$725B In, 142,000 Out: The Layoffs Are Funding the Data Centers

The two biggest tech stories of 2026 are usually told separately. One is the largest concentrated infrastructure build in the industry's history. The other is a sustained wave of layoffs at companies posting record profits. Told apart, each is confusing. A profitable company cutting staff makes no obvious sense, and a $725 billion spending commitment during a period of cost discipline makes even less. Told together, both become straightforward, and the straightforward version is not comfortable.

The Numbers, Placed Side by Side

Start with the scale, because it is difficult to hold in mind. $725 billion in a single year, from four companies, up 77% from the prior year. For comparison, that figure exceeds the annual GDP of most countries and represents capital commitments larger than the entire US interstate highway system cost in inflation adjusted terms. It is being deployed into buildings, chips, and power infrastructure on a timeline of roughly thirty-six months.

Meanwhile, 81,747 tech workers lost their jobs in the first quarter alone. The categories being eliminated are consistent across companies: customer support, quality assurance, content moderation, and middle management. The categories in shortage are equally consistent: machine learning engineers, AI safety researchers, and data infrastructure specialists. Same industry, opposite directions, at the same moment.

The Accounting Story Nobody Puts in the Press Release

Here is the mechanism, and it is less about AI capability than most coverage suggests. Salaries are operating expenses. They hit the income statement in full, in the quarter they are paid, and they reduce reported earnings dollar for dollar. Data centers, servers, and chips are capital expenditure. They are recorded as assets and depreciated across a useful life of several years, which means a dollar spent on infrastructure damages this quarter's earnings far less than a dollar spent on people.

A company that moves $10 billion from payroll to capital expenditure has not reduced its spending. It has changed how that spending appears. Operating margin improves, earnings per share holds up, and the enormous investment shows up in a cash flow statement fewer people read closely. This is entirely legal, entirely standard, and entirely rational under the incentives public companies face. It is also a substantially different story from "AI made these roles unnecessary."

None of this means AI capability is irrelevant to the cuts. Some of those support and QA roles genuinely are being automated, and the automation is real enough that it would have happened at some pace regardless. But the timing, the concentration, and the simultaneity with a record capital cycle suggest the accounting treatment is doing at least as much work as the technology. When 45 CEOs give the same reason in the same year, the reason is partly a description and partly a narrative that markets reward.

This connects to the circularity we examined in the AI circular deal loop, where the same dollars move between a small set of counterparties and each hop is booked as revenue. Capitalised labour cost is a related move: real activity, accounted in the way that reads best. Neither is fraud. Both make the underlying economics harder to read from outside.

The Skills Gap Is Not a Skills Gap

Roughly 275,000 AI roles sit open while laid-off workers cannot fill them. This gets described as a skills gap, which implies a training problem with a training solution. The description is too flattering to everyone involved.

A content moderator and a machine learning engineer are not separated by a course. They are separated by several years of mathematics, a different educational track, and in many cases a different immigration status and geography. Treating those as adjacent because both are "in tech" is like treating a radiographer and a surgeon as interchangeable because both work in a hospital. The people being cut and the people being hired are largely different populations.

The more useful framing is that the industry is not shedding workers, it is re-sorting them, and the re-sort is brutal for anyone in the middle. This is the same dynamic we traced in the missing rung, where junior developer roles fell roughly 20% since 2024 while seniors at the same companies received raises. Compression at the bottom, scarcity at the specialised top, and a thinning middle. That shape describes several previous technology transitions and it is rarely temporary.

What This Looks Like From Inside a Company

The macro numbers obscure how ordinary this feels at close range, and the ordinariness is part of why it goes unchallenged internally. There is no meeting where someone proposes trading headcount for depreciable assets. There is a budget cycle, a margin target handed down, and a set of teams asked to deliver the same roadmap with fewer people because the infrastructure commitment is already signed and is not up for discussion.

The teams that lose people are the ones whose output is hardest to attribute to revenue. Internal platform groups, developer experience, documentation, QA, and support all share the property that their absence is felt gradually rather than immediately. That makes them cheap to cut this quarter and expensive over three, which is a trade organisations make repeatedly because the cost shows up under a different heading than the saving.

Then the AI narrative gets applied afterwards, and often sincerely. A support organisation that shrank by 40% did deploy an agent, and the agent does handle a meaningful share of tickets. Whether it handles enough to justify the cut is a question nobody measures rigorously, partly because the honest answer would be awkward and partly because, as we covered in the 95% problem, most enterprise AI pilots deliver no measurable profit and loss impact at all. A cut justified by a deployment that did not deliver is still a cut, and the deployment rarely gets audited once the headcount is gone.

The version of this worth watching for in your own organisation is when the AI capability arrives after the reduction rather than before it. That ordering tells you which one was the driver.

Two Job Market Numbers That Both Look True

Indeed Hiring Lab reported software development postings down 36.4% against February 2020 and down 6.7% year over year. The Bureau of Labor Statistics projects software developer, QA analyst, and tester employment growing 15% from 2024 to 2034, with about 129,200 annual openings. People pick whichever number supports their prior and stop there. Both are measuring accurately, and the difference between them is worth understanding.

Postings measure hiring intent right now, in a specific channel, heavily weighted toward the companies that post publicly and toward the roles that are competitive to fill. They are a fast, noisy, cyclical signal. Employment projections measure the total stock of jobs across an entire economy, including the very large number of software roles at companies that are not technology companies, and they are slow and structural.

A world where hiring at large technology employers contracts sharply while software employment across banks, hospitals, logistics firms, and manufacturers keeps growing is fully consistent with both numbers. That is probably the world we are in. It also implies that the geographic and sectoral distribution of software work is shifting away from the concentration that defined the last decade, which is a bigger deal for careers than the aggregate count.

Four Signals Worth More Than the Headlines

Layoff announcements and capex totals are lagging, dramatic, and mostly useless for anticipating what happens next. Four quieter indicators carry more information, and all four are public.

Depreciation schedules in quarterly filings. The useful life assigned to AI hardware determines how much of this cycle's cost lands in which year. Extensions to those schedules are a way to soften near-term earnings impact, and they are disclosed. When a company lengthens the assumed life of its servers, that is worth more attention than any press release.

Utilisation language on earnings calls. Watch for the shift from talking about capacity being built to capacity being used, and for whether anyone quantifies it. Sustained vagueness on utilisation across several quarters is the single most informative pattern available to an outside observer.

Grid interconnect queues. These are matters of public record in most jurisdictions, they move slowly, and they bound what can actually be built regardless of capital. A datacentre with financing and no power connection is an announcement, not a facility.

Job postings by function rather than headcount totals. Aggregate headcount tells you almost nothing during a re-sort. The composition tells you everything: which functions a company is quietly rebuilding after cutting them is the clearest available evidence about whether the automation thesis held. Support organisations that shrank in one year and started hiring again in the next are giving you an answer that no earnings call will.

What This Means If You Write Software

The useful conclusions are unglamorous and mostly about positioning rather than skills acquisition.

First, proximity to revenue matters more than it did. Roles that are visibly load-bearing for a product customers pay for survive cost cycles that internal-platform and support-adjacent roles do not. This is not a statement about which work is valuable; plenty of eliminated roles were valuable. It is a statement about which work is legible to a finance function under pressure.

Second, the concentration risk is real. A career built entirely inside large technology employers is exposed to a spending cycle that four companies control. The sectors that keep hiring software people through this are the ones that were never counted as tech, and they pay less, move slower, and are considerably more stable. That trade looks different at thirty than it did at twenty-five.

Third, the specialisations in genuine shortage are shortages for a reason: they are hard, and the barrier is not a weekend bootcamp. Someone who spends two years actually building production AI systems, with the evaluation and reliability work that entails, ends up in a different market than someone who adds a certification. The premium is on demonstrated systems, and it shows up in the fastest-growing titles rather than in the fashionable ones.

Fourth, and most importantly, the Jevons dynamic we covered in why AI made engineers more valuable has not been refuted by these numbers. Engineering was the most resilient job function of 2025 and "AI Integration Engineer" was the fastest-growing title. Cheaper software production expands the amount of software worth producing. The catch is that expansion is slower than a layoff announcement, and it lands in different companies and different places than the contraction did. Both things are happening; they just do not happen to the same people on the same schedule.

If You Are In the Middle of This

Structural analysis is cold comfort to someone whose role was eliminated last month, so a few things that are practically true rather than merely accurate.

The market is bifurcated rather than closed. Generalist applications into large technology employers are competing against an unusually deep pool, and the response rate reflects that. The same person applying into the sectors that quietly kept hiring, financial services, healthcare systems, industrial firms, public sector, sees a completely different funnel. The work is less fashionable and the process is slower. It is also where a large share of the 129,200 projected annual openings actually sit.

Demonstrated systems beat credentials by a wide margin right now. A candidate who can describe an agent they took to production, the evaluation harness they built to prove it worked, and the two things that broke in month two is interviewing against a field of people listing tools. That is a lower bar than it sounds and almost nobody clears it, because most people who have done the work describe it as configuration rather than engineering.

And the timing genuinely is not a referendum on your ability. A capital allocation decision made in a boardroom about depreciation schedules is not a judgment about whether you were good at your job. Both things get communicated in the same email, which is one of the crueller features of how this is being run.

What Would Change the Picture

Three developments would meaningfully alter this, and they are worth watching more closely than the monthly layoff headlines.

The first is depreciation catching up. Capitalised spending delays the earnings impact; it does not remove it. Several years of $700 billion annual builds produce a depreciation charge that eventually lands on income statements whether or not the revenue arrived. If utilisation of that capacity disappoints, the accounting advantage reverses and the pressure returns with interest.

The second is power. Data centers already consume roughly 6% of US electricity, and as we covered in the ratepayer revolt, more than thirty states are contesting the buildout as bills rise in hotspot regions. Capital is abundant; grid interconnects and local political consent are not. The binding constraint on this cycle may end up being electrical rather than financial.

The third is whether the capacity produces returns that justify it. That question gets its first properly audited answer as the largest AI companies move toward public markets, a transition we looked at in the IPO race. Until then, the honest position is that an enormous bet is being placed, part of it is being funded by people who used to work at the companies placing it, and nobody outside those companies can yet say whether it pays.