The numbers landed in the same week and they are not consistent with each other.
Microsoft, Alphabet, Amazon, and Meta finished Q1 earnings the week of April 27 with a combined 2026 AI-capex guide of about $725 billion — a 77% jump on 2025’s $410 billion record. Bloomberg first crossed the $700B threshold on April 30; by Sunday the consensus number had settled near $725B as Meta raised its full-year guide to $125–$145B, Amazon to $200B, Microsoft to ~$190B, and Alphabet to $180–$190B.
Layoff trackers, by Sunday, had crossed 95,878 workers impacted across 249 events for 2026 year-to-date — a 864-per-day pace. One alternative tracker had already passed 113,000. Tom’s Hardware noted that nearly 50% of the Q1 affected positions were explicitly cut for AI/automation reasons, and the Q1 base for those cuts was around 81,000.
The third number, the one that makes the first two impossible to reconcile cleanly, is the open-postings number. As of January 2026 the US alone had 275,000 active job postings requiring AI skills, with global demand outpacing supply roughly 3.2 to 1.
The Invezz writeup on May 4 framed the question this way: “Is Big Tech’s $725B AI splurge being funded by mass layoffs?” The answer is partly yes, partly more interesting than that, and the more-interesting part is the labor-market mismatch under the headline.
What the $725B is buying
The four-firm capex stack is not buying the same thing it was buying in 2024. About 75% of the 2026 spend is identifiably AI-specific: GPU clusters, custom silicon (TPU, MTIA, Trainium, Maia), low-latency networking fabric, AI-optimized server racks, and the data-center power systems that wrap them.
That is the line item where headcount cost shows up most directly as a substitute. A senior platform engineer earning $400K fully-loaded is roughly equivalent in P&L impact to one H200 GPU operating for a year on full duty cycle, including power. A 1,000-engineer cut is roughly equivalent to ~6,000 GPUs for a year, which is enough silicon to train a frontier model.
This is the math at the core of the May 1 Mag-4 piece and the same math the Workman Tesla ex-HR LinkedIn essay put numbers behind. CFOs are not, in general, claiming the cut workers are being replaced by AI. They are claiming AI capex must be funded somehow, and human labor is the only line flexible enough to cut fast.
The May 1 piece covered the same-quarter coincidence at $650B. The May 4 number is $725B because Meta and Amazon both revised up after their Q1 prints. The math has gotten worse, not better.
What the 95,878 is doing
The Washington Post May 1 read on the same data is correct that the cuts are not all about AI substitution. About half are pandemic-overhiring unwinds, austerity-mode cost discipline, or capex reallocation. The AOL/CEOWORLD May 3 framing of the cuts as “AI washing” is also partly right — the press-release mix overweights AI for stakeholder optics.
What none of those framings explain is who is being cut and who is being hired. The composition matters more than the gross number.
The cuts skew toward:
- Customer-service operations. Salesforce’s headcount went from 9,000 to 5,000 customer-service humans on Benioff’s own statement.
- Senior product and platform engineering. The Microsoft 8,750 buyout offer targeted Rule-of-70 senior tenure.
- Recruiting, HR, and corporate communications. Meta’s 8,000 cuts are 35–40% concentrated in those functions.
- Marketing and content production. Disney’s 1,000-person consolidated marketing reorg and Snap’s 1,000 AI-cited cuts on April 15.
- Legacy database administration and on-prem support. Oracle’s 30,000 was largely here.
The 275,000 open AI-skill postings skew toward:
- ML / AI research engineering (~$300K–$700K base).
- AI infrastructure / GPU systems engineers (~$250K–$500K).
- AI safety and AI governance specialists (this category alone is up 150% in posting volume year-over-year).
- Applied scientists for foundation-model evals, RLHF, and post-training (this is where Meta, Anthropic, and OpenAI are competing on retention bonuses).
- Robotics-data and embodied-AI annotation engineers (a category that did not exist as a major posting line two years ago).
These are not the same humans. A 22-year-veteran Oracle on-prem DBA does not retrain into an RLHF post-training scientist by attending a six-week bootcamp. The AI-safety governance role requires a completely different toolset than the recruiting role being eliminated. The Microsoft 8,750-person buyout pool is, by structure (Rule of 70), the senior cohort whose specialization is in what AI replaces, not what AI requires.
This is not a skills gap. A gap is something you bridge. This is a skills chasm.
Why the Q1 productivity print does not save you
Q1 2026 BLS nonfarm productivity grew at roughly the post-2010 trend, not at an AI-displacement-shaped rate. The May 4 AOL/CEOWORLD framing leans on this print as evidence that AI is not the cause of the cuts.
The lag is real. The 1990s computer-on-every-desk substitution did not show up in BLS productivity until 1995, nine years after widespread workplace adoption. The 2026 productivity print is too early to read AI’s signal.
What is not lagging, and what shows up cleanly in the data, is the occupational composition shift. Software-developer headcount in Q1 2026 contracted 3.5% YoY against a growing economy. Customer-service rep headcount contracted faster. Recruiter and HR-coordinator headcount contracted fastest. Meanwhile, ML-engineer postings grew at the highest YoY rate the BLS occupational employment statistics has ever recorded for any single occupation.
Productivity statistics aggregate the two and cancel them. The composition shift is what the labor market is actually doing.
The finance question and the answer that does not square
The Invezz piece’s central frame — “is the $725B being funded by mass layoffs?” — is the wrong question. The right question is whether the $725B is being funded by the right layoffs.
The Mag-4 free cash flow available for AI capex is roughly $400–$450B in 2026 on consensus estimates. The capex is $725B. The gap, on the order of $275–$325B, is being financed three ways:
- Operating margin expansion. This is where headcount reduction shows up — labor cost is a flexible cost, AI capex is a fixed cost, and CFOs trade one for the other.
- Debt issuance. Microsoft, Alphabet, and Amazon all issued investment-grade debt at favorable spreads in Q1.
- Reinvested AI revenue. Anthropic’s $30–40B run rate inside the Mag-4 P&Ls and Azure-AI revenue (~$37B annualized at Microsoft) buy back into capex.
Of those three, the layoff-funding share — column one — is the one that is generating the largest political and labor-market externality. And the externality is not the gross headcount cut. The externality is that the cut workers and the new AI-postings demand are not the same population, and the chasm between them is what creates the residual unemployment that does not show up in the productivity statistics.
This is the Falk-Tsoukalas Pigouvian-tax argument in slightly different language: a $725B capex flow that produces 95,878 displaced workers and 275,000 unfilled postings is generating a private gain to the four firms (operating leverage) that does not internalize the public cost (frictional unemployment, retraining shortfall, lost-decade earnings for the displaced cohort).
What LostJobs is watching for the next 30 days
- The Microsoft May 7 buyout details release. This is the first earnings-aftermath data point on what the senior-cohort exit profile actually looks like — how many of the 8,750 took the offer, what the median tenure was, what the severance package looked like, and crucially what the next-job destination of the takers ends up being.
- Meta’s May 20 reduction in force. The first time the post-AI-washing Sunday-business-press cycle’s framing will be tested against an actual press release. If Meta walks away from the AI-efficiency framing, that is a signal. If it leans in, that is a different signal.
- The H1B FY27 cap data. Specialty-occupation H1B is the principal pipeline for AI-research engineers in the US. The Q3 cap-subject data drops mid-summer; if the Mag-4 share crosses 40%, the labor-market read is that the US AI-skills supply is being satisfied externally rather than from the laid-off cohort.
- The first university CS-program enrollment number for fall 2026. The Harvard 77% under-hiring study and the Goldman Sachs Gen-Z 16,000-hire pause suggest the entry-level pipeline is collapsing faster than the AI-skill pipeline is forming. Fall 2026 enrollment is the leading indicator on whether the chasm widens or stabilizes.
The dry coda
The 2026 AI-capex number is now $725 billion, up 77% on 2025. The 2026 layoff number is 95,878 and counting at 864 per day. The 2026 unfilled-AI-postings number in the US alone is 275,000 and rising.
If you divided the capex flow evenly across the laid-off workers, each would get $7.56 million. If you divided the unfilled postings evenly across the laid-off workers, each would get 2.87 jobs to apply for. Neither calculation is what is actually happening, because the two populations are different people and the unfilled jobs are not the jobs the laid-off workers can do.
The headline is not that Big Tech is cutting workers to fund AI. The Washington Post got that part. The headline is that Big Tech is cutting the wrong workers to fund AI, the right workers do not yet exist in sufficient supply, and the gap between those two facts is now the largest single labor-market dislocation in the US since the 2008 financial crisis.
We will know the shape of it by year-end. We already know the shape of the question.