Google AI Updates: April 23, 2026

1. Google Splits the 8th-Gen TPU Into Separate Training and Inference Chips

Google. At Cloud Next, Google introduced the 8th-generation TPU as two distinct silicon products — TPU 8t for training and TPU 8i for inference — the first time Google has shipped dedicated parts for each side of the workload. TPU 8t claims up to 2.7x performance-per-dollar over Ironwood at scale; TPU 8i targets roughly 80% performance-per-dollar at low-latency MoE serving. TPU 8i pairs 288 GB of HBM with 384 MB of on-chip SRAM (3x the prior generation), and a single TPU 8t superpod contains 9,600 chips delivering 121 exaflops on two petabytes of shared memory. Both parts reach GA later this year. Source

2. Workspace Intelligence Becomes a Unified Semantic Layer Across Gmail, Docs, Sheets, and Chat

Google. Workspace Intelligence ships as a cross-app context layer that indexes a user’s meetings, email, and files to ground Gemini output in their actual writing style and business data. Gmail gets AI Inbox and AI Overviews that synthesize across threads; Chat adds an “Ask Gemini” that queries projects spanning Workspace; Docs generates infographics and triages comments; Sheets gains a conversational build-and-edit mode. The pitch is that outputs should “sound like you” because the model has ingested enough of your prior work to match tone and structure. Source

3. Gemini Enterprise Agent Platform Targets IT Builders, Not Business Users

Google. Google’s new enterprise agent-building tool leads with Agent Designer (a visual flow canvas now in preview) and Agent Engine Sessions plus Memory Bank for persistent context across interactions — both now GA. TechCrunch flags the interesting choice: unlike Microsoft’s Copilot Studio and Salesforce’s Agentforce, which target business users, Gemini Enterprise is pitched at IT and technical staff, betting that serious production agents require engineering ownership rather than a low-code builder. Source

4. Chrome Enterprise Gets ‘Auto Browse’ for Gemini-Driven Task Automation

Google. Chrome picked up a Gemini-powered auto-browse mode for enterprise users that executes multi-step browser workflows — filling forms, extracting data across pages, compiling research — under admin controls. The move puts Chrome in direct competition with standalone agent browsers like Perplexity Comet and Arc Max, but with distribution leverage Google has that the startups don’t: existing managed-Chrome deployments already on employee machines. Source

5. Gmail AI Overviews Summarize Across Multiple Emails at Work

Google. Gmail for Workspace adds AI Overviews that pull instant summaries across a thread or a set of related emails, answering questions like “what’s the status of project X” without opening each message. The feature is distinct from per-email Smart Summary introduced previously — it’s explicitly multi-email and surfaces in the standard Gmail UI once enabled by a Workspace admin. Source

6. Google Maps Gets Street View AI Synthesis and Satellite Analysis Tools

Google. Google announced generative AI additions to the Maps platform including Street View synthesis for film and location scouting, plus satellite analysis tools that collapse what Google describes as weeks of manual geospatial work into minutes. Cities and planners are the named early adopters. The upgrades ride the same Gemini infrastructure powering Workspace Intelligence, reusing multimodal grounding rather than standing up a separate vertical model. Source

7. Mira Murati’s Thinking Machines Lab Signs Multibillion-Dollar Deal With Google Cloud

Google. Google Cloud disclosed a single-digit-billion-dollar compute deal with Thinking Machines Lab that moves some of the startup’s training and inference onto Google’s A4X Max instances, each backed by four NVIDIA GB300 Blackwell Ultra GPUs. Google says the instances deliver a 2x speedup on training and serving over the prior generation. The deal reinforces Thinking Machines’ reinforcement-learning-heavy approach — the technique behind its Tinker product — and gives Google Cloud a flagship frontier-lab customer alongside Anthropic. Source

8. Google Cloud Commits $750M to Partner-Led Agentic AI Delivery

Google. Google Cloud committed $750 million to accelerate agentic AI development at partner firms, funding joint engineering, go-to-market resources, and customer transformation programs. The fund pairs with the Gemini Enterprise Agent Platform launch: Google is explicitly routing serious agent delivery through SIs and ISVs rather than trying to meet enterprises head-on with its own professional services. Source