Memory shortage drives Nvidia AI server prices up by more than 15 percent

Nvidia AI servers are set to become more than 15 percent more expensive due to a persistent shortage of memory components. Bloomberg reports this, citing people familiar with the matter. The price increases apply to systems featuring Vera Rubin and Grace Blackwell chips scheduled for delivery in early 2027, and affect major customers including Microsoft, Google, Oracle, Meta and Amazon, as well as AI labs such as OpenAI and Anthropic.

The cause lies with the three dominant memory chip manufacturers: Samsung, SK Hynix and Micron. Their capacity for both conventional DRAM and the specialised HBM (High Bandwidth Memory) required by modern AI accelerators is falling short of demand. Server DRAM prices roughly doubled in the first quarter of 2026, according to Counterpoint Research, which recorded a quarter-on-quarter increase of 80 to 90 percent for DRAM, NAND and HBM in that same quarter.

The exact price increase for a specific Nvidia server platform depends on the memory intensity of the configuration. Memory now accounts for approximately 25 percent of the material costs for high-end AI server racks. GPUs were previously responsible for more than 80 percent of those costs; in more recent systems that share has fallen to around half, precisely because memory prices have risen so sharply.

How acute the shortage is

By October 2025, SK Hynix had already sold out its entire memory production capacity for all of 2026. Samsung and SK Hynix raised their supply prices for HBM3E, the memory type used in advanced AI chips, by nearly 20 percent ahead of 2026. Contract prices for conventional DRAM rose 90 to 95 percent quarter-on-quarter in the first quarter of 2026; a further increase of 58 to 63 percent is expected for the second quarter.

Consumers are feeling it too. A standard 32 GB DDR5-6000 memory kit cost around $392 in August of this year, compared with $110 to $140 a year earlier, more than a threefold increase in twelve months.

New production capacity for memory chips requires years of investment and construction. Deloitte expects that significant new capacity will not become available until 2029 or 2030. Gartner estimates that the shortage will persist at least until the first half of 2027.

Billions in extra costs for data centre builders

For parties building large-scale AI infrastructure, the higher memory prices add up quickly. Building an AI data centre with a capacity of one gigawatt could cost at least five billion dollars more than previously budgeted due to the price increases, according to the sources cited by Bloomberg.

Amazon Web Services has already raised its GPU rental prices by 20 percent, passing part of the increased procurement costs on to business customers of its cloud services. Whether other cloud providers will adjust their rates accordingly has not been confirmed at this time.

Nvidia itself operates with a gross margin of approximately 75 percent. The company recently signed a six-billion-dollar technology licensing agreement with AI startup Poolside and announced it would invest one billion dollars in the company, indicating that Nvidia is also seeking to broaden its market position through software licensing and partnerships.

Dependence on a limited supplier base

The memory landscape is dominated by three players: Samsung, SK Hynix and Micron. Large tech companies investing billions in AI infrastructure are entirely dependent on that same small group of manufacturers for a critical component. Near-term alternatives are scarce; HBM production requires specialised expertise and capital-intensive fabrication lines that new entrants cannot quickly establish.

This is relevant in the European context: the EU has committed approximately twenty billion euros to a series of AI gigafactories, and a French consortium bid ten billion dollars for a single site. But even large-scale European investments in computing capacity will not resolve the memory problem as long as memory production remains concentrated in Asian factories.

For Dutch and European data centre operators, cloud providers and AI startups that purchase or lease computing power, the situation means that higher infrastructure costs over the next two years are a realistic prospect. Investors in AI infrastructure projects would be wise to factor in memory availability and price trends as a separate risk item, alongside the more visible GPU scarcity of recent years.