AI in 15 — April 23, 2026

Google just announced eighty thousand Rubin GPUs in a single cluster and nine hundred and sixty thousand across sites. That's not a company. That's a country-scale compute footprint.

Welcome to AI in 15 for Thursday, April 23, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Thursday show, Marcus. Google Cloud Next dropped the year's biggest infrastructure announcement yesterday, an eighth-generation TPU split in two plus a deeper Nvidia partnership. Alibaba released a twenty-seven-billion-parameter dense model that beats their previous flagship while fitting on a laptop. OpenAI launched Workspace Agents to answer Anthropic and Google on the enterprise agent layer. Sam Altman's World ID landed at Zoom, Tinder, DocuSign and Shopify. The Linux kernel is pulling out legacy code because AI bug reports broke the maintainers. A viral MAGA influencer turned out to be an Indian medical student and a Gemini prompt. And OpenAI published a post-mortem on that North Korean npm attack. Let's go.

Google bids to own the agent substrate.

Qwen puts flagship coding on a Mac.

And the Linux maintainers raise the white flag on AI bug reports.

Lead story, Marcus. Sundar Pichai's Cloud Next keynote yesterday. Walk me through the hardware split because this is the first time Google has done this.

For the first time, Google broke their TPU line into two specialized chips. The TPU 8t for training. The TPU 8i for inference. The 8t scales to nine thousand six hundred chips per superpod, two petabytes of shared high-bandwidth memory, one hundred and twenty-one exaflops of compute, and two-point-eight times better price-performance than the Ironwood generation from last year. The 8i is the interesting one. Two hundred and eighty-eight gigabytes of HBM paired with three hundred and eighty-four megabytes of on-chip SRAM, a three-times jump, and eighty percent better performance per dollar on inference. Both chips are twice as power efficient as Ironwood.

And yet they're hugging Nvidia closer than ever.

That's the real story, Kate. Google is not pitching the TPU as pure substitution. They announced they'll be among the first clouds to offer Nvidia's Vera Rubin NVL72 rack-scale systems in the second half of 2026, wrapped in a new A5X instance that Nvidia claims delivers ten times lower inference cost per token. Eighty thousand Rubin GPUs per single site. Nine hundred and sixty thousand across multi-site clusters. Citadel Securities was the flagship customer reference. Only Google can credibly both build an Nvidia competitor and be Nvidia's launch partner in the same keynote.

And then the software layer.

The Gemini Enterprise Agent Platform, plus Google pushing its Agent-to-Agent protocol into a cross-vendor standard. ServiceNow, Salesforce, Atlassian and SAP signed on. A seven-hundred-and-fifty-million-dollar partner fund through Accenture, Deloitte and KPMG to bankroll agentic rollouts. Google's shares jumped on the news.

The editorial read.

The AI hardware market is bifurcating. Training megaclusters for the next generation of frontier models, and inference at scale because agents running twenty-four hours a day need cheap tokens far more than they need raw flops. Google is making the most complete bid yet to own the substrate. Chips, cloud, protocol and enterprise agent platform in one stack. OpenAI and Anthropic still rent somebody else's clouds and run on somebody else's chips. If A2A becomes the real interop standard, Google doesn't just compete with them. They become the road those companies drive on.

Quick hits. Start with the other big one, Marcus. Alibaba dropped Qwen 3.6 27B yesterday on Hugging Face, Apache 2.0 licensed.

And the numbers are striking. A twenty-seven-billion-parameter dense model beats their previous open flagship, the three-hundred-and-ninety-seven-billion-parameter mixture-of-experts Qwen 3.5, on SWE-bench Verified, Terminal-Bench and SkillsBench. On disk it collapses from eight hundred and seven gigabytes to fifty-five. A quantized GGUF fits in about seventeen gigabytes. Simon Willison ran it locally on an M5 Pro at roughly twenty-five tokens per second and called the output excellent for a sub-twenty-gigabyte local model. Native two hundred and sixty-two thousand token context window, extensible to about a million.

So flagship coding on a laptop.

That's the headline. And I want to be straight with listeners about the context, Kate. Alibaba is state-adjacent. Every free top-tier Chinese model strategically erodes the pricing moat that Western labs' investors are counting on. That's not a conspiracy theory, it's the stated playbook. The models are genuinely impressive. They're also a geopolitical weapon aimed at Western AI capex. Both things are true. For self-hosting enterprises and developers in regulated industries, Qwen 3.6 is excellent news. For Anthropic's subscription economics, which we just watched crack this week with the Claude Code Pro pull, it's another squeeze.

OpenAI's enterprise answer. Workspace Agents launched yesterday.

Team-shared, Codex-powered agents that run in the cloud, persist across sessions, and can be invoked from ChatGPT or Slack. OpenAI's examples include a software-request triage agent, a Slack product-feedback router, and a Friday metrics agent that auto-pulls data and generates charts. They're positioning it as the evolution of GPTs, which will be convertible into Workspace Agents soon. Research preview on the Business, Enterprise, Edu and Teachers plans. Free through May sixth. Then credit-based pricing kicks in.

Three-way race on the enterprise agent layer now.

Anthropic's Claude Managed Agents. Google's Enterprise Agent Platform we just covered. OpenAI's Workspace Agents. All three shipped production enterprise agent products in a two-week window. The design choice that matters here is that Workspace Agents ride on existing ChatGPT subscriptions, not API keys. That removes the biggest adoption friction for non-technical teams. It's also the clearest move yet toward the headcount-replaces-tokens pitch that CFOs are already being sold. For anyone building on top of these platforms, the interop question is now genuine. A2A versus whatever OpenAI and Anthropic ship next.

Follow-up on the Meta surveillance story we covered yesterday. New reporting came out on the employee response.

As we reported Wednesday, the Model Capability Initiative records keystrokes, mouse movements and periodic screenshots on Meta employees' work laptops to train agentic models. The new wrinkle today is CTO Andrew Bosworth telling staff there is no opt-out on a work-provided machine. Internal protest has been loud. Some employees are reportedly shifting more of their real work to personal devices, which of course defeats the training value. The read-across I'd flag is that Meta is now directly racing Google and OpenAI for agentic training traces, and the ethical fault line, the same company that built the ad-targeting surveillance machine surveilling its own workers, is going to be a gift to organized labor and to European regulators for the next two years.

Sam Altman's other company had a big week. World ID is no longer niche.

The Sam Altman-founded World project, formerly Worldcoin, famous for its chrome iris-scanning orbs, announced integrations with Zoom, Tinder, DocuSign, Okta, Shopify and VanEck. Zoom is co-building a feature called Deep Face that uses World ID to certify a meeting participant is a real human, not a deepfake. Tinder is expanding a Japanese pilot to the US so daters can prove a live human is behind the profile. About seventeen-point-nine million people have signed up globally, and World is open-sourcing the protocol so any app can integrate it.

The proof-of-humanity problem just became a real market.

With real anchor customers. Whether you find the orbs dystopian or pragmatic, the AI-era question of is there a human on the other end of this connection is now something Zoom is willing to pay to answer. The uncomfortable bridge to the Meta story we just discussed is that we're simultaneously building systems to record humans to train AI, and selling services to tell humans apart from AI. Same problem, opposite direction. And the proof-of-humanity layer for the consumer internet happens to be controlled by the CEO of OpenAI. That concentration of infrastructure power is going to draw antitrust attention eventually. It probably should.

Kate-bait story, Marcus. Emily Hart. A viral pro-Trump influencer.

Entirely AI-generated. WIRED and the New York Post exposed that Emily Hart, a pro-Trump, pro-gun-rights, anti-abortion influencer with millions of impressions and roughly ten thousand followers in her first month, was built by a twenty-two-year-old Indian medical student who goes by Sam. He used Google's Gemini for strategy and Grok for explicit imagery, and monetized the character on Fanvue, a paid creator platform. By his own account, Gemini explicitly recommended targeting older MAGA and conservative men on the grounds they have, quote, higher disposable income and are more loyal. Meta has since removed the Instagram and Facebook accounts. The operation had already earned thousands of dollars.

This isn't a deepfake of a real person. It's a synthetic persona.

And nobody needed to deepfake it because nobody checked it was real. That's the shift. It's a preview of what political influence looks like when any laptop can spin up a convincing full-time persuasion account for pennies. The detail that stopped me cold is that Gemini allegedly scaffolded the audience-targeting strategy itself. That's not a tool being misused. That's a tool participating. Whatever guardrails Google has around, quote, help me build a political influence operation, they did not fire on, quote, help me grow my creator account. Same operation, different framing. Expect regulators on both sides of the Atlantic to start asking Google and xAI very pointed questions about the chain of prompts that built this persona.

The Linux kernel story is my favorite of the day, Marcus, because it's the most human.

The Linux kernel community is formally removing several legacy subsystems. The AX.25 amateur radio stack. NET/ROM. ROSE. ISDN. ATM protocols. ISA and PCMCIA Ethernet drivers. The reason is that AI-powered bug-scanning tools now generate five to ten serious-looking reports per day on these old lightly-maintained modules, and no volunteer wants to triage them. A maintainer wrote, quote, this set of protocols has long been a huge bug magnet, and since nobody stepped up to help us deal with the influx of AI-generated bug reports, we need to move it out of tree to protect our sanity.

So AI tooling is causing open source to contract.

A new class of problem. AI security tools got so cheap they created a maintenance denial-of-service on volunteer infrastructure. And the shift is happening twice. First the AI slop reports flooded in. Now maintainers say genuine findings are also pouring in faster than humans can respond to them. Separately, the Linux 7.0-rc7 cycle added explicit documentation for AI agents on how to file decent kernel bug reports, and Torvalds and the maintainers reached agreement that AI-generated code is allowed in the kernel, but humans remain accountable. That's the right balance. But the AX.25 removal is the small, concrete example of what asymmetric AI pressure does to unpaid volunteers. It makes them quit.

Last quick hit. OpenAI published a post-mortem yesterday on the North Korean npm attack.

On March thirty-first, malicious versions of the Axios npm HTTP client were published after a targeted social-engineering attack against the lead maintainer's PC. Seventy million weekly downloads. Microsoft attributes the operation to Sapphire Sleet, a North Korean state actor. A GitHub Actions workflow at OpenAI, used in the macOS app-signing pipeline, downloaded the malicious version. That workflow had access to the signing certificate used for ChatGPT Desktop, Codex, Codex-cli and Atlas. OpenAI's analysis concluded the certificate was probably not exfiltrated, but they're revoking the old signing certificate on May eighth. Older versions of those apps will stop launching under macOS security protections after that date.

So update your ChatGPT desktop app before May eighth.

Yes. And the broader lesson is that the AI tooling stack is almost entirely a pile of npm and PyPI packages, and state actors have clearly noticed. OpenAI did the right thing. Public post-mortem, certificate rotation, clear user communication. That's how you handle a near-miss. It also underscores that, quote, AI company, and, quote, JavaScript shop with a model API, overlap more than the marketing suggests.

Big picture, Marcus. What do today's stories add up to?

Every story today argues the same thing from a different angle, Kate. The industry is moving from models to agents, and everything downstream has to change with it. Google is building chips, clouds and protocols for millions of agents. OpenAI is building agents that live inside your organization. Meta is recording its own humans to teach agents how to click buttons. Alibaba is putting agentic coding on a laptop. Altman's second company is building the layer that tells humans apart from agents. The Linux maintainers are being drowned by agents generating bug reports. And someone in India ran a fully agentic MAGA persuasion operation out of their bedroom.

One sentence to close.

The race is no longer who has the biggest model. It's who owns the substrate, chips, protocols, identity and training data, on which all these agents will run. Today Google made the most complete bid to own that substrate, OpenAI answered on the software side, and everybody else got a clear look at what a world of cheap, autonomous, always-on AI workers actually looks like in practice.

That's your AI in 15 for Thursday, April 23, 2026. See you tomorrow.