Thomas Kurian’s Google Cloud strategy is to sell an integrated enterprise AI system rather than isolated model access. The stack joins Google’s custom TPU infrastructure, Gemini models, Vertex AI development tools, data services, security, and workplace applications. Its commercial test is whether customers can move from experiments to governed production workloads without accepting excessive platform dependence.

This analysis uses public material checked September 14, 2026. Google adoption and performance figures are treated as company-reported unless a filing provides them.

Kurian’s role in the stack

Kurian remains CEO of Google Cloud. His public keynote for Google Cloud Next 2026 describes an organization centered on Gemini Enterprise, agent development, data infrastructure, security, and eighth-generation TPUs. The keynote transcript is direct evidence of his product priorities. It is not an independent assessment of product quality.

His operating challenge has two parts. Google must translate research and infrastructure developed across Alphabet into products that enterprise teams can buy and govern. It must also make those products work with data, applications, and models that do not originate at Google. The second requirement is especially important for large organizations that already use several clouds and software vendors.

Infrastructure and models are sold together

Google Cloud positions TPUs as purpose-built AI infrastructure and Vertex AI as the control surface for models and applications. At Next 2025, Kurian announced Ironwood, Gemini 2.5 support, an Agent Development Kit, the Agent2Agent protocol, and distributed-cloud options. His Next 2025 post also reported four million Gemini developers and a 20-fold annual increase in Vertex AI usage. Those adoption figures are Google’s claims and do not reveal paid production usage, workload size, retention, or customer return.

The integrated design can reduce procurement and integration work. It can also make comparisons difficult because the model, accelerator, data layer, orchestration service, and monitoring tools may be bundled. A buyer should compare the complete task, not a token price or a single synthetic benchmark.

Enterprise data is the strategic anchor

Google’s advantage is not limited to models. BigQuery, databases, Workspace, cybersecurity products, and Google’s network give Cloud several routes into a customer’s data and work. Agent products extend that position by letting software read context and act across systems.

That creates a governance obligation. An agent that can retrieve a document, call an API, and alter a business record has a larger failure surface than a chatbot that only drafts text. Buyers need identity boundaries, tool-level authorization, trace retention, evaluation sets, approval checkpoints, and incident procedures. NIST’s AI Risk Management Framework gives teams a vendor-neutral way to organize those controls.

Revenue validates demand, not every AI claim

Alphabet reports Google Cloud as a business segment, but it does not provide audited standalone revenue for Vertex AI, Gemini Enterprise, TPUs, or individual agent products. Alphabet’s 2025 Form 10-K should be used for consolidated Cloud revenue, operating income, capital expenditure, and risk disclosures. Product keynote figures should not be substituted for audited segment data.

This distinction matters when assessing Kurian. Cloud growth can reflect infrastructure, databases, security, Workspace, committed contracts, and AI services together. It shows that Google Cloud has commercial scale. It does not isolate the return from a particular AI feature.

Interoperability needs proof at exit

Google describes its platform as open and multicloud. The useful test is whether a customer can change a model, move data, export evaluations and traces, preserve tool definitions, and reproduce security controls elsewhere. Standards such as Agent2Agent may improve communication between systems, but a protocol alone does not eliminate differences in identity, billing, state, observability, or policy enforcement.

Procurement teams should test one representative workload through setup, operation, failure, and migration. Record task success, latency percentiles, human-review time, hallucination or tool-error rate, data movement, and total monthly cost. Run the same evaluation after changing the underlying model. That exposes which layers are actually portable.

Assessment

Kurian has positioned Google Cloud around a coherent premise: enterprises will buy AI as infrastructure plus data plus governed execution. Google’s research, chips, models, and existing cloud services make that premise credible. The unresolved question is customer economics. Public evidence does not show that an integrated Google stack is cheaper or safer for every workload.

Google’s own materials support the product direction and reported adoption. Alphabet filings support the financial context. Claims about superiority, production outcomes, or lock-in reduction remain workload-specific and should be verified under the customer’s own controls.