55%的老板后悔因人工智能而裁员。扎克伯格不后悔。

Key Takeaways

• The 55% Number Is Real: Forrester’s Predictions 2026 report and Orgvue’s survey of 1,163 senior leaders both arrive at the same figure. A majority of bosses who cut humans for AI now regret it.
• Two Layoff Waves, Opposite Endings: Mid-market firms that fired humans to deploy AI are quietly rehiring them. Hyperscalers firing humans to build AI will not.
• Layoffs Are the Financing: Meta, Microsoft, Amazon, and Alphabet plan up to $725 billion in 2026 capital expenditure. Cutting payroll is how that gets funded without breaking the income statement.
• Klarna Already Pivoted: After claiming its AI was doing the work of 700 employees and shrinking its workforce from 3,800 to roughly 2,000, Klarna’s CEO admits the cuts went too far. Service quality fell and humans are coming back.
• The Asymmetry Won’t Hold: Builders amortize chips over years on the balance sheet. Buyers see ROI miss in a single quarter. That is why one wave reverses and the other doesn’t.

Two Layoff Waves Pretending to Be One

On April 23, 2026, Meta announced it would cut roughly 8,000 jobs starting May 20, plus another 6,000 unfilled requisitions. Around the same time, Meta raised its 2026 capital expenditure guidance to $125 to $145 billion, almost all of it earmarked for AI infrastructure. By May 8, the framing had hardened in financial commentary: the layoffs were a line item in the company’s AI bill. Microsoft launched its first-ever voluntary buyout on April 23, opening the door for up to 8,750 US employees, about 7% of its 125,000-person American workforce, to take a paid exit. Amazon has shed about 30,000 corporate roles across two rounds since October 2025 (roughly 14,000 in late 2025, then 16,000 in January 2026) while spending $44.2 billion on capex in a single quarter, up 77% year over year.

Add Alphabet’s roughly $180 to $190 billion of planned 2026 capex, and the four largest hyperscalers are pointing about $725 billion at chips, fiber, and concrete in a single year while collectively shedding tens of thousands of corporate roles. Industry-wide tech layoffs in 2026 have already passed 90,000.

That is one layoff story. Here is the other.

Forrester’s Predictions 2026 report, published in October 2025, found that 55% of employers who made AI-attributed layoffs already regret them. A separate survey of 1,163 senior leaders by Vitreous World, commissioned by workforce-analytics firm Orgvue between February and March 2025, came in at the exact same number. Fifty-five percent admitted “wrong decisions” on making employees redundant to bring AI into the workforce.

Klarna, the Swedish buy-now-pay-later firm, marketed in 2024 that AI was doing the work of roughly 700 employees, having shrunk its overall headcount from about 3,800 to 2,000 between 2022 and 2024. By mid-2025, CEO Sebastian Siemiatkowski conceded the cuts went too far. Customer satisfaction had fallen. Klarna started hiring humans again. IBM CEO Arvind Krishna told the Wall Street Journal that despite widespread AI adoption, IBM’s overall headcount actually rose. AI, in his framing, “gives you more investment to put into other areas.”

These are the same headline (“AI is replacing workers”), but they end in opposite places. One wave is structural and won’t reverse. The other is already reversing in real time.

Why the Numbers Look Identical and the Outcomes Don’t

The mainstream framing collapses both waves into one narrative: the great AI replacement. The framing is wrong, and it confuses two completely different uses of layoffs.

The hyperscalers are not firing customer service reps because a chatbot replaced them. They are firing recruiters, mid-level program managers, and overlapping infrastructure teams to free cash for capital investment. That cash buys Nvidia GPUs, Corning fiber, gas turbines, and land. The new AI capabilities are sold to other companies, who then use them to fire their customer service reps. Both groups appear in the layoff count. Only one is making a productivity bet that has to clear in twelve months.

Meta is the cleanest case study. The company’s planned 2026 AI capex of $115 to $135 billion is roughly four to five times its annual cash compensation expense. Cutting 8,000 jobs trims compensation by something like $1.5 to $2 billion at fully loaded cost. That is a rounding error against the AI bill. The financial point of the layoff is not the dollars saved. It is the signal: when public-market investors see capex go up and headcount go down in the same quarter, they read it as discipline, not desperation. The stock holds.

Klarna had no such option. Its AI deployment was supposed to deliver near-term unit-economics improvement in customer service. When customers escalated and revenue at risk grew, the productivity hypothesis broke in a single quarter, and the only response was to hire humans back. There is no four-year amortization schedule for service quality.

The Accounting Trick: Opex Becomes Capex

This is the part the headlines miss. In US Generally Accepted Accounting Principles (GAAP - the standard rulebook for how American companies report finances), salaries are operating expenses (opex). They hit the income statement immediately and dollar-for-dollar reduce reported profit. Servers, fiber, and data center construction are capital expenses (capex). They are depreciated over a multi-year schedule, often five to seven years for IT equipment under IRS Publication 946 guidelines, and only a fraction hits the income statement in any given year.

A company that takes the same dollar amount off its payroll line and puts it into a data center has, on paper, made the same economic decision. But on the income statement, profit jumps. Earnings per share (EPS) goes up. The stock rallies. The actual cash leaving the building did not change.

This is one reason hyperscalers will not reverse course even if the AI productivity claims are exaggerated. The capex is already on the balance sheet. Reversing means writing it down, which is the opposite of the EPS-friendly trade they just made. So the layoffs stick. The chips are bought. And the question of whether any of it produces revenue commensurate with the spend is deferred to a future quarter.

The site has covered the back half of this trade in two prior pieces. The capex unlock explained how rate cuts created the financing window. The dark silicon thesis explained that a meaningful share of those chips will sit in warehouses, depreciating, with power and packaging bottlenecks preventing them from being turned on. This article completes the loop. The layoffs are the line item that pays for the chips that may never compute.

What Klarna Saw That Meta Hasn’t

The buyer-side reversal is not a fringe story. It is the larger share of the dataset.

Vitreous World’s Orgvue survey of 1,163 enterprise leaders found that 55% admit redundancy decisions tied to AI were wrong. Forrester’s same-number finding suggests it is not survey noise. Independent reporting puts the rehire rate higher: roughly a third of firms that did AI-driven layoffs rehired between a quarter and half of those roles, and another third rehired more than half. Forrester forecasts that half of AI-attributed layoffs will be quietly reversed by 2027, often offshore or at lower pay. Reporting and analyst commentary across the same period note that the majority of companies that cut jobs citing AI did not see aggregate returns improve.

Why does the customer side reverse so fast? Three reasons.

The productivity claim is testable on a short clock. A chatbot either resolves tickets within service-level agreements or it doesn’t. Net Promoter Scores and churn move in weeks, not years.

Second, the substitute is not actually a substitute. Forrester’s own analyst notes describe AI handling routine queries adequately and failing at “human-to-human trust” interactions, which happen to be the exact moments where high-value customers decide whether to renew. Cutting those humans is cutting the people who keep the revenue, not the cost.

Third, there is no balance sheet to hide behind. A mid-market firm cannot capitalize the cost of a fired customer service team into a multi-year asset that pretends the savings are real. The income statement tells the truth in 90 days.

Klarna saw all three at once. Customer satisfaction dropped, the firm was preparing for a US initial public offering (IPO), and the AI deployment quietly inverted. Siemiatkowski’s pitch shifted from “AI is doing the work of 700 people” to a hybrid model with AI handling routine queries and humans escalating. The 700 number got quietly recategorized as redeployment, not pure replacement.

The 1990s ERP Rhyme

This pattern is not new. Look at the 1990s wave of enterprise resource planning (ERP - integrated business software like SAP and Oracle’s Financials). Vendors pitched corporate clients on automating their back offices. A wave of accountant, payroll clerk, and procurement specialist layoffs followed. The vendors got the multi-year contracts. By the early 2000s, many of those eliminated roles had quietly returned, often relabeled as “ERP analyst,” “process owner,” or “systems accountant.”

The structural lesson held then and holds now. The technology builders win the contracts. The technology buyers eat the rehire bill. The headline-level “automation replaces jobs” story turns out to be partially true, partially reversed, and partially repackaged. Aggregate employment in those functions did not collapse. It moved.

The Foxconn data point hints at where this wave is moving. The Nvidia partner reported April 2026 revenue up 29.74% year over year on AI server demand. Somebody is still hiring, and that somebody is the manufacturing supply chain that builds the chips the hyperscalers buy with the savings from the layoffs. The total labor pool didn’t disappear. It got reallocated.

The Steelman: Maybe the Hyperscalers Are Right

It is worth taking the builder side seriously. Microsoft’s voluntary buyout is structured for employees whose age plus years of service total 70 or more. That targets senior, expensive staff and reads more like an early-retirement program than a desperation move. Meta’s reorganization into “AI pods” is a real structural shift, not just headcount theater. Amazon is shedding corporate roles while continuing to hire engineers and developers in strategic AI and cloud areas. The composition is changing, not just the count.

Azure, AWS, and Google Cloud have all reported AI-attributable revenue growth that, if it sustains, justifies a meaningful chunk of the capex on a discounted-cash-flow basis. If only half the GPUs end up productive, the math can still pencil because the half that works carries margin that traditional cloud workloads never approached.

The honest answer is that the builder thesis is not yet falsifiable on a one-year window. The buyer thesis already is. That asymmetry is the entire article.

What 2027 Probably Looks Like

The forecast lines up cleanly. Forrester expects roughly half of AI-driven layoffs to reverse by 2027, often at offshore rates. Hyperscaler capex slows from its 2026 rate of change but does not contract, because the contracted obligations to chip vendors and utilities don’t allow contraction without write-downs. The Klarna playbook (cut humans, deploy AI, partial reverse to hybrid model) becomes the standard mid-market template. Foxconn keeps growing. Nvidia keeps shipping. The displaced workers reabsorb at lower wage levels in the rehired roles, which is how Forrester gets to “lower pay.”

The political consequences will catch up later. Hyperscaler corporate cuts are happening alongside supply-chain hiring growth and quiet rehires on the buyer side; the net effect on aggregate tech employment is unlikely to be the wholesale collapse the layoff headlines suggest. The distribution shifts more clearly than the total: fewer salaried mid-level corporate roles, more contract and offshore service roles, and a small number of very large capex-intensive employers consuming a growing share of compute and energy. That is a different labor market than the one being framed in the layoff headlines.

The headline says AI is taking the jobs. The data says AI is taking some of the jobs, hyperscalers are taking some of the jobs to fund the chips, and a meaningful share of both will quietly come back at lower pay. The 55% regret figure is the number that catches up with the narrative. Zuckerberg won’t read it as a warning. Klarna already did.