AI chip demand strains memory supply through 2027; Inversion eyes ASML with X-ray lithography

The week's dominant story was a memory supply crisis reshaping AI infrastructure economics. According to Counterpoint Research data, 2027 DRAM and HBM capacity is already fully allocated, with major buyers receiving only 60%–70% of their requested volumes and often paying deposits upfront. Adata Chairman Simon Chen estimated that HBM and AI servers could consume nearly 70% of DRAM capacity in 2027, while SK Group Chairman Chey Tae-won forecasted AI chip demand to rise 60%–100% that year. This tight supply has created an unexpected winner: Samsung and Micron, which derive most revenue from conventional DRAM and NAND, are capturing larger price gains than HBM specialists. Samsung's DRAM average selling price jumped 44%–46% and its NAND 67%–69%, while Micron's rose over 60% and 80%, respectively. SK Hynix, the HBM leader, posted smaller gains of about 30% and 55%. The competitive dynamic reflects a profitability squeeze in HBM itself: because DDR5 commands US$20/GB versus roughly US$12–US$16/GB for HBM3E, HBM's heavier wafer use is eroding its margin advantage. Three- to five-year agreements with more than ten major customers could temper ASP growth from 2H26 through 2027, further pressuring specialists. On the manufacturing front, Intel accelerated its Ohio fab build with overtime bonuses to keep its US$28 billion-plus complex on track for 2031. EMIB-T (a next-generation advanced packaging technology) yields reached 90%, though substrate yields remained near 50%; mass production is targeted for 2027 with potential costs 40–50% below TSMC's CoWoS offering. TSMC itself is developing an "EMIB-like" approach as AI demand strains its CoWoS capacity. In parallel, MediaTek confirmed that its second-generation data-center ASIC remains on schedule for 2028 mass production, with Intel's EMIB packaging already at strong yield levels. The company's first-generation chip is set to ramp in 4Q26 and generate US$2 billion in revenue. MediaTek also raised its 2027 AI-accelerator SAM (serviceable addressable market) to US$80 billion and its share target to 15%–20%, approving a US$5 billion flexible financing plan on top of more than US$7 billion in cash to secure capacity and support key suppliers. On lithography, Inversion Semiconductor unveiled a particle-accelerator platform designed to surpass ASML's EUV limits: a system 1,000 times smaller than conventional accelerators, capable of 500 million pulses per second, delivering twice the transistor density, and achieving triple wafer throughput. The US Department of Energy awarded Inversion US$750,000 for an AI digital twin developed with Lawrence Berkeley National Laboratory, on top of US$500,000 from Y Combinator and Entrepreneur First. Inversion targets a photomask metrology product within two years, though full commercialization would require competing against ASML's US$150–400 million tools and an ecosystem spanning TSMC, Intel, Samsung, and Zeiss. Finally, the Trump administration is weighing a ban on new Chinese optical transceiver models to extend AI restrictions into data-center networking. Zhongji Innolight, which holds about 27% of the global data-center optical module market and supplies Nvidia and Google, faces the greatest exposure. The move could open opportunities for Lumentum, Coherent, Applied Optoelectronics, Nokia, Ciena, and Cisco, but risks tighter supply, higher costs, and slower qualification as AI networks shift from 800G to 1.6T, potentially slowing deployment and raising infrastructure costs across the industry.