Morgan Stanley: over half of AI GPU servers shipped 2026-2028 may lack power to plug in, as GPU rental rates stay flat
2026-09-29
Ed Zitron's September 29, 2026 newsletter reports that Fidelity's Director of Global Macro, Jurien Timmer, called the AI trade "dead money for more than three months," citing flat-to-down token expenditure and flat-to-down H100/A100 GPU rental rates. Per the Financial Times' Bryce Elder, Morgan Stanley estimated that more than half of GPU servers sold between 2026 and 2028 might not have anywhere to be plugged in, and Jefferies described the gap between planned capacity and physical execution as the central issue, citing power, labor, transformers, cooling, and backup generation constraints.
Significance 3: A named-analyst estimate (Morgan Stanley, via the Financial Times) that over half of near-term GPU shipments may lack usable power is a concrete, attributed data point extending an already-tracked capex-execution-gap pattern; held at 3 rather than 4 because it arrives via Zitron's critical framing of secondhand FT/Morgan Stanley/Jefferies reporting rather than the primary research note.
Implications · machine-drafted, not owner judgment
If accurate, a structural power/execution bottleneck this large means a meaningful share of already-committed hyperscaler AI capex can't translate into usable capacity on the timeline investors are pricing in — extending, with a specific number attached, the same supply-side gap already visible in the Oracle Stargate force-majeure and Crusoe turbine-abandonment items this system tracks. For the owner, it's another reminder that 'AI buildout is unstoppable' vendor framing deserves the same scrutiny as any other vendor claim not allowed to say 'don't buy.'
- Whether Morgan Stanley or another bank publishes the underlying research note directly
- Hyperscaler capex guidance in the next earnings cycle addressing power/execution constraints explicitly
Sources
- curator Dead Money retrieved 2026-09-29