procurement moment, since price dynamics can differ by chip generation and supplier.
- Keep an eye on how memory chip prices develop further (for instance via market trackers like TrendForce), so your own IT budget planning doesn't rest on outdated price assumptions.
The real value of this analysis isn't a prediction of when memory chip prices will normalize again - that can't be credibly forecast from the outside. It's making visible a cause that sits behind many seemingly unrelated price increases of recent months, and treating your own IT budget planning not vendor by vendor in isolation, but with this shared structural cause in view.