Global Thematics note, November 8, 2026: Morgan Stanley's post-election update on AI capex, data center builds and its "Buy the Drop" stock list

Morgan Stanley thinks the market has badly overreacted to the US midterms. In a cross-team thematic note published on Sunday, the bank argues that the election results will delay part of the American data center build-out by roughly a year rather than shrink it, and that the resulting sell-off in AI-related stocks has created buying opportunities across chips, power and software. The headline numbers are unchanged: hyperscaler capital spending of about $1.6 trillion in 2027 and $1.8 trillion in 2028. What changes is the timing, and Morgan Stanley's conclusion that a slower, longer cycle tilts the field even further toward the largest, best-capitalized companies.

The note lands after a brutal week. Data center builders and power-equipment names fell by as much as 24% in the session after the vote, with Vistra down 14% and Vertiv down 13%, after Texas elected a new governor who campaigned against the data center build-out and local voters rejected data center projects in 40 of 41 ballot measures, more than half of them in Ohio. Morgan Stanley calls the move an "earthquake event" in market terms, but sees it as a re-timing story.

5 GW slides from 2028 to 2029, and 2027 is untouched

The core forecast change is narrow. Morgan Stanley assumes newly elected governors will not interfere with projects already under construction, so 2027 completions stay where they were. The damage falls on 2028, when the bank expects some Texas and Ohio projects to stall while governors and legislators write new requirements for developers. It moves 5 GW of US projects from 2028 into 2029, which it describes as about 20% of US projects and 10% of global projects planned for that year.

The new US path is a plateau followed by a jump. Morgan Stanley now models 20.7 GW of new US critical IT capacity in 2027, 19.9 GW in 2028 and 27.6 GW in 2029. The 2028 figure is 4.9 GW, or 19.8%, below its prior estimate of 24.9 GW, while 2029 rises by the same 4.9 GW from 22.7 GW. That turns 2028 into a year of 4% decline in US additions, followed by 38% growth in 2029. Total US critical IT capacity reaches 85.3 GW in 2027, 105.3 GW in 2028 and 132.8 GW in 2029. Globally, the bank still sees data center additions rising 9% in 2028 and 20% in 2029, and it expects about 20 GW of capacity that cannot be built in the US to spill over into Asia and EMEA.

The bank's own power model shows why the delay matters. Its Powering AI analysis already showed a shortfall of more than 30% in powered data center capacity through 2028, and the post-election update raises the 2028 gap by 15% to 44 GW. Demand is not the constraint in Morgan Stanley's framework. The bottleneck is shifting to the powered shell, the building with grid or on-site power ready for servers.

Why Morgan Stanley calls it a shift rather than a cut

The bank's argument rests on one judgment: the requirements new governors are likely to impose, from community benefits to labor and grid-cost rules, are achievable for serious developers. It points to Louisiana as a state where data center development has been structured to work for local constituents, and it expects hyperscalers to redirect new builds toward friendlier states such as Tennessee and Louisiana. It also flags the possibility that some states keep issuing permits in exchange for a higher data center tax or a formal community benefits package. That would raise build costs and lower returns on capital, but Morgan Stanley argues scaled players such as Amazon, Alphabet and Meta can absorb it far more easily than smaller developers.

The bank is candid about how it could be wrong. Labor requirements tougher than expected, bans on on-site fossil fuel generation along the lines of Virginia's recent approach, and stricter emissions limits could all cut deployments below its projections rather than merely delay them. That risk is real. Morgan Stanley's 20% assumption sits well below other estimates of exposed capacity: Goldman Sachs estimates that states likely to add restrictions hold about 45% of planned US capacity, Bloomberg puts the share in states whose new governors want a pause at about 40%, and Wells Fargo estimates Texas alone holds about 27% of planned US gigawatts. Georgia's and Michigan's governors-elect have also signaled pauses or moratoria, which go beyond the Texas and Ohio cases the note focuses on. Morgan Stanley's forecast works only if most of those states end up with rules rather than bans.

Capex held at $1.6 trillion and $1.8 trillion, with a downward bias in 2028

Morgan Stanley keeps its hyperscaler capex forecasts unchanged for now. Its model has aggregate spending across eight companies rising from $481 billion in 2025 to $951 billion in 2026, $1.64 trillion in 2027 and $1.76 trillion in 2028. The reasoning is that the hyperscalers will keep building as much capacity as they can to scale products for billions of users, because large language models remain compute-hungry while still earning returns on invested capital above 30%.

The company-level split shows how concentrated the spending is. For 2027, the bank models $375 billion at Alphabet, $355 billion at Amazon, $269 billion at Microsoft, $236 billion at SpaceX, $225 billion at Meta, $97 billion at Oracle, $46 billion at CoreWeave and $37 billion at Nebius. For 2028, it has Alphabet at $400 billion, Amazon at $365 billion, Microsoft at $323 billion, Meta at $250 billion and SpaceX at $240 billion, while Oracle slips to $93 billion and CoreWeave and Nebius barely grow.

Two details deserve more attention than the headline. First, even before any election effect, growth was already set to slow sharply: from 98% in 2026 and 73% in 2027 to just 7% in 2028. Second, Morgan Stanley admits that delays in project approvals, on top of existing labor and materials constraints, put a downward bias on its 2028 number. For suppliers whose valuations assume continued double-digit growth in spending, a flat 2028 is a very different world from the one priced a month ago.

A longer cycle has benefits, and they accrue to whoever already has capacity

Morgan Stanley argues that stretching the build-out over more years is not all bad. A slower cycle means a smoother rebalancing of supply and demand along the data center supply chain, a better credit outlook, and higher lease rates for compute. The bank also expects a "high grading" of AI capex: when US power is scarce, the most creditworthy customers, mainly hyperscalers, are the ones that secure it. Capacity that is already built and running becomes more valuable, which again favors the largest platforms.

The same logic shapes its view on financing. Morgan Stanley still expects higher AI-related debt issuance in 2027, led by the hyperscalers, but says new construction loans and neoclouds will face more scrutiny. Tenant quality, a credible path to energization and contractual protections become more important. Projects already backing outstanding bonds look less exposed to new state restrictions, although execution risk remains, and slower competing development should support stabilized, cash-flowing assets. The bank prefers colocation over hyperscaler data center asset-backed securities.

Compute efficiency becomes the release valve

If power and buildings stay scarce for longer, the way to keep scaling AI products is to need less compute per user. Morgan Stanley says it is highly focused on large efficiency gains from the leading labs in 2027. It notes that Meta and Alphabet have both stressed the need for smaller models that are just as capable, with Meta targeting a 10-fold improvement in compute efficiency in 2027 after the high compute intensity of its Muse model. The gains come from distillation, more targeted pre- and post-training, better model architectures, curated training data and smarter inference routing.

Hardware is moving in the same direction. Alphabet's TPU 8t and 8i chips offer up to 2.7 times better training performance and 80% better inference price-performance, respectively, than the prior generation. Morgan Stanley's conclusion is that the efficiency race, like the power race, favors companies with the money and the willingness to build their own first-party models.

Bucket one: the hyperscalers and Nvidia

The bank's first and highest-conviction group is the hyperscalers, which it argues are advantaged by scale, access to capital and the ability to exploit advances in model technology, including recursive self-improvement, to generate more revenue and profit per watt of data center power. Morgan Stanley analyst Brian Nowak rates Amazon, Meta and Alphabet Overweight, Adam Wood rates Microsoft Overweight, and Adam Jonas rates SpaceX Overweight. Nvidia is not a hyperscaler, but Joe Moore also rates it Overweight because it drives the increase in tokens and revenue per watt that matters most in a power-constrained world.

Bucket two: a selective "buy the drop" in the supply chain

Morgan Stanley is notably more cautious on the upstream supply chain than its headline suggests. With 2028 growth now close to flat, it warns that suppliers could go through a classic down-cycle of slowing orders, margin normalization and earnings downgrades even if AI demand itself stays intact. Pricing power falls and inventory can build as the bottleneck moves to the powered shell. How hard each segment is hit depends on how undersupplied it was in 2027 and how much new supply arrives in 2028. The bank calls the resulting risk an air pocket in 2028, and its conclusion is that valuations already reflect much of it, with margin sustainability as the main risk.

Within that frame, the bank highlights five areas that have de-rated enough to look attractive. In data center real estate, it favors Equinix and the US powered shell providers. Among those, it sees TeraWulf, Riot Platforms, Hut 8 and Applied Digital as less exposed to Texas, while Cipher Mining, Galaxy Digital and MARA Holdings face a greater risk of delay or a moratorium.

In behind-the-meter power generation and independent power producers serving data center developers, Morgan Stanley stays positive on Bloom Energy, Solaris Energy Infrastructure, Liberty Energy, Vistra, Talen Energy, Constellation Energy and BorgWarner even if data center growth disappoints. It sees GE Vernova and INIO as more sensitive to oversupply risk, though at valuations that limit the downside. In energy, it names Williams, Clearway Energy, HA Sustainable Infrastructure, Ormat Technologies and Fervo across natural gas, renewables and geothermal.

The fourth area is international. Morgan Stanley sees developers in the fastest-growing data center markets, particularly in Asia-Pacific, as beneficiaries of the difficulty of building AI and power infrastructure in the US. It names Gulf Developments, Keppel, Adani Power, Yangtze Power, Samsung SDI and Hokkaido Electric as key winners.

In semiconductors and hardware, compute remains the highest-conviction bucket, with a preference for the largest, best-capitalized leaders: Nvidia, Broadcom, TSMC, MediaTek and Samsung. The bank also flags Cambricon and CXMT as potential share gainers if some token demand moves to Chinese cloud providers that have ample power. It is more selective on memory. In broader hardware, it prefers areas that have de-rated, including hard disk drives through Seagate and Western Digital, and optical networking through Lumentum, Coherent, Arista Networks and Keysight. The clearest warning is for Asian cloud chip suppliers to US hyperscalers, such as Aspeed and Montage, which the bank sees at risk over the next 2 to 4 quarters as investors focus on 2028. Its valuation work also shows server components looking the most stretched on a combination of valuations and profit margins versus history.

Bucket three: alpha moves to software and AI adopters

The bank's longer-term call is that leadership in AI stocks will rotate from chips and infrastructure toward software and the companies that use AI. It points to the mobile internet era as a template: between 2010 and 2016, chip designers such as Qualcomm and Arm led first, device and infrastructure makers such as Samsung and Apple took over next, and software and services platforms such as Google and Amazon delivered the strongest relative performance last. Morgan Stanley says it is now seeing value creation at the application layer, including productive enterprise use cases, and prefers early software enablers, particularly in cybersecurity and infrastructure software, alongside companies leading in AI adoption.

The analogy is useful but imperfect. In the mobile cycle, the infrastructure phase did not run into a political and physical cap on construction. If the US build-out is delayed for longer than Morgan Stanley expects, software companies benefit only if efficiency gains let AI products scale on the capacity that already exists, which is why the bank's emphasis on compute efficiency and its software call are really the same bet.

The weak points in the bull case

The note's strongest point is that the AI demand story did not change on election night. Its weakest point is that the call depends on a policy outcome that is still unknown. The new governors have not yet taken office, Morgan Stanley's 20% delay assumption is about half of what Goldman Sachs and Bloomberg consider exposed, and the bank concedes that bans on on-site gas generation or strict emissions rules would turn a delay into a reduction. Even in its base case, the supply chain faces a flat 2028, with the risk of a traditional down-cycle in orders and margins, while hyperscaler capex growth was already set to slow to single digits that year.

What Morgan Stanley gets right is the direction of the competitive effect. Scarce power, higher local costs, tighter financing for new projects and a premium on existing capacity all push in the same direction: toward Alphabet, Amazon, Meta and Microsoft, which can pay up, relocate and build their own models, and away from leveraged developers and suppliers that need the build-out to stay on its original schedule. The signals to watch are the first formal data center rules from the governors-elect in Texas, Ohio, Georgia and Michigan, Nvidia's results later this month, and whether hyperscalers start naming new sites in states such as Tennessee and Louisiana.