Memory used in AI data centers has become so scarce that Samsung and SK Hynix hold less than 10 days of inventory, according to industry tracking cited in a September 11 briefing, and Nvidia says memory pricing has turned extreme and is headed even higher into next year.
The shortage sits at the center of the AI build-out. High-bandwidth memory, the stacked DRAM that sits next to GPUs in AI accelerators, has absorbed production capacity that would normally serve PCs, phones and servers. As hyperscalers race to expand, the conventional memory market is left fighting for whatever remains.
Nvidia flagged the squeeze directly. The company told investors that memory pricing has turned extreme and that its own price increases are already executed, according to Yahoo Finance. Nvidia’s latest results showed revenue above $96 billion, with memory commitments soaring to $160 billion, a signal that the chipmaker is locking in supply years ahead rather than buying quarter to quarter.
Ten days of stock
Inventory below 10 days leaves almost no buffer. If a fab hiccups or demand spikes, buyers face allocation rather than spot purchases. Samsung and SK Hynix have both shifted production toward HBM4, the newest generation of stacked memory, which consumes more wafer capacity per gigabyte than standard DRAM. Kioxia, a major NAND flash maker, separately denied reports of joint production talks with SK Hynix this week while seeking price stability to sustain AI investment.
The demand side shows no sign of slowing. TSMC posted record monthly revenue in August, up more than 53 percent year over year, and held a 72.5 percent share of the foundry market in the second quarter, driven by advanced AI processors. Broadcom’s AI semiconductor revenue rose 221 percent to $16.7 billion in its latest quarter, and the company introduced a new Ethernet platform to support large-scale AI networks. Nvidia reported a 27 percent increase in Grace Blackwell shipments, and AMD lifted its 2027 data center outlook toward $70 billion in AI sales, citing a total addressable market its chief financial officer says could reach $3 trillion.
OpenAI has moved to secure its own supply. The company confirmed it is working with Samsung on next-generation chips, with general manager of OpenAI Korea Harrison Kim telling Reuters the two companies have made significant progress on production and research, though he did not specify Samsung’s exact role. Samsung’s sixth-generation HBM4 is reportedly designated for OpenAI’s Jalapeño inference accelerator, developed with Broadcom and taped out at TSMC. The relationship dates to an October 2025 letter of intent under which Samsung estimated OpenAI’s memory demand could eventually reach 900,000 DRAM wafers per month, an extraordinary figure against current industry output.
Anthropic is following a similar path, with reports that the Claude maker is in talks with Samsung for 2nm foundry capacity and advanced packaging while building an in-house custom chip team.
Consumers start paying
The shortage is already reaching retail shelves. Apple chief executive Tim Cook told the Wall Street Journal in June that price increases on its products are unavoidable because of rising memory and storage costs, saying the situation has become unsustainable. The company’s first foldable iPhone went on sale this month at $1,999, a price shaped in part by component costs, and preorders opened Friday as the foldable race with Huawei’s Mate XT2 heats up.
PC builders face the same math. Random-access memory prices have climbed through the summer, and IFA 2026 coverage in the trade press repeatedly flagged the RAM shortage as the story of the show. Builders who spec’d systems in the spring are paying noticeably more for the same memory today.
The memory makers are responding with capacity, but slowly. Micron has leaned into AI memory, and analysts at Morningstar argue the strategy could pay off for investors. New fab capacity takes years to qualify and ramp, which means the tight conditions look set to persist through 2027 even if demand cools. Samsung has also revealed a three-phase HBM roadmap, dubbed zHBM, that would stack DRAM directly on top of the processor, a sign the suppliers are racing to differentiate rather than simply add volume.
What it means for the AI race
Memory has become a strategic input on par with GPUs. The companies that control HBM supply, effectively Samsung, SK Hynix and Micron, now sit in the middle of every major AI deal, and their allocation decisions shape who can build what. Nvidia’s $160 billion in memory commitments and OpenAI’s wafer reservation estimates show buyers are no longer waiting for the spot market.
Wall Street has noticed. Intel shares jumped 9 percent on reports of Nvidia’s $30 billion investment and the pricing environment, and Broadcom forecast AI chip revenue reaching $230 billion by fiscal 2028 on the back of customer deployments. The money flowing into memory and custom silicon is starting to rival the GPU market itself.
For everyone else, the arithmetic is less friendly. Data center operators, PC makers and phone vendors will pay more for memory through next year, and some of that cost lands in consumer prices. Smaller AI startups without committed supply face real allocation risk, which could consolidate the field further toward the hyperscalers who signed early.
The AI boom is often described as a compute story. Right now it is just as much a memory story, and the shelf is empty. When Nvidia, the largest buyer in the market, says pricing is extreme and headed higher, the rest of the industry should plan accordingly.