Google’s AI division has gone through a fresh wave of personnel upheaval, with ABMedia reporting that several core researchers, including a co-lead of Gemini, have left the company. A new episode of the All-In Podcast argued that the latest AI brain drain may reflect more than a fight for talent. At the center of the debate is a capital allocation question: whether Google is better off spending tens of billions of dollars to chase the world’s top model, or putting that money into data centers and renting out AI compute.
David Friedberg described the trade-off in financial terms. In his framing, data center capex is “high alpha, low beta,” while frontier model research is “high alpha, very high beta.” The distinction is simple: both can produce large upside, but infrastructure offers returns that are easier to model, while cutting-edge AI research comes with much higher uncertainty.
SemiAnalysis says Gemini is slipping while GCP grows
SemiAnalysis recently published a piece titled Gemini is Cooked but GCP is Cooking. Its argument was that Google’s frontier model position is under pressure as the company reshuffles AI leadership and loses prominent researchers. At the same time, Google Cloud and TPU-related business are expanding quickly.
The report pointed to a particularly awkward dynamic. Google is supplying large amounts of AI compute to outside customers such as Anthropic. That creates a situation where Google sells TPU capacity to Anthropic, Anthropic uses Google compute to build Claude, and Claude then competes with Gemini.
That leads to the central question raised by SemiAnalysis: should Google reserve scarce compute for DeepMind and Gemini training, or sell that capacity to outside customers willing to pay a premium?
All-In: data centers may be the better bet
Friedberg said Google plans to spend about $200 billion in capex this year, with a large share going toward AI infrastructure and data centers.
His case is that demand for AI compute remains extremely high, and Google already knows how to run large-scale data centers. If the company puts money into infrastructure, it can more easily find enterprise customers to lease that capacity, and the return on that investment is easier to predict.
Frontier models, by contrast, also require tens of billions of dollars, but the business outcome is far less certain. OpenAI, Anthropic, Chinese AI companies, and open-weight models are all moving fast. A model that ranks first today could be overtaken within months. From a boardroom perspective, Friedberg said he would allocate more capital to compute infrastructure than to model development.
Google Cloud and DeepMind are competing for the same TPU pool
Altimeter Capital founder Brad Gerstner said Google has a built-in channel conflict. Google Cloud wants more compute because it can lease that capacity to customers such as Anthropic. DeepMind researchers want the same resources to train Gemini and compete with Anthropic.
Gerstner’s view is that the balance currently appears to be tilting toward the infrastructure side of the business. He suggested that this may help explain why top AI researchers are leaving.
For Google shareholders, putting the next $10 billion into data centers that can generate steady revenue may be entirely rational. For researchers who want to build superintelligence, cure cancer, or push the scientific frontier, the picture looks different. If the company’s most important resources are gradually moving toward infrastructure, leaving to start a new venture may become more attractive.
Friedberg added that in the current AI capital market, top researchers may be able to raise billions of dollars with little more than a pitch deck.
Sacks sees a two-tier AI market
David Sacks went a step further and said the companies at the front of frontier intelligence may be narrowing to OpenAI and Anthropic.
He outlined a two-layer market structure:
- The first layer consists of the leading frontier models. Because high-end intelligence remains scarce, these models can charge a premium.
- The second layer consists of a large number of models trailing the frontier by roughly 6 to 12 months. That segment is moving toward commoditization.
In that second layer, Sacks said, demand will still be large, but as more open-weight models emerge, enterprise spending may shift away from the model itself and toward compute, inference, and deployment services.
He compared the setup to Apple and Android: Android has scale, but Apple captures more profit through a premium experience.
Google may not need to win the model war
Jason Calacanis was less convinced by the two-player framing. He argued that Google’s biggest advantage is not benchmark leadership. It is distribution through Search, Android, Chrome, Gmail, and YouTube. Even if Gemini is not the top model, Google can still push AI products directly to billions of existing users.
That is where the All-In discussion and the SemiAnalysis report overlap most clearly. Google is different from OpenAI and Anthropic because it is not just a frontier lab. It also controls TPU capacity, Google Cloud, data centers, and major distribution channels.
That gives Google an unusual strategic option. It can support Gemini, sell compute to Anthropic, and host a range of open-weight models at the same time. Enterprises may also end up mixing different models for different tasks.
Under that setup, Google has more than one way to make money. If Gemini wins, Google benefits. If Claude wins, Google can still profit by renting out the compute behind it.