TechWithAdit (TWA)

Anthropic just paid $1.5 billion to settle the first major AI copyright case in US history. A federal judge approved the settlement yesterday — the largest known US copyright payout ever — establishing that training AI on books is fair use, but downloading 7 million pirated books to a server library is not. Every AI lab with a pending copyright lawsuit now has the first real data point for what resolution looks like.

On the ground, TCS is acquiring land in Vizag and Pune to build India's OpenAI data centres — the India AI ambition is moving from press releases to concrete foundations. The Trump administration's AI standards chief resigned yesterday, the third leader of the agency in four months, as US AI executives openly alarm about the competitive threat from Chinese open-weight models that now account for 45% of US company token use.

Tata Consultancy Services has acquired land in Visakhapatnam and Pune to build the physical infrastructure for its OpenAI-linked AI data sites, according to a report today. The land acquisitions mark the first concrete on-the-ground steps of the TCS-OpenAI partnership announced in February 2026, when OpenAI signed on as the first customer of TCS's HyperVault data centre business, committing to an initial 100MW of capacity with a roadmap to 1 gigawatt. TCS has committed up to $7 billion to the HyperVault buildout, backed by a $1 billion investment from TPG, and is in active discussions with multiple other hyperscalers to bring additional customers into the same infrastructure.

The choice of Vizag and Pune is strategic. Vizag offers lower land costs, proximity to a new submarine cable landing station, and state government incentives tied to Andhra Pradesh's AI corridor push. Pune gives TCS proximity to its largest engineering talent pool and to the dense cluster of enterprise clients already running TCS-managed IT environments.

The insights: India's AI infrastructure story has been discussed for 18 months. Land acquisition is the moment it becomes irreversible. For every operator thinking about where to locate AI compute for India-facing workloads, this buildout will matter — TCS is positioning HyperVault as the sovereign, regulated alternative to hyperscaler cloud in India.

The Wall Street Journal is reporting that top American AI executives are escalating their warnings about the competitive threat from Chinese open-weight models, as internal company data shows Chinese alternatives — including Kimi K3, DeepSeek, and Alibaba's Qwen series — now account for roughly 45% of US company token use, a figure that has roughly doubled in six months. The alarm is directed specifically at the open-weight format: models whose weights are publicly downloadable, run on any hardware, carry no export controls, and price in at a fraction of what US frontier labs charge. The Trump administration was reportedly weighing a ban on Chinese open-weight models over the weekend before backing down after immediate backlash — including from former AI czar David Sacks, who called it protectionism dressed as security policy.

The economics underneath this are unignorable. Kimi K3's latest pricing is $3 per million input tokens and $15 per million output tokens — a fraction of Claude Fable 5 or GPT-5.6 Sol. The frontier capability gap has also narrowed: Kimi K3 topped the Frontend Code Arena leaderboard, and multiple Chinese models are now within striking distance of US frontier performance on specialist benchmarks.

The insights: The US-China AI competition has moved from chip export restrictions to model-level market share — and the score is moving fast. For operators evaluating AI providers, Chinese open-weight models are no longer an emerging option. They are an active choice that nearly half of US companies have already made. The question is not whether to evaluate them. It is whether your governance and security framework permits it.

Chris Fall, the director of the Center for AI Standards and Innovation (CAISI), resigned Monday, the Commerce Department confirmed. He was appointed just three months ago — the second director since David Sacks stepped down from his White House AI and crypto czar role in March and was never formally replaced. The director before Fall, Collin Burns, lasted less than a week before being reportedly pushed out because of his previous work at Anthropic, at a time when the administration was battling the company over Fable 5 export controls. No reason was given for Fall's departure. Arvind Raman, former dean of engineering at Purdue University, has stepped in as acting director. The agency is responsible for developing AI testing and evaluation capabilities and supporting standards for advanced AI systems.

The timing is pointed. Fall's exit comes as the administration is wrestling with how to respond to Kimi K3's benchmark performance, reportedly weighing whether to ban Chinese open models — a move that would require exactly the kind of policy coherence that three leadership changes in under a year makes very difficult to achieve.

The insights: The most important AI testing and safety institute in the US government has now had three directors in under a year. At a moment when Chinese open models are challenging US frontier performance and the administration is considering sweeping regulatory action, the governance vacuum at the top of the federal AI apparatus is not a bureaucratic footnote. It is a structural risk.

Google is developing a new custom AI chip specifically designed to reduce the cost of running Gemini inference workloads, TechCrunch reports. The chip targets the inference layer — the part of AI that runs in production at scale, serving actual user queries — rather than the training phase. Google already has its TPU (Tensor Processing Unit) architecture for training, but the new chip is distinct, targeting efficiency specifically for deployed Gemini models. The development arrives as the economics of AI inference come under increasing pressure: Chinese models are pricing inference dramatically below US frontier rates, and every major lab is racing to reduce per-token costs before the market normalises at lower price points.

The insights: Google building a Gemini-specific inference chip is the confirmation that the inference cost war is structural, not cyclical. OpenAI has Jalapeño. Anthropic is in talks with Samsung. Meta targets September production. Now Google is building its own. Nvidia's inference business is surrounded on every side.

Operator's Spotlight Read

A federal judge in San Francisco granted final approval Monday of Anthropic's $1.5 billion copyright settlement with a group of authors who accused the company of using pirated books to train Claude. US District Judge Araceli Martinez-Olguin signed off on the deal — the largest known settlement of a US copyright case and the first major AI copyright case to reach final resolution — rejecting objections from some authors who argued the amount was too low. The settlement will pay $3,000 per work across an estimated 500,000 works, shared among the authors and publishers who hold rights to them. Potential damages, had the case gone to trial, could have run into the hundreds of billions of dollars.

The legal history matters for understanding what the ruling actually means. The original judge, William Alsup, ruled in 2025 that Anthropic had made fair use of the authors' work to train Claude — a significant win for the entire AI industry. But he also found that Anthropic violated the authors' rights by downloading and storing more than 7 million pirated books in a centralised library that was not necessarily used directly for AI training. That distinction — training on books is fair use, storing pirated copies is not — is now the legal framework that every AI company operates within. Anthropic's own statement was precise: "We reached this settlement in 2025, after the court's landmark ruling that training AI on books is fair use under copyright law — which remains the law today."

The settlement closes one case and opens the broader question. There are dozens of similar lawsuits pending from authors, journalists, and news organisations against OpenAI, Google, Meta, and others. None of them have settled yet. The Anthropic case establishes the price of resolution at $3,000 per copyrighted work — and gives every party in every other case a number to negotiate from.

The insights: The copyright era for AI training is not over — it is being priced. The Anthropic settlement sets a reference rate that will now anchor every negotiation, every demand letter, and every courtroom argument in the twenty-plus AI copyright cases still pending. For operators building products on top of AI models trained on public data, the legal exposure of your AI stack is now quantifiable. That changes how enterprise clients evaluate vendors, how insurers price AI liability, and how regulators think about AI training data standards. The question is no longer whether AI companies owe something for the content they trained on. It is how much — and now there is a number.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit