Every hyperscaler keynote carries the same posture, dressed in different vocabulary. This week it was Google’s turn. Cloud Next 2026 was a demonstration of vertical ambition: new silicon at the bottom of the stack, a new agent platform in the middle, and $750 million in partner capital to fill in the edges. Read the announcements below with one question in mind: what layer of the enterprise stack is Google not yet claiming?

The Lead: Google Builds the Full Agentic Stack, Chip to Identity to Capital

Google used Cloud Next 2026 to announce an architecture, not a product, pairing eighth-generation Tensor Processing Units with a Gemini Enterprise Agent Platform and a $750 million partner fund that covers the entire vertical stack for agentic AI.

Starting at the silicon layer, Google unveiled two purpose-built variants of its eighth-generation TPU. The TPU 8t handles training, designed for the sustained, memory-intensive compute of teaching a model. The TPU 8i handles inference, optimized for the continuous loop workloads that define agentic AI: an agent calling a tool, receiving a result, deciding what to do next, calling another tool, thousands of times per task. The split is not incidental. Training happens once. Inference runs forever. If agents are the future of enterprise software, inference silicon is the new gold mine, and Google wants to own the mine.

Above the silicon sits the Gemini Enterprise Agent Platform, an evolution of Vertex AI that adds four things worth understanding. First, natural-language agent creation, meaning operators can describe what they want an agent to do in plain English rather than code. Second, a central agent registry, a searchable directory of every agent running in an enterprise, each issued a cryptographic identity so the system can verify who is acting, not just what is being done. Third, a Memory Bank, a persistent cross-session context layer that lets agents carry knowledge from one task forward into the next. Fourth, Workspace Intelligence, which embeds agents directly into Google Docs, Gmail, and Meet.

Read that list slowly. Chips. Identity. Memory. Collaboration. That is a complete operating system for agentic work, and it is all running on Google infrastructure, verified by Google credentials, remembered in Google storage. The $750 million partner fund, the largest single partner commitment from any hyperscaler, is the distribution mechanism. Google’s 120,000-member partner ecosystem becomes the sales force for this stack. The partners get capital. Google gets lock-in at scale.

The TWO angle: the Gemini Enterprise Agent Platform does not merely centralize resources. It enframes them. It converts colleagues into a registry. It converts memory into a stored asset. It converts identity into a cryptographic token. None of that is evil. All of it conditions operators to see human work through the lens of extraction, supply, and dispatch. That conditioning is what you should be watching this quarter, not just the benchmark scores.

Source: The Decoder.

Today’s Movers

OpenAI retired Custom GPTs for Business, Enterprise, and Education tiers and replaced them with Workspace Agents, a new category of long-running autonomous workers embedded directly inside Slack, Salesforce, Google Drive, Notion, and Atlassian Rovo. Powered by Codex, the agents maintain persistent memory across 30-day windows using Redis-backed storage, meaning they carry context from one task forward without a user re-establishing background. Operator angle: an agent that remembers 30 days of context and operates autonomously across your CRM, document store, and communication platform is not a productivity tool. It is a participant with standing access, and most enterprise security frameworks were not written with that participant in mind. Source: OpenAI.

Anthropic quietly removed Claude Code from the Pro plan’s feature list, triggering immediate backlash on Reddit and Hacker News before head of growth Amol Avasare clarified via social media that the update was an A/B test affecting roughly 2 percent of new prosumer sign-ups. Existing Pro and Max subscribers were unaffected, but Avasare acknowledged that subscription usage patterns have shifted dramatically since Claude Code, Cowork, and long-running agents launched, and that current plan structures were not designed for the level of consumption those products generate. Operator angle: flat-rate subscriptions were priced for episodic, single-turn interactions. Agents running in continuous loops are a fundamentally different cost structure, and every provider will face this reckoning. Source: The Register.

Tencent unveiled Hy3 Preview, a 294-billion-parameter mixture-of-experts model that activates 21 billion parameters per forward pass, built from scratch in approximately six weeks of training beginning February 2026. The model was deployed directly into Tencent’s production applications including Yuanbao, CodeBuddy, and WorkBuddy, where it reduced agent latency by 54 percent and end-to-end task duration by 47 percent. Operator angle: Western frontier labs typically measure training runs in months, followed by extended safety evaluation. Tencent compressed the entire cycle by treating production deployment as part of the evaluation itself, a different philosophy about acceptable risk that is producing competitive results at scale. Source: South China Morning Post.

NEC Corporation announced a strategic collaboration with Anthropic to jointly develop secure, industry-specific AI solutions for Japan’s finance, manufacturing, and local government sectors, with plans to deploy Claude Cowork to approximately 30,000 engineers globally. The partnership centers on embedding Claude within NEC’s BluStellar Scenario framework and establishing an internal Center of Excellence. Operator angle: Japan’s enterprise AI adoption has moved cautiously, favoring domestic or deeply localized solutions. NEC’s bet is that Anthropic’s safety posture is the differentiator regulated Japanese industries will accept where they might not accept a less governance-forward provider. Source: NEC.

Hugging Face released ml-intern, an open-source agent built on its smolagents framework that autonomously performs the complete post-training research cycle: browsing arXiv, reading methodology sections, traversing citation graphs, discovering datasets, running training scripts, and iterating on evaluation results without human direction. Early PostTrainBench results show ml-intern outperforming Claude Code on scientific reasoning and Codex on a healthcare evaluation, with a 32 percent performance improvement extracted from a 1.7-billion-parameter Qwen model. Operator angle: the work of surveying literature, identifying promising techniques, and running ablations has historically required a skilled researcher. ml-intern does not replicate human insight, but it removes the lowest-value portions of that workflow, which accelerates the arms race between closed and open frontier capabilities. Source: MarkTechPost.

Court documents released via a Democracy Forward lawsuit revealed SweetREX, an AI tool built by DOGE staffer Christopher Sweet, pitched to federal agency employees as a solution for identifying and eliminating federal regulations at speed. Programmed to evaluate regulations against nine criteria including constitutional concerns, costs to private enterprises, and race-based classifications, SweetREX drafts notices of proposed rulemaking and processes hundreds of thousands of public comments in under 30 minutes. Operator angle: a system that generates regulatory responses to 100,000 public comments in 30 minutes is not reading those comments in any meaningful sense. Public comment exists as a legal check on executive action, and automating the response layer does not eliminate that legal obligation. SweetREX will likely become one of the cases that defines what adequate review means in a world of agentic AI. Source: Jacobin.

SAP and Google Cloud announced a joint multi-agent marketing system at Cloud Next 2026 that automates campaign planning, content generation, audience segmentation, and performance analysis inside SAP’s enterprise data environment. Gemini orchestrates the workflow through specialized sub-agents, each handling a discrete stage with context passed forward at each handoff, extending Google’s agent ecosystem through SAP’s Fortune 500 installed base. Operator angle: this is the business model for Google’s $750 million partner fund made visible. SAP brings the customer relationships and the data. Google brings the orchestration layer. Expect similar announcements with Salesforce and ServiceNow before the quarter closes. Source: e-commerce.news.

The United States Congress introduced a comprehensive federal AI regulation bill this week addressing liability frameworks, transparency requirements, and safety standards for high-risk AI systems, entering a policy environment already complicated by state-level action and White House deregulation pressure. Connecticut’s AI bill cleared the state Senate in the same week, adding to the patchwork of state-level frameworks that federal legislation is, in part, designed to preempt. Operator angle: for enterprise operators, the practical posture is to build for the stricter standard and document the reasoning. Liability frameworks being drafted will shape how courts evaluate AI-related harms when cases arrive, and cases are arriving faster than legislation. Source: Today’s US.

One Tool Worth Knowing

ml-intern is Hugging Face’s new open-source research agent that runs the full LLM post-training loop: literature review on arXiv, dataset discovery, training script execution, and iterative evaluation. It is available as both a CLI and a web app, with early users receiving $1,000 in GPU compute and Anthropic credits. The early PostTrainBench results (outperforming Claude Code on scientific reasoning and Codex on healthcare evaluation, plus a 32 percent improvement extracted from a 1.7B Qwen model) are notable, but the operator-relevant feature is the autonomy. ml-intern is a research workflow that runs while you sleep.

Practical note for this week: if your team does any fine-tuning work or evaluates frontier model techniques, point ml-intern at a real research question you have been meaning to investigate and compare its output against what a skilled researcher would produce in the same window. The comparison tells you where the automation already exceeds human throughput and where judgment is still required.

Wisdom Speaks

“And they said, Go to, let us build us a city and a tower, whose top may reach unto heaven; and let us make us a name, lest we be scattered abroad upon the face of the whole earth.” Genesis 11:4, KJV

The biblical warning about Babel is not about ambition. It is about a particular form of ambition: building so total a structure that dependence on anything outside it becomes unnecessary. “Lest we be scattered,” the builders said. The fear underneath Babel was not failure. It was the vulnerability of remaining distributed, remaining dependent on relationships and exchange that could not be controlled. Google Cloud Next 2026 is a Babel moment, not because Google is wicked, but because the structural logic is identical. Chips, identity, memory, capital, partners, all consolidated under a single architectural umbrella designed to eliminate the friction of working across providers.

“The essence of modern technology is by no means anything technological. It is a way of revealing, a challenging which puts to nature the unreasonable demand that it supply energy which can be extracted and stored as such.” Martin Heidegger, The Question Concerning Technology (1954)

Heidegger adds a second layer to the diagnosis. The Gemini Enterprise Agent Platform does not merely centralize resources. It enframes them. It converts colleagues into registry entries, memories into stored assets, identity into cryptographic tokens, and human attention into dispatched tasks. Nothing in that list is inherently wrong, but the frame changes what we notice and what we fail to notice. When every interaction is routed through a system optimized for extraction and dispatch, the relationships that exist outside the system become progressively harder to see. Operators who understand this are not anti-technology. They are exercising discernment, the capacity to distinguish between what a tool does and what it does to the person using it.

The question for this week is not whether to build with these platforms. It is whether the people using them are growing in judgment, or merely in throughput.

---

Yesterday’s digest: SpaceX’s $60 Billion Cursor Option: The AI Coding Stack Consolidates, on SpaceX moving to own the developer distribution layer through Cursor and the Colossus compute cluster. Earlier this week: The Borrower and the Lender, on Amazon’s $100 billion compute commitment to Anthropic and what it means when a frontier lab’s infrastructure is rented from its largest investor. Monday: The House Divided, on the NSA deploying Anthropic while the Pentagon simultaneously listed it as a supply-chain risk. Today’s Gemini Enterprise Agent Platform announcement is the same consolidation story seen from the top of the stack: while others rent compute or acquire distribution, Google is building the layer that all of them will eventually need to route through.