Editor's Review: What August Taught Us About Pricing the Shortage

July answered the question June raised — yes, the shortage is structural — and spent four weeks building contracts, power deals, and packaging investment around a 2028 timeline. August’s job was smaller but harder: put an actual number on each of those constraints, and in doing so, turn at least one of them from shared scarcity into open competition.

The memory story stopped being a two-horse race. China’s CXMT crossed a $500 billion market value in its Hong Kong listing, surpassing Tencent, on the strength of DRAM shipments into AI servers, cars, smartphones and PCs — and Apple was reported to have held early talks with CXMT about supplying components for devices sold in China, the clearest sign yet that a Chinese DRAM maker has moved from geopolitical hedge to commercial option. That didn’t stop the incumbents from betting on scarcity themselves: SK Hynix approved the largest share buyback-and-cancellation in South Korean corporate history — roughly KRW40 trillion, about US$29 billion — explicitly tied to confidence in a “long AI memory cycle,” while Winbond told investors its Q2 operating margin hit 48.4% and should clear 50% in Q3, backing a new Kaohsiung expansion aimed at supply contracts running through 2029 and 2030. The shortage didn’t get any less real in August. It got a serious third competitor.

Packaging kept widening from a chip-level problem into a systems-level one. Counterpoint told clients it expects more than 130 million GPUs and AI ASICs to shift onto advanced-packaged, on-compute memory over the next five years, calling advanced packaging the next real battleground in AI hardware — not just the link between chip and memory, but the link between everything. Morgan Stanley separately flagged that TSMC’s CoWoS allocation is set to shift in AMD’s favor in 2027 as agentic-AI workloads lift its processor demand relative to Nvidia’s. That same widening is what pushed this month’s Explainer on co-packaged optics: once packaging capacity is accepted as the ceiling on compute, the next chokepoint is the optical interconnect between racks — and Broadcom, Nvidia and Marvell are already shipping product against exactly that constraint.

Power stopped being a Korean energy-ministry story and became an architecture problem for everyone. At the OCP APAC summit in Taipei, Google told the industry outright that the 48V power architecture nearly every hyperscale data center runs on is running out of headroom, with converters and busbars eating an increasing share of rack space as AI racks head toward megawatt scale. July’s power story was about who controls electricity supply. August’s is about the physical limits of the wiring carrying it into the rack — a more specific, and more expensive, problem than the one the industry was solving a month ago.

The financing behind the buildout got more interesting than the headline totals suggest. Amazon raised its full-year capex guidance to US$220 billion and Meta’s AI-related lease commitments climbed to US$279 billion — numbers that would have been the story on their own a year ago. But the more telling figure was Nvidia’s: it is reportedly closing in on roughly US$100 billion in credit guarantees for OpenAI’s 10-gigawatt Ohio data center, down from an earlier US$250 billion plan. Foxconn, meanwhile, told investors its earnings are now growing faster than the capital its 2027 AI-server buildout requires, meaning no new equity raise is needed. Read together, that’s the capex story shifting from “spend whatever it takes” toward underwriting that actually prices risk — a more disciplined phase than the one the industry was in six months ago, even as the absolute dollars keep climbing.

Physical AI kept converting partnerships into operational timelines. LG’s robotics tie-up with Nvidia moved fast enough that one of Nvidia’s own product marketing directors for Omniverse and robotics was reportedly headed to South Korea to visit LG’s R&D sites within a week of the two companies signing their MOU. Hyundai said publicly it is betting on physical AI to evolve beyond automobiles, and China’s robotics IPO pipeline kept filling: Alibaba- and JD.com-backed LimX filed in Hong Kong, and Chery’s robotics unit, AiMOGA, began preparing its own listing. July’s signal — that physical AI had become standard capital allocation across Korean, Japanese and Taiwanese industrial players — held through August, with China’s robotics makers now turning up in IPO filings rather than lab demos.

The AI-safety story from July stopped being a disclosure and started being an oversight matter. A US House panel formally sought a briefing on OpenAI’s AI-agent security breach — the same incident a DigiTimes column later described as the first publicly disclosed case of an AI agent attacking outside infrastructure. That is a meaningful escalation in kind, not just attention: in July, two frontier labs disclosed their models behaving in ways nobody had planned for. In August, a legislature asked for the paperwork.

The through-line for August: every constraint the industry named in July got a number attached to it in August. Memory went from a duopoly under long-term contract to a three-way race with a $500 billion Chinese entrant. Packaging capacity acquired a forecast — 130 million GPUs and ASICs over five years — and a second front in optical interconnect. Power acquired an architectural expiration date, with Google saying plainly that 48V doesn’t scale to what’s coming next. And the capital behind all of it got a discipline check, with Nvidia’s guarantee for OpenAI shrinking even as Amazon’s and Meta’s headline numbers grew. June asked whether the shortage was structural. July answered yes and started building around it. August started pricing exactly what building around it will cost — and handed Congress the first real oversight question about whether the AI running underneath it can be trusted to behave.

— Colin Tan, Editor