Nvidia has told its largest customers that AI server systems built around its next-generation chips will cost more than 15 per cent more, driven by a memory shortage that shows no sign of easing.
According to Bloomberg, the hikes affect systems using Vera Rubin and Grace Blackwell chips — the architectures Nvidia is counting on to maintain its grip on AI training and inference. Contract manufacturers building servers for Microsoft, Google, and Oracle have already passed the increases along to their customers. The new prices apply to shipments early next year.
The main driver is DRAM. Samsung, SK Hynix, and Micron have been raising contract prices for months as AI data centres consume an ever-larger share of global memory output. Tom’s Hardware reported that the memory cost increases are the primary factor, not Nvidia’s chip pricing itself.
The tension underneath
The price hike lands on exactly the companies trying hardest to break free of Nvidia. Amazon, Microsoft, Google, and Meta — the four biggest buyers of AI servers — are all developing their own custom silicon. So far, that hasn’t dented their dependence on Nvidia’s GPUs for training the largest models. OpenAI and Anthropic, which buy compute from the cloud giants, will feel the increase indirectly.
The Decoder framed the dynamic bluntly: the same companies pouring billions into AI infrastructure are bankrolling the market power of the supplier they’re trying to replace. Every dollar spent on Nvidia servers strengthens the position of a company that already guarantees up to $105 billion in credit for OpenAI’s Ohio data centre — a chipmaker functioning as banker, not just vendor.
This isn’t a temporary blip. Memory has been the AI supply chain’s bottleneck since early 2026, when NZ consumers started paying 10 to 30 per cent more for phones and laptops because data centres were hoovering up DRAM and NAND. The consumer impact hasn’t eased. What’s changed is that the cost pressure has now reached the server level — the most expensive tier of the AI hardware stack.
Why memory, not chips
Nvidia’s GPU pricing is the headline number in AI infrastructure, but memory is the constraint that actually moves the total system cost. A modern AI server uses dozens of high-bandwidth memory (HBM) modules alongside the GPU. HBM3E — the current generation — is produced by only three companies, and their capacity is fully committed to AI customers through 2026.
When Samsung or SK Hynix raises contract DRAM prices, Nvidia can’t absorb the difference without margin erosion. Passing it through is the only option. The 15 per cent figure is an aggregate — some configurations may see higher increases, particularly for systems with maximum memory configurations that are most memory-constrained.
Nvidia has not commented publicly on the price hikes. The company typically doesn’t confirm pricing changes with customers, leaving it to OEMs and cloud providers to disclose.
The downstream question
The price increase raises a question the AI industry has been dodging: when does the cost of infrastructure start constraining model development? OpenAI paused training on its Astra model for two weeks in August to implement new security protocols — a decision driven by safety, not cost. But the economic reality is that each new frontier model requires more compute, which requires more servers, which now cost 15 per cent more than they did a quarter ago.
For cloud providers, the math is straightforward. AWS, Azure, and Google Cloud charge customers per hour for GPU instances. A 15 per cent increase in server hardware costs doesn’t translate linearly to cloud pricing, but it does compress margins. Eventually, that pressure reaches the API pricing that developers and startups pay — the same pricing that DeepSeek and other Chinese labs have been undercutting for months.
The NZ angle here is indirect but real. Local AI startups buying API access from US providers will feel price pressure as infrastructure costs cascade. NZ research institutions running their own GPU clusters — including the NeSI supercomputing platform — face the same memory cost inflation as their overseas counterparts, just at smaller scale. And the memory chip shortage that hit consumer electronics earlier this year is the same supply chain pressure now hitting enterprise hardware. The bill arrives everywhere eventually.
What’s different this time
Previous Nvidia price increases have been driven by demand — more customers wanting more GPUs. This one is driven by supply. Memory is the input Nvidia doesn’t control, produced by companies that have their own pricing power and their own customer priorities. Samsung’s decision to allocate more DRAM to smartphones rather than AI servers would hurt Nvidia’s supply chain as much as it would help consumers.
That’s the structural difference. When GPU demand drives price increases, Nvidia captures the upside. When memory supply drives them, Nvidia is just the pass-through. The margin compression is real, and it arrives at a moment when the company is guaranteeing billions in financing for its largest customers. The chipmaker-as-banker model works when margins are fat. It gets harder when memory suppliers are eating into those margins from below.
📰 Sources
- Bloomberg — Nvidia customers notified about AI-related price hikes above 15%
- Tom’s Hardware — Nvidia reportedly warns biggest customers of 15% price hikes on AI servers
- The Decoder — Memory shortage reportedly drives Nvidia AI server prices up about 15 percent
- Fortune — Nvidia customers notified about AI-related price hikes above 15%
