Morgan Stanley keeps its $1.6tn AI spending forecast

Featured photograph: Jenix89, 2007, public domain. Compressed for the web and cropped for display.

Morgan Stanley is maintaining its forecast of roughly $1.6 trillion in hyperscaler capital expenditure for 2027, even as it expects some US data-centre construction to move back a year. The bank’s post-midterm argument is that a slower buildout can extend the AI investment cycle without ending it.

Excerpts of its 8 November report, Buy the Drop: AI Investment Opportunities Post the US Midterms, were shared by investor Jonah Lupton. Zhitong’s account of the report also gives an unchanged forecast of about $1.8 trillion for 2028. These are analysts’ projections, not newly announced spending commitments from the companies involved.

The change is in the construction calendar

The excerpts describe a partial shift in building activity from 2028 to 2029. Zhitong reports that Morgan Stanley models around five gigawatts of IT capacity moving back: roughly one-fifth of the US buildout previously expected for 2028, or 10% globally. The bank expects projects already under construction to be less exposed to the new approval risks.

That distinction separates the amount companies want to invest from the date a facility can start serving customers. Buying accelerators, financing a campus and obtaining an energised grid connection are different steps. A delay at the last step can change when equipment is delivered, when rental income begins and how long a developer must finance unfinished capacity.

Morgan Stanley’s public commentary supports this interpretation. In a 6 November podcast, its global head of fixed-income research, Andrew Sheets, said the election did not change the bank’s broad AI view. He argued that a longer buildout could ease concerns about financing supply and mismatches between demand and capacity.

The accounting perimeter matters

The $1.6 trillion headline also needs a definition. In an October briefing, Hamilton Lane reproduced Morgan Stanley’s earlier estimate for investment by hyperscalers and neocloud customers. That calculation added $195 billion of lease-financed investment in 2027 and $246 billion in 2028 to headline capital expenditure.

A rented data centre can supply computing capacity without appearing as an equivalent cash capital purchase in the tenant’s accounts. Comparing a lease-inclusive forecast with a simple sum of companies’ reported cash expenditure can therefore produce a misleading gap.

The publicly available excerpts and reporting on the November note do not reproduce its complete accounting model. The earlier figures establish that this forecast has used a broader perimeter; they do not establish that every assumption or company estimate remained identical in the new report.

Watch delivery, not just the spending total

The bank’s thesis still has a practical test. Zhitong reports that wider approval delays could put its 2028 expenditure forecast under pressure. A postponement can preserve an intended project while changing its cost, financing needs and economic return.

These constraints already have concrete counterparts. Sulat’s review of data-centre approvals in Southeast Asia found that announced capacity and power allocations did not remove the need for environmental and building permissions. The public records behind OnePWR’s generator-platform sale show another route companies are exploring to obtain power, with deployment details still unresolved.

The useful measures now are projects receiving permits, securing finance and connecting to electricity on their revised schedules. Morgan Stanley’s forecast describes the scale of the ambition. Those milestones determine how much usable computing capacity the spending ultimately buys.

Reporting note: Sulat reviewed the shared report excerpts, Zhitong’s account and Morgan Stanley’s public commentary. The full report was not available for inspection.