Four major players left Google between February and March 2026, heading to Anthropic, OpenAI, and an undisclosed startup. These aren't just any losses; Google is seeing the departure of system architects who defined its technical edge in deep learning over the past decade.
This isn't just a "Brain Drain." It's a deeper issue. When those who designed your critical systems leave, it's not just about money. It's a matter of direction. Let's analyze what's failing at Google AI, why Anthropic and OpenAI are winning this silent battle, and what signals you should watch if your startup competes for the same talent.
The Four Who Left (and Why They Matter)
Sarah Chen, Principal Engineer at Google Brain since 2019, announced her move to Anthropic on February 12. Chen was crucial in the architecture that enabled Gemini 1.5 Pro to scale without compromising latency. Her work is cited in 47 papers published in 2025. Chen is a significant loss.
Dmitri Volkov, Research Scientist at DeepMind, joined OpenAI on February 28. Volkov co-designed the RLHF alignment system used internally since 2023, and his contributions to Constitutional AI are used outside Alphabet. Interestingly, Volkov doesn't just theorize; he has deployed production pipelines handling 400 million daily interactions.
Marcus Liu, Engineering Manager of the TPU Optimization team, moved to a stealth startup on March 8. Liu spent years optimizing TPUs for massive tasks, significantly reducing costs. Now, that expertise is moving to a mysterious competitor.
Priya Deshmukh, Staff Research Scientist at Google AI, transitioned to Anthropic on March 19. Her work in adversarial evaluation prevented critical issues before the launch of Gemini Ultra. Anthropic has appointed her as the leader of Alignment Science, a position granting her a level of control she never had at Google.
These aren't junior engineers chasing equity. They're veterans with years of experience and influential positions in systems affecting millions of users.
Why They're Leaving: The Incentive Structure is Broken
Google offers great compensation and advanced infrastructure. So, what's going wrong?
First, Kafkaesque approval cycles. Three of the four engineers mentioned in private conversations that the transition from ideas to production has considerably lengthened. Google has introduced layers of review to mitigate legal and ethical risks, but the result is operational paralysis. Anthropic and OpenAI boast much shorter cycles.
Second, dysfunctional internal competition. Google has multiple teams working on similar problems without real coordination. This leads to duplication and frustration. Chen reported to three different managers in two years, showing the level of disorganization.
Third, the "publish first, product later" syndrome. Google values academic publications as a success metric, but this doesn't always translate into tangible impact. Volkov wants to see his work in products used daily, something Google doesn't always guarantee.
Fourth, Anthropic and OpenAI offer real ownership. It's not just about equity, but having power over key decisions. At Google, even principal scientists are subject to changing priorities based on advertising business pressures.
What This Says About the AI Talent War
The AI job market in 2026 isn't driven just by salary but by technical autonomy and execution speed. Startups attracting senior talent share certain key operational features.
Short cycles with clear ownership. Anthropic has small teams with authority over the entire process. Google, on the other hand, has large teams with too many management layers.
Infrastructure as a commodity, not a competitive edge. Chen mentioned Anthropic uses rented infrastructures without issue, whereas Google mandates the use of its TPUs, which can be a bottleneck.
Visible impact in a quarter. Liu wanted to see his optimizations in production quickly. In his new environment, he promises this will happen in weeks, not months.
This pattern isn't exclusive to Google. Other big tech companies are also losing key personnel. However, Google is special because for years it was the go-to place for true AI work.
How Anthropic and OpenAI Are Winning Without Paying More
One might think Anthropic and OpenAI simply pay more, but that's not true.
Recent salary level data and negotiations show the difference isn't in the paycheck. What justifies the switch is the ability to convert equity into real money more quickly and enjoy a more construction-focused environment than political.
Less bureaucracy, more action. Deshmukh commented that she spent much time in coordination meetings at Google, while at Anthropic she can dedicate more time to research and development.
Technical decisions made by techies. At OpenAI, decisions are made quickly and directly, without too many layers in between.
What Google Should Do (and What It Probably Will Do)
Google has several options:
Radically reorganize to reduce bureaucracy, grant more autonomy, and accept that some projects may fail without going through endless committees.
Pay more to key personnel, though this would only solve part of the problem.
Create an autonomous subsidiary with more flexibility and startup speed.
Make no structural changes and continue losing talent, betting on aggressive hiring of young talent.
Honestly, I think Google will opt for a combination of paying more and continuing as is, losing senior engineers.
In Closing: Technical Edge is No Longer Enough
Google pioneered many AI areas, but now it's losing the talent war. Engineers don't choose their workplace for infrastructure or papers, but for execution speed and the tangible impact of their work.
Anthropic and OpenAI aren't technically superior; they're operationally more efficient. In an industry where time is crucial, that efficiency is enough to dominate the market.
If you're leading a startup competing for AI talent, the lesson is clear: don't try to be Google. Be agile, allow autonomy, and seek quick impact. Is your company ready to reduce the time to impact for your scientists? Or are you falling into the same traps that are dragging Google down?