Andrew Ng is right: there is no AI jobpocalypse. The Jevons paradox, BLS projections, and CEO behavior all point the same direction.
TL;DR
- Andrew Ng's call: software engineering hiring stays strong despite being the sector most affected by AI tools.
- US BLS projects 15% software developer growth from 2024 to 2034, vs. 3% for all occupations. AI is cited as a demand driver.
- Jevons paradox and Bessen's ATM-teller research show cheaper tools historically expand employment, not shrink it.
- For builders: learn AI tools to compound your skills, then build distribution before it commoditizes.
Most 2026 tech layoffs framed as AI efficiency are not about AI replacing workers. They're a mix of post-COVID over-hire correction, slowing revenue growth, and the need to fund $700 billion in AI capital expenditure. Andrew Ng's argument that there is no AI jobpocalypse is supported by US Bureau of Labor Statistics projections of 15% growth in software developer employment through 2034. Historically, when a productive input gets cheaper, total consumption expands. That's the Jevons paradox, observed since 1865 and confirmed in ATMs, spreadsheets, and compilers. AI is making building cheaper. The lesson for developers: learn the new tools, then learn distribution.
I keep getting DMs from senior devs panicking about the layoffs. The memos all say AI. The framing all says "efficiency." Most of them are reading the memo wrong and acting on the wrong lesson.
A friend got laid off from a Series B last month. His memo cited "AI-driven productivity gains." He spent the next two weeks trying to ramp up on Claude Code at speed because he thought he was behind. The real reason his role got cut? His company missed its Q4 revenue target and the AI line read better in board decks than the slowdown line did.
The 2026 layoff story is the cleanest example I've seen of a press release winning over a spreadsheet.
What companies are saying vs. what their CEOs admit
Andy Jassy was unusually honest on Amazon's Q3 2025 earnings call. Asked about the largest layoff in Amazon's 31-year history, he said the cuts "were not really financially driven, and it's not even really AI-driven, not right now at least. It's culture."
Three months later, Beth Galetti's formal layoff memo at Amazon talked about reducing layers, increasing ownership, and removing bureaucracy. AI was not mentioned once. By spring 2026, the continued cuts (now 30,000 corporate roles since October 2025) had been absorbed into the broader industry narrative of AI-driven efficiency.
The CEO of the company doing the largest cuts said publicly it wasn't AI. The press and the market treated it as AI anyway.
The script is consistent across announcements. Meta cut 8,000 in April 2026 to "offset the other investments we're making." Block cut 4,000 with Jack Dorsey citing intelligence tools paired with smaller and flatter teams. Snap cut 16% citing rapid advancements in AI. Salesforce cut customer support from 9,000 to 5,000 with Marc Benioff saying AI agents handle 50% of interactions. Microsoft offered buyouts to 8,750 US employees.
Through April 2026, AI has been cited as a factor in 49,135 announced job cuts, per Challenger, Gray & Christmas. The narrative is dominant. The math doesn't support it.
Andrew Ng named the mechanism directly in his May 2026 Batch letter: "Businesses have a strong incentive to talk about layoffs as if they were caused by AI. Talking about how they're using AI to be far more productive with fewer staff makes them look smart. This is a better message than admitting they overhired during the pandemic when capital was abundant due to low interest rates and a massive government financial stimulus."
That sentence describes most of the layoffs of 2026.
The three things driving 2026's layoffs
Strip away the AI narrative and three things are happening at once across the companies announcing the largest cuts.
COVID over-hiring is still being unwound
Amazon hired aggressively from 2019 to 2022, growing global headcount from 798,000 to 1.6 million. Meta doubled. Microsoft, Google, Salesforce all hired into pandemic-era demand assumptions that didn't survive 2023.
Block expanded headcount aggressively through 2021 to 2023, building parallel teams across Square and Cash App. The 40% cut Dorsey announced in February 2026 is mostly Block returning to roughly its 2020 size. The "AI-native, flatter teams" language is the public-facing wrapper around what is, structurally, a duplicate-org cleanup.
Marc Andreessen, hardly a layoff skeptic, attributed recent cuts to "higher interest rates and a complete loss of discipline in hiring during the pandemic. The hiring binge that companies went on in COVID was just wild." This is the same Marc Andreessen whose firm is one of the loudest voices on AI replacing work. Even he won't credit AI for the current wave.
"We over-hired" is a flat story. "AI made us more efficient" is a forward-looking transformation story. Same cuts, different press release.
Revenue is slowing and Chinese competition is squeezing margins
Some of the most aggressive layoffs aren't at hyperscalers building data centers. They're at companies losing market share or facing structural revenue pressure.
PayPal's cuts followed slowing revenue growth, stalled active-user counts, and competition from Stripe, Apple, Visa, and Mastercard. Coinbase rode the 2021 crypto boom, cut during the 2022 winter, rehired into the next cycle, then framed 2026 cuts around "AI-native teams." The underlying driver is the volatility of crypto demand, not a productivity unlock.
Chegg cut 45% of its workforce in October 2025 because students stopped using it. They use ChatGPT instead. That is a real AI-driven layoff, but in the inverse sense: AI killed the product, not the headcount of an "efficient" company.
The macro backdrop matters. US GDP grew just 0.5% in Q4 2025 before rebounding to 2.0% in Q1 2026. The Conference Board's Leading Economic Index declined 0.6% in March 2026. The Challenger, Gray & Christmas tracker tells the cleanest version of the story: the most-cited reason for 2026 layoffs is "market and economic conditions" at 53,058 cuts, more than double the AI count of 21,490 in the same period.
Then there's China. DeepSeek V4 Pro is priced at $1.74 / $3.48 per million input/output tokens. Claude Opus 4.7 sits at $5 / $25. GPT-5.5 at $5 / $30. RAND research puts Chinese model costs at one-sixth to one-fourth of comparable US systems. When the cost of your most strategic capability is being undercut 4 to 6 times by an open-source competitor, you cut headcount somewhere. You don't blame DeepSeek in your layoff memo. You say "AI efficiency."
AI capex is eating the room
This is the story most companies don't want to tell directly: they need to fund $700 billion of capex, and headcount is the easiest line to cut.
The four largest hyperscalers (Amazon, Microsoft, Alphabet, Meta) are projected to spend $725 billion on capex in 2026, up 77% year over year. Roughly 75% is AI-specific. Capital intensity at hyperscalers is now 45 to 57% of revenue, ratios that look like utility companies, not software companies.
Meta is the cleanest case. The company is planning $125 to $145 billion in 2026 capex, per the January 29 earnings call. The 8,000 layoffs free roughly $2.4 billion in annual run-rate operating expense. That is 1.7% of the capex bill. Even fully replacing the workforce with AI would save about $27 billion, a fraction of the $145 billion infrastructure spend.
Meta's Q1 2026 still printed $56.3 billion in revenue (up 33% year over year), 41% operating margins, and $10.44 EPS. This is not a company in distress. The cuts aren't about AI productivity. They're about creating room on the income statement for a capex bill growing roughly twice as fast as revenue.
Larry Page reportedly told colleagues: "I'm willing to go bankrupt rather than lose this race." That is the actual posture inside hyperscalers. The AI-framed layoffs are the public face of that posture.
There's a fourth incentive Ng called out that's worth reading directly. AI companies anchor their pricing to salaries rather than SaaS norms. A SaaS tool charges $100 to $1,000 per user per year. If an AI tool can replace a $100,000 employee or make them 50% more productive, charging $10,000 looks reasonable. By anchoring to salaries, AI vendors capture much more revenue than traditional SaaS pricing would allow. That commercial logic depends on the layoff narrative being true. The incentive to keep the narrative alive is direct and financial.
Is software engineering finished? History says cheaper tools grow employment
In 1865, British economist William Stanley Jevons noticed something counterintuitive about coal. As steam engines became more efficient, total coal consumption rose instead of falling. Cheaper coal per unit of output made coal-powered production viable in more industries, expanding total demand faster than efficiency reduced it. He called this the paradox of efficiency. Microsoft CEO Satya Nadella invoked it explicitly when DeepSeek's pricing hit the markets in early 2025.
The textbook case in employment economics is bank tellers and ATMs. From 1988 to 2004, ATMs cut the number of tellers needed per US bank branch from 20 to 13. The intuitive prediction was teller employment would collapse. It didn't. Cheaper branch operations let banks open 43% more branches in urban areas. Total teller employment rose. Economist James Bessen documented this for the IMF in 2015, and the pattern has become the standard reference for thinking about automation and jobs.
The same pattern shows up with spreadsheets, where the prediction was that VisiCalc would kill accountants. The reality: financial analyst jobs grew because cheaper analysis made more analysis worth doing. With compilers, where assembly programmers were supposed to be displaced. The reality: total developer count grew by orders of magnitude because cheaper code made more code worth writing. With electrification, where the same panic played out across factory work.
The mechanism is consistent. When a productive input gets cheaper, the supply of work that input can support expands faster than the input becomes redundant. People build things that weren't worth building when the input cost was higher.
The US Bureau of Labor Statistics, working from the assumption that AI will accelerate over the next decade, projects software developer employment to grow 15% from 2024 to 2034, against 3% for all US occupations. Their report names AI explicitly as a demand driver: "Demand for software developers, software quality assurance analysts, and testers is projected to be strong due to the continued expansion of software development for artificial intelligence, Internet of Things, robotics, and other automation applications." About 129,200 openings are projected per year over the decade.
There's a Jevons split inside the BLS data worth noticing. The narrow category "computer programmers" (repetitive coding work) is projected to decline 6%, with the explicit reason being "computer programming work continues to be automated." The broad category "software developers" (designing, integrating, shipping software) grows 15%. The narrow, repetitive task gets automated. The broader role expands. This is exactly the pattern Bessen described for tellers in 1988.
Andrew Ng's Batch letter argues the same point at a higher level. Software engineering is the sector most affected by AI tools. Hiring remains strong. US unemployment is 4.3%. His prediction is what he calls an "AI jobapalooza": more good AI engineering jobs, in companies that aren't traditionally software employers, with skill mixes that look different from 2018.
Why this time could still be different
This time might be different. Three reasons to take seriously.
First, speed. The ATM-to-teller transition played out over forty years. AI's transition into coding has taken about three years. Even if the long-run equilibrium is more developers, the transition is happening on a timeline that gives workers little room to retrain.
Second, completeness of automation. Bessen's bank teller story has a sequel most quotations miss. Teller employment did eventually decline, not from ATMs, but from mobile banking after 2010. When automation went from partial (ATMs handled some tasks) to nearly complete (mobile banking handled them all), the Jevons effect stopped protecting jobs. The question is whether AI's coding capability will graduate from partial automation (helps engineers be faster) to complete automation (does the job end-to-end). Today it's clearly partial. The question is for how long.
Third, the macro data isn't yet showing the productivity boom that would generate Jevons-style demand expansion. Torsten Slok, chief economist at Apollo, summarized this in a phrase that's now widely quoted: "AI is everywhere except in the incoming macroeconomic data." Stanford's "Canaries in the Coal Mine" study from November 2025 found employment declining for workers whose jobs may be affected by AI. The specific roles named were software developers and customer-service representatives. These are exactly the roles Jevons should be protecting.
So the long-run pattern says Jevons holds. The short-run data is mixed. Builders should adapt now rather than wait to find out which way the next five years go.
What this means for developers in 2026
Here is where most takes on these layoffs go wrong.
The instinct, especially for engineers, is to read the layoff memos at face value: "AI is taking our jobs. Better get good at AI fast." Half of that is wrong. Half is right.
Half is wrong. AI is not currently replacing software engineers at scale. The cuts aren't because AI engineers write 10x more code. They're because companies over-hired, growth is slowing, and someone needs to absorb $700 billion in capex.
Half is right. AI is making building cheaper. Not because it replaces engineers, but because it compresses the time from idea to working prototype. I can spin up a working React Native app with auth, theming, i18n, and Redux Toolkit using my own expo_boilerplate plus Claude Code in an afternoon. Two years ago that was a weekend. Five years ago it was a small team's first sprint.
The Jevons reading: cheaper building means more total building. That expands demand for the work AI can't yet do well: system design, integration, debugging, judgment calls, shipping, and getting users.
The bottleneck has moved. Building used to be the moat. The number of people who can ship a working product has exploded. The cost of building has collapsed. Distribution didn't get cheaper. Attention didn't get cheaper. Trust didn't get cheaper. The audience you spent five years building is still worth what it was worth in 2020. The newsletter with 10,000 engaged readers is still rare.
This is the inversion that matters more for your career than any layoff announcement. The work didn't get harder. It shifted. The skill stack that paid in 2018 (deep technical specialization) pays less now. The skill stack that pays in 2026 is technical work plus distribution work.
What to learn this week: AI tools and distribution
If I were starting over today as a senior engineer reading the layoff news, this is what I'd learn first, in this order.
Use AI as a builder, not a topic to study
Stop reading about AI and start using it. Pick one workflow you do every week and rebuild it with Claude Code or Cursor. Measure the time saved. Notice where it breaks. The point is not to become an "AI engineer." The point is to compound your existing skill stack with tools that make you 3 to 5 times faster on the boring parts.
Ship one real thing per month
Not a tutorial project. A real thing you put online with your name on it. The boilerplate I open-sourced on GitHub was a forcing function for me: every time I built a CasaInnov client project, I extracted the reusable parts and pushed them back. Two years of that is now a credible authority signal anyone can clone.
Pick one distribution channel and commit publicly
Newsletter, LinkedIn, Twitter, YouTube, Reddit, GitHub. Pick one. Get good at it. The cost of building a 5,000-person audience in your niche is one consistent post per week for two years. That sounds boring. It is boring. It also outperforms 95% of what your peers are doing.
Write what you wish someone had written for you
I started Code Meet AI because I kept losing days to integration problems nobody had documented well. Hermes failing on cold start after Expo SDK upgrades. Claude Code hallucinating React Native imports that don't exist. Generative UI patterns that work on web but break on mobile. The writing is now its own moat.
The combination of these four habits is, in my opinion, more resilient than any specific technical skill. Skills depreciate. A distribution channel and a track record of shipping compound.
If you want a head start, my expo_boilerplate is MIT-licensed and built for exactly this: TypeScript, auth, theming, i18n, Redux Toolkit, feature-first architecture, Cursor and Claude rules already wired in. Clone it, change three things, ship something this weekend.
The 2026 layoffs are not the signal most people are reading them as. The current cuts are about COVID over-hire correction, slowing growth, and AI capex pressure. The economics history says cheaper tools grow employment, not shrink it. Andrew Ng calls it an AI jobapalooza. The Bureau of Labor Statistics is projecting 15% growth.
The work shifted. Adapt to where it went.
FAQ
Will AI replace software engineers?
Probably not net replace, based on Bureau of Labor Statistics data and historical precedent. BLS projects 15% growth in software developer employment from 2024 to 2034, with AI explicitly named as a demand driver. The Jevons paradox, cheaper inputs expand total consumption, has held for two centuries across electrification, spreadsheets, and ATMs. The transition may displace specific narrow roles (BLS projects "computer programmers" to decline 6%) while expanding the broader category of software work.
What did Andrew Ng say about AI and jobs?
In his May 2026 Batch letter, Ng predicted there will be no AI jobpocalypse. His evidence: software engineering hiring remains strong despite being the sector most affected by AI tools, and US unemployment sits at 4.3%. He attributes the panic narrative to AI labs wanting to sound powerful, AI companies anchoring pricing to salaries rather than typical SaaS norms, and businesses preferring "AI efficiency" to admitting they over-hired during the pandemic stimulus era. He predicts an "AI jobapalooza" instead.
What is the Jevons paradox and why does it matter for developers?
Jevons paradox is the 1865 observation that when a resource gets used more efficiently, total consumption of that resource often rises rather than falls. Applied to software: AI making coding cheaper doesn't necessarily reduce demand for code. It expands what's worth building. The bank teller case is the canonical example. ATMs cut tellers per branch from 20 to 13, but banks opened 43% more branches, so total teller employment grew. The pattern holds until automation gets complete enough to handle the whole job, which AI hasn't yet for software.
Should I learn AI to avoid being laid off?
Yes, but not because AI is taking your job. Because AI tools make you 3 to 5 times faster on boring parts of the work, which is now table stakes for senior engineers. The bigger lever is distribution. AI commoditized building. Distribution didn't get cheaper. A small audience and a track record of shipping are now more durable than any specific technical specialization.
Where do I start this week?
Three concrete moves. Audit your AI tooling (set up Claude Code or Cursor with project rules and verification gates). Pick one distribution channel and schedule three posts. Ship one open-source artifact with your name on it. The combination compounds faster than any individual technical certification.
I write Code Meet AI, a weekly newsletter for engineers shipping AI features in production. No hype, no thought-leader cadence. Real builds, honest takes, what's working in mobile-AI right now.
The boilerplate is at
The React Native foundation I use on every production project — open-sourced. Feature-first architecture, TypeScript strict, auth, i18n, theming, Redux Toolkit, and Expo SDK 54 with the New Architecture. Structured so Cursor, Claude Code, and Antigravity generate consistent code without hallucinating your patterns. Want the full version? RevenueCat, Firebase, U-AMOS 2.0 memory bank, and AI Pro features are in the paid tier.
Most React Native starters give you a blank canvas. That's fine for a side project — it's a liability on production work or when you're using AI coding tools. After 7 years of shipping React Native apps — enterprise clients, health tech, coaching platforms — I kept rebuilding the same foundation from scratch. Authentication, onboarding, theming, i18n, state management, folder structure, TypeScript config. Every time. This is that foundation, extracted and open-sourced. Three reasons…
chohra-med
/
expo_boilerplate
AI-first React Native + Expo boilerplate. Feature-first architecture, TypeScript, auth, i18n, theming, Redux Toolkit — with Cursor/Claude rules included. Lite version of AI Mobile Launcher.
MobileLauncher — React Native Boilerplate
→ AI Mobile Launcher — aimobilelauncher.com
Why this boilerplate?
MIT-licensed. Clone, fork, ship.