AI just got $535 billion more expensive to build, and harder to buy.

In this edition:

• This week: Anthropic charged a premium for refusal, OpenAI signed a $500B Ohio compute lease, and Apple shipped Siri AI on a Gemini foundation

• Under the radar: xAI was caught training on Claude outputs and continued after detection

• What's on the calendar: Anthropic's IPO window opens, Oracle earnings, and Tesla's first major EU FSD approval

THE WEEK IN AI

THE WEEK IN ONE SENTENCE

The supply side of the AI stack committed roughly $535 billion to building more of it this week, while the demand side got a frontier model engineered to refuse work, a consumer product running on a competitor's foundation model, and the first EU order forcing AI chatbots into a gatekeeper messaging platform. Capacity is being financed at a scale that assumes buyers will pay any number, and buyers are getting unfamiliar prices.

THREE SIGNALS

01 • Model

Anthropic shipped its best model and engineered it to refuse the work that makes it best

Anthropic released Claude Fable 5 on Tuesday, the public version of its Mythos-class model, and held the uncapped Mythos 5 back for a roster of corporate partners, "select biology researchers," and the United States government. Fable 5 runs on the same weights as Mythos and sells for $10 input and $50 output per million tokens, roughly twice the rate of Opus 4.8. When a user asks Fable about cybersecurity, biology, or chemistry, the model refuses and routes the request to Opus 4.8, which by Anthropic's own data, fires in fewer than 5 percent of sessions.

The fourth guardrail is the one buyers should know about. The System Card discloses that requests aimed at building competing frontier models, including pretraining pipelines and ML accelerator design, are silently degraded through prompt modification, steering vectors, or parameter-efficient fine-tuning. Anthropic estimates that this affects 0.03 percent of traffic across fewer than 0.1 percent of organizations. Unlike the cybersecurity and biology routing, this intervention is invisible to the user. The model still answers, just worse, and the buyer cannot tell.

What I keep coming back to is that refusal has become the product feature. At twice the going rate, Fable 5 does not charge a premium for capability. It charges for proximity to capability, for the right to operate one classifier away from a frontier the buyer is not permitted to cross. That is a different pricing logic than the one this category has been running on.

AI READY PRO · FREE UNTIL FRIDAY

I will get you AI trained in 30 days for free

After several months of running it quietly with top AI operators and teams, I’m excited to launch AI Ready Pro today to our newsletter subscribers like you.

It’s a 30-day personalized AI training program that is personalized to you and where you are in your AI journey. It’s the culmination of 100’s of hours spent teaching AI to top Fortune 500 AI teams and operators (and working with Mark Cuban).

Each day, you receive a 10-15 minute exercise to complete to improve your AI skills.

This isn’t vague AI theory or a way to pitch you a tool. It’s 30 days of learning to use AI in your real work. So you come out the other side with the skills to use AI to actually improve you and your teams output.

Until this Friday (May 15), it’s free for newsletter readers who complete the extended assessment. This assessment will help us personalize the learning experience to you.

If you lead a team, AI Ready Team is open today too, also for readers first. It’s the same engine, but to train your whole organization. You’ll get a detailed view of:

• Your team’s AI readiness level + detailed strengths & weaknesses

• Ranked list of AI automation opportunities

• Get tactical advice on how to advance AI efforts in your org

Leading a team? Take the Team assessment instead.

02 • Infrastructure

OpenAI and Broadcom locked in $535 billion for more compute

OpenAI is reported to be in talks to lease a 10-gigawatt Ohio data center campus with Nvidia credit backing. The site would sit on federal land under a 20-year OpenAI lease, with a build-out cost of at least $500 billion and phase one online in 2028. Nvidia hardware credit routes into the lease, which makes OpenAI the financing counterparty for a single piece of land and a 20-year power purchase.

The same day, Broadcom announced a $35 billion AI data center fund co-anchored by Apollo and Blackstone, sized to finance more than 20 gigawatts of AI capacity through 2028 using Broadcom-designed chips. First commitments are tied to projects for Anthropic and OpenAI. The fund routes AI infrastructure debt onto investment-grade balance sheets, so Apollo and Blackstone get a long-dated, contract-backed asset, the labs get capacity they did not have to raise equity to underwrite, and Broadcom gets the silicon order. The structure is starting to look like project finance for power plants, not like cloud capacity buys. If the Apollo and Blackstone tranches price at investment-grade spreads when the first phases close over the next two quarters, the lab-side cost of compute can drop independent of any chip generation, and the price the buyer pays for inference is what holds or moves.

03 • Platform

Apple shipped a multi-provider AI stack, and the EU named the gatekeeper question

On Monday at WWDC, Apple released iOS 27 with Apple Foundation Model Cloud Pro behind Siri AI. Apple confirmed to CNBC's Kif Leswing that the most capable cloud-side model is built on the Gemini foundation and training data, with Apple-led pretraining, post-training, and reinforcement learning on top. OpenAI and Anthropic remain available behind ChatGPT and Claude integrations. That is the first time a major consumer AI surface has shipped with a competitor's foundation model under the brand.

Two days later, Apple told Reuters that Siri AI will not ship in the EU at launch because the Digital Markets Act would force it to give rival assistants deep access to user data. The Commission says nothing in the DMA blocks Siri AI, and the exclusion is Apple's choice. The same week, the EU ordered Meta to allow rival AI chatbots on WhatsApp while a market-abuse investigation continues, the first interoperability ruling on AI distribution against a gatekeeper messaging platform.

Together, they set the regulatory frame for next year. The first frontier consumer AI product runs on a competitor's foundation model and cannot be sold inside the EU under the current Apple legal posture, while the first AI-distribution interop order lands on a gatekeeper messaging platform. Procurement teams that own platform contracts should expect the gatekeeper question before the model question on the next renewal.

UNDER THE RADAR

The most under-reported story this week is the reason Anthropic added the silent fourth guardrail to Fable 5. Reporting in The Information says Anthropic detected xAI training on Claude outputs, confronted the company, and the activity continued. The system-card disclosure follows that detection. The new guardrail is the response.

That puts a different frame on it. The labs are now policing each other inside the product surface, not at the API layer or in the courts. The instrument is a silent classifier and a steering vector built into the model the buyer is using for legitimate work, on the same system every Fable 5 customer is on, with no way for the buyer to see when the classifier fires.

What I keep hearing from operators running internal AI engineering teams is that the procurement diligence question stopped being "what does this model do" twelve months ago. It is now "what does this model refuse to do, and how do I know." Fable 5 turned that into a written specification.

QUOTE OF THE WEEK

"Market pressure will force Anthropic to reverse Fable restrictions. Nerfing coding is where the market says no."

The line lands because Fedus is not the critic Anthropic would expect, and the question the week leaves open is which of those two forces, the system card or the procurement contract, gives first.

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

• Anthropic's public IPO window opens as early as Monday, June 16. Confidential S-1 filed two weeks ago. If both Anthropic and OpenAI move from confidential to public filings within the same quarter, the public AI-pure-play category arrives faster than the late-2026 consensus had it.

• Oracle reports earnings on Tuesday. The next enterprise AI cloud spend read after Google's $920 million per month SpaceX compute deal closed last week. Watch for any commentary on AI workload demand relative to AWS and Azure.

• Tesla received FSD Supervised approval in Denmark. On Wednesday, the first major EU green light after France, Spain, and Portugal. The next two regulator decisions are the ones worth tracking — the country-by-country approval pattern is the operational map for European FSD rollout.

REPLY

Hit reply and tell me whether your company's next AI contract is being scored on the model, on the platform that owns the distribution, or on the gatekeeper that owns the rail. I read every reply.

Have a good weekend,
Haroon