Product2 publishersIndependently confirmed2 min readPublished Updated
Nvidia's server partners are passing the DRAM bill straight to Microsoft, Google and Oracle
Contract assemblers have warned of increases above 15% on Vera Rubin and Grace Blackwell systems shipping from early next year. Every AI capacity budget written before that is now short.
The Product Desk · Product desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Nvidia's biggest customers have been told that servers built around its AI chips will cost more than 15% more in many cases, on systems shipping from early next year.
- Vera Rubin and Grace Blackwell systems are covered, with the size of the rise set by chip generation and memory configuration.
- The warnings came from the contract manufacturers that assemble servers for operators including Microsoft, Google and Oracle.
- Amazon Web Services has already raised its GPU prices by 20%.
- Nvidia reports quarterly earnings next week.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost Every $1bn of planned server purchasing needs another $150m found somewhere, and it comes out of the buyer's capex rather than the chip vendor's margin.
- decision Anyone who got next year's capacity plan approved at old prices now chooses between going back for more money and ordering fewer racks.
- exposure Public procurement reprices slowest, so Europe's gigafactory awards carry a gap that lands on appropriations instead of on a supplier.
- precedent Component pass-through has already cleared the consumer tier, which makes a further round easier for suppliers to justify than to argue about.
Memory is the binding constraint, and it sits with three suppliers. Samsung, SK hynix and Micron make most of the world's DRAM, and output is rising without catching up to what AI infrastructure is absorbing [6]. That shapes how the increase arrives at a buyer. It is not a flat vendor uplift but a function of chip generation and memory configuration [2], so nobody can take one percentage and run it across a plan. Each configuration reprices on its own terms, and the denser the memory, the less last year's quote is worth.
Then there is the question of who is making the call. Nvidia runs a gross margin of about 75% and is the most valuable listed company on the market, and it is not absorbing this [5]. The contract manufacturers are the ones explaining the number to the customer [15], and Nvidia did not respond to The Next Web's request for comment [4].
The arithmetic is easy to do and nobody has published it. The EU has committed around EUR 20bn to a set of AI gigafactories, and a French consortium has bid $10bn for one site [12][13]. If that entire European commitment were spent on server hardware, which it will not be, 15% of it is EUR 3bn [18]. Treat that as a ceiling rather than a forecast. The real figure is smaller, and it is still not appropriated anywhere, because as The Next Web puts it, those bids were built on last year's hardware prices [17].
Commercial buyers are standing in the same queue. Nebius is tripling Nvidia capacity at its Finnish data centre on economics that have moved underneath it [14], and that comes on top of the conditions that were already slowing builds: project delays, labour shortages, tighter capital markets and local opposition [8].
The earnings call is where this gets tested. The question is not whether orders are holding but how much of the memory cost stays inside a 75% margin, and on The Next Web's reading of who is currently paying, the answer so far is none of it [10].
What to watch
- Whether Nvidia quantifies memory cost exposure or trims gross margin guidance on next week's earnings call.