UBS Sees $4.1 Trillion AI Infrastructure Buildout Through 2028

UBS projects hyperscaler capital expenditures will total about $4.1 trillion from 2026 through 2028, a spending wave that would put annual outlays at $1.619 trillion by the final year. The wager is not simply on more chips and data centers; it is on enough durable AI demand to keep an enormous new base of computing capacity productive.

The spending curve keeps rising as growth moderates

The forecast moves from $492 billion in total hyperscaler capex in 2025 to $1.009 trillion in 2026, $1.447 trillion in 2027 and $1.619 trillion in 2028. That three-year projection exceeds the roughly $1.292 trillion UBS estimates hyperscalers spent over the preceding six years.

The important distinction is between the rate of increase and the level of spending. UBS expects the annual total to continue climbing through 2028 even as the pace of growth slows, leaving the industry with a much larger recurring capital requirement if its forecast proves right.

Cloud revenue is being recycled into capacity

For Amazon, Alphabet and Microsoft, UBS projects 2026 capex equal to 102% of their combined cloud revenue. The ratio is forecast to ease to 99% in 2027 and 94% in 2028, but that decline does not indicate shrinking investment: the projected dollar totals still rise each year.

That comparison is a measure of investment intensity, not a claim that the three companies are spending more cash than their entire businesses generate. Their operations extend beyond cloud infrastructure. Still, it shows how central AI capacity construction has become to the use of cloud-generated revenue.

UBS’s largest projected spenders, 2026–2028

• Alphabet: about $938 billion.
• Meta: about $683 billion; Microsoft: about $672 billion; Amazon: about $628 billion.
• SpaceX: about $335 billion; Oracle: about $276 billion; CoreWeave: about $130 billion; Nebius: about $93 billion.

The commitment can arrive before the asset does

Reported capex is only one lens on the buildout. A Wall Street Journal analysis, based mainly on filings current through June 2026, estimated roughly $3 trillion in off-balance-sheet obligations across nine large technology companies: Microsoft, Amazon, Alphabet, Meta, Oracle, Nvidia, Broadcom, AMD and SpaceX.

The estimate consists of about $1.9 trillion in purchase commitments and $1.2 trillion in signed leases that have not yet begun. Such arrangements need not appear as balance-sheet liabilities before the underlying asset or service goes live, so they can make the future financial load harder to compare across companies.

Disclosure is uneven. Nvidia quantified $30 billion in cloud-service commitments and Oracle disclosed $10 billion, while Meta did not quantify the cloud-capacity share of its reported $238 billion in commitments. That leaves meaningful uncertainty around company-by-company comparisons, even where the aggregate direction is clear.

Utilization is the return-on-capital question

The projected buildout spreads beyond the largest cloud platforms to neocloud providers, Oracle, SpaceX and newer entrants. That broadens the potential market for chips, networking, power equipment and data-center operations, but it also expands the volume of capacity that must find paying workloads.

The eventual constraint is economic rather than architectural: capacity must be utilized, customers must pay for AI services, and revenue must cover depreciation, electricity, financing and operating costs. One projection cited alongside the commitment estimates puts cumulative depreciation at more than $520 billion over three years for Microsoft, Oracle, Meta and Alphabet as infrastructure comes online.

UBS’s numbers frame the next phase of the AI race: less a question of whether companies will build, and more whether the resulting capacity can sustain returns before depreciation and committed payments become a drag on cash flow.