Morgan Stanley: AI Sector Pullback Is Technical, Not Fundamental
0xBroomberg
Morgan Stanley's 120-page July 27 report concludes the recent global AI selloff was driven by crowded positioning and deleveraging, not a breakdown in fundamentals, and quantitatively rebuts the market's three key concerns.
Companies are capping AI token budgets — does that mean demand has peaked?
Morgan Stanley calls this concern unfounded: a single enterprise AI call saves roughly $55 in labor costs, while an Agent-driven task costs just $2–5 — an ROI above 10×.
This means → at 10× ROI, adoption is no longer a budget decision — it is a competitive-survival decision.
On the supply side, the firm's Intelligence Factory model shows Blackwell-based data-center token margins at about 58%; Rubin and Feynman generations lift that to roughly 80% and 90%, respectively.
In plain terms = hyperscalers can cut token prices by about 75% without sacrificing margins — lower prices and higher profits can coexist.
China can train competitive models at lower cost — is US capex overkill?
The concern was sparked by the release of Kimi K3: if compute efficiency is rising fast, is the US hyperscalers' trillion-dollar-plus annual AI capex headed for an ROI haircut?
Morgan Stanley invokes the Jevons Paradox — efficiency gains cut unit cost, which unlocks more use cases, more users, and higher call frequency, ultimately raising total compute consumption rather than reducing it.
A single data point captures the supply-demand gap: Google executives say the company may need to double compute every six months — a 1,000× increase over five years. NVIDIA's 2025–28 AI chip sales CAGR is roughly 140%, yet extrapolated over five years that covers less than 10% of Google's demand projection alone.
This reflects a supply-demand mismatch far wider than the market currently prices.
Power shortages — speed bump or dead end?
Morgan Stanley frames the physical constraint as "3P": People, Power, and Politics. Grid-interconnection queues in some regions have stretched to 5–7 years; the House is reviewing a bill to codify hyperscalers' voluntary power-cost sharing commitments.
The firm classifies these as "speed bumps," not structural barriers, and points to on-site self-generation as the core solution.
Its estimate: US data-center power demand runs to about 68 GW from 2026–28; subtracting 15 GW under construction and 15 GW of contracted grid capacity leaves a potential gap of 38 GW.
This means → "Time to Power" — how fast a site can get electricity — is the most mispriced variable in the market right now.
Bitcoin mines to data centers — the undervalued "Powered Shell"?
The report identifies two de-bottlenecking paths: Bitcoin mines already hold substantial grid capacity and land, convertible to data-center use for a potential 10–19 GW; gas turbines and fuel cells can add 20–28 GW, arriving one to three years faster than new grid connections.
Even after probability-weighting, the base case still shows a ~1 GW net gap; the bear case widens it to 11 GW.
Morgan Stanley argues that what hyperscalers lack most is not capex but "Powered Shell" — physical space that already has electricity.
What are these transition companies worth?
Current Powered Shell providers trade at an EV/Watt — enterprise value divided by available power capacity, measuring "how much each watt of electricity is worth" — of just $2–4. Names include TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, and Galaxy Digital.
Mature data-center operators such as Equinix and Digital Realty trade at 20–25× EV/Watt. Morgan Stanley assigns these transition companies a discounted target of 15× EV/Watt.
This means → if the valuation gap narrows, the upside is several multiples — but only if the power shortfall is filled within the expected window.
Content is for reference only, not financial advice.