Why Google's Compute Surplus Can't Buy AI Leadership

A Quarter of the World’s Compute. Nowhere to Go.

Google controls roughly 25 percent of global AI compute capacity as of late 2025. It is the single largest owner on Earth — more TPU instances, more GPU clusters, more data center square footage dedicated to machine learning than any other company. And yet its newest flagship model, Gemini 4 Argon, scores dead even with OpenAI’s GPT-6 Astra at 53 points on Artificial Analysis benchmarks, while trailing Anthropic’s Claude Opus 5.5 by five full points. Five points is an eternity in this race. It is the difference between first place and irrelevance when the market consolidates around whoever produces the most capable model first.

The paradox is now visible enough that even Google’s most sympathetic analysts have stopped treating it as a training problem. This is not about having fewer chips. It is about what happens when the chips you own get redirected before they ever reach your best models.

The Internal War Nobody Announced

Dylan Patel, founder and CEO of SemiAnalysis, laid out the picture most bluntly on his Big Technology Podcast on October 9. Google’s total compute is growing every year. DeepMind’s share of it is not. Over the past six months, the proportion of Google’s compute dedicated to training and running its own models has actually contracted relative to what it was two or three years ago. The raw numbers are still impressive — DeepMind trains on hundreds of thousands of TPUs — but the trajectory matters more than the absolute figure.

Patel framed the dynamic as a tug-of-war between two executives whose incentives now pull in opposite directions. On one side sits Thomas Kurian, head of Google Cloud, whose job is to rent that compute to anyone who can pay. His division posted $32 billion in annual revenue last year and is accelerating. On the other was Demis Hassabis, co-founder of DeepMind, who stepped back from operational leadership in August to focus on research strategy. According to Patel, Kurian is winning. The Cloud division’s board-level influence has grown alongside its profitability, while DeepMind’s operational voice has simultaneously diminished.

The prediction is stark: by the end of 2026, DeepMind will have less compute per capita than either OpenAI or Anthropic if the current allocation pattern holds. The gap could widen further by 2027. Patel projected that OpenAI and Anthropic are each adding between five and ten gigawatts of new capacity that year, while Google may only add a few gigawatts for its own internal use. The rest — the vast majority — flows outward to paying Cloud customers.

In other words, Google’s compute surplus is becoming someone else’s competitive advantage by design. Every contract signed with an external customer is a model training run that does not happen internally.

Selling Fire to Your Enemies

The business logic behind this allocation is not hidden. Google Cloud is one of the most profitable segments of Alphabet’s empire. Renting TPU and GPU capacity to external customers — including AI startups that may one day challenge Google’s own models — generates immediate revenue with known margins. Deploying that same capacity internally generates longer-term strategic bets with uncertain returns and no guarantee that the resulting model will close the benchmark gap.

When quarterly earnings call questions arise, the Cloud division’s growth metrics answer directly. DeepMind’s benchmark scores do not. Kurian’s numbers appear on the income statement. Hassabis’s appear in research papers that rarely move stock prices.

The irony is almost literary. OpenAI and Anthropic are not Google’s partners here. They are its future competitors. And Google is financing their compute expansion while simultaneously restricting its own. Every gigawatt sold to an external customer is a gigawatt not available for Gemini. Every enterprise contract signed in Tokyo or Frankfurt is a training run denied in Mountain View.

This is not unique to Google. Almost every major cloud provider faces the same tension between rental revenue and internal R&D. Amazon AWS walks it carefully. Microsoft Azure leverages its partnership with OpenAI to manage the conflict. But Google’s position is distinctive because it is both the landlord and the tenant in its own market. The conflict is internal rather than market-driven, which makes it harder to spot, harder to quantify, and harder to resolve through normal corporate governance.

The second-order effects are already visible. Several mid-tier AI startups that began as Google Cloud customers have quietly shifted portions of their infrastructure to competing providers — a sign that even Google’s clients recognize the strategic paradox of dependence on a company that competes with them in AI model development.

The Search Fear Factor

There is a second layer to the problem that Patel did not emphasize enough but which sources close to Google’s product teams have raised privately. Google has become visibly cautious about model hallucination in search — an area where its brand is most exposed. When Google’s own search product surfaces confident but incorrect answers, the backlash is immediate and global. It has been for years. The introduction of AI Overviews in late 2024, which featured several widely reported factual errors, triggered congressional scrutiny and a wave of user skepticism that persists today.

That fear appears to shape how aggressively Google trains its models. Conservative guardrails, narrower exploration of model capabilities, and a tendency to prioritize reliability over raw reasoning performance can all depress benchmark scores even when the underlying compute is available. OpenAI and Anthropic, operating outside a search product that depends on daily accuracy reputations, can afford different trade-offs. They can push harder on reasoning, take bigger swings on capability, and accept higher hallucination rates in exchange for superior performance on evaluated tasks.

The cancelled Gemini 3.5 Pro launch — initially previewed by Sundar Pichai in June and then quietly pulled — suggests the caution may have gone too far. A model ready for announcement but retracted for performance reasons is an unusual signal from a company that once led with confident releases. Internal sources described the cancellation as a rare moment of self-correction, but also as evidence that Google’s risk calculus around public launches has tightened significantly.

The Talent Ripple

A third consequence of the compute squeeze is harder to measure but potentially more damaging over time. When DeepMind researchers know their projects will compete for allocation against Cloud revenue commitments, they adjust their behavior. Some shift toward applied work with clearer commercial timelines. Others leave. Senior staff attrition at DeepMind accelerated through 2024 and 2025, with former employees citing unclear strategic direction and resource uncertainty as primary factors.

Talent exits compound the compute problem. Experienced researchers carry institutional knowledge about training methodologies, infrastructure optimization, and model architecture that no amount of new hardware can replicate. Each departure makes the remaining team less efficient, which means the compute that does reach DeepMind produces less than it would have a year earlier.

What Comes Next

The structural problem will not fix itself through engineering. Until Google’s leadership reconciles the Cloud revenue model with its AI ambition, DeepMind will continue to compete with fewer resources than its rivals, regardless of how many data centers Alphabet builds. The question is not whether Google can build enough compute — it can and is — but whether it will allocate enough of it to the right problems at the right time.

Patel’s 2027 projection is worth watching closely. If OpenAI and Anthropic each add five to ten gigawatts while Google adds only a fraction of that internally, the compute gap will widen faster than any benchmark can explain. Argon’s 53-point score is not a failure of effort. It is a reflection of priority.

Alphabet’s board faces a choice that gets simpler the longer it is avoided. It can either insulate DeepMind’s compute allocation from Cloud revenue pressures — treating foundational model development as a strategic monopoly rather than a cost center — or it can accept that Google will remain a powerful but second-tier player in generative AI while its Cloud division profits from the very companies it competes against. There is no clean third option. The status quo is already a decision, and its consequences are accumulating.