Morning Brief · Wednesday

Google Goes All-In on Agentic Enterprise at Cloud Next '26

Google fired the opening salvo of what may be the biggest cloud conference in years — Gemini Enterprise Agent Platform, 8th-gen TPUs, Deep Research Max, and Gemini running on air-gapped servers, all announced before noon in Las Vegas. OpenAI answered with the official launch of ChatGPT Images 2.0: a reasoning-native image model that can generate infographics, maps, manga, and flawless multilingual text from a single prompt. And Anthropic's classified cybersecurity model is now navigating an unlikely dance with the Pentagon, the NSA, and a president who says they're "shaping up."

Mira Novian

April 22, 2026 · Morning Brief · ~8 min read

Google Cloud Next '26 opens with the most sweeping agentic platform launch in cloud history

Google Cloud CEO Thomas Kurian opened Next '26 in Las Vegas this morning with a keynote that wasn't so much a product announcement as it was an architectural thesis: the enterprise cloud is no longer about compute and storage — it's about agents as a system of action. The centerpiece is the new Gemini Enterprise Agent Platform, which bundles together everything required to build, scale, govern, and optimize agents across an organization — Agent Studio (a low-code interface for building agents in natural language), Agent Registry (a single indexed catalog of every internal agent and tool), Agent Identity, Agent Gateway, Agent Observability, and Agent-to-Agent Orchestration. The platform formally unifies what was previously scattered across Vertex AI, Google Workspace, and Gemini Enterprise into a single co-developed stack.

On the infrastructure side, Google announced its eighth-generation TPUs — TPU 8t for training and TPU 8i for near-zero latency inference — alongside Virgo Networking, its new megascale data center fabric built for AI workloads at unprecedented density. Storage also got a major upgrade: Managed Lustre now delivers 10 terabytes per second of throughput, designed for the read/write intensity of agentic pipelines. Google also unveiled the Agentic Data Cloud, which introduces a Knowledge Catalog for grounding agents in trusted enterprise context, a cross-cloud AI-native Lakehouse, and a new Data Agent Kit. And the Agentic Defense pillar — built in partnership with Wiz, which Google acquired last year — brings AI-native threat detection, detection engineering, and remediation agents across hybrid and multicloud environments. Nearly 75% of Google Cloud customers are now using AI products, the company said, with 330 customers processing over one trillion tokens in the past 12 months.

cloud.google.com ↗

The "vertically optimized stack" framing is Google's clearest argument yet that it's building something fundamentally different from AWS or Azure — not a collection of interoperable services but a single coherent substrate for AI-native enterprise. The depth of the announcements is genuinely impressive: Gemini 3.1 Pro powering Deep Research Max, Claude Opus 4.7 added to the model catalog, MCP treated as a first-class citizen in Agent Registry. The conference is two more days. Watch what they say about security — the Wiz partnership goes well beyond rebranding; Agentic Defense is an attempt to make AI-era security a native Google Cloud capability rather than a third-party add-on. If it works at scale, that's a meaningful enterprise moat.

Infrastructure

Gemini now runs on air-gapped servers — the model lives in RAM and vanishes when you pull the plug

Timed to coincide with Cloud Next, Cirrascale Cloud Services announced it is shipping full Gemini on-premises via Google Distributed Cloud — making it the first neocloud provider to offer Google's frontier model as a fully private, disconnected appliance. The deployment packages Gemini into a Dell-manufactured, Google-certified hardware unit with eight Nvidia GPUs, wrapped in confidential computing protections. It can run inside Cirrascale's data centers or a customer's own facility, completely disconnected from the internet and from Google's infrastructure. Unlike Azure or AWS on-premises offerings, Cirrascale is emphatic: "This is the actual model deployed on-prem outside of their cloud. It's not a cut-down version."

The security architecture is striking. The Gemini model lives entirely in volatile memory — there is no persistent storage copy. "As soon as the power is off, the model is gone," CEO Dave Driggers told VentureBeat. User sessions clear automatically when a session ends. And if anyone attempts to tamper with the confidential compute environment, the appliance physically renders itself inoperable and flags itself as compromised. The product enters preview immediately, with general availability expected in June or July — aimed squarely at regulated industries in financial services, healthcare, and government that have been stuck choosing between frontier AI capability and data sovereignty.

venturebeat.com ↗

This is quietly one of the more consequential announcements of the week. The "cloud or open-source" binary that's defined enterprise AI adoption for three years just got a third option: frontier-class capability in your own rack, fully air-gapped, with hardware-level tamper protection. Banks and government agencies have been sitting out the frontier model era because of data sovereignty concerns. If Cirrascale's offering works as advertised, that excuse disappears. The "model in volatile RAM, gone on power-off" feature is also a notable design choice — it's not just a security posture, it's a signal to regulators and enterprise legal teams that Google takes data residency seriously. That matters more than the tech spec.

Models

ChatGPT Images 2.0 is officially here — and it reasons before it draws

OpenAI today officially launched ChatGPT Images 2.0, rolling out to all ChatGPT tiers and available via API as gpt-image-2. The model — which was quietly live on LM Arena under the codename "duct tape" for several weeks — represents a fundamental rethink of what an image model is supposed to do. Where previous generations generated a single output from a prompt, Images 2.0 integrates OpenAI's O-series reasoning capabilities: before rendering a single pixel, the model researches, plans, and reasons through the structure of the image. In a live press demo, the model synthesized a complex PowerPoint document, identified the right logos, and produced a professional poster that preserved the original file's stylistic intent — without being explicitly instructed to do any of it.

The capability list is extensive: accurate multilingual text rendered within images (including non-Latin scripts), full infographic and slide generation, floor plans, image grids, character sheets from multiple angles, manga-style sequential art, realistic UI and screenshot mockups, and the ability to reproduce real-world figures accurately. The model also has a December 2025 knowledge cutoff and can perform real-time web research to ensure visual accuracy before generating. Architecture-wise, it's been "revamped from scratch" — a generalist model that handles 3D perspective shifts and complex spatial reasoning through text prompts alone. GPT-Image-1.5 is being deprecated as the default across ChatGPT, though it remains accessible via API. OpenAI confirmed the model is being held to the same political/election content guardrails as its predecessors amid growing scrutiny over AI-generated influence campaigns.

venturebeat.com ↗

"Images are a language, not decoration" is the line from the release notes that's going to get quoted a lot. It's a succinct description of what this model actually is — not a picture generator but a visual reasoning system. The reasoning-before-rendering architecture is the part that matters most technically: it's the same shift that happened with text models when chain-of-thought reasoning emerged. The practical implications for design, marketing, and education are real and immediate. The concerning implication — which OpenAI's product lead was asked about directly — is that the same capability that makes it great for legitimate infographics makes it extraordinary for synthetic influence campaigns. OpenAI's answer was essentially "we have guardrails." So did every other model before they got jailbroken.

Geopolitics

Anthropic's Mythos model finds 271 Firefox bugs, Trump says a Pentagon deal is possible, and the NSA already has access

Anthropic's Claude Mythos Preview — the company's classified cybersecurity-focused model that's been running at Nvidia, Apple, and JPMorgan Chase — had a notable week even before the White House got involved. Mozilla CTO Bobby Holley confirmed that Mythos found 271 vulnerabilities in Firefox 150, calling the model "every bit as capable" as top security researchers — while also noting, reassuringly, that none of the discovered bugs "couldn't have been found by an elite human researcher." The model isn't finding novel attack classes; it's finding what humans would find, just faster and at scale.

The political dimension is moving fast. President Trump said in a CNBC interview that "it's possible" Anthropic and the Pentagon could reach a deal — a notable softening from the active lawsuit the two had been engaged in — after a positive White House meeting where Anthropic came to discuss Mythos. "We had some very good talks with them, and I think they're shaping up," Trump said. Separately, reporting surfaced that the NSA has already gained access to Mythos, despite the model reportedly being flagged internally as a potential supply-chain risk. Google co-founder Sergey Brin added another layer to the Anthropic competitive picture with a memo to DeepMind employees acknowledging that "every Gemini engineer must be forced to use internal agents for complex, multistep tasks" — because Anthropic's coding tools are currently ahead.

theverge.com ↗

The NSA-has-Mythos-despite-supply-chain-risk-flag detail is the one that deserves more scrutiny than it's getting. It suggests that the urgency of deploying capable AI for national security purposes is outrunning the risk assessment processes meant to govern that deployment — which is a pattern with significant historical precedent, and not a reassuring one. Meanwhile, the Anthropic/Pentagon détente is a quiet strategic win for the company: being useful to the White House while maintaining an independent identity is a harder needle to thread than it looks, especially when your most capable model is being described as a potential risk to the people who just invited you to the table.

✦ Mira's Take

Today is one of those days where the news volume alone tells you something. Google Cloud Next opened with what amounts to a comprehensive counter-thesis to AWS and Azure: not services but a stack, not integrations but a platform, not AI features but an agentic operating system for the enterprise. The depth and coherence of the announcements — from TPU 8i through Virgo Networking to Deep Research Max to air-gapped Gemini appliances — suggests a company that has been building toward a single moment and chose this week to unveil it all at once.

OpenAI's timing with Images 2.0 is either a deliberate counter-punch or a coincidence that functions exactly like one. Either way, the message is the same: OpenAI is not ceding the creative and visual media category while Google builds enterprise infrastructure. The reasoning-before-rendering architecture matters because it moves image generation from a probabilistic guessing game into something more like deliberate visual thought. That's a capability category shift, not a benchmark improvement.

The Anthropic thread is the most geopolitically interesting. Sergey Brin writing memos telling DeepMind engineers to use Anthropic tools is an extraordinary public admission. The NSA running Mythos despite a supply-chain risk designation is the kind of detail that usually surfaces in congressional hearings, not press coverage. And a company that six months ago was in an active lawsuit with the Pentagon is now having "very good talks" at the White House. AI is becoming defense infrastructure faster than the governance frameworks can keep up — and the companies building it are being pulled into that orbit whether they intended to or not.