Agent / TPU / Ecosystem / Security / Openness: What Exactly Is Google Cloud Betting On? — Notes from Google Cloud Next
Summary
- Google Cloud Next offered no knockout move in the monthly model race; the core signal is that AI competition has become a full-stack ecosystem battle. From AI Hypercomputer and Agentic Data Cloud to Agentic Defense, the model platform and Agentic Task Force, Google is betting on an enterprise infrastructure stack; 庄明浩’s post-event take was: “There was no killer move, but the contest seems to have entered an all-around ecosystem battle” (其实没有大招，但比赛似乎已经进入全方位的生态的较量).
- Enterprise AI demand is still rising steeply, with Google’s first-party API calls reaching 16B tokens per minute. The figure was 10B last quarter and 7B in Q3 2025, a 2x increase in six months; 75% of Google’s new internal code is now AI-generated, up from 50% last quarter. 庄明浩 set these figures against the judgment that “The agentic enterprise is real,” underscoring the rapid growth in AI usage.
- TPU 8 formally separates training and inference into TPU 8t and TPU 8i, reflecting a view that inference could account for more than 50% of future compute demand. Thinking Machines said TPU doubled its training efficiency; Anthropic’s early alignment with Google shows that TPU’s value extends beyond chip performance to model customers, Cloud delivery and long-term ecosystem investment.
- Google has elevated security into a full layer independent of data and models; 庄明浩 reads the $32B Wiz acquisition as filling out its security product line and raising security’s strategic priority. Deployments at vertical-agent companies in legal, finance and traditional IT show that enterprise customers care particularly about data isolation, permissions, privacy, governance and complex workflows; these may be the areas where vertical agents remain temporarily insulated from foundation models.
- Google Cloud is deliberately turning “open” into a competitive strategy rather than insisting on a closed Google stack. Apple, Microsoft 365, NVIDIA Rubin and AWS all appeared in the keynote: Gemini Enterprise can connect to Microsoft 365, data services can span AWS and Microsoft’s cloud, and Thomas said next-generation Apple AI models will be provided primarily by Google Cloud and could surface in a new Siri.
- Anthropic’s view of software is more radical than “AI makes developers more productive”: AI capabilities 2x every seven months, shifting software value from implementation to defining problems. As implementation, validation, deployment and maintenance are compressed, the scarcest step becomes identifying, defining and framing the problem; Anthropic disclosed that Claude Code revenue has reached $500M, up 100% in 2026 from the prior year, even though only four months of 2026 had elapsed at the time, with “be water” as its product metaphor.
- Investment and M&A are a parallel thread in Google’s AI strategy, delivering financial returns while buying ecosystem access, industry know-how and product gap-fillers. 庄明浩 cited Anthropic, Wiz, Producer.ai and three funded vertical-agent companies; on SpaceX, he kept the uncertainty explicit—if its “legendary” IPO valuation reaches $2T and Google’s stake exceeds 5%, the position could be worth $100B. “That’s crazy.”
Deep dive
1. Google Defines the Agentic Enterprise as a Reality Already Here
Google Cloud CEO Thomas opened with a clear thesis: “The agentic enterprise is real.” The conversation over the past two years was still whether AI and agents could be put to work; the enterprise narrative has now shifted to how to build, scale, govern and optimize them.
Gemini Enterprise is positioned as a new enterprise AI-agent platform organized around four pillars: build, scale, govern and optimize. Its underlying infrastructure is AI Hypercomputer, its data layer is Agentic Data Cloud, and security, models and tasks are each treated as distinct layers rather than features of an isolated chat product.
庄明浩 summarized Google Cloud’s capabilities in five layers: AI Hypercomputer, Agentic Data Cloud, Agentic Defense, Agentic Platform and Models, and Agentic Task Force. The most counterintuitive point is that the third layer is security, not models.
2. 16B Tokens per Minute Put Quantitative Weight Behind the Reality of Enterprise AI
Pichai disclosed that Google’s first-party API calls have reached 16B tokens per minute, versus 10B last quarter and 7B in the preceding quarter; usage rose 2x over the six months from Q3 2025 to Q1 2026.
The share of new code generated by AI inside Google rose from 50% last quarter to 75%. 庄明浩 emphasized the speed of growth in both figures and linked them to his view that enterprise AI has entered the real world.
The model lineup also led 庄明浩 to conclude that Google currently looks strong in multimodality: Gemini 3.1 Pro, Gemini 3.1 Flash Image (Nano Banana 2), Lyria 3, Veo 3.1 Lite and Anthropic’s Opus 4.7. He suggested that the main competitive arena could rotate among language, coding and multimodality.
3. TPU 8 Bets on More Than Training: Inference Demand Is Structurally Rising
Google unveiled TPU 8t and TPU 8i for model training and inference, respectively, with the two produced by different foundries. The split reflects a view that inference could account for more than 50% of compute demand in the next few years, potentially requiring more compute than training.
Mira Murati’s Thinking Machines was presented as a customer case study: training efficiency doubled after the company adopted TPU. Two chips were raised onstage, while TPU and Google GPU racks with blue lights drew heavy applause.
Anthropic’s session covered the path from small-scale to large-scale clusters, including training scale-up, performance optimization and configuration. 庄明浩 candidly called it the day’s “hardest session to follow” and did not turn technical details he had not understood into conclusions.
On Acquired, Google infrastructure head Amin and DeepMind chief scientist Jeff reviewed more than a decade of TPU development. The journey included failures and raised further questions around power, data-center construction, advanced manufacturing and packaging, and foundries. What 庄明浩 took away was not a single technical breakthrough, but the organization, culture and strategic choices that sustained the investment over time.
4. Security Becomes an Independent Product Layer—and the Hardest Part for Vertical Agents to Be Absorbed by Models
Agentic Defense sits above the data layer and below the model layer, signaling that security is not an add-on. Google showed both Gemini-native security products and a Wiz co-founder onstage; Wiz also evaluated agent risk through red, blue and green teams across different tiers.
庄明浩 sees the $32B Wiz acquisition as filling out Google Cloud’s existing security product line while further elevating security’s strategic standing. The prominence of security in the keynote was one of the clearest signs of where resources are being directed.
Legal, finance, banking and government customers impose demanding requirements around data, permissions, privacy and workflows. Harvey, Rogo and Mechanical Orchard show that the value of vertical agents lies not only in the interface or prompts, but in embedding models into real systems with heavy constraints; those same layers may provide temporary protection from full replacement by foundation models.
5. Vertical Agents Seek Room to Survive Through Model Routing, Complex Delivery and Industry Talent
A broad market consensus holds that vertical agents will eventually be subsumed by foundation models. The experience of the three companies suggests otherwise for now: a single model provider struggles with complex tasks, while an intermediate layer performing routing-like work still has practical value.
The panel, hosted by Crystal Huang of Google Ventures, featured legal agent Harvey, financial research tool Rogo and Mechanical Orchard, which upgrades the legacy computer systems of large banks, government agencies and other institutions. All three companies are Google investments.
Outcome-based pricing remains difficult to implement. The three panelists agreed that the current model is mainly a base subscription plus add-ons, with some experimentation around charging for outcomes, while relying on complex workflows and delivery execution.
The real scarcity is people who understand both the business and the boundaries of AI: they have vertical-industry know-how and know what models can and cannot do. The same talent gap is a common bottleneck in building RL environments and agent harnesses.
6. Anthropic Says Software’s Value Triangle Is Inverting
Anthropic’s provocative claim is that AI capabilities 2x every seven months, faster than Moore’s law. The drivers are increasingly complex problems, longer runtimes and more sophisticated agent frameworks.
AI can affect enterprises through three routes: making employees smarter, speeding up workflows and creating transformative products. The first two include enterprise search, research and analysis, document generation, faster software development and workflow automation; the third means building new applications, revenue streams and even industry shifts around AI.
Traditional software puts a small amount of effort into identifying and defining the problem, with most of the cost going to implementation, validation, deployment and maintenance. As AI coding improves, that triangle will invert, shifting most of the value to identifying, defining and framing the problem—and deciding to solve it.
Software, too, will be like water: foundation models call different tools within workflows and repeatedly loop back toward the objective. Claude Code is treated as an example of this logic; Anthropic disclosed that product revenue has reached $500M, up 100% in 2026 from the prior year, even though only four months of 2026 had elapsed at the time.
7. Openness, Investment and M&A Together Form Google’s Ecosystem Bet
The keynote featured not only Google but also Apple, Microsoft 365, NVIDIA Rubin and AWS. Thomas said next-generation Apple AI models would be provided primarily by Google Cloud and could surface in a new Siri; Gemini Enterprise can connect to Microsoft 365, while data services can span AWS and Microsoft’s cloud.
庄明浩 extended the upside case for Google’s investments to SpaceX. He said the “legend” of SpaceX reaching a $2T IPO valuation was possible, and Google’s stake could exceed 5%; if that happened, an early investment of a few hundred million dollars—or even close to $1B—could be worth about $100B. All of this remains hypothetical.
Anthropic received Google investment early and began experimenting with TPU, giving Google visibility into the value of the model ecosystem while potentially generating valuation upside. Harvey, Rogo and Mechanical Orchard, meanwhile, bring vertical-industry know-how into Google’s investment and model ecosystem.
M&A is being used to fill product gaps: Wiz strengthens security, while Producer.ai adds AI arrangement tools for professional musicians and was linked by 庄明浩 to improvements in Lyria 3. His conclusion was that financial returns are ancillary; ecosystem construction and model evolution matter just as much.
There were two bursts of applause. One came during a demo of Gemini helping the U.S. ski team improve its technique, when 庄明浩 said that, if he remembered correctly, the athlete featured should have been the three-time Olympic champion 肖恩·怀特. The other came when Thomas discussed Gemini’s work with NASA to improve America’s latest space shuttle.
庄明浩 offered no single answer for the direction of Google’s bets. His post-event verdict was: “There was no killer move, but the contest seems to have entered an all-around ecosystem battle” (其实没有大招，但比赛似乎已经进入全方位的生态的较量).