AI infrastructure spending has entered a new phase, says GlobalData, with demand outpacing semiconductors, memory, and power capacity despite record earnings across the technology ecosystem.
This earnings season reveals that capital investment, geopolitics, and supply-chain resilience – not compute demand alone – are increasingly determining the pace of AI deployment and reshaping competitive dynamics across the global semiconductor value chain.
An analysis of GlobalData’s Company Reports Database reveals that SK Hynix’s net income jumped 397,5% year-on-year, Samsung’s 486,7% – both riding an HBM shortage that has sold out DRAM through 2026. Micron posted a staggering 1 398,3% net income surge on 345,7% revenue growth, confirming that AI accelerators are as memory-hungry as they are compute-hungry, with supply still lagging.
Nvidia’s net income more than tripled (up 210,6%) on data-centre demand, though operating cash flow growth of 83,6% trailed capex growth – down 70,9% QoQ but up 43,2% YoY – showing how much cash is being redirected into supply-chain commitments, including securing extra TSMC capacity for H200 production.
Murthy Grandhi, company profiles analyst at GlobalData, comments: “TSMC’s 77,4% net income growth against a 33,8% YoY capex rise reflects its race to add advanced-node capacity, which is running short of AI demand. Constellation Energy’s 1 247,5% net income growth underscores that power, not chips, is now the binding constraint – a point hyperscalers have made explicitly on recent earning calls.”
Two outliers break from the boom.
Intel’s net loss narrowed 278,1% YoY on a genuine turnaround: its 18A node hit high-volume manufacturing in January 2026 with yields improving about 7% monthly; its 9,9% CHIPS Act equity stake has appreciated on paper; and Nvidia, Microsoft, Amazon and Apple have all opened foundry talks – though Intel does not expect meaningful foundry revenue before 2027.
Marvell’s net income fell 80,6% YoY despite 27,6% revenue growth, showing custom-silicon suppliers can see margins squeezed by ramp costs even amid strong AI demand.
Hyperscalers’ capex is the real story.
Microsoft, Alphabet, and Meta each posted capex growth above 65% YoY, matching public guidance: Alphabet has raised 2026 spending to $175-billion to $190-billion; Meta to $125-billion to $145-billion; Microsoft toward $120-billion to $190-billion. Combined, the four hyperscalers are on pace for roughly $700-billion to $725-billion in 2026 capex – up about 77% from 2025 and the largest peacetime capital cycle in corporate history, funded even as free cash flow tightens.
“Geopolitics is now embedded in these numbers,” continues Grandhi. “In January 2026, the Trump administration invoked Section 232 to impose a 25% tariff on select advanced AI chips, with a two-phase plan that could raise rates further while rewarding domestic production. Separately, the Bureau of Industry and Security shifted H200 and MI325X export reviews for China from presumptive denial to case-by-case approval, capped near 50% of prior US sales – a partial reopening that still left buyers like ByteDance, Alibaba, and Tencent able to secure only a fraction of the over 2-million H200s they’d ordered. TSMC, Samsung, and SK Hynix lost blanket Validated End-User exemptions on 1 January 2026 and now need annual US licences to ship equipment into their China fabs.”
Layered atop this is Pax Silica, the US-led pact signed by two dozen nations in December 2025 to lock allied countries – including India, Japan, South Korea and the Philippines – into a trusted semiconductor and critical-minerals chain outside China’s orbit.
India is a direct beneficiary. Micron’s Sanand, Gujarat assembly-and-test plant, inaugurated February 2026, is now packaging DRAM and NAND for Micron’s global supply chain – an India-origin option arriving just as DRAM prices rose roughly 90% in Q1 2026.
The Tata Electronics–PSMC wafer fab at Dholera has passed the halfway mark, targeting trial production by December 2026 at the 28nm node – mature by global standards, but enough to cut India’s reliance on imported logic and power chips.
Says Grandhi: “GlobalData sees none of this capex as optional now: HBM and advanced-node capacity are sold out; tariffs and export licensing are permanent cost variables; and Pax Silica signals self-reliance blocs – not free trade – will govern chip supply chains for the rest of the decade. The risk isn’t slowing AI demand, but power, memory and licensing constraints capping deployment speed.”