AINEWS
Memory (SK Hynix), advanced chips (Intel), and hyperscale data centers (Meta) all committed major investment in one week, sharing the same demand signal.
The AI buildout is no longer a story about GPUs. Across a single news cycle, the loudest capex announcements came from memory, foundry, and data-center operators: SK Hynix ordering equipment for a new Korean fab called Y1, Intel committing roughly $5.7B (€5B) to expand its Leixlip campus in Ireland. Read together, the three moves say the same thing: the chokepoint in AI infrastructure has moved upstream of the GPU.
SK Hynix has begun equipment orders for the Y1 fab at its Yongin cluster, with initial capacity of 20,000 wafers per month and the first cleanroom's fit-out pulled forward by roughly three months. English-language coverage from Korea JoongAng Daily, Cleanroom Technology, and The Economic Times all point to a single driver: AI demand has run past what existing memory lines can ship, so the new fab has to come online faster. The same pressure shows up across the broader Korean memory market, where Samsung and SK have continued to expand capacity even as spot prices for HBM (high-bandwidth memory, the DRAM stacked onto accelerators) and high-density DRAM stay elevated. Memory has been the segment with the tightest supply in AI training racks, and the ramp at the existing fabs has not kept pace with accelerator shipments, which is what makes a new fab like Y1 the operative move.
Editor's note: The 36kr Chinese-language digest for this section cites "2026年5月提前至2026年2月" (May 2026 to February 2026), dropping a "7" in both date references. SK Hynix's Feb 25, 2026 newsroom statement and all five independent English-language Korean press sources — Korea JoongAng Daily, Cleanroom Technology, The Economic Times (manufacturing), TheLec, and Dong-A — describe the acceleration as May 2027 to February 2027. The draft follows the English-press framing consistent with SK Hynix's own corporate disclosure.
Intel's announced €5B investment in Leixlip upgrades Fab 34 with Intel 3 process tooling for Xeon 6 and next-generation server processors. The bet has the same shape as SK Hynix's: the bottleneck is no longer at the GPU rack. It is in the surrounding compute, memory bandwidth, and packaging capacity that has to keep up with HBM and accelerator volumes. Foundry capacity at Intel 3 specifically is one of the tightest in advanced packaging, and the Leixlip upgrade lets Intel route more server-class silicon to the same customers already buying its foundry output for AI training systems.
Meta's Hyperion super-cluster in Richland Parish, Louisiana, is now planned at 5GW with total investment projected above $50B, part of a multi-year race among Meta, Microsoft, Google, and Amazon to bank AI compute against a single demand signal. Each of those hyperscalers has, by now, anchored its own data-center buildout to AI training and inference workloads, and each one is hitting the same constraint: the power grid and the memory supply chain cannot scale at GPU-purchase speed.
SoftBank founder Masayoshi Son said that, by 2040, the world will need $5T per year in AI infrastructure spending across data centers, power, and humanoid robotics, framing the spend as underwriting a shift toward an "ASI-driven" human-centered work model, according to a 36kr digest of the remarks. The figure is a 15-year forward projection, not a near-term forecast, and is best read as Son's framing of the buildout's ceiling rather than its base case. That SoftBank, one of the more aggressive backers of AI companies, is publicly anchoring a 15-year demand curve on the same constraint is, in itself, a marker of how the discussion has shifted from chips to infrastructure.
These four moves are not a coordinated plan. They are the same structural pressure showing up in four different stacks: the AI workload has outgrown the parts that scale easily (GPUs, model weights, software) and is now dragging the slower parts (memory fabs, advanced packaging, hyperscale power) into the same acceleration curve. The capex announcements land in the same news cycle because they are responding to the same pressure, and the pressure is now visible in fab timelines, capex guidance, and power-purchase agreements rather than in GPU allocation alone.
SK Hynix pulled the Y1 cleanroom fit-out forward by three months; Intel's Leixlip upgrade depends on EUV and Intel 3 yield; Meta's Hyperion requires a power grid that does not yet exist at 5GW. If any of those slips, the constraint shifts downstream, and the next capex wave, at the power and grid layer, becomes the story to track.