Why the AI Race Has Grown Too Big to Stop

The United States is placing an AI wager so large that slowing down now can look more dangerous than pushing ahead. In the second quarter of 2026, Microsoft, Google, Amazon and Meta together spent about 165 billion dollars in cash on property and equipment — more than the same four companies spent in all of 2023. Under normal conditions, when costs swell and financing tightens, firms pull back. The data point the other way: the buildout is accelerating, and the question has shifted from whether the next dollar pays to what happens if you stop while rivals keep spending.

Three years ago most people met AI as a chat box, an image generator and a slightly better search — a software revolution. Today the frontier has moved into the physical world. The four hyperscalers are buying huge volumes of chips, building data centers, locking in power and reserving capacity that will not come online for years. Like breaking ground for a factory, once the lease is signed, GPUs are ordered and power is secured, you are no longer experimenting with a product; you are installing an industrial system whose commitments stretch far beyond the next quarter.

The pace shows in the numbers. The group spent roughly 147 billion dollars in 2023, rose to about 376 billion in 2025, and had already reached nearly 295 billion by the halfway mark of 2026. Not all of that is AI, but company disclosures point to the driver: Microsoft says roughly two-thirds of its recent capital spending went to short-lived assets, mainly CPUs and GPUs; Google says most of its technical infrastructure investment goes to servers, data centers and networking, with much of that capacity not yet online. Microsoft disclosed about 329 billion dollars in additional leases, largely for data centers not yet commenced, and Meta carries hundreds of billions tied to future capacity — not money already spent, but physical capacity contracted years ahead.

From software story to power and concrete

At this point we are no longer talking about a better chatbot. One Ohio campus is being planned alongside about 10 gigawatts of new generation — add cooling, fiber, transmission and transformers, and AI becomes a matter of concrete, copper and megawatts. Models make headlines, but the real bottleneck is connecting power to the grid, securing chips and removing heat. That is why the story should be read as infrastructure and logistics, not just code.

Why can spending more still look rational as costs climb? Because in a race where the frontier moves, standing still has a cost. Imagine Microsoft builds tens of billions of capacity, then Google leaps ahead on model quality or cost efficiency; Microsoft's servers do not vanish, but their competitive value slips. Economists call this the Red Queen effect — named for the Alice Through the Looking-Glass character who must run to stay in place: as the frontier advances, the economic value of existing capital erodes for anyone who stops. The model built this year by economist Yukon Zhang around AI formalizes that intuition, and field signals fit it: Google says it remains supply-constrained while Microsoft tells investors it is committing at significant scale on an accelerated schedule and in advance of fully developed revenue streams — capital before certainty, because waiting carries competitive risk.

A second lens comes from the Bank for International Settlements (BIS — the central bank for central banks). Its Working Paper No 1367 models the same race from another angle and finds competition alone can push firms to invest well beyond what would be optimal for the economy as a whole. The force that extends the boom may also be the force that eventually creates excess capacity. That is why the usual bubble check — does the next dollar make financial sense in isolation — misses the point in a race; management must ask what happens if a rival spends that dollar and you do not. A traditional return-on-capital lens can therefore keep investment alive longer than expected, even as overbuild risk rises.

From boardroom to borders

So far the story stayed in the boardroom, where budgets can be cut, debt trimmed and projects canceled. The stakes change when governments recast compute, chips and energy as national and economic security infrastructure. Washington and Beijing have both signaled that AI is no longer just another market but strategic infrastructure. The U.S. AI Action Plan released in July 2025, titled Winning the Race, frames the contest as one for global leadership and lays out infrastructure pillars — data centers, semiconductor manufacturing, energy and advanced computing capacity. A White House directive instructs federal bodies to guarantee that the national security enterprise can tap advanced computing and hardened facilities built for next-generation workloads at very large scale. The question is no longer only whether Microsoft earns a return on its next data center, but whether the country has enough compute, chips and power to keep an edge.

Beijing's language differs but the direction rhymes. Its latest five-year plan documents and the AI Plus initiative call for expanding the national computing backbone, lifting high-performance computing resources, pushing breakthroughs in core AI technologies and diffusing AI Plus across the economy. State media explainers foreground the same idea: a national compute grid, industrial integration and model capability. Both capitals are, in effect, building the physical and technical capacity they think an AI-shaped economy will require, worried about dependence on another power for the intelligence layer.

That shift invites a distinction between too big to stop and too big to fail. The argument is not that Washington must guarantee investments or protect shareholders; individual projects and even firms can still fail and loans can still sour. The claim is that the broader buildout can continue anyway once AI crosses the strategic threshold — the system reorganizes to carry it. It is like a highway network: a single contractor may go bust while the road still gets built, because the need is defined at national scale.

How constraints are being widened, not accepted

Once the race hits a constraint, the response so far has been to widen the constraint rather than abandon the race. The first phase was funded largely from corporate cash flow and balance sheets; as scale grew, that wall appeared and the financing model evolved. Wall Street stepped in to broaden the base. The clearest public example is Meta's roughly 27 billion dollar Hyperion data center, structured as a joint venture where outside investors hold the majority and supply additional financing — a model detailed in coverage of Blue Owl's partnership. The broader push, associated with figures like Larry Fink and Nvidia, is to move AI infrastructure financing beyond Big Tech balance sheets and securitize it, that is, to package future lease and usage cash flows into bond-like instruments, much as mortgages were once packaged — when capital grows scarce, expand the capital base.

The same pattern is visible in power and chips. Data centers demand enormous electricity and grid interconnection is a binding constraint. The Department of Energy is working to accelerate large-scale generation and transmission, regulators are revising how very large loads connect, and campuses like the Ohio project are co-planned with generation — about 10 gigawatts on site, on the order of a mid-size country's installed capacity, which gives a sense of scale. On semiconductors, the bottleneck has turned from corporate supply chain to industrial policy. TSMC's planned U.S. buildout, described as up to 165 billion dollars across dozens of facilities, is the emblem of that shift: chips are no longer just a product but a capacity policy.

Why is America willing to go this far for one technology? Not just for tax receipts or corporate profits, but for productivity. In a highly leveraged economy, growth lifts incomes and profits, which lifts tax receipts and makes debt easier to carry. A traditional growth tailwind is fading: the Bureau of Labor Statistics (BLS) projects U.S. employment to grow about 3.5 percent over the coming decade, versus roughly 11 percent the prior decade, while federal debt is still projected to rise relative to the economy. If the labor force will not expand as fast, the other path to more output is more output per worker — where AI enters as an economic lever: same hours, more done, lower unit cost, faster iteration and businesses that were previously uneconomic.

Early evidence is encouraging but not yet proof at economy-wide scale. One randomized study across thousands of knowledge workers found those given generative AI saved roughly two hours per week on email alone; other work finds stronger productivity growth in industries adopting AI faster, though researchers remain cautious on causality. The productivity question therefore becomes whether AI can generate enough real output to justify the committed capital. If yes, it could be a generational boom in productive capacity; if the payoff arrives slowly, the technology can still transform the economy while those financing the first wave absorb losses. What is being built is infrastructure; when it pays back remains uncertain.

History suggests both can be true at once. Over the last three hundred years, about six major technological revolutions have followed a similar four-phase cycle, roughly on a fifty-year rhythm. The most recent analogue is the telecom and internet buildout: carriers spent aggressively ahead of demand, then capital spending collapsed when expectations broke, WorldCom failed and shareholders were wiped out, yet internet adoption kept rising and the overbuilt fiber lived on under new owners at different prices. Technology won while the original capital structure did not — the key distinction for AI, in this view. The author's investor playbook follows from it: first, be cautious about picking a single winner early, as installation-phase leaders often fall; second, own what every contender must consume regardless of winner — electricity, chips, copper, concrete, cooling, a heads-you-win-tails-you-win logic; third, track the shift from speculative capital (a promise about the future) to productive capital (an asset producing measurable value). The signal he watches is the actual quarterly cash the four big builders spend on property and equipment; a single weak quarter matters little, but if that spend falls year over year for several quarters while future capacity commitments and outside financing are not replacing it, the thesis has changed. Until then, assuming the race ends just because it looks expensive misses the incentives that can keep it running more expensive for longer — the real question is whether anyone can afford to be first to stop.