In a wide-ranging interview published by Stratechery, Google Cloud CEO Thomas Kurian laid out the company's thesis for the emerging era of AI agents: that building and deploying agents at enterprise scale requires a unified stack spanning custom silicon, frontier models, data infrastructure, and cybersecurity — and that Google is uniquely positioned to deliver all of it. The interview, conducted on April 15 ahead of Google Cloud Next, covered agents, TPU strategy, competitive dynamics with Anthropic, and how Google balances internal and external demand for compute.
The conversation arrives at a moment when the narrative around AI agents has shifted from conceptual demos to operational deployments. Kurian cited customers including Citigroup, Comcast, Walmart, and the U.S. Food and Drug Administration as examples of organizations running complex, multi-step agent workflows on Google's platform. According to Stratechery's reporting, Google now generates 16 billion tokens per minute, up roughly 40% from late 2025. The core editorial question, however, is not whether Google Cloud is growing — it clearly is — but whether its integration-first strategy can outpace more focused competitors in the agent platform race.
The Integration Thesis Under Scrutiny
Kurian's argument rests on a deceptively simple claim: every Google product runs on the same version of Gemini, on the same day, using the same harness. The feedback loop between Google Cloud's enterprise customers and DeepMind's model development team is, he says, tight — with cloud engineers sitting alongside DeepMind CEO Demis Hassabis's researchers, feeding real-world agent workflows back into Gemini's reinforcement learning pipeline. This is the kind of structural advantage that sounds compelling on paper but is notoriously difficult to execute inside a company of Alphabet's scale.
Stratechery's Ben Thompson pressed precisely on this tension: does having so many internal and external stakeholders pulling on the same model and infrastructure create coherence or confusion? Kurian's response was emphatic — the harness is unified, the priorities are clear, the company is not distracted. Yet the question lingers. Anthropic and other frontier labs have earned developer mindshare in recent months through a more singular focus on model capability and developer experience. Google's counter-pitch is that raw model quality is necessary but insufficient; what enterprises actually need is the orchestration layer — identity, security, memory, tool use, and data context — wrapped around the model. The introduction of what Google calls "Knowledge Catalog," a global dictionary for enterprise data objects, exemplifies this approach: it is not a model improvement but an infrastructure improvement designed to make models more useful in constrained, real-world environments.
TPUs, Anthropic, and the Compute Balancing Act
Perhaps the most revealing portion of the interview concerned Google's evolving TPU strategy. Kurian announced two new chip families — TPU 8t for training and TPU 8i for inference — and disclosed that Google is now placing TPUs in third-party data centers, not just its own cloud regions. This move expands Google's total addressable market into segments like capital markets (where firms such as Citadel Securities are replacing algorithmic trading workloads with TPU-based inference) and national laboratories running energy simulations.
The competitive dynamics here are layered. Anthropic, one of Google Cloud's largest tenants, runs significant inference workloads on TPUs. When Thompson asked whether Anthropic winning deals at Gemini's expense is still a win for Google, Kurian framed the company as a platform player that monetizes multiple layers of the stack — chips, cybersecurity, data infrastructure — regardless of which model sits on top. This is a pragmatic stance, but it also reveals a structural tension: Google is simultaneously competing with and supplying its most capable rival in the model layer. The analogy to cloud-era dynamics, where AWS hosted Netflix while building Prime Video, is imperfect but instructive. Google's ability to maintain credibility as both a model provider and a neutral infrastructure supplier will be tested as agent deployments scale and enterprise customers make longer-term platform commitments.
The broader strategic picture that emerges from Kurian's remarks is one of deliberate breadth. Google Cloud is not trying to win on any single axis — not just models, not just chips, not just enterprise tooling — but on the compound effect of owning and integrating all of them. Whether that compound advantage materializes faster than the focused intensity of smaller competitors remains the central question of the agentic era. As enterprise AI moves from pilot programs to production workflows, the answer may depend less on which model scores highest on benchmarks and more on which platform can reliably orchestrate the messy, multi-system reality of how large organizations actually operate.
With reporting from Stratechery
Source · Stratechery
