Apple sues OpenAI, Nadella calls out distillation hypocrisy, and New York freezes data centers

Apple's trade secrets lawsuit rocks OpenAI, Nadella calls out AI labs' data double standard, NY enacts a data center moratorium, and Soofi S drops a strong open 30B model.

The big picture

The through-line today is accountability — who gets to benefit from data, and who gets to say what’s proprietary. Apple is suing OpenAI for allegedly stealing hardware secrets while Satya Nadella is calling out the entire closed-model industry for a different kind of hypocrisy. On top of that, New York just enacted the country’s first statewide data center moratorium, which could rewrite where the next generation of AI infrastructure actually gets built. Busy day.

Apple vs. OpenAI: The trade secrets lawsuit is genuinely wild

Apple filed a blockbuster lawsuit against OpenAI alleging that the company conspired with former Apple employees to steal confidential documents, spy on hardware prototypes, and trick a trusted partner into revealing a proprietary manufacturing technique. The central figure is Tang Tan, a 24-year Apple veteran who served as VP of Apple Watch before departing to join OpenAI. The complaint also alleges that OpenAI’s hardware team asked job candidates to literally bring unreleased Apple components and product samples to interviews, and that employees joked internally about unauthorized access to Apple systems. The Verge TechCrunch Ars Technica

If even a fraction of this holds up in court, it is a massive problem for OpenAI — not just legally but reputationally, at exactly the moment it is trying to position itself as a trustworthy enterprise partner. The hardware angle is telling: OpenAI has been vocal about building its own devices, and Apple’s allegations suggest the methods of getting there were considerably less clean than a normal talent poach. The ask-candidates-to-bring-prototypes-to-interviews allegation, in particular, reads less like aggressive recruiting and more like industrial espionage. Worth watching very closely — this isn’t a thin IP dispute, it’s a trade secrets complaint with named individuals and specific alleged acts.

Nadella’s distillation broadside and what it means for your AI stack

Microsoft CEO Satya Nadella publicly called out OpenAI and Anthropic for what he labeled a “reverse information paradox”: these labs train on vast swaths of public internet data under fair-use arguments, learn continuously from customer interactions, but then prohibit anyone from using their model outputs to train competing models (a practice called distillation — essentially teaching a smaller model by having it imitate a larger one). Nadella’s proposed remedy is that companies should own and control their own learning infrastructure rather than handing that advantage to the model providers. The Decoder TechCrunch

It’s worth naming the obvious: Microsoft sells exactly the kind of on-premises and sovereign AI infrastructure Nadella is recommending. So this is competitive positioning dressed as ethical critique. But the underlying argument isn’t wrong — there is a genuine asymmetry when labs hoover up the world’s text and then lock down what flows out of their models. For developers, the practical implication is real: if you’re fine-tuning on GPT or Claude outputs today, you’re operating in a gray zone that OpenAI and Anthropic’s ToS actively prohibit, even though those same companies built their advantage on data they didn’t create. Whether courts or regulators eventually force a change here is unclear, but Nadella’s public framing makes it a mainstream conversation now.

New York freezes large data centers — and the infrastructure implications are serious

Governor Kathy Hochul signed the nation’s first statewide data center moratorium, blocking new environmental permits for any facility over 50 megawatts for up to a year. The explicit rationale is giving the state time to develop regulations protecting residents from rising energy costs and environmental impact. A separate, stricter bill passed by the state legislature (using a lower 20 MW threshold) still awaits her signature. The Verge

Fifty megawatts is a meaningful number — that’s roughly the scale where hyperscale cloud regions start. This doesn’t freeze smaller colocation or edge facilities, but it effectively tells AWS, Google, Microsoft, and any AI-first infrastructure company that New York is off the table for major expansion for at least 12 months. Given how fast compute demand is outrunning grid capacity, expect other states to watch this closely. If it becomes a model for other blue-state jurisdictions, the geography of AI infrastructure could shift substantially toward the Sun Belt and Midwest, which have more permissive energy environments. The energy cost argument is the one with the most political traction — nobody campaigns on cheaper server farms, but “your electricity bills are going up because of data centers” is a message that lands.

Soofi S 30B: Europe’s open-model moment

A German research consortium released Soofi S 30B-A3B, an openly licensed language model trained entirely on Deutsche Telekom’s cloud infrastructure in Munich. The architecture is a mixture-of-experts hybrid (meaning only a fraction of its 31.6 billion parameters are active per token, keeping inference fast without sacrificing capacity), and the training data was deliberately skewed toward German text. The result reportedly tops all fully open competitors on both German and English benchmarks. The Decoder

This is the kind of release that matters more than its benchmark numbers suggest. Non-English open models have historically been underserved — most of what ships from US labs is English-first with multilingual as an afterthought. A model that actually prioritizes German (and apparently doesn’t sacrifice English performance to do it) is exactly what European enterprise developers have been waiting for. The sovereign infrastructure angle (Deutsche Telekom’s Munich cloud, no US hyperscaler involvement) will also resonate with organizations that have compliance constraints around data residency. Watch for the model to gain adoption in German-speaking markets fast.

Funding and valuations: Video generation and agent tooling attract serious capital

PixVerse, a Singapore-based AI video generation startup, closed an extended Series C that values it above $2 billion after raising $439 million. The company is targeting world model capabilities and international expansion with the cash. TechCrunch The Decoder

Video generation is already crowded — Runway, Sora, Kling, Veo — so a $2B bet on yet another entrant is either investor confidence in the market size or a fear-of-missing-out reflex. PixVerse has built real user traction, particularly in Asia, but the question of whether there’s room for multiple large independent video generation companies is genuinely open. At $439M, they have the runway to find out.

Separately, Nous Research, the team behind the Hermes series of open-weight agent-tuned models, is in talks for a new funding round of at least $75 million at a $1.5 billion valuation, led by Robot Ventures with participation from Union Square Ventures. TechCrunch

Nous Research is genuinely interesting to developers because Hermes models are among the best open-weight options for building agentic workflows — they’ve been fine-tuned specifically for tool use and instruction following, not just chat. A $1.5B valuation is a big number for a company whose primary output is fine-tuned open-weight models, but it reflects the bet that the infrastructure layer for AI agents is where the durable value accrues.

Research and safety: The interesting and the inconclusive

Rich Sutton — 2024 Turing Award winner and one of the architects of modern reinforcement learning (RL is the training approach where an agent learns by trial and error with reward signals) — has founded Oak Lab in Toronto to build AI agents that learn continuously from their environment. He’s publicly characterized current deep learning methods as “weak and inefficient.” The Decoder

Sutton is not a person whose opinions you dismiss. His “Bitter Lesson” essay (scale and search beat hand-crafted representations every time) shaped a decade of ML research. If he thinks the current LLM paradigm is weak, that’s worth engaging with seriously. Whether Oak Lab produces results is another matter — it’s early-stage and thin on details — but the directional bet on continuous online learning versus static pre-training is a real research frontier.

MIT Technology Review has a measured look at Anthropic’s latest research publication, which the piece treats with appropriate skepticism — noting that Anthropic, now valued at nearly $1 trillion, has a habit of publishing research that is genuinely novel but also imperfectly interpreted. The review distinguishes between what the research actually shows and what enthusiastic summaries tend to claim. MIT Technology Review

Separately, a new Anthropic study looked at how Claude’s expressed values shift across languages — finding more warmth in Hindi responses and more rigor in Russian ones, among other patterns. It’s interesting applied sociolinguistics, and it raises real questions about what “alignment” means when the model’s behavior is language-dependent. The methodological questions the study itself raises are as notable as its findings. The Decoder

More than 200 economists and AI researchers including 16 Nobel laureates signed a coordinated statement warning that the window to prepare for AI’s economic impact is closing. The comparison to the Industrial Revolution is doing a lot of work here, and notably the statement doesn’t propose concrete policy measures. Labor market studies haven’t yet shown significant AI-driven employment effects, which makes the alarm feel ahead of the evidence — even if the underlying concern is legitimate. The Decoder

On the security side, Ars Technica has a solid piece on “context bombing” as a defensive tactic against AI hacking agents. The idea is to overwhelm an attacking agent’s context window with noise, causing it to shut down before it can complete harmful actions — essentially using prompt injection offensively as a defense. Ars Technica It’s clever and very early-stage, but the broader implication — that AI-vs-AI security dynamics are arriving — is something production system builders should be thinking about now.

Developer tools and the coding agent productivity signal

Simon Willison published a genuinely fun link: DOOMQL, a Doom-like game where SQLite is the actual game engine (handling movement, collision, rendering via a recursive CTE ray tracer) rather than a backing store. It was built using GPT-5.6 Sol and the SQL involved is legitimately impressive. Simon Willison’s Weblog Not practically useful, but a nice demonstration of what’s possible when you treat SQL as a general computation substrate.

More substantively, Willison also shared a GitHub code-frequency chart for his Datasette project showing the largest spike in contributions in the project’s history, aligning with his recent use of Claude Opus 4.8, GPT-5.5, and GPT-5.6 Sol as coding agents. Simon Willison’s Weblog The chart is anecdotal but it mirrors what a lot of developers are reporting: the current generation of coding agents isn’t producing marginal gains, it’s enabling a qualitatively different output rate for solo developers and small teams.

Willison also posted a clean TIL on cache-friendly usage of uvx in GitHub Actions, using a UV_EXCLUDE_NEWER environment variable to pin tool resolution to a specific date and use it as a cache key. Simon Willison’s Weblog Thin on narrative, but immediately useful if you run Python tooling in CI.

OpenAI published a new prompting guide aimed at everyday users rather than developers. The framework is four optional building blocks — goal, context, format, and constraints — and the core message is to describe the output you want rather than the steps to produce it. First time OpenAI has unified Chat and Codex guidance in a single document. The Decoder Not a developer document, but useful for teams onboarding non-technical colleagues to AI tools.

Quick hits

Waze is rolling out new AI-powered features driven by Google’s Gemini assistant, part of Google’s broader push to embed Gemini across its product portfolio and compete with Apple Maps. TechCrunch
Anthropic is now offering rupee-denominated subscription plans for Claude in India, its biggest market outside the US — a localization move that signals serious commercial intent in the region. TechCrunch
Apple’s iOS 27 first public beta is out; the Siri AI upgrades are apparently changing how users interact with the phone day-to-day, though the overall release is refinement-focused rather than feature-heavy. The Verge
Google DeepMind launched ATL Saathi, a Gemini-powered tool for Indian educators running school robotics labs. Google DeepMind
Sam Altman dismissed the idea of space-based data centers publicly; TechCrunch notes most infrastructure experts already agree with him. TechCrunch
Ars Technica has a solid explainer on world models — the promise, the limits, and what’s still genuinely unsettled in the research. Ars Technica