How much revenue does AI infrastructure need by 2030?

By Azeem Azhar, Exponential View.

This email is an excerpt from Exponential View's AI Investment Brief, which analyzes what level of revenue the AI infrastructure layer (hyperscalers and neoclouds) would need to justify the current pace of capital investment heading into 2030. Azhar poses the central question of the AI cycle: in his high-growth scenario, AI infrastructure would need $1.3 trillion in annual infrastructure revenue in 2030 to achieve a 15% return on invested capital (ROIC). For the customers of that infrastructure to earn gross margins of 30-50%, revenue at the 'downstream' layer would have to land between $1.9 trillion and $2.6 trillion.

The analysis starts from the context of hyperscaler CapEx, which has amounted to close to $2 trillion over the past five years. On top of that comes a new category of companies, the 'neoclouds', which are adding further capital spending. These companies' own CapEx guidance for 2026 adds up to $880 billion, and analyst projections point to more than $1 trillion a year from next year onwards, which would push hyperscalers' free cash flow into negative territory for the first time.

Azhar qualifies that headline as potentially misleading, however, since AI CapEx is only one part of these companies' total CapEx. Before the 'ChatGPT moment', these firms were spending $160 billion a year on capital expenditure. That is why, under his methodology (which strips out non-AI CapEx), they estimate AI-specific CapEx in 2026 at $560 billion, 64% of the total reported by hyperscalers. Analyst estimates for combined hyperscaler and neocloud total CapEx average around $1.07 trillion a year in 2027 and 2028.

From there, the report models three scenarios out to 2030: a low case, a mid case and a high case. In the mid, plateau case, annual AI CapEx would exceed $800 billion in 2030. In the low case, growth would peak in 2028, putting the 2030 figure at $600 billion. In the high case, with continued commitment to AI infrastructure, spending could top $1.9 trillion within four years. Azhar points to an earlier piece, 'Making sense of the AI CapEx logjam', to explain why much of the capacity under construction generates no revenue until it is operational, which means the investment bill and the productive asset base grow at different rates.

The analysis then focuses on the high-investment case. By 2030, they forecast $3.5 trillion of net AI capital 'in service' (that is, already operational, neither queued for construction nor depreciated). The cost of maintaining that capital breaks down into depreciation, operating expenses and financing costs. The largest component is compute (chip) depreciation, at $555 billion; hence the productive lifespan of AI chips will remain under scrutiny. The report models a current useful life of 5-6 years, up from 4-5 years just a few years ago. Depreciation of the 'shells' (buildings/physical infrastructure) and OpEx together add roughly another $216 billion to the required revenue threshold.

The third component is financing cost: a 10% weighted average cost of capital (WACC) on that in-service capital stock would require more than $350 billion a year simply to cover the cost of capital. On top of that, for every additional 5 points of required ROIC, hyperscalers and neoclouds would need an extra $177 billion in revenue. Adding it all up, in the high case Exponential View estimates an annual hurdle rate of $1.3 trillion for the infrastructure layer.

According to their own model of AI economy revenue, the infrastructure layer currently generates $176 billion in annualized revenue. Reaching the $1.3 trillion threshold would mean multiplying that figure by 7.4 times (a compound annual growth rate of 59%) from today's annualized revenue. For reference on the recent trajectory, growth was 500% in 2024 and 202% in 2025; for 2026 they expect annual growth of 187%, closing the year with an annualized growth rate of 135%.

Azhar adds a nuanced reading: these are early days in the cycle, and hyperscalers will likely accept lower margins —and therefore a lower hurdle rate— in order to keep growing, something they have already demonstrated by being willing to sacrifice cash flow for growth (he cites Google's cloud, which operated at negative margins until 2023). He also notes that each hyperscaler's recent growth is correlated with the increase in its CapEx, which is why he sees the surprise as more likely to come 'on the downside' — that is, with hyperscalers accepting lower revenue coverage than the model suggests.

Finally, he argues that if there are obstacles that could leave chips idle, infrastructure builders will be more willing to take on higher costs at the 'shell' and OpEx layers in exchange for greater certainty and faster time to service. In that context, he expects the winners to be builders with innovative solutions —such as modularized data centers— capable of bringing compute capacity into service sooner, even if that means higher construction or operating costs.

The email is a promotional excerpt from the full report, which is part of Exponential View's paid 'AI Investment Brief' service, and invites subscribers to read the full analysis and to join a conference call on Wednesday, September 30 with an exclusive launch offer.