April 23 Roundup: Google sells the agent stack, OpenAI scales Codex through consultants, Anthropic buys a decade of compute, and AI governance turns physical

1. Google turned Cloud Next into a full-stack pitch for the age of agents

Google spent yesterday telling enterprise buyers that the AI market is moving out of experimentation and into scaled deployment. Reuters captured the shift crisply from Las Vegas: Google is “deepening a push into enterprise software,” while CEO Sundar Pichai and Cloud chief Thomas Kurian framed AI agents as the centerpiece of monetization rather than an experimental add-on. Kurian put it bluntly in his keynote: “The experimental phase is behind us, and now the real challenge begins.” That line matters because it is the clearest possible signal that Google wants to be judged less like a research lab and more like a systems vendor.

The substance behind that message was just as important. Google unified major AI products under the “Gemini Enterprise” banner, repositioned Vertex AI around custom agent building, and announced new governance and security features for agents. Reuters also noted that Google is seeing a “sudden explosion in users building their own custom AI agents,” which helps explain why the company is emphasizing control planes, governance, and integration instead of just raw model quality. This is Google reading the enterprise mood correctly: CIOs do not buy possibility; they buy repeatability.

“There’s definitely a strategic shift as the models become much more sophisticated,” Kurian told Reuters. The primary use case of Vertex AI recently shifted from “old-style machine learning” to a sudden explosion in users building their own custom AI agents.

2. Google’s TPU split shows inference is now its own strategic war

Alongside the platform story, Google used the moment to make a more technical but equally consequential move: it split its eighth-generation TPU family into separate chips for training and inference. CNBC reported that Google is no longer treating those workloads as close cousins. Instead, it is designing dedicated hardware for each. Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, said, “With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving.”

This matters because agent-heavy systems put unusual stress on inference. It is not enough to train a great model once. Enterprises need low-latency responses, persistent memory, routing logic, and cost discipline across thousands or millions of repeated calls. Google says its new TPU 8i is tuned for that exact job and offers 80% better performance for speedy inference tasks than the prior generation. In parallel, Reuters reported that the TPU 8t is designed for training large language models in giant pods that can scale to 134,000 chips.

“Both have been sort of architected and designed end-to-end for what’s called the age of agents,” Google VP Mark Lohmeyer told Reuters, describing the new TPU family.

3. OpenAI is scaling Codex from product adoption into channel-driven enterprise transformation

OpenAI’s latest Codex update is less flashy than a new model launch, but it may be one of the week’s most important commercial stories. In its own announcement, OpenAI said more than 4 million developers now use Codex every week, up from 3 million earlier in April. But the real development is the company’s push to operationalize Codex inside large organizations through Codex Labs and partnerships with global systems integrators including Accenture, Capgemini, Cognizant, Infosys, PwC, CGI, and TCS.

The message is unmistakable: OpenAI no longer wants to rely only on bottom-up enthusiasm from developers. It wants enterprise rollout, workflow redesign, and repeatable deployments that can move from pilot to production. The company said enterprises are using Codex not only for code review, testing, and feature development, but also for “briefs, plans, checklists, drafts, and follow-ups.” In other words, Codex is becoming a work engine, not just a coding copilot.

“Our professionals are using Codex to move from static requirements to working solutions in hours, not weeks,” Accenture Chief AI Officer Lan Guan said in OpenAI’s announcement.

4. Anthropic and Amazon just made compute procurement look like geopolitical strategy

Anthropic’s expanded Amazon deal is so large that it changes how people should think about AI infrastructure. In the company’s own write-up, Anthropic said it has secured up to 5 gigawatts of new capacity for training and deploying Claude and will commit more than $100 billion over the next decade to AWS technologies. The deal includes significant Trainium2 capacity coming online in Q2, nearly 1 gigawatt of combined Trainium2 and Trainium3 capacity by the end of 2026, and broader international inference expansion through AWS.

This is not a procurement note. It is a balance-sheet declaration that the frontier model race is increasingly constrained by power, chips, and cloud relationships. Anthropic also said its run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025, while admitting that unprecedented consumer growth has strained reliability and performance for multiple product tiers.

“Our users tell us Claude is increasingly essential to how they work, and we need to build the infrastructure to keep pace with rapidly growing demand,” Anthropic CEO Dario Amodei said.

5. Private, air-gapped frontier AI is becoming a real category

One of the most interesting deployment stories yesterday came from VentureBeat: Cirrascale is expanding its partnership with Google Cloud to deliver Gemini on-premises through Google Distributed Cloud as a fully private, disconnected appliance. The company says regulated enterprises and government agencies can run Gemini on a Dell-manufactured, Google-certified hardware appliance with eight Nvidia GPUs, confidential computing protections, and the ability to operate fully disconnected from both the internet and Google’s cloud.

VentureBeat quoted Cirrascale CEO Dave Driggers saying, “It is full blown Gemini. It’s not pulled. Nothing’s missing from it.” The article also described a security posture in which the model resides entirely in volatile memory, meaning “as soon as the power is off, the model is gone.” That is a stark reminder that model access and model custody are now product design questions.

“The move signals a deepening shift in the enterprise AI market, where the most capable models are migrating out of hyperscaler data centers and into customers’ own racks,” VentureBeat wrote.

6. The policy fight is getting more specific: industrial AI wants exemptions, and ambient AI wants acceptance

The policy thread yesterday was less about sweeping new laws and more about narrowing conflicts. Reuters reported that German Chancellor Friedrich Merz said industrial AI should face a lighter regulatory burden in the European Union than consumer AI. “I will push to ease the regulatory burden in the EU on AI and, where possible, to exempt industrial AI from the current regulatory straightjacket,” he said at Hannover Messe. That suggests policymakers increasingly want to separate factory-floor optimization from higher-risk consumer or general-purpose systems.

At the same time, Peter Diamandis’ latest essay, How AI Will “Feel” in 2 Years, argued that AI is moving from “something you use” to “a ubiquitous, always on, always enabling ambient intelligence layer that orchestrates your life.” Whether or not one agrees with his timeline, it is a useful counterweight to the enterprise stories. The governance question is not just how to regulate labs. It is how much agency, surveillance, and autonomy users and institutions are willing to normalize.

“The shift isn’t from an ‘okay AI’ to a ‘better AI.’ It’s from AI you talk to, to AI that acts on your behalf, before you even think to ask,” Diamandis writes.

Sources: Reuters on Google Cloud Next, CNBC on Google’s new TPUs, OpenAI on scaling Codex to enterprises worldwide, Anthropic on its expanded Amazon compute partnership, VentureBeat on private Gemini deployments, Reuters on Germany’s industrial AI stance, Peter Diamandis on ambient AI.