The government just became an AI counterparty

In this edition:

• This week: Fable 5 returned mid-contract with government-negotiated restrictions, OpenAI offered Washington a $42.6B equity stake, and Anthropic is in talks to design its own chip

• Under the radar: Anthropic embedded tracking code in Claude Code in March, and disclosed it this week after developers found it

• What's on the calendar: The July 7 Fable 5 credits-only cutover, Crunchbase's $510B H1 venture record, and AI Engineer World's Fair sessions circulating from OpenAI, Anthropic, and DeepMind

THE WEEK IN AI

THE WEEK IN ONE SENTENCE

The US government renegotiated a frontier AI model's deployment terms, OpenAI offered Washington a $42.6 billion equity stake, and Anthropic started designing its own chip. Three moves that look less like a product week and more like the early formation of a regulated industry.

THREE SIGNALS

01 • Policy

The government is now a negotiating party in your AI vendor relationship

Claude Fable 5 returned globally on Wednesday after Anthropic announced it had reached an agreement with the US government following the model's export-control pause. The terms: new cybersecurity classifiers, a 50% usage cap through July 7 after which Fable moves to credits-only pricing, and early access for government partners to future unreleased models. One more detail buried in the announcement: coding and debugging tasks now fall back to Opus 4.8. No timeline on "near term."

This is the first commercial AI model to be pulled, renegotiated with a government, and redeployed with capability restrictions. The mechanism matters more than the specific restriction. A company that bought Fable 5 seats for its engineering team had those seats redefined, mid-contract, by a negotiation it was not party to. The government did not need to contact customers. It contacted the vendor.

I keep hearing a version of this question from operators who have started building on frontier models: what happens to my workflow if the government changes what my vendor is allowed to ship? This week is the first time that question has a documented answer, and operators who are building on frontier infrastructure should treat the Fable 5 case as the template for how the next one will go.

TEAM READINESS AUDIT

Get your team’s AI readiness score. Leave with an AI readiness brief.

A 15-minute assessment that turns your answers into a real decision artifact: your readiness level, your six-axis shape, where you're strongest, where you're constrained, and what not to build yet.

02 • Market

OpenAI proposed giving the US government a 5% stake in itself

CNBC reported Wednesday that OpenAI proposed giving the US government a 5% equity stake worth roughly $42.6 billion at its current $852 billion valuation. Sam Altman also floated extending the proposal: Washington holding 5% of every leading US AI developer, including Anthropic, Google, and Meta.

What gets glossed over in the coverage is that this is not a revenue-sharing arrangement or a regulatory concession. Equity means governance rights and economic interest. A government with 5% of OpenAI is not a regulator standing outside the company. It is a shareholder.

Gergely Orosz, among others, flagged the obvious question: what does the government get besides the stake, and what does OpenAI get in exchange? That is still not public. The Fable 5 timeline runs alongside this proposal, and the two are probably connected, though neither Anthropic nor OpenAI has said so directly. What is visible is the shape of the deal: labs offer access and equity, and in exchange the government lifts export restrictions and stays out of the capability decisions, mostly.

03 • Infrastructure

Anthropic is in talks to design its own chip, on Samsung's 2nm process

The Information reported Thursday that Anthropic is in early talks with Samsung to develop a custom AI chip on Samsung's 2nm process. Amazon, Anthropic's largest cloud infrastructure partner, already runs its own Trainium chips. Google has TPUs. Microsoft has Maia. OpenAI has been reported to be working on custom silicon for over a year. Anthropic is the last major frontier lab without it.

The case for a custom chip is straightforward: inference at scale is the largest cost driver once a model is trained, and the margin between Nvidia pricing and custom silicon can be substantial at the volumes the frontier labs are running. Samsung's 2nm process is competitive with TSMC's, and Anthropic's existing relationship with Amazon gives it a distribution channel for whatever chip it designs, since Amazon could run it on Bedrock.

The talks are early. Custom chip development takes three to four years and hundreds of millions in engineering investment before a single inference runs on production hardware. But the pattern across the industry is now complete. Every lab that expects to be running at frontier scale in five years is designing its own silicon.

UNDER THE RADAR

The story that got buried this week: Anthropic quietly embedded tracking code in Claude Code in March, and disclosed it this week after developers found it.

The code, as described by Jim Clyde Monge, checked whether users were routing Claude through proxies. It also scanned for Chinese time zones, proxy hostnames, and domains linked to Chinese companies or resale services. The signal was passed not through a visible telemetry field but hidden in the system prompt through changes like date separator formatting and Unicode apostrophes. Anthropic said it was an anti-abuse experiment targeting unauthorized resellers and model distillation, and that the behavior is being rolled back.

What I keep hearing from procurement teams is that trust in AI vendors is still running on goodwill, not on contracts. Most enterprise AI agreements do not specify what vendors can and cannot embed in the tooling. The Anthropic tracking incident is a reasonable case study for what that gap looks like. The disclosed intent was legitimate (stopping resale and distillation), but the method was invisible to buyers. At the moment most of the industry is trying to agree on what responsible AI vendor behavior looks like, an experiment like this sets the conversation back.

QUOTE OF THE WEEK

"Kimi K2.7 Code is the first open-weight model offered as a selectable option in the Copilot model picker"

The line matters because Copilot has been a closed model surface since launch. One open-weight model in the picker is a small change in policy terms and a large one in signal terms: the enterprise coding tool that set the standard for closed AI deployment just opened a door.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

WHAT’S ON THE CALENDAR

• The AI Engineer World's Fair wrapped on Wednesday at Moscone West after three days and 6,000-plus attendees. Sessions from OpenAI, Anthropic, and DeepMind are circulating. Worth scanning for anything on agent deployment in production environments.

• Crunchbase's H1 2026 venture data landed this week: $510 billion invested globally, a half-year record, with OpenAI and Anthropic accounting for $217 billion of it, or 43% of all startup funding. Q2 alone was $205 billion into 5,000-plus startups.

• The July 7 cutover is the next Fable 5 inflection point, when the 50% weekly usage cap ends and credits-only pricing begins. If your team uses Fable, that is the moment to confirm what your contract actually says.

Have a good weekend,
Haroon