Compute & Market Power
Data Center Capex Needs 32% Growth to Hit $3T
Dell'Oro doubled its 2030 data center capex forecast to $3 trillion. Reaching it from 2026's $1 trillion pace needs 31.6% compound growth, every year.
Dell’Oro Group has nearly doubled its outlook for global data center capital spending, and the revision carries a growth rate worth stating plainly. Worldwide data center capex will surpass $3 trillion by 2030, according to Dell’Oro’s August 18 announcement, with research vice president Baron Fung noting the 2030 outlook “has nearly doubled since the January 2026 forecast, reflecting higher hyperscale capex guidance, increased projections for global data center power capacity, and higher commodity costs.”
Set that against the firm’s own current-year figure. Dell’Oro’s data center capex research program describes the market as “on pace to exceed $1 trillion in 2026.” Compounding $1 trillion to $3 trillion over the four years from 2026 to 2030 requires 31.6% annual growth, sustained every year — no pause, no digestion phase, no quarter where a hyperscaler trims guidance. That is the arithmetic the forecast is asking the industry to deliver, and it is a far more demanding statement than the headline number implies.
Where the $3 trillion goes, and what has to hold
Composition matters more than the total. Fung told RCR Wireless that accelerated servers with AI accelerators account for roughly 40% of the projected $3 trillion — about $1.2 trillion of accelerator spending, a figure that flows almost entirely to a handful of chip suppliers and their custom-silicon competitors. The newly added AI-specialized cloud segment, covering model builders and neoclouds, is projected to grow at nearly a 60% compound annual rate, outpacing every other customer group in the report.
Two other lines in the forecast deserve attention because they signal where spending broadens. General-purpose server demand is expected to benefit from growing inference, agentic, and storage workloads, which means the buildout is no longer purely a training story: serving agents at scale pulls conventional compute and storage along with accelerators. And the revision itself incorporates higher commodity costs, meaning some of the doubling is price rather than capacity — a distinction that matters enormously to anyone forecasting how many usable megawatts the money actually buys.
The constraint is not demand and not capital. Fung identified power availability as the biggest limit on the industry’s ability to sustain expansion, with grid expansion and onsite generation becoming increasingly important. That matches what the buildout looks like on the ground, where developers are contracting for generation years ahead of racks — the pattern behind ONEOK turning AI power demand into gas capex and the grid-behavior questions regulators raised in NERC’s assessment of data center load.
Note also what the forecast implies about the near term relative to today’s spending. S&P Global estimates Alphabet, Amazon, Meta, Microsoft, and Oracle will collectively spend roughly $750 billion in capital expenditures in 2026, equal to about 38% of their combined revenue, per Motley Fool’s summary of the estimate. Fung expects the top four US hyperscalers alone to represent about half of global capex. Both statements can be true only if the non-hyperscale half — neoclouds, colocation, telco, enterprise, sovereign programs — grows at least as fast as the giants, which is the least proven part of the forecast.
The operator implication is contract timing, not stock picking
Most readers of this paper do not allocate capex; they buy capacity from people who do. The useful reading is about lead times and pricing power.
If accelerator spending really runs toward $1.2 trillion by 2030 while power remains the binding constraint, then the scarce good is not the chip but the energized, cooled, networked site to put it in. That inverts the negotiation for anyone signing multi-year compute agreements: the vendor’s flexibility on price will track its site pipeline, not its silicon allocation. Buyers should be asking suppliers for interconnection dates and generation contracts, not GPU counts — and should expect the spread between delivery-ready and speculative capacity to widen, which is precisely what makes compute futures a hedging instrument worth pricing.
The forecast’s fragility is worth naming. A 31.6% compound rate has no room for a demand pause, and Fung explicitly flags that growth depends on “the sustainability of investment, power availability, and supply chain conditions,” while noting enterprise investment “remains constrained by uncertain AI returns.” Enterprise hesitancy is the crack in the wall: hyperscalers and model builders can fund each other for a while, but the segment expected to absorb general-purpose servers and storage is the one waiting for proof. If enterprise capex stays flat, the forecast leans harder on neoclouds whose own funding is increasingly leveraged, a dependency visible in today’s brief on the debt stack behind frontier chip supply.
What would change the verdict: two consecutive quarters of hyperscaler capex guidance below consensus, or a public retreat on power procurement in a major market. Either would break the compounding, and a forecast that requires uninterrupted 31.6% growth does not degrade gracefully. Until then, treat $3 trillion as the industry’s stated intention rather than its schedule — and price your own capacity contracts against the constraint that actually binds, which today’s operators are meeting in substations rather than in fabs. Teams weighing whether to own or rent that capability at all will find the same calculus in today’s lead on the cost of training a model in-house.