For two years the argument about artificial intelligence has been about whether too much money is being spent on it. Larry Fink used a stage in Berlin on Wednesday to make the opposite case: that too little is being spent, and that the shortfall will arrive as higher prices.
The BlackRock chief executive told the Berlin Global Dialogue that delays in building data centers raise the cost of computing, and that expensive computing works through an economy like a tax on everything else. “The faster we can build out more data centers, the more compute we have, the faster we can bring down the cost,” he said in an interview with Bloomberg Television at the conference.
His framing moves the risk to an unfamiliar place. Investors spent the autumn marking down companies that borrow against computing contracts, on the theory that supply is running ahead of demand. Fink said the shortage of capacity, not the surplus, is what should worry them. The scarce inputs in his account are not chips but electricity, land, permits, transformers and the crews that install them.
The macro claim underneath is simple enough. Fink put technology capital spending at roughly two thirds of all capital expenditure in the American economy, counting cloud capacity and the buildings that house it. If that share of investment is being slowed by queues and lawsuits, the cost of the output rises, and the cost of the output is now embedded in the price of most services.
He is not a disinterested observer. BlackRock closed its $12.5 billion purchase of Global Infrastructure Partners in 2024, and the vehicle it formed with Microsoft, MGX and Nvidia agreed last year to buy Aligned Data Centers for about $40 billion. The firm earns fees on infrastructure assets and has told clients that private markets are where the computing buildout will be financed. When Fink says the buildout is too slow, he is describing a portfolio as well as an economy.
The bottlenecks he named are the ones operators complain about in private. Grid interconnection is the first. A data center can be designed, financed and built in eighteen months; the substation and generating capacity that feed it are booked years out, and in the largest American markets the queue for a new connection has run to a median of about five years, according to grid operators and the analysts who track the queues.
Community opposition is the second. County boards that once competed for the property tax revenue of a warehouse now hear about water consumption, diesel generators and the effect on residential electricity bills. Several large projects have been withdrawn after zoning hearings, and the pattern has moved from exurban Virginia to the Midwest and to Europe, where the same objections arrive in different languages.
Financing is the third. Much of the capacity under construction is leased to companies that do not carry investment-grade credit, and the debt is priced against a policy rate that has not fallen as fast as the industry assumed. A project that penciled at six percent does not pencil at eight, and the difference is paid in delay.
There is already a public example of what delay costs. Nextdc, the Australian operator, cut its guidance after power delivery slipped at sites it had leased and pre-sold. The company had customers, capital and land. What it did not have was a connection.
Fink’s argument has a weak point, and it is the one his critics raise. Compute demand is not a fixed quantity that either gets served or does not. If capacity is short and prices high, buyers change how they use models: smaller models, more efficient inference, batch processing moved to wherever power is cheap. Efficiency gains of the kind Nvidia advertises, up to 30 percent less power for the same number of tokens, reduce the shortage that Fink describes even as they increase the number of tokens the world can afford.
The other objection is the one markets have been making. Graphics processors lose value quickly, and a facility built for one generation of hardware is worth less when the next generation ships. A company that cannot get power for five years is not necessarily a company that should get power; it may be a company whose business model will not survive the wait.
What would settle the argument is not a speech but a lead time. If transformers can be delivered in eighteen months instead of four years, and if a transmission line can be permitted in three years instead of ten, the constraint moves and the shortage eases. If lead times stay where they are while orders keep climbing, the industry is promising capacity it cannot energize.
Europe is the test case Fink did not dwell on. The continent has the engineering base and, in the Nordic countries, the power. It also has permitting rules that make a five-year American queue look brisk, and a politics in which electricity prices are a household issue rather than an industrial one.
His own institution has spent the past two years warning about the other side of this trade: that a handful of companies now account for an outsized share of global equity value, and that the market is paying for a buildout whose returns are unproven. Both warnings can be true. Capital can be misallocated and insufficient at the same time, in different places, at different points in the chain.
The immediate question for the companies building is narrower. Every quarter that a project sits in an interconnection queue is a quarter of depreciation with no revenue, and the lenders who financed it are counting the same days. Fink’s contribution this week was to say out loud that the schedule, not the technology, is the binding constraint. The grids will decide whether he was early or wrong.
