AI News Roundup – October 28, 2026

TL;DR

• OpenAI plans to buy Google’s AI chips (TPUs), The Information and Bloomberg report. No terms are public, and neither company has confirmed a deal.
• Anthropic expanded its GSA deal: 40% off Claude Enterprise, Claude for Government and the Claude API for every level of US government, with no minimum spend.
• Google launched a preview of its Connected Apps Platform, so any developer can bring their app into Gemini and Google Search through MCP.
• You can now text your ChatGPT dot over iMessage or RCS. It’s a beta for ChatGPT Pro users in the U.S.
• Nvidia released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes 40 languages and labels up to eight speakers.
• Delaware drafted a law that would let AI agents run entire companies. House Oversight Chair James Comer is pushing back.
• Axiom says swarms of AI agents produced the first complete computer-checked proof of the Classification of Finite Simple Groups, the longest proof in math, in 24 days.

---

Is OpenAI really buying Google’s AI chips?

Short answer: That’s what two reports say, but nothing is confirmed. The Information reported on Tuesday that OpenAI plans to buy Google’s Tensor Processing Units (TPUs), citing a person familiar with the matter. Bloomberg reported the same, adding that final terms of any deal aren’t known yet. OpenAI and Google declined to comment to Bloomberg.

What we know	Source
OpenAI plans to buy Google TPUs; no financial terms disclosed	The Information, via Reuters
OpenAI is one of the largest buyers of Nvidia GPUs	Reuters
OpenAI began renting Google’s AI chips in June 2025	Reuters
OpenAI also plans to start using its own chips, built with Broadcom, later this year	Bloomberg
Google TPUs come to Oracle Cloud in 2027, the first hyperscaler outside Google to offer them	Oracle and Google

Reuters notes that Google began selling TPUs directly to some customers this year, and that deals with Anthropic have fueled TPU demand. Last week, OpenAI said its frontier models are coming to Google Cloud. Alphabet reports quarterly results later today.

Why it matters: More chip suppliers means more capacity for the AI you use every day, and less dependence on one vendor. It also shows how rivals like OpenAI and Google now buy from each other. Compare the two labs’ assistants: ChatGPT and Gemini. For yesterday’s Oracle–OpenAI news, see the October 27 roundup.

Sources: Reuters via Investing.com · The Information · Bloomberg via Business Standard · Oracle

---

How much cheaper is Claude for US governments now?

Short answer: 40% off. Anthropic’s expanded agreement with the General Services Administration (GSA) covers Claude Enterprise, Claude for Government and the Claude API. For the first time, the discount applies on every product surface: buying from Anthropic directly today, and through cloud providers in the coming months. It also extends to state, local, tribal and territorial governments.

Detail	Anthropic’s figure
Discount	40%
Minimum spend	None: agencies pay for the usage they consume
Estimated savings	$1 billion
Government organizations using Claude	About 500
Claude for Government weekly active users	Up about 8x since February
Utah Claude Code pilot	71% of 34 developers saved 20+ hours in four weeks

Agencies can buy at the OneGov discount through GSA’s Multiple Award Schedule today. Anthropic is also launching a program to help agencies roll Claude out.

Why it matters: Government is turning into a price war for AI labs. If you work in or sell to the public sector, Claude just got a lot cheaper to pilot. See our Claude and Claude Code guides, and Claude Enterprise pricing.

Sources: Anthropic

---

What is Google’s Connected Apps Platform?

Short answer: A single place for developers to build apps that work inside Gemini and Google Search. Developers connect a cloud-hosted MCP server (or Android MCP via the AppFunctions APIs), then register, test and submit the app in the Google Connected Console. They choose whether it goes live in Search, the Gemini app, or both, and can track traffic, latency and error rates afterward.

• Access: Preview, opening in phases. TechCrunch says it starts with U.S. developers on the waitlist.
• Apps today: Google has launched 50+ Connected Apps. Adobe, Experian, Localiza and Peloton are live in Gemini and come to Search in the coming weeks. Calm is coming to both.
• Also today: Chrome’s AI autofill can now suggest details from your Gmail, like a flight confirmation number. It’s opt-in, rolling out to U.S. users on desktop and in Gemini in Chrome on iOS and Android (TechCrunch).

Why it matters: This is SEO for the agent era. If your app isn’t connected, Gemini can’t use it for people. New to MCP? Read what the Model Context Protocol is, or check Gemini pricing.

Sources: Google · TechCrunch · TechCrunch on Chrome autofill

---

Can you text your ChatGPT dot now?

Short answer: Yes, if you’re a ChatGPT Pro user in the U.S. OpenAI’s always-on agents, called dots, can now be reached by iMessage, or RCS on Android. It’s a beta, and one-to-one only.

Texting your dot	Detail (OpenAI)
Who	ChatGPT Pro users in the U.S. (beta)
Send	Text, photos and voice messages
Replies	Your dot can text you proactively with updates
Approvals	It sends a link to review the request in ChatGPT
Setup	Dot profile → Texting → verify your number → save the contact card
Not supported	Group chats, phone calls, FaceTime

OpenAI says your dot keeps the same context and plugins wherever you reach it, and texting doesn’t give the dot or OpenAI access to your other text conversations.

Why it matters: Always-on agents are moving into the apps you already check all day. You can hand off a task from your messages without opening ChatGPT. Learn more in our OpenAI dots guide and dots pricing explained.

Sources: ChatGPT Learn · 9to5Mac

---

What is Nvidia’s Nemotron 3.5 Transcribe?

Short answer: A free, open-weight speech model that does transcription, speaker labels, translation and audio Q&A in one checkpoint. You can guide it with plain-language prompts, like “Transcribe this recording and label each speaker.”

Nemotron 3.5 Transcribe	Nvidia’s figure
Languages transcribed	40
Translation to/from English	31 languages
Speaker labels	Up to 8 speakers, in 13 languages
Max recording	About 1 hour
Size	30B parameters, about 4B active per token
Open ASR Leaderboard (English)	#1 open-weight model, 5.33% average WER

It works on recorded audio, not live streams. Nvidia also released Granary v2, an open speech dataset covering 463,062 hours in 48 languages. The model is on Hugging Face and ModelScope, with quantized versions for local use “coming soon.” Leaderboard claims are Nvidia’s.

Why it matters: Free, high-quality transcription you can run yourself is good news for meeting notes, podcasts and medical dictation, especially if your audio can’t leave your own servers. Compare hosted options in best AI meeting notetakers, Otter.ai and Fireflies.ai, or read what open-source AI means.

Sources: Nvidia on Hugging Face · TechCrunch

---

Could AI agents legally run a company in Delaware?

Short answer: Not yet, but a draft bill would allow it. The proposal creates an Artificial Intelligence Company (AIC) whose business can be managed by an AI agent, inside a 30-month regulatory sandbox, Ars Technica reports. An AIC could have any lawful purpose except banking.

• Who’s responsible: A single member (a person or company) funds the AIC and must keep it adequately capitalized. The AIC must keep a log of its activities.
• Liability shield: The draft says the member isn’t liable for what the AIC does, with exceptions such as fraud, willful violations of law, or failing to fund it properly.
• Pushback: Rep. James Comer, who chairs the House Oversight Committee, wrote to Governor Matt Meyer on Oct. 26 asking for documents. He warned the plan could shield AI companies from responsibility for reckless harm, and floated federal limits on state AI rules.
• Timing: Spotlight Delaware reported the bill is likely to be introduced next year (via Ars). About two-thirds of Fortune 500 companies are incorporated in Delaware.

Why it matters: Who is liable when an AI agent signs a contract or causes harm is becoming a real legal question. It’s a draft, not law. Background: what an AI agent is and why agents need an escalation policy.

Sources: Ars Technica · House Oversight letter (PDF)

---

How did AI agents check math’s longest proof?

Short answer: With a lot of agents and a proof checker. Axiom says it released the first complete computer-checked proof of the Classification of Finite Simple Groups, which lists the basic building blocks of all finite symmetry groups. The original proof runs about 15,000 pages across 500+ articles by about 100 mathematicians. Mathematicians call it “the enormous theorem.”

Axiom’s project	Detail
Time	24 days
Workers	Engineers and mathematicians directing swarms of dozens of AI agents
Output	About 43.1 million lines of Lean code
Problems found	34 noteworthy deviations from the published record
Biggest fix	A rerouted proof of the Unbalanced Group Theorem, using a new result proved by AxiomProver

Group theorists Richard Lyons and Ronald Solomon were consulted, and Axiom plans to publish an errata list. Axiom notes that humans still check that the final formal statement says what it should. The Lean software checks every step of the proof itself.

Why it matters: AI agents can now check work no human team could fully re-read, and they found real errors doing it. The same idea applies to auditing code, contracts and research. See yesterday’s AI math claims, our guide to AI reasoning models, and AI agent orchestration.

Sources: Axiom

---

What should you watch next?

• Chips: Whether OpenAI or Google confirms a TPU deal, and what Alphabet says about TPU sales on today’s earnings call.
• Agents in your apps: Which apps show up in Gemini and Search as Connected Apps, and whether dot texting expands beyond U.S. Pro users.
• Agent liability: Whether Delaware introduces the AI-company bill, and whether Congress moves to limit state AI rules.

AI Tools Optimizer is for education only. Figures come from company announcements and the press reports linked above; benchmark claims are company-reported. Verify primary sources before making decisions. This is not investment or legal advice.