AI News September 15, 2026: Anthropic Locks In $517B in Compute, Trump Attacks Amodei, 'Project Lily' Exposes ChatGPT Privacy, Z.AI Raises $5B, and Enterprises Pull Back From Claude Fable

While the model-release treadmill paused for the weekend, the AI industry’s bigger forces — capital, politics, and trust — all lurched forward at once.

1. Anthropic’s Compute Commitments Hit $517 Billion — Triple Its Earlier Estimate

The most staggering number of the year in AI infrastructure: Anthropic has agreed to $517 billion in compute commitments covering 14.8 gigawatts of capacity in the 11 months through August 2026 — nearly triple the ~$180 billion in server-leasing spend it had previously told investors to expect through 2029.

Amazon and Alphabet together account for roughly 11GW and more than $300B across ~10-year agreements. And the spending isn’t done: the company reportedly still needs more capacity as it prepares for a long-rumored IPO while reportedly seeking up to $100 billion at a $2 trillion valuation.

The same weekend, The Information identified Anthropic as the previously unnamed GPU customer in Rum Group’s $13.7 billion, six-year compute agreement — spread across three ~$4.57B tranches at a Maysville, Georgia site still under development. Rum Group owns Rumble and hosts Truth Social, giving Anthropic’s infrastructure buildout an unlikely political entanglement.

Why it matters: Anthropic has effectively bet the company on compute, and the $517B figure rewrites the scale of the AI capex race. When one lab’s commitments approach the GDP of a mid-sized economy, “frontier training runs” become a macroeconomic event.

2. Trump Attacks Amodei as the US–China AI Argument Escalates

President Trump lashed out at Anthropic CEO Dario Amodei on Truth Social Monday, calling him a “perfect little angel” and declaring that the only AI “guardrail” the country needs is “a STRONG AND SMART (High IQ!) PRESIDENT” — adding that his administration holds “tremendous CRIMINAL and REGULATORY power.”

The attack follows Amodei’s CBS interview urging the industry to slow AI development, and his Saturday essay pressing the US to maintain chip and equipment export controls against China. Beijing’s response was swift: Foreign Ministry spokesperson Guo Jiakun dismissed the essay as “fearmongering” that “will only disrupt the process of global AI governance.”

Why it matters: The exchange lands days before planned mid-September US–China AI safety talks and the September 24 Trump–Xi summit in Washington. Amodei called the US–China race the “toughest dilemma” in AI; this week it became a three-way argument among Washington, Beijing, and the lab CEOs themselves.

3. OpenAI’s ‘Project Lily’: Contractors Are Reading Real ChatGPT Prompts

A 404 Media investigation exposed “Project Lily,” OpenAI’s internal program hiring hundreds of contractors to read real ChatGPT user prompts — many containing sensitive personal information — and rate the responses on a 1–7 scale. Contractors are recruited via Crossing Hurdles and paid through Mercor, with one worker reporting over $50 per hour.

With ChatGPT surpassing 900 million users and training on by default for consumer plans, the piece notes the practice spans the industry — Anthropic and Google run comparable human-review programs. Opting out of data use, where available, is buried in settings.

Why it matters: “Humans in the loop” has been marketed as a safety feature; Project Lily shows the loop includes your actual conversations. Expect renewed regulatory pressure on default-on training and disclosure — the EU AI Act’s transparency provisions were built for exactly this fight.

4. Z.AI Raises $5 Billion to Fund Next-Generation GLM Models

Z.AI (formerly Zhipu) launched a concurrent $5 billion raise: up to 21.965 million H-shares at HK$714 (~HK$15.68B net) plus RMB20.14B ($3.016B) in zero-coupon convertible bonds due 2027 at a HK$892.50 conversion price.

The company will direct 60% of proceeds to next-generation GLM foundation models, training, and inference — continuing the open-weight push that made GLM-5.3’s released weights a September landmark.

Why it matters: Amid the CISA advisory alleging six Chinese labs distilled US frontier models, Z.AI’s raise shows Chinese labs aren’t slowing — they’re institutionalizing capital access through Hong Kong markets while Western labs lean on private megarounds.

5. Nvidia, Palantir, and Booz Allen Curtail Claude Fable Over Data Retention

The Information reports that Nvidia, Palantir, and Booz Allen Hamilton are limiting internal use of Anthropic’s Fable model after a June policy change gave Anthropic rights to retain usage logs for 30 days to defend against “complex and novel attacks.”

The fallout is concrete: a major US utility pulled a Fable deployment running core power infrastructure over Anthropic’s refusal to sign a non-revocable zero-data-retention clause, and Booz Allen has barred Fable from proprietary cybersecurity client work.

Why it matters: The industry’s most safety-forward lab is learning that trust is asymmetric — a retention policy meant to catch misuse reads to enterprise buyers as a data-exposure risk. Zero-retention clauses are becoming procurement table stakes for regulated AI deployments.

Quick hits: Nvidia debuted the RTX PRO 5500 Blackwell workstation GPU with 84GB GDDR7 (~$10K estimated) amid a memory shortage; X Corp and SpaceXAI dropped their Apple antitrust suit over ChatGPT’s exclusive Siri integration while pressing on against OpenAI; OpenAI’s EMEA policy chief urged the UK to pass binding frontier-AI rules — including for OpenAI itself; Andon Labs launched Pion, letting AI agents run real businesses end-to-end after two years of vending-machine and café experiments; and OpenAI DevDay is set for September 29 in San Francisco, where GPT-6 Astra’s wider release is expected.

$517 billion in compute, a president name-calling a lab CEO, and contractors reading your prompts at $50 an hour — the AI story is now a political economy story. Follow our daily AI news coverage for tomorrow’s update.