The artificial intelligence industry has spent the past two years in a frenzied race to build ever-larger data centers, secure ever-more powerful chips, and train ever-more capable models. But according to Demis Hassabis, the Nobel Prize-winning head of Google DeepMind, the next bottleneck won’t be processors or energy—it will be memory. And the shortage, he warns, could arrive as soon as next year. In a striking admission that underscores the physical constraints bearing down on the AI boom, Hassabis told attendees at a recent event that high-bandwidth memory (HBM)—the specialized, ultra-fast memory chips stacked onto AI accelerators like Nvidia’s GPUs—is poised to become the scarcest and most fought-over resource in the global technology supply chain. Hassabis described memory as “the bottleneck” for the industry heading into 2026, a constraint that could slow the pace of AI development even as demand for more powerful systems accelerates. ## A Bottleneck That Money Alone Cannot Solve The warning from Hassabis carries particular weight given his position. As the co-founder of DeepMind and now the leader of Google’s combined AI research efforts, he sits at the nexus of some of the most ambitious AI projects on the planet. Google has committed tens of billions of dollars to AI infrastructure spending, and Hassabis has direct visibility into the supply chain dynamics that will determine whether those investments translate into actual computing capacity. High-bandwidth memory is not a commodity product that can be quickly scaled up. It is manufactured by a small number of companies—primarily SK Hynix, Samsung, and Micron—using advanced packaging techniques that stack multiple layers of DRAM chips vertically and connect them with thousands of tiny through-silicon vias. The manufacturing process is complex, yields are not always predictable, and expanding production capacity requires years of lead time and billions of dollars in capital expenditure. SK Hynix, the dominant supplier, has been running near full capacity to meet orders from Nvidia, AMD, and other chip designers, and even with aggressive expansion plans, the gap between supply and demand is expected to widen. ## Why Memory Has Become the Chokepoint To understand why memory has emerged as the critical constraint, it helps to understand how modern AI hardware works. The massive neural networks that power systems like Google’s Gemini, OpenAI’s GPT series, and Anthropic’s Claude require not just raw computational power but also the ability to move enormous volumes of data in and out of processors at extraordinary speeds. HBM serves as the bridge between the processor and the data it needs to operate on. Without sufficient memory bandwidth and capacity, even the most powerful GPU sits idle, waiting for data. As AI models have grown from billions to hundreds of billions—and now potentially trillions—of parameters, the memory demands have scaled accordingly. Each new generation of Nvidia’s data center GPUs has required more HBM. The H100, which became the most sought-after chip in the world in 2023, uses 80 gigabytes of HBM3. Its successor, the H200, bumped that to 141 gigabytes of HBM3e. The forthcoming Blackwell B200 will require even more. Multiply those figures by the tens of thousands of GPUs that major AI companies are deploying in single clusters, and the aggregate demand for HBM becomes staggering. ## The Supply Chain Is Already Under Strain Hassabis’s comments reflect a growing consensus among industry leaders that the AI supply chain is approaching a period of acute stress. The concern is not hypothetical. SK Hynix has already sold out its HBM production for 2025, and customers are now negotiating for 2026 allocations. Samsung, which has struggled with yield issues on its own HBM3e products, has been working to qualify its chips with Nvidia but has faced delays. Micron, the third major player, has been ramping production but remains a smaller contributor to the overall HBM market. The concentration of manufacturing capacity in a handful of South Korean and American companies introduces geopolitical risk as well. The memory industry’s dependence on facilities in South Korea—where both SK Hynix and Samsung are headquartered—means that any disruption, whether from natural disaster, geopolitical tension on the Korean Peninsula, or trade restrictions, could have outsized effects on the global AI buildout. ## Big Tech’s Spending Spree Meets Physical Reality The memory shortage warning arrives at a moment when the world’s largest technology companies are pouring unprecedented sums into AI infrastructure. Microsoft, Google, Amazon, and Meta have collectively committed well over $200 billion in capital expenditure for 2025 alone, much of it directed toward data centers packed with AI accelerators. But spending money is only useful if the hardware can actually be procured. If HBM supply cannot keep pace with demand, these companies may find themselves with half-built data centers and unfilled server racks—or, more likely, engaged in an increasingly aggressive competition for limited supply. This dynamic has already played out with GPUs themselves. In 2023 and 2024, Nvidia’s data center chips were so scarce that companies paid premiums, signed long-term contracts, and even acquired smaller firms partly to gain access to their GPU allocations. The memory shortage threatens to replay that scenario at a different point in the supply chain, with potentially even fewer options for workarounds. ## What the Memory Makers Are Doing About It The memory manufacturers are well aware of the demand surge and are responding with aggressive expansion plans. SK Hynix has announced investments of over $75 billion through the end of the decade to expand its HBM and advanced DRAM production capacity, including new facilities in South Korea and the United States. Samsung has reorganized its memory division and is prioritizing HBM yield improvements. Micron is expanding its facility in Boise, Idaho, and has secured CHIPS Act funding to support domestic production. But building semiconductor fabrication and advanced packaging facilities takes time—typically three to five years from groundbreaking to volume production. The investments being made today will not meaningfully increase supply until 2027 or 2028 at the earliest. ## A Sobering Signal From One of AI’s Most Prominent Voices Hassabis’s public acknowledgment of the memory constraint is notable for its candor. Technology executives typically prefer to project confidence about their ability to execute on ambitious plans. For the head of Google DeepMind to identify a specific, near-term supply chain risk suggests that the problem is serious enough that it cannot be managed quietly behind the scenes. The AI industry has so far powered through every bottleneck—first compute, then GPUs, then power and data centers. Memory may be the first one that truly slows the scaling laws that have defined the past decade of progress.