Loading today’s stories
Invest1 distinct publisher2 min readUpdated
Contract builders for Microsoft, Google and Oracle have told customers to expect more than 15% on Vera Rubin and Grace Blackwell systems. On a flat budget, that is roughly 13% less compute.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
Treat the "more than 15%" figure [1] as a rebase of 2027 unit economics rather than a line item to haggle over. On a flat budget, a 15% increase buys about 13% fewer systems [12], and Fortune's sources put 15% at the low end in many cases, with the actual number varying by chip generation and memory configuration [3]. Any 2027 plan denominated in racks rather than in dollars is already short.
Why the cost gets passed through is legible in Nvidia's own margin. The company runs a 75% gross margin [8]. Arithmetic on that: a cost increment equal to 15% of the selling price, absorbed rather than billed, would take the margin to 60%, a 15 point hit [13]. The comparison is loose, because 75% is company-wide while the 15% is measured on complete servers that contract manufacturers assemble and buy DRAM for [4]. It still sets the direction of travel. No vendor surrenders 15 points of margin to protect a customer that has nowhere else to go.
The leverage sits one tier upstream. An accelerator's usefulness depends on how much DRAM is paired with it [9], and Samsung, SK Hynix and Micron account for most of the world's supply, with output increases that have not caught up to demand [10]. Fortune's reading is that Nvidia's failure to hold the line is itself the evidence of how much power that hands the memory makers [11]. Nvidia's other bottleneck has not moved either: it can charge tens of thousands per chip because TSMC's supply still cannot meet demand [14].
That squeeze does not respect the in-house silicon story. Amazon, Microsoft, Google and Meta all have their own chip programs and all remain dependent on Nvidia purchases for their build-outs, and how far they can push toward independence depends on their own access to Samsung, SK Hynix and Micron [15]. A custom accelerator with no DRAM allocation is a slide, not a server.
The caveat is worth stating plainly: this is anonymous sourcing about communications that have not been made public [4], and Nvidia did not respond to Fortune's requests for comment [5]. What supports it is the surrounding pattern. Apple and Qualcomm have publicly said chip shortages forced them to raise prices [6], and Nvidia has already lifted gaming graphics card prices, per Tom's Hardware [16]. Fortune also notes the increases land on build-outs already dealing with project delays, labour shortages and tighter capital markets [18].
For anyone signing 2027 orders, the clause that determines the outcome is the memory pass-through, not the headline price per accelerator.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
The price hikes take effect on systems shipped early next year and will affect systems including those with the flagship Vera Rubin and Grace Blackwell chips.
The size of the increases will depend on the generation of Nvidia chips and the memory configurations.
Nvidia has a gross margin of 75%.
Some of Nvidia's biggest customers have been told that prices of servers containing its AI chips are going up more than 15% in many cases, with memory chip costs soaring.
Companies that build the servers under contract for large data centre operators such as Microsoft, Alphabet's Google and Oracle recently notified their customers of the forthcoming increases, according to people who asked not to be identified discussing communications that have not been made public.
Nvidia representatives did not respond to requests for comment.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single outlet, anonymous supply-chain sourcing
Every specific in the cluster traces to one publisher relying on people who asked not to be identified describing communications that have not been made public. Nvidia did not respond to comment requests, and no contract builder, hyperscaler or memory maker is on record. The supporting economics (75% gross margin, DRAM concentration among three suppliers, TSMC capacity constraint) are conventional and verifiable, but the load-bearing facts - the >15% figure, the early-2027 ship window and the notifications themselves - are uncorroborated.
Notices issued, shipments not yet priced in
There is concrete evidence of a pricing action in motion - notices already sent to customers of contract builders, plus separately reported gaming GPU increases and Apple and Qualcomm passing chip costs through - but the increases only bite on systems shipping in early 2027. No order volumes, signed contracts, cancellations or substitution decisions are reported, and hyperscalers are described as still dependent on Nvidia, so realised market impact is prospective rather than observed.
Slightly ahead of what is verified
The reporting is hedged where it should be - 'more than 15% in many cases', variation by generation and memory configuration - and its economic framing is grounded in publicly known facts about DRAM concentration and TSMC constraints. The overstatement is modest and definitional: a precise-sounding percentage, a named ship window and derived consequences such as roughly 13% fewer systems per fixed budget all inherit their certainty from a single anonymous notification that the vendor would not confirm.
Unnamed supply-chain sources, days before earnings
The claims come from anonymous participants in a negotiation about prices, disclosed a week before Nvidia's fiscal Q2 report - a window in which buyers, builders and suppliers all have reasons to shape expectations about who bears memory costs. Nvidia's silence and its 75% gross margin mean the framing of 'cannot absorb the cost' is contestable and commercially useful to whichever party leaked it. Offsetting this, the source is a general-interest business publisher with no product to sell, and no vendor marketing material is involved.
Plausible mechanism, thin verification
Directionally the story fits verifiable structural facts: DRAM output from three dominant suppliers has not caught up with AI demand, prices have risen sharply, and accelerator value scales with paired memory. That makes a pass-through highly plausible. But with one publisher, no named source, no vendor confirmation and no absolute price baseline, the exact magnitude, scope across configurations and treatment of in-flight orders remain unsettled.
product
Nvidia's server partners are passing the DRAM bill straight to Microsoft, Google and Oracle1 distinct publisher
build
SemiAnalysis to software teams: your token cost starts at the fab, not the price list1 distinct publisher
invest
Morgan Stanley puts a price on the zoning board: local opposition moves into the capital model1 distinct publisher
product
Samsung's July foundry price rise moves the AI crunch into your bill of materials1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 22, 2026