This article was first published in Shadow Memo

Written by Mo Yingsheng

Just halfway through 2026, layoff figures in Silicon Valley are already enough to write a tragic history of the tech industry. In January, Amazon confirmed approximately 16,000 layoffs; in February, fintech company Block laid off nearly half of its workforce; in March, Meta was reported to plan laying off 16,000 employees.

In the first quarter of 2026, the tech industry laid off about 81,700 people, marking the highest single-quarter record in two years.

In the same quarter, the four largest hyperscale cloud service providers invested $725 billion in AI infrastructure, a 77% year-over-year increase.

The money didn't disappear; it just changed hands—flowing from payrolls to GPU clusters.

In this swiftest capital migration in human history, the first to fall were precisely those who taught AI to write code.

When Capital Flows from Payrolls to GPUs

The capital that once supported engineering hiring in 2021 now powers GPU clusters. This sentence encapsulates the most brutal structural transformation in the tech industry over the past two years.

In 2024, the global tech industry laid off 153,000 people; in the first three months of 2025 alone, the figure exceeded 114,000. In the first half of 2025, 77,799 tech jobs globally were marked as 'directly eliminated due to AI-driven layoffs.'

An estimated 230,000 layoffs globally were publicly attributed to AI-driven workforce reductions.

Behind these numbers lies a fundamental shift in enterprise IT budget allocation: traditional software development investment dropped from 67% in 2020 to 38% in 2025, while AI infrastructure spending surged to 41%.

This is not a cyclical market correction akin to the dot-com bubble burst. Salesforce's CEO explicitly told investors he didn't need as many people.

Meta framed layoffs as reallocating compensation budgets toward AI; Oracle eliminated 30,000 traditional roles to fund data center construction. Workplace discourse shifted from 'financial discipline' to 'technological substitution.'

After laying off 10,000 employees last year, Oracle planned to cut another 20,000–30,000 starting in March, representing about 18% of its workforce.

Forty-seven database administrators were replaced by AI, with only three senior engineers retained for oversight. While net profit soared 95% year-over-year, 'indiscriminate layoffs' occurred regardless of performance or rank.

Under the brutal reality of 'humans inferior to AI,' tech companies now view humans not as core assets generating revenue but as historical burdens desperately needing shedding.

A Moody's Analytics report revealed two possible paths: either AI drives productivity to soar, reducing unemployment to 3.8%, or the bubble bursts, leaving 4.6 million jobless.

Technologists predict AI will contribute 3%-30% annualized productivity gains and trigger mass unemployment, while economists forecast just 0.07%-0.9% with a smooth labor market transition—a 40-fold discrepancy.

Regardless of which path materializes, programmers stand at the forefront of the storm.

Tech Giants' 'Code Automation'

From 15% to 75% in Just Two Years

In April 2026, Google CEO Sundar Pichai announced a staggering figure at Google Cloud Next: nearly 75% of new code internally was generated by AI and reviewed by human engineers.

Two years prior, this figure stood at just 15%. It crossed the 25% threshold in October 2024, climbed to 50% by autumn 2025, and now, just six months later, has surged toward three-quarters.

Industry analyses previously projected AI code generation might reach 30%-40% by the end of 2026; Google's execution speed shattered these expectations.

Alphabet no longer positions AI as a supplementary tool atop engineering work but as the default engine behind most new code creation.

Google is not alone.

Tencent announced at its AI Industry Application Conference in June that most code is now AI-generated, with engineers focusing more on architectural design while delegating coding tasks to AI.

Data shows over 90% of Tencent's engineers use AI programming assistant CodeBuddy, with 50% of new code AI-assisted in 2025. During code reviews, AI participation reaches 94%.

Meta set an even more aggressive roadmap: by the first half of 2026, 65% of engineers will use AI to write over 75% of their code.

Zuckerberg bluntly stated, 'Projects that once required large teams can now be accomplished by a single talented person plus AI.' He mandated AI proficiency assessments for all employees; those unable to use AI face imminent departure.

Snap announced at least 65% of new code would be AI-generated. Its stock jumped nearly 8% that day.

Expanding the perspective from Silicon Valley giants to the entire industry, the data becomes even more staggering. Global code quality platform Sonar's 2026 Developer Survey Report revealed 72% of developers use AI programming tools daily, with AI-generated or -assisted code accounting for 42%—a sharp rise from 6% in 2023.

Global developer adoption of AI tools climbed from 76% in 2024 to 84%. GitHub Copilot's cumulative users surpassed 20 million.

Tech giants are handing keyboards to AI at speeds exceeding 'tripling every three quarters, quintupling in two years.' The very identity of programmers is being rewritten, with roles rapidly shifting from 'code writers' to 'code reviewers.'

Token Economics: Programmers Become AI Fuel

If code automation represents the 'overt path' of AI replacing programmers, token consumption is a far more concealed yet devastating 'covert path.'

This year's AI boom stems from a token explosion, with surging volumes and prices. Anthropic's annualized revenue skyrocketed from $14 billion last year to $47 billion.

Tokens are no longer just units for model invocation; they've become employees' 'second salary.'

NVIDIA CEO Jensen Huang proposed a ruthless logic in a March 2026 podcast: if an engineer can spend half their annual salary on token budgets, it proves they're efficiently driving AI to work for them.

Thus, big tech began tying token consumption to employee performance, creating new class divisions among programmers: 'token middle class' and 'token poor.'

Those using fewer tokens face reprimands, even absurd scenarios where programmers' monthly token expenses far exceed their salaries.

A backend engineer in Seattle described this dysfunctional ecosystem: 'Our internal performance metrics now heavily favor AI. The most direct indicator is code volume—we have a leaderboard refreshing daily, showing everyone's lines of code. Performance is directly tied to this.'

'If you don't use AI, you can't compete.' 'We have two overachievers who generate ten times more code annually than average. They rely entirely on AI, submitting 500 PRs in three months and setting the entire department ablaze.'

However, the token frenzy soon collided with reality.

According to developer productivity platform Entelligence.AI's survey, every $1 invested in AI tokens incurs $0.44 in bug-fixing costs and $0.27 in code rewriting costs, with another $0.11 lost to review and merge delays.

This means nearly 80% of spending vanishes as invisible overhead.

More cruel ly, model evaluation and risk research institute METR's controlled experiments showed developers using AI programming tools actually experienced a 19% decline in real work efficiency .

Engineering efficiency and developer productivity platform DX's survey of over 450 enterprises and 120,000 developers revealed that despite 93% using AI programming tools, real work efficiency gains stalled at just 10%.

Microsoft quietly canceled most Claude Code licenses, while Amazon shut down its internal token consumption leaderboard. Silicon Valley's shift from 'all-in on AI' to covertly applying brakes unfolded in just months.

But it was too late; programmers had already become the first sacrificed link in the token economic system.

The Alienated Programmer

From Creator to Overseer

Has AI reduced programmers' workloads?

The answer is quite the opposite.

An open-source project maintainer lamented: 'I used to handle 20–25 code PRs weekly. Now that number has exploded to over 100, mostly AI-generated. Yet I must meticulously review each request.'

Some compare today's AI revolution to the sweatshops of the early Industrial Revolution: 'Steam engines boosted productivity exponentially but didn't liberate workers—instead, they enabled more extreme exploitation. Machines ran nonstop, so workers had to too, working longer hours.' 'We're now like workers tending machines—they can't stop, so we can't rest.'

While AI multiplies code generation efficiency, review and validation efficiency hasn't kept pace. 'It's like a factory replacing a part-stamping machine with one ten times faster but keeping just one quality inspector at the end of the line. Production soars, the inspector's workload doubles, yet defect rates remain unchanged. Ultimately, the person bearing all review pressure collapses.'

AI-driven productivity gains haven't translated into free time for employees but into higher corporate expectations.

Before AI, a software engineer submitting 20 code pull requests weekly was standard; with AI assistance, theoretical output capacity rises to 50, so companies set 50 as the new standard.

Worse, AI is eroding training pipelines for junior developers. Microsoft Azure CTO Mark Russinovich and Developer Community VP Scott Hanselman argued in a peer-reviewed paper that AI coding tools with autonomous decision-making capabilities are creating structural crises in software engineering.

While AI dramatically boosts senior engineers' efficiency, junior developers lack the judgment to guide, validate, and integrate AI outputs.

The result: companies hire senior engineers while automating junior roles, collapsing the talent pipeline for next-generation senior engineers.

The data is alarming. Since 2022, hiring for junior developers has dropped 67%. Harvard research found that after GPT-4's release, employment rates for 22–25-year-olds in AI-related fields like software development fell about 13%.

Stanford economists' 2025 study also showed a nearly 20% decline in software developers aged 22–25 since 2022. Over the past three years, large tech companies reduced hiring of new graduates by 50%.

A 2025 MIT study found adults outsourcing coding tasks to ChatGPT exhibited reduced brain activity and worsened memory—a phenomenon researchers dubbed 'cognitive debt.'

'Programming ≠ software engineering.' The judgment required to identify defects in AI-generated code is precisely the skill junior developers should cultivate through real-world production work.

When AI eliminates entry-level jobs junior developers rely on for learning, the pyramid's base crumbles.

Even those at the pyramid's peak lose direction. Django co-founder Simon Willison, who has coded for 25 years, openly admitted losing all ability to estimate project timelines.

Programmers are transforming from 'code writers' into 'AI overseers,' endlessly reviewing, revising, and repeating—trapped in a cycle of generation, review, and regeneration. Workloads have increased tenfold while professional fulfillment approaches zero.

After the Bubble Bursts: Who Pays for Excess?

What is the essence of the AI bubble?

The five AI giants committed $680 billion in capital expenditures for 2026. The top ten tech companies issued $120 billion in debt in 2025, soaring 167% year-over-year.

But where is the money going? Model training, data centers, and securing photolithography machine capacity. Where does it come from? Borrowed funds. Can it be repaid? Highly uncertain.

More dangerously, capital circulates idly within the ecosystem. NVIDIA invests in Oracle, which uses the funds to buy NVIDIA chips; Microsoft injects $13 billion into OpenAI, which pays Azure for cloud services.

If any intermediary link fractures, this circular financing collapses instantly.

While retail investors crazy bought in, the Magnificent Seven executives net-sold $8.4 billion in stocks over the past year. Zuckerberg sold, Bezos sold, Huang sold. Current stock P/E ratios have surged to 20x, just 4 points below the 24x peak during the 2000 dot-com bubble burst.

Moody's outlined a bubble-bursting scenario: In a 2026 quarter, if AI revenue growth falls short of expectations, panic selling could trigger a 25% stock market crash, erasing $20 trillion in value.

The burst of the bubble is a risk for the future, but unemployment is the pain of the present.

In January 2026, the U.S. saw 108,000 layoffs in a single month, breaking the highest record since the 2009 financial crisis. Those precisely targeted by AI are precisely the middle class with wages in the 60%-80% percentile—accountants, programmers, and junior analysts.

Gartner even predicts that by 2027, half of the companies that lay off employees due to AI will rehire employees for similar positions. This phenomenon of “laying off and then rehiring” is known in the industry as a “boomerang.” But how many of those laid off will return?

Anthropic’s report clearly states that the profession of programmer will not disappear, but those programmers who “only know how to write code” will gradually be eliminated by the market.

AI is rewriting the rules of the game in software development, compressing projects that originally took months to complete into just two weeks. But during these two months, a full development team is no longer needed.

The founder of a Silicon Valley startup announced on social media: “I fired the entire development team and replaced them with o1, Lovable, and Cursor. Now, I deliver products 100 times faster, and the code is 10 times cleaner.”

This is the reality programmers face: you teach AI to write code, and then AI replaces you.

In Closing

Programmers were once the builders of the digital age, the creators of internet miracles. They used code to build Google’s search box, Meta’s social graph, and Tencent’s instant messaging.

And now, they are using the same code to train AI that will replace them.

This is not alarmist.

From Google’s 75% to Tencent’s “most code generated by AI,” from Meta’s 65% of engineers using AI to write 75% of the code to Snap’s 65% of new code generated by AI, the numbers don’t lie.

The first victims of the AI bubble are precisely those who embraced AI first.

They thought AI was a tool, only to find that they were the tool, used to train AI, review AI, pay for AI’s Token bills, and ultimately be replaced by AI.

When capital flows from payrolls to GPUs, when programmers go from creators to supervisors, when 22-year-old graduates find that the programming skills they learned over four years are already obsolete.

What we are witnessing is not just a technological transformation, but the collective disorientation of a professional group.

The AI bubble of 2026 will eventually fade. But not all of those swept away by the bubble will return.