The AI Soft Landing Scenario

The AI Soft Landing Scenario

key points

For three years the AI build-out has blown through every forecast placed in front of it, with spending growth overshooting consensus so badly that investors, chipmakers and datacenter builders were repeatedly caught flat-footed, scrambling for capacity and components they had not budgeted for. The numbers for 2027 and 2028 suggest that run of upside surprise is ending — gently, if investors are lucky.

Wall Street's base case has hyperscalers' AI infrastructure spending roughly doubling in 2026 to about $800bn, then rising a further 54 per cent to $1.2tn in 2027, before decelerating sharply to just 12 per cent growth, or $1.4tn, in 2028. The rate of change, not the spending itself, is the story: capex is decelerating from exponential to merely large.

That matters most for the "neoclouds," GPU-cloud specialists that lack hyperscaler-grade balance sheets and have leaned on high yield debt to fund buildouts. Bonds issued at a high-single-digit coupon by one prominent neocloud now yield close to 15 per cent, with its credit default swaps trading at distressed levels and the stress rippling across sector equities. Investment grade issuers borrow in an entirely different universe, and that gap could widen into something structural: a well-capitalized operator may find it cheaper to acquire a cash-strained rival than keep underbidding it, seeding consolidation among the weakest high yield names.

Slower neocloud growth also has a hardware angle. 2026's great surprise was the memory squeeze: DRAM contract prices rose roughly 90-95 per cent quarter-on-quarter in early 2026, and spot commodity memory chips went from about $12 to a record $42.50, as datacenters competed with laptops and phones for scarce fabrication capacity. A capex slowdown should let that supply chain exhale.

The rate path is the hinge. Markets currently price at least 50 basis points of further hikes before policymakers contemplate easing, with projections holding rates near 4.5 per cent through 2027 rather than cutting. The optimistic logic: cooler component demand eases a genuine contributor to goods inflation, giving the Fed room to resume cutting in the second half of 2027 into 2028, once those hikes are banked. Lower rates compress credit spreads, easing refinancing pressure on the neoclouds that need it most — and narrowing the window in which they get bought rather than refinanced.

None of this requires the AI trade to end, only to mature. A deceleration from doubling to low double-digit growth sounds dull next to 2026's fireworks, but a maturing cycle should also be a more predictable one, with outcomes landing closer to initial expectations rather than blowing through them. After a year of surprises, markets would welcome nothing more than being right the first time. Investors should position for deceleration, not denial — and for a widening gap between cheap and expensive funders that consolidation, not just credit markets, may ultimately close.

— Peter Wilke, CFA – Head of Tactical Asset Allocation, Global Asset Allocation

See more: Is AI Crowding Out the Rest of the Economy?