Google Cloud Next 2026 Pinned Gemini 3.5 Pro to $7/$21, Locked Anthropic, Apple and CoreWeave into TPU v7 Ironwood

Google used Cloud Next 2026, which opened at the Moscone Center in San Francisco on August 15, to take three coordinated shots at the AI infrastructure stack: it opened Gemini 3.5 Pro preview at $7 per million input tokens and $21 per million output tokens, signed Anthropic, Apple and CoreWeave as the first three TPU v7 Ironwood customers, and rolled out a Vertex AI overhaul that puts agent-evaluation and adversarial-testing tooling into the default workflow. The combined play, set out in Sundar Pichai and Thomas Kurian's 105-minute keynote, is a multi-year bet that the AI infrastructure market is now a three-vendor race — Google, NVIDIA and AMD — not the two-vendor race the industry had been pricing for the last two years.

What Gemini 3.5 Pro's Price Actually Means

The $7/$21 price point lands Gemini 3.5 Pro squarely in the middle of the new frontier-tier pricing band. Google's own AA benchmark index scores Gemini 3.5 Pro at 68 — the highest score in the public aggregator as of mid-August, ahead of Anthropic's Claude Opus 5 and OpenAI's GPT-5 family on the same aggregator's read. By pricing Gemini 3.5 Pro between the cheap-and-fast tier (DeepSeek, Mistral) and the premium tier (Claude Opus 5, GPT-5), Google has explicitly targeted the enterprise buyer who needs a frontier-tier model but does not want to pay the premium-tier premium. The output-token price of $21 is the operational headline: long agentic loops, which produce more output tokens than chat, are exactly where the price gap matters most.

TPU v7 Ironwood and the Three Customer Commitments

The TPU v7 Ironwood announcement — Google's seventh-generation Tensor Processing Unit — came with three named enterprise commitments that, taken together, are the largest single TPU-generation launch in Google's history. Anthropic committed to one million TPU v7 chips over the multi-year ramp, Apple committed to use TPU v7 across Apple Intelligence training and inference workloads, and CoreWeave committed to deploy 500,000 TPU v7 chips across its neocloud footprint. The three customers were not just contracted; the keynote included an on-stage handoff between Pichai and Anthropic's head of compute. For Google, this is the first TPU generation where the customer mix rivals the captive internal demand that defined TPU v4 and v5.

The Vertex AI Agent Stack

Alongside the model and chip announcements, Google shipped a Vertex AI overhaul that bakes agent-evaluation, adversarial-testing, and incident-reporting workflows directly into the platform. The new Vertex AI Agent Engine — available in preview on August 15 and in general availability in October — provides managed agent runtimes, evaluation harnesses, and a red-team toolkit that enterprise customers can run against their own agentic deployments. The August 2 EU AI Act general-application deadline and the February 2027 first-reporting deadline are clearly the demand drivers; several enterprise customers, including two large European banks, cited EU AI Act compliance as the reason they accelerated adoption of the new Vertex tooling.

Why This Reshapes the AI Stack Map

The Cloud Next 2026 announcements move Google from 「fast follower」 to 「structural counterweight」 in the AI infrastructure market. NVIDIA's dominance in GPU-based AI compute remains intact for workloads optimized for that path, but the TPU v7 customer commitments show that for the next generation of frontier-model training, three of the largest AI compute buyers are hedging with custom silicon. For the broader market, the practical implication is that the TPU-vs-GPU choice is now a first-class procurement decision rather than an experimental one. Cloud Next 2026 also introduced a tighter integration between Gemini 3.5 Pro and Workspace, putting AI agents inside Docs, Sheets and Meet with an audit-log trail designed to satisfy regulated-industry buyers.

What to Watch Through Year-End

Three checkpoints follow. TPU v7 Ironwood general availability is scheduled for October, with the first customer shipments following within four to six weeks. The Vertex AI Agent Engine GA release will land alongside TPU v7, and the first set of enterprise compliance attestations built on top of it are expected before the EU AI Act's February 2027 reporting deadline. And the next OpenAI devday, also in October, will set the competitive counterpoint: whether OpenAI's GPT-5 generation has closed the agentic and pricing gap that Gemini 3.5 Pro now leads.