Google used the opening of its Cloud Next conference to stake a claim as the most structurally positioned company in the AI stack, unveiling two new custom chips, an agentic version of Chrome for enterprises, and a multi-billion dollar compute deal with Mira Murati's Thinking Machines Labs. Taken together, the announcements push Google into direct competition with every layer of the AI supply chain at once.
The centerpiece is a split TPU lineup: the TPU-8T aimed at training workloads and the TPU-8I designed for inference, the dominant cost of running AI in production. Google claims the new silicon is three times faster at training and delivers 80 percent better performance per dollar than comparable Nvidia systems, with the ability to scale more than a million TPUs in a single cluster. The numbers are the company's own, with no independent benchmarks yet available.
Cluster size has been a persistent bottleneck for rivals. When xAI wired together 200,000 Nvidia GPUs earlier this cycle, it had to engineer much of the surrounding infrastructure itself. Google is going directly at that problem with a design built for million-chip deployments. The company is not abandoning Nvidia, however, and will resell the Vera Rubin platform through Google Cloud later this year, a hedge that contrasts with Amazon Web Services' all-in bet on its own Trainium silicon.
Key facts
- 01Google. A key thread of reporting in this story.
- 02TPU. A key thread of reporting in this story.
- 03Thinking Machines. A key thread of reporting in this story.
The second announcement turns Chrome into what Google is calling an AI coworker. A new feature called auto browse, powered by Gemini, runs inside Chrome for Workspace and reads context across open tabs to automate tasks such as entering CRM data, comparing vendor quotes, summarizing candidate portfolios and writing up competitor research. Users can save frequent workflows as skills and trigger them with a forward slash. The rollout starts with US Workspace customers, and Google is committing that enterprise prompts will not be used to train its models.
“Google is basically the only company really doing the full stack.”— Jaeden Schafer
The approach keeps a human in the loop on every action, which leaves space for more autonomous rivals. Anthropic's Claude-based coworker tools, which reach into the local desktop and file system and can run skills without repeated approvals, still look ahead on pure automation. Google appears to be aiming at a more cautious enterprise buyer willing to trade speed for oversight.
The third leg is a single-digit billions deal with Thinking Machines Labs, the startup led by former OpenAI executive Mira Murati that is raising at a $12 billion valuation. The agreement gives Thinking Machines access to Nvidia's GB300 systems on Google Cloud along with training and deployment services for its first product, Tinker, a tool for building custom frontier models. The company runs heavy reinforcement learning workloads that are compute-intensive, and one of its founding researchers has publicly credited Google with getting the operation running quickly.
The pattern across the three announcements is a three-layer strategy: silicon at the bottom through TPUs and resold Nvidia chips, compute hosting for frontier labs in the middle where Anthropic already runs on TPUs and Thinking Machines now joins, and an agent layer on top inside Chrome and Workspace. Microsoft leans heavily on OpenAI at the middle layer but is weak in silicon. Amazon has tied itself to Anthropic. Nvidia owns the chip layer but does not touch applications. Google is attempting all three at once.
The counter-case is that Gemini still trails GPT-5.4 and Claude Opus 4.7 on the benchmarks consumers watch most closely, and the headline TPU performance claims remain unverified by independent labs. Distribution and integration, however, may matter more than leaderboard position over the next year if enterprises can run inference more cheaply on Google's stack and pipe it through tools their employees already use.
Regulatory risk looms over the strategy. Google is making these moves while the Department of Justice's search remedies case remains live, and turning Chrome into an agent layer that pulls from Workspace data is the kind of cross-product integration that tends to attract antitrust attention, similar to the scrutiny Microsoft has faced over its OpenAI alliance. The next signals to watch are independent benchmarks on the TPU-8T and 8I, whether auto browse moves from pilots into broad enterprise rollouts, and whether additional frontier labs follow Thinking Machines onto Google Cloud.
“Google is basically the only company really doing the full stack.”
“Google's claims is that it is three times faster at training and 80% better at performance per dollar against the NVIDIA alternatives, and the ability to scale more than a million TPUs in a single cluster.”
“The customer really doesn't care whether Gemini tops the Elo leaderboards.”
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
