The Lyceum: AI Daily — Jun 25, 2026

The AI cold war just got a paper trail. Anthropic formally accused Alibaba of running 28.8 million queries through 25,000 fake accounts to clone Claude — and handed the receipts to the U.S. Senate. Meanwhile, OpenAI built its first chip, and Qualcomm spent $3.9 billion buying the software layer that could make Nvidia's lock-in optional. Today isn't about who has the best model. It's about who controls the infrastructure underneath — and who's trying to take it without paying.

A genuinely quiet 24 hours on the model front — no major US, Chinese, or open-source lab shipped a new model overnight. The most recent tracked releases all predate this window. For the record, the latest items from the past several days:

• Gemini 3.1 Pro (Google): Updated Pro model rolling out across AI Studio, Vertex, and the Gemini app.
• Claude Sonnet 4.6 (Anthropic): Now the strongest of the Sonnet tier, available across all plans.
• Grok 4.2 (xAI): Latest iteration in the Grok series.
• Qwen 3.5 397B (Alibaba): First open-weight release in the Qwen 3.5 line — notable given today's distillation accusations.
• MAI-Image-2 (Microsoft): Photorealistic image model that debuted #3 on the text-to-image Arena leaderboard; live in Copilot and Bing.

Model distillation is industrial espionage with a chatbot interface: instead of breaking into a lab, you ask the AI millions of questions, record every answer, and train your own cheaper model on the transcripts. You get the knowledge without paying for the research.

Anthropic has accused Alibaba of the largest known attempt to illicitly extract capabilities from Claude. The campaign ran from April 22 to June 5, generating more than 28.8 million exchanges through nearly 25,000 fraudulent accounts, allegedly run by operators affiliated with Alibaba and its Qwen AI lab. It targeted Claude's most commercially valuable skills — software engineering and agentic reasoning, per BusinessToday. For scale: earlier campaigns Anthropic flagged from DeepSeek (150,000 exchanges), Moonshot AI (3.4 million), and MiniMax (13 million) combined are still smaller than Alibaba's alleged operation.

Anthropic sent the June 10 letter to Senators Tim Scott and Elizabeth Warren ahead of a Banking Committee hearing on AI. It warned that distilled models typically strip away the safety guardrails baked into the original — capability transfer without the safety layer, per Global Banking & Finance.

If this succeeds as a policy lever, API access starts being treated as an export surface, not just a product feature. The signal to watch: whether Congress turns Anthropic's quiet ask — formal industry-government threat-intelligence sharing — into actual legislation, or whether Alibaba's silence holds and this stays a one-sided accusation.

Every time you use ChatGPT, OpenAI pays Nvidia. That's been the arrangement since 2022. Wednesday, OpenAI took its first real step toward changing the math.

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first inference chip — an accelerator purpose-built to serve AI responses as fast and cheaply as possible, as opposed to the general-purpose GPUs used to train models from scratch. It's an ASIC, optimized for one job: running trained models for live users. Per Broadcom, it went from initial design to manufacturing tape-out in nine months — possibly the fastest advanced-semiconductor cycle ever — with OpenAI's own models used to accelerate the design. The chip helped design itself.

Engineering samples are already running workloads in the lab, including GPT-5.3-Codex-Spark, and OpenAI says early testing shows performance-per-watt "substantially better" than state-of-the-art — though, per OpenAI's own materials, these are vendor lab numbers, not independent benchmarks. OpenAI's reported economics are brutal, and a chip that cuts inference cost per token is the most direct path to a viable P&L ahead of an IPO.

The real test isn't the benchmark — it's whether Jalapeño ships at scale by its targeted end-of-2026 deployment before Nvidia's next generation resets the bar, per CNBC. Watch the deployment date. If it slips, so does the IPO narrative.

Nvidia's real moat isn't its chips — it's CUDA, the software that makes those chips easy to program. Write your inference stack for CUDA and switching hardware means rewriting everything. Qualcomm just spent $3.9 billion on the startup trying to make that lock-in irrelevant.

Qualcomm agreed to acquire Modular in an all-stock deal, expected to close in the second half of 2026, per The Tech Portal. Modular's platform runs models across CPU, GPU, NPU, and custom ASIC architectures without rewrites for each, per the company's own announcement. This is not a random startup: Modular's CEO is Chris Lattner, creator of LLVM and the Swift language — the compiler infrastructure underneath most modern software.

The strategic logic, per Qz: if you can't out-chip Nvidia, out-software them. Qualcomm is betting that as AI spreads to phones, cars, and sensors, running models across heterogeneous hardware beats any single chip's raw speed.

The risk: Modular's developer community built its identity around vendor neutrality — and just got bought by exactly the kind of chip vendor it stayed independent from. Watch whether that community stays engaged after close, or quietly migrates to alternatives. That's the tell for whether the moat-buster becomes just another captured moat.

Most Western autonomous-vehicle programs are private companies racing to commercialize. China just made it a state directive.

China's Ministry of Transportation, with four other departments, issued guidance this week calling for development and testing of "end-to-end" large models for intelligent driving — a single AI that takes raw sensor data in and outputs driving commands directly, with no hand-coded rules in between. It's the same approach Tesla's Autopilot team has pursued, and the one most researchers think is necessary for true autonomy.

The significance isn't the technology — it's the mandate. When five Chinese ministries align on a specific AI architecture, state resources, regulatory fast-tracking, and procurement follow. This is how China turns a research preference into industrial policy, and that tends to produce real deployments faster than Western regulators allow. Watch the follow-on implementation guidance for testing corridors or deployment timelines; that's the difference between a press release and a road map.

The most consequential agentic capability isn't the one that makes headlines — it's the one embedded in the cheapest, fastest model and quietly becomes the default.

Google shipped computer use — the ability for an AI to actually operate software, clicking buttons and filling forms — as a built-in tool inside Gemini 3.5 Flash, its cheap high-volume tier, replacing a separate computer-use model, per The Next Web. Flash is what developers reach for when they need something fast and cheap at scale. Putting computer use there means it stops being a premium demo and becomes infrastructure.

Anthropic's Claude has had computer use for months. But Google's distribution is the differentiator — Gemini is embedded in Workspace, Android, and Chrome, so screen-driving agents could reach enterprise users with zero added integration. The question was never whether agents can use computers. It's which agent gets there first at the scale of a billion users. Watch Workspace adoption signals next quarter; that's where you'll see if this is real or just shipped.

Broadcom's sales and AI chip forecast missed Wall Street's expectations, sending shares lower, per Reuters. That doesn't mean AI demand is collapsing — Broadcom's AI revenue grew 143% year-over-year in the last quarter, and the stock still fell on the session.

What it means: investors are starting to care less about the "AI exposure" label and more about whether spending converts into predictable revenue. Broadcom sits in the custom-silicon and networking layer that keeps giant clusters connected — and it's the same company building Jalapeño with OpenAI. A miss there is a read-through on hyperscaler spending pace and selectivity. The market is moving from "build everything" to "show me the utilization." Watch whether that selectivity spreads to other AI infrastructure names next earnings cycle.

Nvidia announced that version 3 of its Metropolis "video search and summarization" Blueprint is now generally available, adding "agent skills" that let developers build systems which watch video feeds and trigger follow-up actions, per its developer forum. The blueprint bundles recommended models, a reference pipeline, and infrastructure templates so customers can stand up multi-camera analytics with far less bespoke engineering.

This nudges camera analytics from simple search toward off-the-shelf, agent-like stacks that watch, decide, and act across physical spaces. City operators, logistics firms, and retailers now have a vendor-blessed recipe for things like cross-store theft detection — long before regulators have engaged with how "AI guardians" over public space should be constrained. It's a product move, not an independent breakthrough, but a clear window into how Nvidia expects agents to live inside real-world infrastructure. Watch which sectors adopt it first; that's where the regulatory fight lands next.

Alibaba allegedly logging 28.8 million chats with Claude under 25,000 fake names; a chip that helped design itself shipping with caveats in the fine print; a robotaxi getting a facelift while the lawyers argue about kill switches. Somewhere in Hangzhou, a model is being trained on answers it was never supposed to keep — and the only person who noticed timed the heist to the week before the export ban that would've stopped it anyway.

Stay sharp out there.

Know someone who still thinks the AI race is about who has the smartest model? Forward this — it's about who owns the floor they all run on.