Bloomberg reported at the weekend that some of Nvidia's biggest customers have been told what next year's servers will cost, and it is more than 15% above this year's in many cases. The increase lands on systems shipped from early next year and covers the flagship Vera Rubin and Grace Blackwell machines, according to people Bloomberg describes as familiar with the process, who asked not to be identified because the communications are not yet public.
How much any given system goes up by depends on the chip generation and the memory configuration. The contract manufacturers that assemble servers for the large data centre operators, Microsoft, Google and Oracle among them, have already told their own customers what is coming. Nvidia did not respond to Bloomberg's requests for comment.
Who has the leverage now
An accelerator is only as useful as the memory beside it, and memory is where the squeeze is. Samsung, SK Hynix and Micron between them make most of the world's dynamic random access memory, the working memory those chips are paired with, and they have not caught up with AI demand despite raising output, which has driven the price of a commodity component up sharply and handed the three of them unusual influence.
Nvidia is not a company that normally has to take that. It keeps about 75 cents of every dollar it sells and charges tens of thousands of dollars a chip, because TSMC, the contract manufacturer that makes them, still cannot meet demand either. Apple and Qualcomm have both said recently that chip shortages are forcing them to charge more, and Tom's Hardware reported earlier this month that Nvidia has raised the price of its gaming graphics cards too.
How long the squeeze runs
Not briefly, on the industry's own account. SK Hynix's chief executive has forecast that 2027 will be the worst supply year the memory industry has had, with demand running ahead of what the company can make well into the next decade even as it expands capacity. Nvidia's own chief executive said last month that AI memory shortages would continue for several years and that SK Hynix would remain his largest supplier. UBS expects the market to stay undersupplied until at least the second quarter of 2028.
The spending carries on regardless. Micron plans to invest more than $250 billion in the United States through to 2035, up from $200 billion, and Bank of America expects the large cloud operators to spend about $851 billion on capacity this year and $1.15 trillion next.
Nvidia reports its fiscal second quarter on Wednesday.
