Big Tech keeps spending after DeepSeek

Meta raised its 2025 capex forecast to $60–$65 billion in late January, and Microsoft said annual infrastructure spending would approach $80 billion. DeepSeek had just knocked a hole in AI valuations. The biggest technology companies used their earnings calls to say, without much hedging, that the buildout would carry on regardless.

A cheap model, a costly Monday

The scare arrived quickly. DeepSeek, a Chinese lab backed by the quantitative hedge fund High-Flyer, released its R1 reasoning model on January 20th, and its engineers had already said the final training run behind the underlying V3 model cost less than $6 million in rented computing time. On January 27th Nvidia's shares fell by about 17 per cent, the largest one-day loss of market value any American company had suffered, and power producers and networking suppliers were dragged down with it. The timing was awkward: only a week earlier the White House had hosted the launch of Stargate, a venture of OpenAI, SoftBank and Oracle promising up to $500 billion of data-centre investment.

The shock owed as much to where the model came from as to what it cost. American export controls had barred Nvidia's best chips from China, and DeepSeek said it had trained V3 on H800s, a pared-back part Nvidia had designed for the Chinese market before the rules were tightened again. The cost figure also left a good deal out. It covered one final training run, not the earlier experiments, the salaries or the hardware the lab owns, so the true bill was certainly larger. Markets did not wait for the footnotes, and DeepSeek's chatbot climbed to the top of Apple's American app store over the same weekend.

Cheaper queries, more servers

Hyperscaler bosses read DeepSeek as a case of the Jevons Paradox. If better software cuts the marginal cost of running an inference query by an order of magnitude, they argue, enterprise demand will grow far faster than it shrinks. Cheap intelligence lets more companies deploy applications, which should mean more servers, more networking gear and more dedicated power from the utilities. Satya Nadella, Microsoft's chief executive, invoked Jevons by name on social media on the day of the sell-off, and the earnings calls that followed repeated the argument at length.

The idea is old. William Stanley Jevons observed in 1865 that more efficient steam engines had raised Britain's consumption of coal, because cheap power found uses that dear power never could. The logic holds only when demand is elastic: a tenfold fall in the price of a query must bring more than a tenfold rise in queries before total spending on compute goes up. For chatbots sold at a flat monthly fee that is doubtful. For software agents that fire off thousands of queries to finish a single task, the hyperscalers have a better case.

Where the money lands

The spending flows down a familiar chain. A large slice goes to Nvidia for accelerators, and the rest is shared among makers of networking equipment and custom chips, builders of data-centre shells and the utilities that must connect them. Electricity has become the scarcest input. Microsoft signed a 20-year agreement last September to buy the output of a Three Mile Island reactor that Constellation Energy plans to restart, a sign that the largest buyers now secure power years before the chips are delivered.

DeepSeek changes the mix within that chain more than its size. Google, Amazon, Microsoft and Meta all design their own accelerators, with help from firms such as Broadcom and Marvell, and a world in which cheaper, smaller models do most of the work favours those in-house chips over Nvidia's premium parts. Inference also rewards capacity close to customers rather than a handful of giant training clusters. The builders of shells, the suppliers of cooling and the utilities get paid either way, which is why the sell-off in power stocks looked overdone to the companies actually signing the contracts.

Buybacks pay the bill

The bet is not free. Pouring most of their operating cash flow into assets that depreciate quickly leaves less for share buybacks and drags down return on invested capital for now. Depreciation is the slower-acting cost. A data-centre building lasts decades, but the accelerators inside it are overtaken by new chip generations every year or two, so much of this year's capex becomes a charge against profits within a few years, long after the decision to spend has been taken.

Accounting softens the blow for a while. Several of the hyperscalers have lengthened the useful lives they assume for servers in recent years, which spreads the charge over more years and flatters reported margins. If the economic life of an accelerator is closer to the cadence of Nvidia's product launches than to those book lives, earnings are overstated today and a write-down or a quicker depreciation schedule awaits. Investors watching free cash flow rather than earnings per share will see the cost first.

Echoes of the fibre glut

The nearest precedent is not encouraging. In the late 1990s telecoms carriers laid far more fibre-optic cable than traffic could fill, on the reasoning that internet demand would keep compounding. Demand did grow enormously, but only after bandwidth prices had collapsed and Global Crossing had gone bankrupt. The difference now is who pays. The carriers borrowed heavily, whereas Meta, Microsoft and their peers are funding the buildout mostly from cash flow, so an overshoot would dent shareholder returns rather than threaten solvency.

The optimists have a fair point of their own. Unlike the fibre boom, today's capacity is being used as soon as it is switched on, and Microsoft has said that shortages of data-centre space, rather than a lack of customers, have held back its cloud growth. Idle capacity would be the warning sign, and so far there is little of it.

Open weights, thinner margins

The strongest objection concerns who captures the gains, rather than whether demand grows. DeepSeek released R1 under a permissive open-source licence, so any company can run a capable reasoning model without paying a frontier lab. If Jevons holds, query volumes rise, but the price of each answer falls towards the cost of electricity and hardware. The hyperscalers would still sell the computing, as landlords do, while the premium they hoped to earn on their own models shrinks. That is a business with utility-like returns attached to technology-like depreciation, which is a worse combination than the capex guides imply.

What would change the call

Alphabet and Amazon report in the coming days, and a cut to either spending plan would show that the Jevons argument is not universally held. The hourly price that cloud providers charge to rent accelerators is the best running gauge: if it falls faster than usage grows, the paradox is failing. Nvidia's own results, expected in late February, will show whether its largest customers are ordering more chips or merely squeezing more work out of the ones they already have.

None of this changes the incentives at the top of these companies, where the cost of missing the next platform looks larger than the cost of idle servers. Big Tech would rather overbuild than fall behind, and shareholders will probably fund that choice through thinner buybacks and weaker returns on capital for the next year or two, whatever DeepSeek does to the price of inference.