AI spending binge leaves hyperscalers exposed

Wall Street’s rush to treat hyperscalers as the safest way to profit from artificial intelligence is running into a harder question: whether all that spending will generate returns fast enough to justify the cost. Microsoft, Amazon, Alphabet and Meta are pushing capital expenditure to extraordinary levels as they race to build data centres, secure chips and expand cloud capacity, but investors are facing a business model that demands relentless outlay long before the payoff is fully visible.

That strain is becoming more visible in market pricing. Reuters reported on March 30 that the top five big technology groups are expected to spend about $630 billion in 2026 and $800 billion in 2027, with nearly 70 per cent of operating cash flow already being absorbed by capital expenditure. Analysts cited by Reuters expect almost 90 per cent of operating cash flow to be consumed over two years, while projected debt issuance by hyperscalers has been lifted to $175 billion in 2026 from $121 billion in 2025. The message for investors is straightforward: even companies with formidable balance sheets are no longer funding AI expansion out of effortless surplus.

Alphabet’s latest guidance underlined how quickly the numbers are moving. Reuters reported on February 4 that the Google parent projected $175 billion to $185 billion of capital spending in 2026, far above expectations, as it sought to ease cloud capacity constraints and defend its position in generative AI. Amazon went further, projecting roughly $200 billion of capital expenditure for the year, up from $131 billion in 2025, a scale that unsettled investors enough to send its shares sharply lower after the announcement. Meta, meanwhile, said in January that it expected 2026 capital expenditure of $115 billion to $135 billion, a 73 per cent jump tied to Mark Zuckerberg’s push for what he calls superintelligence.

None of that means demand for AI infrastructure is imaginary. Cloud growth remains strong, and companies continue to report that enterprise customers want more computing power than providers can readily supply. Yet the investment case is becoming less clean than the market assumed during the early phase of the boom. Investors buying hyperscalers today are not simply backing AI adoption; they are also accepting execution risk in power procurement, construction, hardware delivery, labour, financing and pricing. Reuters Breakingviews argued last week that the immediate problem may not be a collapse in demand but the difficulty of translating giant budgets into functioning data centres on time and at acceptable cost.

Energy is another fault line. Reuters, citing S&P Global, reported on March 31 that planned AI infrastructure spending of $635 billion this year faces a serious test from higher energy prices and geopolitical instability. Data centres are becoming increasingly exposed to electricity costs and supply bottlenecks at the same time as oil and gas markets have turned less predictable. Reuters also reported in March that the Electric Power Research Institute expects data centres to account for 9 per cent of US electricity demand by 2030, up from 4 per cent in 2025. That creates a second layer of uncertainty for equity holders: the economics of AI are no longer just about software demand, but about physical infrastructure in an energy-constrained world.

There is also the issue of where value will settle. Hyperscalers have the scale to finance the build-out, but scale does not guarantee the best shareholder outcome when an industry is still in its capital-hungry phase. Heavy spending can support revenue growth while depressing free cash flow, squeezing margins and making companies more sensitive to bond yields. Reuters noted that the Roundhill Magnificent Seven ETF has fallen about 20 per cent from its October high, reflecting a market that is no longer willing to cheer every extra billion spent on AI. Microsoft has already faced pressure to show returns on its cloud and AI investments after slower cloud growth coincided with record expenditure.

For investors, that points to a more selective approach. The strongest gains in an infrastructure cycle do not always accrue to the largest builders. Suppliers with pricing power, software groups with lighter capital needs, and specialised firms that monetise AI demand without carrying the full weight of hyperscale construction may offer cleaner exposure. Hyperscalers are unlikely to disappear as AI winners, but right now they look less like effortless growth machines and more like utilities in the middle of an expensive build-out.