Morgan Stanley Raises U.S. AI Power Gap Estimate to 38GW, Names Nvidia as Top Semiconductor Pick
nashnova research
Morgan Stanley raised its 2026–2028 U.S. data-center net power gap from 33GW to 38GW, warning that roughly a third of the shortfall is not priced in; Nvidia and Broadcom remain its top semiconductor picks, while ASICs face share pressure in a power-constrained world.
How did they arrive at 38GW?
The report stacks the gap in three layers. Layer one — the "raw gap": U.S. data-center power demand of 116GW, minus capacity under construction and available grid, leaves 70GW.
Layer two factors in five "time-to-power" solutions. Low / mid / high scenarios yield 60 / 43 / 23GW; Morgan Stanley calls the mid case — 43GW — the most likely outcome.
Layer three adds Asia-Pacific overflow, landing at 38GW — the new headline number.
This means → even after every plausible fix is counted, more than 30% of U.S. data-center power demand remains unmet.
Why did demand jump from 97GW to 116GW?
Three model revisions drove the leap: higher utility-power estimates for chips already purchased — now including storage, networking, and other rack-level components.
Historical non-generative-AI data-center power was also revised up, and a 75% load factor was properly incorporated into the methodology.
In plain terms = the old model counted GPU power alone; the new one adds the full ring of "supporting electricity" around each chip, so the total naturally jumped.
Five gap-closing solutions — which works, which hits a wall?
Behind-the-meter turbines and generators carry the highest success probability (90%) and can close 22GW at mid-case — but the bottleneck is labor, not hardware capacity. The existing workforce supports only 10–20GW of new gas capacity per year.
Asia-Pacific overflow is estimated at 6GW for 2026–2028. Asia has roughly 30GW of spare power, but grid constraints limit near-term serviceability — having power is not the same as being able to connect it.
Grid-level peak shaving is ruled out: utilities manage grids for worst-case conditions and will not raise data-center interconnection limits based on demand-response promises.
This reflects a ceiling set not by technology but by labor and grid access — two physical constraints that money alone cannot quickly solve.
Why is Vera Rubin's rack power pegged at 234kW?
Morgan Stanley assumes Vera Rubin NVL72 runs each GPU at 2,300W thermal design power. Seventy-two GPUs total 165.6kW; add 68.4kW for CPUs, networking, and other rack components — roughly 234kW per rack.
Four reference points back the figure: Nvidia's own roadmap (reaffirmed at OCP last week) marks NVL72 at ~240kW; MGX 1.2 supports up to 250kW; Schneider Electric quotes 227–246kW; Pegatron specs its NVL72 Max-P at 228kW.
In plain terms = one rack draws as much power as roughly 80 average U.S. households — that is why the power gap keeps widening.
Which chips win? Why might ASICs lose ground?
Top picks are Nvidia (NVDA.US) and Broadcom (AVGO.US), on conservative guidance, chip-deployment visibility, and Nvidia's higher tokens per watt.
In a power-constrained environment, ASICs — chips custom-built for a single customer — deliver fewer tokens per watt and are less fungible across clients, putting them at a structural disadvantage versus GPUs.
This means → when power is scarce, whichever chip produces more results per kilowatt-hour wins the order — and that is Nvidia's GPU today.
If construction pauses, who gets hit first?
The "bullwhip effect" — demand swings amplified down the supply chain — hits the tail end hardest: memory, optical modules, and other downstream components whose pricing is currently propped up by shortage.
In a bear case, power-management and analog chips may be over-ordered, concentrating risk in the same segment.
Beneficiaries of "time-to-power" acceleration include behind-the-meter suppliers Bloom Energy (BE.US) and Solaris (SEI.US), plus energized-site operators TeraWulf (WULF.US), Hut 8 (HUT.US), and others.
Disclosure note: Morgan Stanley reports business ties with many companies named above, holds >1% of common stock in Nvidia, Broadcom, Bloom Energy and others, and has managed public offerings for several — a potential conflict of interest.
市场有风险，内容仅供研究参考，不构成投资建议。