Physix Frontier
Chip Talent War: AI Faith Is Shifting
Just like during the height of the gold rush, the most profitable people weren't those digging for gold, but those selling shovels. Now, the shovel sellers have started poaching each other's talent.
Since the beginning of this year, the talent war in the chip industry has intensified. TSMC, Samsung, and Intel are hiring aggressively, especially for technical experts in advanced processes and packaging. Salaries have doubled overnight, with signing bonuses exceeding the average for Silicon Valley software engineers. The essence of this scramble isn't a battle over technological routes, but a bet on "future computing infrastructure"—whoever controls more talent gains an advantage in the next round of chip density competition.
At the same time, another piece of news is brewing: The AI hype is cooling down. Multiple venture capital firms have lowered valuations for large model startups, monthly active user data for previously hyped "AI-native apps" is dismal, and the financing environment has tightened noticeably.
"We are experiencing a typical 'trough of disillusionment' following the 'peak of inflated expectations.'"
— A partner at a top-tier USD fund, speaking at a closed-door meeting
Viewing these two events together is interesting. On one hand, the hardware foundation layer is frantically grabbing talent, with capital sparing no expense; on the other, the application layer is deflating bubbles, and investors are becoming cautious. This tells the market: AI faith is shifting anchors, from "telling stories" to "looking at infrastructure."
Temperature Difference Between Two Markets
Comparing talent flow in these two sub-sectors reveals the temperature difference:
| Sector | Talent Demand | Salary Growth | Financing Heat |
|---|---|---|---|
| Chip Design/Manufacturing | Extremely scarce, especially for <3nm processes | 30%-50% annual growth, stock options for senior engineers | Government subsidies + industrial capital, continuous reinforcement |
| Large Model Applications | Approaching saturation, too many PMs | Flat or declining, options hard to cash out | Financing rounds lengthened, valuations corrected |
This temperature difference isn't accidental. It reflects a shift in underlying logic: The boundaries of AI are no longer determined by algorithmic breakthroughs, but by the physical limits of chips. As large models move from "parameter races" to "inference cost-efficiency," whoever can run faster and more energy-efficiently on chips will truly succeed in deployment.
The image above comes from an MIT TechReview report, showing a real scene from a chip design lab—the people in cleanroom suits are becoming the scarcest assets in the entire AI ecosystem.
Essence of the Talent War: Betting on "Can It Be Built"
This round of chip talent wars is completely different from internet companies fighting for programmers ten years ago. Internet companies fought for "traffic-driven" operational capabilities, while chip talent fights for "physical limit" engineering capabilities. Improving yield by 1% in a 3nm process can bring billions of dollars in value. Such talent requires over a decade of accumulation and cannot be rushed.
Therefore, TSMC building factories in Arizona, Intel poaching talent in Europe, and Samsung setting up R&D centers in Silicon Valley—these actions appear commercially competitive but are actually battles for control over "manufacturing capability." Behind this control lies the security of the AI computing foundation.
[!tip]
If you focus on hard-tech investment, keep an eye on this metric: Chip design talent inflow/outflow ratio. When a company sees net engineer outflows for three consecutive quarters, it indicates problems with its technical route or management, serving as an early warning sign before financial reports.
AI Recession: Not Bubble Burst, But Gear Shift
The cooling of AI hype doesn't mean AI lacks value. On the contrary, the recession indicates the market is re-evaluating using a "deployment index." Previously, a PowerPoint presentation could secure funding; now, quantifiable ROI data is required.
Here is a key comparison: Large Model Companies vs. Chip Companies.
The biggest problem facing large model companies currently is "homogenization": GPT, Claude, Llama, Gemini—the differences in core capabilities are shrinking, user switching costs are extremely low, leading to fierce price wars and dismal gross margins. Chip companies, especially those offering "custom AI chips + advanced packaging" solutions, have much deeper moats—because replacing a chip solution requires redesigning the entire inference stack, which is prohibitively expensive.
So, talent flowing from the application layer to the hardware layer is voting with their feet: Over the next decade, the core incremental value of AI may come from chips, not models.
Actionable Advice for Readers
If you are a tech professional or investor, the most rational choice right now is not chasing the next funding round of any large model company, but focusing on these three directions:
1. Learn chip design or EDA tools: There is a huge shortage of basic software talent, and salary growth is stable.
2. Focus on the intersection of "Chips + Applications": Chip companies specializing in AI inference acceleration, such as RISC
Original Link: https://www.technologyreview.com/2026/07/29/1140884/the-download-chip-talent-battle-deflating-ai-hype/