$950 billion. That is the value of semiconductor and AI infrastructure commitments signed at the San Francisco AI Summit on July 25, where South Korea's three largest conglomerates, Samsung Electronics, SK Group, and Hyundai, locked in long-term supply and construction agreements with Broadcom, Nvidia, Microsoft, and Anthropic. The number is not a government pledge. It is contractual capex locked against real delivery schedules.
The Memory Bandwidth Choke Point
SK Hynix sits at the center of this deal. The Korean chipmaker produces the majority of the world's high-bandwidth memory (HBM): the specialized DRAM stacks that sit directly on AI accelerator dies and move data at speeds standard DDR memory cannot approach. As model sizes scale and inference workloads compound, memory bandwidth, not raw compute, has become the binding constraint.
The deal effectively locks AI chip supply through 2030. That timeline matters. HBM production cycles run 18 to 24 months. Fab expansions committed today determine AI training capacity in 2027 and 2028. Samsung and SK Hynix are betting that AI accelerator demand will not slow. So are Broadcom, Nvidia, and Anthropic. The combined $950 billion gives that bet physical form.
The bearish case is straightforward: long-cycle infrastructure bets have historically misfired when demand forecasts overshoot. Semiconductor over-ordering contributed to the 2022 chip glut. If AI inference scaling hits an efficiency wall before 2030, these commitments become stranded assets.
The Power Equation
One week before the summit, Fluidstack, the AI infrastructure startup building Anthropic's physical data center network, closed an $830 million Series A at a $7.5 billion valuation. Leopold Aschenbrenner's Situational Awareness fund is among the institutional backers. Fluidstack holds a $50 billion supply agreement with Anthropic and is expanding data center capacity across three continents.
The power math is stark. The International Energy Agency projects global data centers will consume over 1,000 TWh in 2026, equivalent to Japan's entire annual electricity demand. Morgan Stanley estimates a 49-gigawatt power deficit in U.S. grid capacity by 2028. Individual hyperscale AI data centers now draw 100 to 750 megawatts per site.
Bitcoin Mining's Grid Dilemma
Bitcoin miners were the original industrial power consumers. Today they compete for the same interconnection queues, the same utility contracts, and the same grid expansion corridors as AI data centers. CleanSpark currently holds 1.8 gigawatts of power capacity, a meaningful position. But bidding competition from hyperscalers with trillion-dollar balance sheets introduces structural cost pressure that Bitcoin mining companies did not face three years ago.
The pressure is not symmetrical. AI developers need stable, always-on power and will pay premium rates to secure it. Bitcoin miners have always operated as marginal load, arbitraging low-cost surplus power. Capital flooding into AI infrastructure now competes directly for that surplus, raising the floor price for miners.
Bitcoin's hard money thesis holds regardless of who builds the grid. But the miners who survive the next difficulty cycle will be the ones with the lowest locked-in electricity costs. The $950 billion AI build-out makes cheap power scarcer. Not more abundant.
Follow @UTXOMacro for daily Bitcoin, macro, and AI breakdowns.