AI's 'Power Wall' Day Bifurcated the Chip Trade — Memory Held, Power Equipment Cleared

Morgan Stanley's October 5 chip-supply-chain note did two things at once: it put a number on the power gap and it segmented the supply chain by exposure. The headline — a 34% net shortfall through 2028, or 32 GW of missing capacity even after behind-the-meter generation and fuel cells — reframes a problem investors had been calling 'AI capex sustainability' as a concrete binding on the buildout.

The day's closing prices draw a clean line through the AI compute supply chain. Memory and server integrators with order books tied to active builds held the bid; power-equipment makers and the capex-paying hyperscalers got the multiple compression.

Micron Technology closed at $1,086.47, up 3.9% on the day, riding a fiscal2026 revenue ramp from $13.6B in Q1 to $54.2B in Q4 — a 4x arc inside one fiscal year that left trailing gross margin at 80.7%. The point isn't that memory is immune to a slowdown; it's that every shipped HBM stack still needs DRAM regardless of how many GW of gas turbines arrive on time. The Morgan Stanley note flagged memory as one of the most exposed components on the secondary side, but the tape isn't pricing that risk yet because HBM allocation is still sold out into 2027.

On the other side, GE Vernova dropped to $993.74 with trailing EV/EBITDA near 81x and a P/E around 29.5 — a multiple built on the assumption that every gigawatt of AI demand would translate into turbine, grid and Electrification backlog. The same logic now risks a haircut if data-center builds defer until power clears. Vertiv trades on a similar story for cooling and power-management hardware. The capex-paying hyperscalers absorbed a smaller move, but the direction was the same: Meta Platforms -1.5%, Alphabet -0.5%.

Vistra closed up 3.5% to $166.12 — one of two AI-power names that actually gained on October 7. The reason sits in the company's 8-K and press release dated January 9, 2026: Vistra signed 20-year PPAs with Meta Platforms covering 2,609 MW of nuclear capacity from its Perry, Davis-Besse and Beaver Valley plants, plus 433 MW of uprates. Power that is already contracted to a hyperscaler, already sited on an operating nuclear plant and already cleared by PJM doesn't need to wait for the 32 GW to land.

Stijn Van Nieuwerburgh's BPEA paper 'Financing the AI Buildout,' posted by Brookings on September 23, 2026, gave investors the macro scaffolding that Morgan Stanley's note then activated. The paper estimates a 1 GW AI campus costs about $41B and projects a 183 GW U.S. buildout through 2032, totaling $10.3T of investment and averaging 3.6% of GDP annually. About 68% of that campus cost is IT and GPUs; roughly 32% is real estate and power infrastructure.

Days to quarters: the tape's message is that the market is now willing to pay a premium for power-locked order books and a discount for power-dependent multiples. Vistra, Constellation Energy and the nuclear-PPA cluster should continue to hold the bid as long as the 32 GW shortfall stays priced. GE Vernova and Vertiv stay range-bound until hyperscaler capex commentary confirms whether2027 orders slip or hold. NVIDIA, Broadcom and the HBM complex (Micron Technology, Super Micro Computer) sit in between — order books intact, but the secondary-component risk flagged by Morgan Stanley is the canary.

October 7, 2026 was the day the AI capex trade stopped hiding on capacity. The market finally had to acknowledge that the $700B+ 2026 capex envelope was no longer just a chip story — it was a power story, and the power isn't all there. What the tape produced was a clean two-factor sort: operators whose megawatt output is already under contract held the bid, and operators whose multiples assumed multi-year backlogs got marked down. The NVIDIA/Broadcom mega-cap insulation thesis held, but only because their chips actually ship and their order books already extend into 2027. The risk that the next 18 months will test is whether Morgan Stanley's 32 GW shortfall is a transient or a structural constraint — and the answer will separate the AI trade into two cohorts: those that own power and those that have to wait for it.