Google Cloud’s Kurian: TPU Business Now Twice the Size of the Next Largest Hyperscaler’s Silicon Franchise
Goldman Sachs Communacopia + Technology Conference, September 8, 2026
Thomas Kurian, CEO of Google Cloud, used his appearance at Goldman Sachs' Communacopia + Technology Conference to lay out the clearest picture yet of how Alphabet's enterprise arm is scaling toward its nearly $100 billion annual revenue run rate, and the numbers he shared point to a business accelerating on multiple fronts simultaneously. New customer acquisition is growing roughly twice as fast as it was a year ago, deals over $100 million are up 2x both quarter-over-quarter and year-over-year, and customers who commit $100 in spending typically end up spending more than 50% above that commitment. For a division whose scale increasingly rivals the largest pure-play cloud franchises, this is a meaningfully bullish data set.
TPU Silicon Business Is Now Bigger Than Any Rival Hyperscaler's
The most striking disclosure of the session was Kurian's claim that Google's accelerator and CPU silicon business is "more than twice that of the next largest hyperscaler." This is the first time Google Cloud has quantified its silicon scale relative to competitors, and it reframes the TPU story from a cost-saving internal tool into a standalone franchise with real external monetization. Kurian detailed three go-to-market models for this hardware: cloud subscription, outright capital purchase for on-premise deployment, and a neocloud structure, exemplified by a partnership with Blackstone. The capital purchase model is particularly capital-light for Google, since customers absorb the data center power and space costs. Kurian said payback periods on AI servers broadly run under two years, and on Google's own TPU silicon specifically, that payback period is cut in half. He added that the majority of infrastructure contract value is now locked into long-term, five-year committed agreements, a detail that should reassure investors worried about the durability of AI infrastructure demand.
Performance Claims Get More Specific
Kurian quantified Google's silicon performance advantage with figures investors will want to track against rival hyperscaler disclosures: 2.7 times better price performance for training, 80% better price performance for inference, and 30% better price performance for CPU workloads. He tied this directly to why Google's newest model, Gemini Flash 3.8, undercuts competitors on a cost-per-intelligence basis, stating plainly, "it's our tokens, real intelligence per dollar, it's because we can optimize that whole stack." The demand base for this infrastructure is broadening beyond AI labs into capital markets firms doing quantitative research and high-performance computing users migrating off legacy solid-state and computational fluid dynamics workloads. Named references included Deutsche Börse, Pfizer, and the U.S. government's Genesis mission for energy research, underscoring diversification beyond the hyperscaler-versus-hyperscaler AI lab narrative.
Gemini Enterprise Adoption Metrics Get Sharper
Kurian disclosed that Gemini Enterprise, the company's agent platform, is used by more than 90% of the Fortune 100, separately from the 90% of the Fortune 100 that also use Google for cyber defense. Roughly 80% of Google Cloud customers now use its AI products, and critically, those AI product users consume 1.8 times as many products as non-AI customers, evidencing real cross-sell traction rather than just AI adoption for its own sake. Kurian estimated the five-year lifetime value of a customer using the Gemini portfolio in the cloud at 1.5 times that of a non-Gemini customer, giving investors a concrete framework for how AI adoption is expected to compound cloud economics rather than simply add a new product line.
The Wiz Rationale, Explained in Full for the First Time
Kurian gave the most detailed public explanation to date of the strategic logic behind the Wiz acquisition, tying it directly to the emergence of AI-driven cyber threats. "We saw in 2023 that as models learn to code... they could become extremely proficient in finding vulnerabilities," he said, explaining that Wiz was acquired specifically because it connects across multiple clouds, not just Google's own, to build a risk score identifying which applications are most likely to be compromised. Kurian described a four-part workflow: Wiz prioritizes risk, Gemini scans for vulnerabilities, a new product called CodeMender actually repairs the code, and Wiz then re-tests the fix. He was blunt about the limits of AI alone in this domain: "You can only defend a threat from an AI model by using a combination of a security platform and an AI system. An AI system by itself cannot protect you, a security platform by itself cannot protect you." This is a notable admission that pure-model security plays are structurally incomplete, a competitive point aimed squarely at security vendors without an underlying platform.
Multi-Model Strategy Defended on Efficacy Grounds, Not Just Optionality
Kurian offered an unusually concrete justification for why Gemini Enterprise deliberately supports multiple third-party and open-source models rather than pushing customers exclusively toward Gemini. Using cybersecurity as the example, he noted that "no single model finds all the issues from a cyber point of view" and that no model finds a superset of what other models catch, meaning defenders need multiple passes with different models, including open-source ones, because "many times the attackers are going to use that." This is a more defensive-sounding rationale than the usual "customer choice" framing vendors typically offer, and it implicitly concedes that Gemini alone is not sufficient for enterprise security use cases, a nuance investors should weigh against Google's broader push to position Gemini as a frontier model leader.
Forward-Deployed Engineers and the Accenture Partnership
Kurian was careful to distinguish Google Cloud's use of forward-deployed engineers from a services-heavy model, noting that Google's services revenue as a percentage of total business remains the smallest of any hyperscaler. The forward-deployed teams serve three functions: identifying where model capability needs to be pushed forward with top customers in specific industries, building the technical tooling to capture those advances, and creating training and certification programs to scale adoption through partners. The clearest evidence of that partner-led scaling strategy came with confirmation that Accenture is building a large Gemini Enterprise business group, announced the same morning as the conference. Kurian also outlined a three-pronged partnership architecture spanning eight targeted industries, a country-by-country systems integrator and AI specialist network, and data partnerships with providers like Bloomberg, FactSet, and MSCI to make proprietary datasets reasoning-ready within Google's platform.
The Citigroup Wealth Advisor as a Proof Point for Vertical Integration
To make the case for Google's full-stack strategy tangible, Kurian pointed to Citigroup's wealth advisor build on Gemini Enterprise, which combines a real-time avatar interface running on TPU infrastructure, a reasoning layer pulling accurate financial data, security protections, and cost-efficient token serving at scale. The example was deployed to argue that vertical integration is not just an efficiency story but a monetization diversification strategy, allowing Google to capture revenue whether customers value the silicon, the model, the data platform, or the security layer. It is a helpful, concrete illustration of an otherwise abstract "full stack" pitch that has sometimes lacked specificity in prior investor communications.