Chinese AI heats up the chip wars, MCP gets easier, and Claude Code ships 65% of its own PRs

Nvidia faces AMD pressure from Microsoft and Anthropic, MCP usability improves, Google's Frozen v2 chip targets 10x TPU efficiency, and Anthropic's $1.5B settlement closes.

The big picture

The dominant thread today is geopolitics bleeding into AI infrastructure at every layer — Chinese models are rattling US policy, Alibaba’s TTS is topping leaderboards, Nvidia’s pricing power is softening under AMD pressure, and there’s a serious proposal circulating to reshape US copyright law around distillation. The other thread worth tracking: the tooling for building AI systems is quietly maturing, with MCP getting friendlier and Claude Code’s team dropping genuinely surprising metrics about how AI now writes production code.

The US-China AI war is no longer just about benchmarks

Two pieces this week make the same argument from different angles: stop acting surprised by Chinese AI. The Verge’s piece on China’s AI competitiveness points out that every time a Chinese lab ships something capable — Deepseek in January, now two more frontier-competitive models at the World AI Conference in Shanghai — US media reaches for the Sputnik metaphor and markets wobble. The pattern itself is the story at this point. Treating each Chinese release as a “surprise breakthrough” suggests the people writing those headlines haven’t been paying attention to Qwen, Kimi, or the pace of Chinese open-source releases over the past 18 months. The Verge

Ben Thompson’s proposal, surfaced by Simon Willison, is the most concrete policy idea to come out of this week’s hand-wringing. Thompson argues the US should pass a law making AI training data collection explicit fair use, while simultaneously prohibiting terms of service that ban distillation — the practice of training a smaller model on the outputs of a larger one (think: teaching a student model by having it answer questions alongside a teacher model). His logic: labs already trained on unlicensed data, so their anti-distillation ToS is hypocritical; and since stopping distillation is technically near-impossible to enforce anyway, the US might as well lean into it and let American researchers build freely on frontier models. Separately, Willison flags that Alibaba’s reversal on releasing Qwen 3.8 Max as open weights may trace back to a Xi Jinping speech explicitly endorsing open source and open collaboration. If that’s accurate, Chinese AI openness is now literally state policy. Simon Willison’s Weblog

The domestic policy side is a mess. Trump’s AI czar revolving door continues — the director role for the Center for AI Standards and Innovation (CAISI) has turned over again after David Sacks departed. MIT Technology Review reports that current and former Trump AI advisors were publicly attacking leading US AI companies over the weekend. It’s hard to form coherent AI policy when the people responsible for it are fighting each other in public. TechCrunch / MIT Technology Review

OpenAI’s own position in this debate is awkward. TechCrunch has a sharp piece on OpenAI’s fear of open-weight models — and the tension it creates: you can’t simultaneously lobby for restricting Chinese open-weight models and claim to be pro-open AI development. The piece correctly identifies that OpenAI’s anxiety is partly about business model survival. Open weights commoditize the thing they’re trying to sell. TechCrunch

Nvidia’s monopoly is showing cracks

Microsoft is expanding Azure AI infrastructure using AMD’s new Helios platform, with deployment targeted for the second half of 2026. That’s a direct competitive challenge to Nvidia’s H100/H200 dominance in cloud AI training and inference. More pressure: a public GitHub profile apparently shows Anthropic testing AMD hardware too. Neither of these is a done deal — AMD has been “about to compete” with Nvidia for two years now — but having both Microsoft and Anthropic evaluating alternatives publicly is the kind of signal that erodes pricing power even before a single Helios cluster ships. The Decoder

Google’s longer-term bet is even more aggressive. Reports say Google is developing a chip codenamed “Frozen v2” that bakes the Gemini model architecture directly into silicon — meaning the hardware is purpose-built for Gemini’s specific computational patterns rather than being a general-purpose accelerator. Internal sources claim it could be 6 to 10 times more efficient than current TPUs, with a 2028 ship date. If that number is anywhere close to accurate, Google’s inference cost per token would drop dramatically, giving it room to undercut OpenAI and Anthropic on API pricing. That’s the real strategic play: not being the best model, but being the cheapest to run at scale. Both TechCrunch and The Decoder covered this one. The Decoder / TechCrunch

Nvidia also released Cosmos 3 Edge on Hugging Face — a world model for physical AI and robotics at edge scale — though details in the post were thin. Hugging Face Blog

MCP gets friendlier, Claude Code gets surprisingly autonomous

The Model Context Protocol — the standard that lets AI models connect to external tools and data sources without every developer writing custom integration code from scratch — is getting a usability overhaul. TechCrunch describes the update as making MCP meaningfully easier to work with for developers who aren’t deep in the weeds of the spec. MCP matters because it’s becoming the de facto interoperability layer for agentic applications: if you’ve built anything that connects an LLM to a calendar, a database, or internal APIs, you’ve either implemented something like MCP or wished you had. Reducing friction here has a real multiplier effect across the ecosystem. TechCrunch

The Claude Code fireside chat transcript published by Simon Willison — from a session he hosted with Cat Wu and Thariq Shihipar from Anthropic’s Claude Code team at the AI Engineer World’s Fair — is worth reading in full if you build with coding agents. The headline stat: Claude’s collaborative Slack integration (called Claude Tag) now handles 65% of product engineering pull requests for the Claude Code team itself. That’s not a benchmark number — that’s production code landing in a real repo. The team also shared that the Claude Code system prompt recently shrank by 80%, because newer models like Fable 5 and Opus 4.8 actually perform worse when you load them with examples and lists of prohibitions. This is a real and underappreciated point: prompting patterns that worked well in 2024 can hurt you on frontier models in 2026. The team only ships features that demonstrate retention with Anthropic employees first — a sensible dogfooding-first release strategy. Simon Willison’s Weblog

OpenAI published a safety and alignment post covering what they’ve learned from deploying long-horizon (long-running, multi-step) AI models — the kind used in agent workflows. They documented specific failure modes they’ve observed and the safeguards added iteratively. It’s more substantive than most AI safety blog posts because it’s grounded in actual deployment data, even if it’s light on technical specifics you’d need to reproduce their findings. OpenAI

One smaller observation from Simon Willison worth noting: he’s been collecting anecdotes about people using coding agents to reverse-engineer home devices — automating smart plugs, extracting undocumented local APIs, that sort of thing. His point is that the barrier was never purely technical; it was the cost-benefit calculation of writing code that might break and need constant maintenance. Coding agents collapse that maintenance cost, making it worth attempting things that previously weren’t. This is a real behavioral shift and probably underestimates what hobby-level automation is about to look like. Simon Willison’s Weblog

Anthropic’s copyright settlement and the AI music crisis

Anthropić’s $1.5 billion copyright settlement received final court approval. This is the largest settlement in AI training data litigation so far, and while it closes this specific case, TechCrunch is right to note that it doesn’t establish a clear legal precedent about whether training on copyrighted data is permissible — courts have been reluctant to issue broad rulings that would clarify the law for the whole industry. Expect more cases, more settlements, and continued ambiguity. TechCrunch

Sony Music filed a new lawsuit against Udio, the AI music generator, listing more than 30,000 specific songs it claims were infringed — Elvis Presley to Beyoncé to Harry Styles. Sony explicitly says this list is only a portion of the alleged infringement, discovered after gaining access to Udio’s training data through discovery in a prior 2024 lawsuit. The scale here is striking: this isn’t a sample complaint, it’s a detailed catalog that suggests Sony spent serious legal resources going through Udio’s training set song by song. The Verge

Meanwhile, Deezer reports that more than half of its daily uploads in June were AI-generated — over 90,000 AI tracks per day. That’s not a future problem. That’s a current content moderation and rights management crisis. TechCrunch

Robotics and physical AI: data volume beats model size

Xiaomi’s new robotics research, Xiaomi-Robotics-1, makes a clean empirical point: when training robots to move, more data outperforms bigger models — and the scaling gains from data haven’t plateaued yet. They collected over 100,000 hours of motion data using camera-equipped handheld grippers (operated by humans, not robots), which is a clever way to get high-quality demonstration data without needing a fleet of expensive robot arms running 24/7. Absolute success rates are still low, which is honest of them to admit, but the data-scaling finding aligns with what researchers have been seeing in language models and suggests robotics may be early on a similar curve. The Decoder

Hugging Face published details on Grabette, an open system for recording robot manipulation data — essentially open-source tooling for the data collection problem Xiaomi just demonstrated is the bottleneck. Timing feels deliberate. Hugging Face Blog

Gritt exited stealth with $34 million to automate construction tasks, starting with solar plant assembly. Construction robotics is a crowded pitch right now, but the solar installation angle is specific enough to be credible — it’s repetitive, physically demanding, and geographically distributed in ways that favor a robotic approach. TechCrunch

Firefighting drones are being tested in California through an XPRIZE competition, targeting early-stage wildfire suppression before fires grow beyond manual control. Genuinely useful application, still early-stage technology. Ars Technica

AI video and the “what is a photo” debate opens up

District 9 director Neill Blomkamp released “Nightborne,” a 13-minute sci-fi horror short generated entirely with Seedance 2.0, a text-to-video model. He directed frame by frame through prompts and has founded a new studio, Barley Studios, to produce a full feature. This is the most technically ambitious AI film from a credible director to date — Blomkamp has a visual language and demonstrated taste, which makes this a better test of what current video generation can do than anything a tech company demos. Whether the resulting film is actually watchable as cinema is the real question, and reactions will probably hinge entirely on that. The Decoder

Adobe’s Project Indigo camera app, originally built to give iPhone photos a more natural, SLR-like look, is now adding an “AI Playground” suite with generative editing tools — notably not using Adobe’s own Firefly models. Adobe is positioning this as an experiment with an opt-out button, tested with a small percentage of users, no sign-in required. The interesting detail is the non-Firefly model choice: it suggests Adobe is either hedging on Firefly’s capabilities for this use case or experimenting with third-party integrations. Either way, bolting generative AI onto an app that was specifically designed to feel less artificial is a tension worth watching. The Verge / TechCrunch

Alibaba’s Qwen Audio 3.0 TTS Plus has taken the top spot on Artificial Analysis’ Speech Arena leaderboard, supporting 16 languages with natural-language style control (you can tell it to sound angry, for instance, using tags). The catch: at 16 characters per second, it’s substantially slower than rivals Sonic 3.5 and Simba 3.2. If you’re building real-time voice applications, speed matters more than leaderboard rank. For offline or batch TTS work, Qwen Audio 3.0 TTS Plus is worth evaluating. The Decoder

YouTube updated its monetization policies to more explicitly define which AI-generated and low-quality videos are ineligible for ad revenue. No dramatic surprises in the policy itself — it’s clarification, not a new ban — but it signals YouTube is trying to get ahead of the AI slop problem rather than wait for advertisers to complain. TechCrunch

Quick hits

X relaunched a fully rebuilt Android app after a year-long development effort — unrelated to AI but notable for anyone building on X’s platform. TechCrunch
Halliday Gen 2 smart glasses replace the finicky sliding display window from Gen 1 with a more conventional but apparently functional AR display. Better, but the smart glasses market remains a “not yet” category. The Verge
MIT Technology Review has a sponsored piece on advanced materials enabling next-gen AI hardware — worth skimming if you’re into the semiconductor supply chain, thin on editorial content. MIT Technology Review
SpaceX’s Nasdaq-100 inclusion and what it means for index funds — tangentially AI-adjacent given Elon Musk’s portfolio, but mostly a finance explainer. The Verge