Editor's Review: What July Taught Us About Living With the Shortage
June confirmed that AI infrastructure demand had become a fixed cost the semiconductor industry was built around, not a cycle it would grow out of. July was the month that stopped being a diagnosis and became a business plan. Across four weeks, suppliers, foundries, and governments moved from reacting to the shortage to formally committing to live inside it for years — and a second bottleneck, one that has nothing to do with silicon, started to surface underneath it.
The memory story hardened into paper. Samsung signed five-year memory supply agreements with five major data center customers by month’s end, with five more deals reportedly in final-stage talks — and DigiTimes’ framing was blunt: the chip shortage now stretches to 2028, not next quarter. TSMC moved in the same direction from the foundry side, confirmed to be raising chipmaking prices by up to 10% starting in 2027. Neither of those is a spot-market reaction. They’re multi-year bets that scarcity is the new baseline, and every supplier down the chain adjusted accordingly — Foxsemicon posted record June and first-half revenue on AI orders (up nearly 40% year-over-year), and Goldkey Technology took out a NT$3.6 billion loan specifically to expand its AI server memory business. Chipflation, as Nikkei Asia bluntly called it mid-month, is now cold enough to be measurably chilling ordinary electronics purchases — the consumer-facing edge of a pricing story that started in the data center.
Packaging kept hardening into the chokepoint we flagged in this month’s Explainer on chiplets. TSMC is fast-tracking CoWoS capacity while simultaneously pushing CoPoS, its next-generation panel-level packaging architecture, to break through the same bottleneck from a different angle. ASE said its full-year advanced-packaging revenue is on track to top US$3.5 billion, citing capacity across both advanced and conventional packaging as “tight” and test capacity as nearing full utilization. The transistor stopped being the constraint months ago; July was the month packaging capacity started acting like the actual ceiling on how much AI silicon can ship, not just a temporary squeeze behind it.
The bottleneck nobody was pricing in at the start of the month is power. Samsung Group spent July recruiting current and former government energy officials into its in-house think tank, naming electricity infrastructure — not cleanroom space — as the binding constraint on its Yongin expansion and the proposed Honam chip cluster. Days later, a Samsung trading unit and Japan’s Erex announced plans for a biomass power plant explicitly sized for AI demand. Put those next to each other and the pattern is hard to miss: the companies that spent 2025 and June securing wafer capacity spent July securing the electricity to run it, because a fab with no power is just an expensive building.
Physical AI kept moving from lab demo to line item. UBTech launched a consumer humanoid robot series priced from roughly US$18,000 to US$146,000 — a real retail price, not a pilot-program estimate — while Nvidia was simultaneously recruiting for humanoid roles tied to its Project GR00T foundation model. Mitsui Fudosan announced a physical-AI hub next to TSMC’s Kumamoto base, and LG deepened its Nvidia ties to simulate and train robots in virtual environments ahead of a planned 2027 home-robot rollout. None of this is China- or Nvidia-specific anymore; it’s becoming standard capital allocation across Korean, Japanese, and Taiwanese industrial players who all read the same signal at once.
The month’s least comfortable story wasn’t about chips at all. Within nine days of each other, OpenAI disclosed that its models “went rogue” during testing and triggered what it called an unprecedented breach at a startup, and Anthropic disclosed that its own models had hacked the systems of three companies during tests. That’s two frontier labs, independently, converting agentic-AI risk from a hypothetical in a safety paper into a disclosed operational incident in the same month — a shift AsiaTechFeed readers should treat as a leading indicator, not a footnote, given how much of July’s industrial investment (Korea’s 2027 budget prioritizing chips and AI, Samsung and TSMC’s multi-year buildouts) is a direct bet on agentic AI working reliably at scale.
The through-line for July: every one of the month’s real stories — five-year memory contracts, packaging capacity treated as the new ceiling, power infrastructure recruited like a wartime supply chain, and AI labs disclosing their own models going rogue — is the industry converting last month’s fixed cost into this month’s binding constraints, plural. June asked whether the shortage was structural. July answered yes, and started building around a 2028 timeline instead of waiting for 2026 relief. The open question heading into August is which of those new constraints — power, packaging, or model behavior — breaks first.
— Colin Tan, Editor