Mountain View - Top AI researchers are quietly slipping out of Google, frustrated that internal battles over server time and custom chips have turned compute into a scarce, career-defining resource. For engineers used to something closer to academic freedom, being told to wait in line for TPUs and large training runs has become a breaking point. Several former employees say they chose to launch startups instead, simply so they could buy or otherwise guarantee the compute they need.
The departures include people who helped build AlphaGo and contributed to Gemini, and they stem from complaints that computing power is steered toward high-priority projects and paying cloud customers, according to the Los Angeles Times. Inside Google, researchers say access to tensor processing units and other accelerators often determines which ideas get funded and which are quietly shelved.
Backlog and a 'Compute Constrained' Reality
Alphabet's own numbers highlight the squeeze. Google Cloud's backlog nearly doubled to more than $460 billion, and CEO Sundar Pichai told investors that "we are compute constrained in the near term," according to Alphabet's Q1 2026 earnings call transcript. Executives have flagged heavy capital spending to add servers and chips, but data-center buildout and chip lead times mean actual supply is poised to lag demand for months.
Startups Offer Predictability and Chips You Can Actually Buy
Former DeepMind researcher Andrew Dai said he left after concluding he could not secure enough compute inside Google to pursue a visual-reasoning project, and other ex-employees say founding companies gives them clearer, paid access to hardware. "No one's going to take it away from you," one founder told the Los Angeles Times, describing the appeal of purchasing compute directly instead of begging for time on internal clusters.
Startups That Scored Compute and Cash
Some startups spun out of Google have quickly landed both money and access. Ricursive Intelligence, formed by former Google researchers Anna Goldie and Azalia Mirhoseini, has raised about $335 million as it builds AI tools for chip design, according to TechCrunch. ReflectionAI, co-founded by Misha Laskin and Ioannis Antonoglou after leaving DeepMind, emerged with substantial early backing and an open-model pledge, as reported by Verdict.
How Google Says It Is Responding
Google says it is formalizing how compute is handed out, creating executive councils and allocation processes to balance Cloud deals, product priorities and long-term research, according to reporting in The Information. The company is also ramping up capital expenditures and preparing to deliver TPU hardware to customers in an effort to ease internal competition for accelerators.
What This Means for the Bay Area
The near-term result is more founders and venture dollars flowing out of the Google campus and into Bay Area startups that can control their own stacks. That shift could reshape which ideas get time and compute, and which remain the risky, long-shot experiments that once had a better shot inside big labs.
The compute debate is likely to be part of the backdrop when Google rolls out its next wave of AI work at its developer conference in Mountain View on May 19–20, where the company is expected to spotlight Gemini and product integrations that depend on expanded hardware capacity, according to Tom's Guide. For now, the shortage is a pointed reminder that in modern AI, chips can shape research as much as bright minds do.
