Morgan Stanley Raises Global AI Capex Forecasts to $1.2T/$1.4T for 2027/2028
Taylor Wilson
Morgan Stanley lifted its 2027/2028 capex forecasts for the five largest hyperscalers by 9%/10%, to $1.23 trillion and $1.396 trillion — driven by its view that AI compute remains supply-constrained, forcing top players to lock in resources early while unit build costs themselves climb roughly 20%.
How much money are we talking about — from $426B to $1.4T?
Combined capex path for the five hyperscalers: 2025 ~$426B → 2026 ~$779B → 2027 $1.23T → 2028 $1.396T.
This means → 2025-to-2027 is the steepest ramp, with spending nearly tripling; 2028 growth slows but the absolute number stays at record levels.
The biggest upward revisions go to Meta (2027/2028 capex raised 29%/22%) and Amazon (raised 15%/29%).
Why isn't this just "buying more GPUs"?
Morgan Stanley raised its per-GW build cost estimate by roughly 20% — the increase is not chips alone.
Costs rose on two layers: inside the rack — memory, CPUs, networking, and other bill-of-materials items; outside the rack — power equipment, building materials, skilled-labor shortages, and self-generation schemes pushing up power-shell costs.
In plain terms = a large share of the capex revision is "the same stuff got more expensive" — storage, networking, power, cooling, and engineering services are all sharing in this cost inflation.
Compute capacity near 120 GW by 2028 — what does that mean?
Combined capacity for the five hyperscalers rises from ~30.5 GW in 2025 to ~116.6 GW in 2028 — nearly a 4× expansion.
Amazon leads at 35.8 GW, Google at 31.6 GW, and Meta ramps aggressively from 3.5 GW to 21.2 GW.
This means → 2028 alone adds 36.2 GW of new capacity, but chips, power, construction, engineering labor, community opposition, and political cycles are all stretching delivery timelines — some data centers may take three years from groundbreaking to go-live, which is exactly why top players are locking in capacity early.
Why are custom chips cheaper — and by how much?
Nvidia GB200 costs roughly $35B per GW to build out, GB300 ~$39B, next-gen Vera Rubin ~$49B, Rubin Ultra ~$50B.
Custom silicon comparison: Google TPUv7 ~$27B per GW, Amazon Trainium3 ~$21B per GW.
In plain terms = building 1 GW of compute on custom chips is 25%-40% cheaper than on Nvidia's latest GPUs — the core economic case for Google's and Amazon's custom-chip programs.
Why is Meta Morgan Stanley's top pick — and what's the risk?
The core thesis: the market has priced in Meta's capex burden but has not fully valued multiple AI revenue optionalities.
Five potential revenue options: Neocloud compute monetization, Meta AI search, API revenue, subscriptions, and advertising upside — if several land, incremental EPS could approach $10.
But the cost is equally clear: the model adds $40B in new debt, cutting 2027/2028 EPS by 3%/7%; 2028 depreciation is projected at $103.8B, and free cash flow runs at -$29.8B in 2027.
What is the market validating next?
Amazon, Google, and Meta currently trade at roughly 16-20× 2028 EPS — the market has accepted the "sustained AI infrastructure spending" narrative.
This means → valuations have already digested the "spending" side; the next key validation is whether this capex converts into sustainable AI revenue.
Put simply = whoever proves first that "the money spent can be earned back" gets valuation support; those who cannot will face multiple compression even if revenue keeps growing.
Content is for reference only, not financial advice.