Server costs in the AI buildout are rising fast because memory has stopped being a side component and started setting the price of the whole rack.
The AI hardware bill is moving in the wrong direction for the companies buying it. Microsoft, Google, Amazon and Meta can still get the Nvidia systems they need, but they're no longer buying into a supply chain where the GPU is the only scarce part. Memory is now doing its own damage.
TrendForce put hard numbers on that shift. The research firm said conventional DRAM contract prices jumped 90% to 95% quarter over quarter in the first quarter of 2026, while NAND Flash contract prices rose 55% to 60% over the same stretch. By late March, TrendForce was already forecasting another 58% to 63% increase for conventional DRAM in the second quarter and a 70% to 75% jump for NAND. That isn't normal inflation. It's a repricing of one of the basic inputs behind AI servers.
Supply just isn't there.
Reuters reported in July that SK Hynix Chief Executive Kwak Noh-jung expects the memory industry to face its worst supply shortage in 2027, with demand likely to exceed the company's production capacity beyond 2030. Micron has also said customers such as Nvidia committed $22 billion to lock in memory supply, according to Reuters. Those are not ordinary short-term orders. They're attempts to secure allocation before the next round disappears.
Nvidia Confirms Vera Rubin Is Shipping but Not the $630 Billion Number Nvidia confirmed Vera Rubin has ramped into full production with shipments starting this fall, but the widely shared figure of 1,000 racks a day and $630 billion in quarterly revenue traces back to no statement from Nvidia or any named analyst. What is verified, from Ming-Chi Kuo's rack estimates to Jensen Huang's own $300 billion and $1 trillion... - vera rubin gpu production timeline - nvidia vera rubin revenue claims
Memory Is Now A Real Line Item #
Nvidia's own hardware shows exactly how much that matters. A Morgan Stanley estimate reported by Tom's Hardware put the cost of a Vera Rubin based VR200 NVL72 rack at about $7.8 million for hyperscalers, nearly double the roughly $4 million estimate for a GB300 NVL72 rack. Memory alone accounts for about $2 million of that, up 435% from the memory cost inside a GB300 system. A few years ago, memory was something buyers noticed after the GPU price. Not anymore.
That's the change hyperscalers have to swallow.
Nvidia has little reason to act as the shock absorber. The company reported a 75.0% non-GAAP gross margin for its fiscal first quarter of 2027 and guided to the same 75.0% non-GAAP margin for the second quarter, according to its own investor materials. If memory inflation moves through system pricing, Nvidia protects the number investors care about most. Its customers get the other side of that trade.
CNBC has reported that Alphabet, Microsoft, Meta and Amazon are expected to spend nearly $700 billion combined this year on AI buildouts. You don't commit that kind of money because the next server is cheap. You do it because falling behind on compute looks worse than paying the higher bill. That is a powerful position for suppliers and a difficult one for buyers, even buyers with trillion-dollar market values.
The Bill Is Reaching Renters #
The pain is already showing up further down the chain. Nebius raised on-demand GPU rental prices by roughly 30% from June 1, with H100 pricing moving from $2.95 to $3.85 per GPU-hour and B200 pricing from $5.50 to $7.15, according to pricing notices tracked by market reports. Amazon followed with an increase of about 20% to EC2 Capacity Block pricing for machine-learning GPU instances from July 1, as Investing.com reported from AWS pricing documentation. Those increases came before the latest round of anxiety over memory costs. That tells you this isn't a one-time shock.
It's the cost curve now.
Nvidia reports fiscal second-quarter results on August 26, with recent Wall Street previews looking for about $92 billion in revenue against Nvidia's own guidance of $91 billion, plus or minus 2%. That number will get the headline treatment. The margin commentary deserves just as much attention. Investors have spent months asking whether AI capital spending has outrun what the technology can return, and higher rack prices make that question sharper, not softer.
SK Hynix is betting $29 billion that the AI memory boom is nowhere near over SK Hynix has announced plans to raise $29.4 billion via a record-breaking Nasdaq ADR listing targeting a July 10 debut, with proceeds funding new chip factories and EUV equipment tied directly to surging AI memory demand. The deal, which would eclipse Alibaba's 2014 record, reflects a company betting that high-bandwidth memory shortages are... - AI memory chip shortage solutions - SK Hynix Nasdaq listing AI demand
Frankly, this is the part of the AI buildout that gets less attention than it should. Everyone talks about GPU supply. Fewer people follow the memory sitting beside those GPUs - DRAM, NAND, high-bandwidth memory - even though it's shaping what hyperscalers pay and how quickly new capacity can come online. Nvidia didn't create the memory shortage on its own. But when Nvidia's platform sits at the center of the AI supply chain, every component shortage inside that platform becomes everyone else's problem.
Also read: OpenAI Joins Forces With Common Sense Media on a California AI Safety Measure • Nearly 6,000 CEOs Say AI Hasn't Boosted Productivity, Yet Layoffs Continue • Fortinet Buys Virtue AI Two Months After Meta Poached Its Founders