VC firm SignalFire's latest "State of Talent Report" contains a counterintuitive statistic: in 2025, overall hiring by major tech companies dropped 25% compared to 2019, yet the share of engineers among all new hires surged from 46% to 55%. NVIDIA CEO Jensen Huang also stated bluntly that after the entire company uses agentic AI, "software engineers are busier than ever," and replacing engineers with AI is "complete nonsense." At a time when the "AI replacement" narrative is rampant, the strong resilience of engineering roles starkly contrasts with the cliff-like decline faced by other foundational positions. What exactly is going on?
Shrinking Tech Giant Headcounts and the Counter-Trend Expansion of Engineers
Over the past two years, layoff news in the tech industry has been almost nonstop. From social media to financial outlets, AI is frequently cited as the primary reason tech companies are reducing headcount. However, SignalFire's tracking of 12 tech giants including Alphabet, Meta, Apple, Amazon, Microsoft, and NVIDIA reveals a fact obscured by emotional narratives: big companies aren't stopping hiring; they are undertaking extreme structural optimization.
In 2025, overall hiring by tech giants dropped 25% compared to 2019—a macro-level signal of hiring pullback. But if you break down hiring by function, the divergence is extreme. Engineering hiring fell only 11%, far below the average. The more critical shift is in proportion: engineers as a share of all new hires rose from 46% in 2019 to 55%. That means for every two hires at a tech giant, more than one is an engineer.
In sharp contrast to the resilience of engineering roles is the shrinkage of junior and non-technical functions. The SignalFire report notes that new graduate hiring at Big Tech accounts for only 7% of hires, plunging more than 50% from pre-pandemic 2019 levels. Meanwhile, non-technical functions such as recruiting, product, and sales continue to shrink. Early-stage startups show the other extreme, hiring 7% more engineers than in 2019.
The underlying logic of this structural optimization is not complicated. Before the explosion of generative AI, tech companies maintained large workforces to support business expansion and feature iteration. Large numbers of junior engineers and execution staff handled basic development tasks, test case writing, and routine maintenance. When AI tools can accomplish foundational coding, copywriting, customer service responses, and even some sales lead screening with extreme efficiency, companies no longer need massive junior execution teams. On the contrary, the need to deeply integrate these AI capabilities into existing business lines and build AI-powered products has caused demand to surge for senior engineers with system architecture skills and experience using AI tools. Asher Bantock, SignalFire's head of research, states that if AI were truly replacing engineers, engineering hiring would be the first to plummet—but reality is the opposite.
For investors and industry observers, the signal is clear: tech giants' capital spending and human resource allocation are shifting from "horizontal expansion" to "vertical deepening." In the past, it was about piling on headcount to grab market share; now, it's about concentrating resources to build the underlying infrastructure and application ecosystem for the AI era. Engineers are no longer mere cost centers but the leverage point determining whether a company survives the AI wave.
The More Efficient, the More Staff Needed: Jevons Paradox Validated in the Code World
Why hasn't the demand for engineers decreased even as AI boosts coding efficiency? SignalFire's analysis suggests we are seeing a classic case of Jevons Paradox.
In the 19th century, economist William Stanley Jevons studied coal consumption and found that improvements in steam engine efficiency did not reduce coal use; instead, because steam engines became more economical and practical, they were applied more widely, leading to a surge in global coal consumption. This paradox is now being perfectly validated in the world of code.
AI-assisted coding tools are driving the cost and time of writing basic code toward zero. Companies haven't laid off programmers and reduced development volume; instead, because "software has become extremely cheap," they are integrating AI across all business lines. Demands that were shelved due to high development costs are being unleashed en masse, and system complexity is rising exponentially. The boundaries of engineers' work are being infinitely broadened.
Jensen Huang elaborated on this shift in recent public remarks. He pointed out that after the entire company adopted agentic AI, the AI is "micromanaging" employees, and software engineers are "busier than ever." AI hasn't replaced engineers; it has pushed them toward higher-order creative and architectural work. Engineers are no longer mere "code typists" but have become "AI foremen" and system architects, needing to handle more agent orchestration, system integration, and code review tasks.
In this process, a key technical concept is reshaping engineers' daily workflows: Harness. In the current AI engineering context, Harness typically refers to "toolchain, scaffolding, or orchestration framework." It is responsible for encapsulating, scheduling, and landing the general capabilities of large models into specific business flows. As the capabilities of large models themselves become commoditized, the main competitive battleground has shifted to the Harness layer outside the model. One of the core tasks of engineers is to build and maintain these Harness systems, ensuring that AI agents can run stably and securely within enterprise environments. This involves a vast amount of interface integration, permission control, exception handling, and context management tasks, greatly increasing system architecture complexity.
This shift is microscopically corroborated by tool popularity and developer sentiment analysis on OmniTools. From 2024 to 2025, views and bookmarks for AI coding and code assistant tools on the site showed a steep upward curve. But the sentiment distribution in user comments is not uniformly optimistic. Positive feedback focuses on "no more hand-writing basic CRUD code" and "prototyping speed doubled." Negative feedback zeroes in on changes in work intensity.
On developer communities like Reddit and Hacker News, many senior developers expressed frustration with Jensen Huang's "engineers are busier" comment. They point out that while AI saves time writing basic code, it brings an endless stream of "crap code reviews." AI-generated code often contains hallucinations and hidden logic flaws, requiring engineers to spend significant effort on complex prompt debugging and system-level bug hunting. Specifically, when AI-generated code snippets are integrated into large projects, issues like context loss or type mismatches often arise. Engineers cannot blindly trust this code; they must review it line by line for logical correctness, security, and performance impact. The cognitive load of such review is far higher than writing from scratch, because you need to understand the AI's "black box" reasoning path.
Furthermore, debugging AI agent workflows is an extremely mentally taxing task. A complex agent may involve multiple dialogue rounds, tool calls, and external API interactions. When an agent behaves abnormally, it's difficult for engineers to set breakpoints and step through as with traditional code; instead, they must analyze large volumes of logs and intermediate states, and even repeatedly modify system prompts to correct the agent's behavioral deviations. The nature of work has shifted from repetitive "manual labor" to high-intensity "mental overdrive." Engineer headcount demand hasn't decreased, but cognitive load per unit time has increased significantly.
Some developers are also pessimistic about the long-term validity of Jevons Paradox. They believe that when AI coding capabilities cross a "singularity"—not only writing code but self-validating, self-debugging, and understanding global business logic—Jevons Paradox will break down, because AI will become the "architect" itself, ultimately reducing absolute demand for humans. But at least in 2025, AI remains a lever that requires human engineers to wield, not an independent creator.
Plummeting Basic White-Collar Jobs: Who Is AI Actually Replacing?
The counter-trend expansion of engineering roles is just one side of the AI employment shockwave. The other side is the tangible replacement of basic white-collar jobs.
Third-party data firm Bloomberry, based on Revealera's tracking of 5 million public posts on the Upwork platform, shows that within 15 months after ChatGPT's release, writing jobs on the platform dropped 33%, translation dropped 19%, and customer service dropped 16%. In stark contrast to these declining categories, back-end development jobs grew 6%, front-end grew 4%, and chatbot development demand surged 2000%. While this third-party analysis doesn't represent the entire job market, it accurately reflects real changes in the most sensitive nerve—the freelance marketplace.
Why are these three roles hit hardest? The core lies in the degree of task structure. The basic tasks of writing, translation, and customer service are highly rule-defined and repetitive. AI handles such information conversion and rule-matching tasks at extremely low marginal cost and with quality already at a usable level. Companies find that using AI for these foundational tasks is far more cost-effective than hiring humans.
This replacement is happening not only on freelance platforms but also quietly within enterprises. Anthropic CEO Dario Amodei has warned that AI could eliminate 50% of junior white-collar jobs within five years. This is not alarmist, but an extrapolation based on current AI agent capabilities.
In concrete business scenarios, AI agents are substantively taking over traditional execution teams. Take the sales domain: OmniTools previously analyzed the case of AI employee Viktor. With no human sales team, this product used AI agents to secure 30,000 enterprise customers and generate $20 million in revenue. It replaced traditional junior sales, SDRs (Sales Development Representatives), and implementation teams. The core tasks of these roles—following standardized scripts to follow up on leads, entering data, answering basic questions—are precisely the AI agent's home turf. Viktor can handle massive volumes of leads around the clock and automatically adjust communication strategies based on customer feedback, with efficiency and consistency unattainable by human teams.
In the administrative and HR fields, a similar erosion is taking place. The rise of office agents like Tencent WorkBuddy has opened a crack in reshaping basic support positions. Process-oriented tasks such as initial resume screening by HR, daily schedule coordination, and reimbursement review are being taken over by office agents. Headcounts aren’t being directly slashed; instead, staff reductions are achieved through natural attrition and “invisible compression.” Employees in these roles are never explicitly told, “You’ve been replaced by AI,” but their job content is gradually stripped away by agents, ultimately rendering the position without value. When 80% of an HR person’s daily work can be done by sending a single instruction to WorkBuddy, the continued existence of that headcount depends solely on whether the remaining 20% of non-standardized tasks are worth retaining a full-time employee.
62% Premium and Elimination: Two Tracks Split Apart by AI
The macro labor market has not collapsed under the impact of AI. Instead, it exhibits what PwC defined in its 2026 Global AI Jobs Barometer as a “two-track labor market.”
PwC data shows that roles requiring AI skills are growing 8 times faster than the overall market (69% vs. 9%). More critically, compensation is diverging—AI skills command a salary premium as high as 62%. This is not a short-term dividend from simple technological upgrades, but a restructuring of the labor value assessment system.
On the first track are those who wield AI as a lever. They could be senior engineers proficient in AI coding tools, or business architects capable of designing complex Agent workflows and Harness systems. AI amplifies their output capacity, allowing one person to complete the workload of several, so companies are willing to pay a higher premium. The Silicon Valley giants mentioned in the SignalFire report that choose to cut overall hiring but protect engineering headcount are essentially paying for this leverage effect. What companies pay for is no longer labor cost by the hour, but an investment in system design capabilities and complex problem-solving skills.
On the second track are those engaged in “clearly defined rule-based” tasks that AI can easily automate. Their work is replaced by AI at extremely low marginal cost, causing their demand to plummet off a cliff. The shrinking of writing, translation, and customer service roles on the Upwork platform is a true portrait of this track. When a task can be clearly broken down into input, processing rules, and output, the irreplaceability of human labor approaches zero.
This is no longer a competition between humans and AI, but a competition between “those who use AI” and “those who do not use AI,” a competition between high-leverage roles and low-leverage roles. The market is being split apart by AI, and the middle ground is disappearing. For industry observers, this means that future enterprise organizational structures will tend toward a “dumbbell shape”: one end consists of a very small number of core architects and strategy makers, the other end is the AI system itself, and the vast middle layer of execution will be dramatically thinned out.
No Empty Consolations: Career Deadlock and High-Intensity Survival in the AI Era
In the face of such polarization, vague “career survival guides for the AI era” or empty consolations like “humans will eventually triumph over AI” are meaningless. We need to confront the harshest realities in today's job market.
First is the career deadlock brought about by the “experience paradox.” The SignalFire report highlights a brutal phenomenon: companies are all hiring senior ICs (Individual Contributors) who can deliver output independently. To save costs and pursue efficiency, they essentially use senior engineers to fill junior positions, or simply let AI handle the basic work. This leaves fresh graduates and career switchers facing a deadlock of “can't get a job without experience, can't get experience without a job.”
The impact of this deadlock on junior developers is fatal. In the past, junior developers became familiar with the codebase and business logic by writing basic code and fixing simple bugs—a traditional “apprenticeship” training model. Now, these basic tasks are taken over by AI, depriving junior developers of the soil needed to accumulate experience through real-world practice. Companies expect new hires to immediately jump into AI code review and Agent orchestration, but these high-order skills precisely require massive amounts of foundational experience as their support. The plunge in new graduate hiring ratios by over 50% is not just a numerical decline; it represents the break in the traditional talent development pipeline. How to break this deadlock is a systemic risk the entire industry must face in the next five years. If a new mechanism for incubating junior talent cannot be established, the industry will face a crisis of bottom-level talent depletion.
Second is the reality that while engineers are being hired in larger numbers, their work intensity is being amplified indefinitely. Jevons Paradox has secured the jobs of engineers but has not guaranteed their quality of life. The shift from “physical labor” to “mental overexertion” is no blessing. Reviewing hallucinated AI code, debugging complex Agent workflows, and handling the exponentially growing demands of system integration—these high cognitive load tasks are reshaping the daily life of engineers. The sense of security in the AI era does not come from ease, but from irreplaceable architectural capability and high-intensity human-machine collaboration. Engineers need to adapt to the role shift from “creator” to “reviewer” and “orchestrator,” which demands higher abstract thinking ability and a system-level perspective.
AI has not killed off engineers; instead, it has turned them into the core assets most heavily relied upon by tech giants. However, it has precisely struck down those basic white-collar jobs with clearly defined rules and high repetitiveness, and in this process has split open a two-track labor market. In this market, the middle state is vanishing, and professional value is being redefined. Understanding this polarization is more important than blind panic or blind optimism.
