Today’s brief tracks the money quietly reshaping AI infrastructure, from a chipmaker investing in its own biggest customer to Washington’s sanctions threat against a Chinese lab, plus a full roundup of new agent tools and industry rumors.
The Circular Money Behind AI Compute: Chipmakers Bet Billions on Their Own Customers
AMD, OpenAI, and TSMC are all pouring money into the same GPU buildout, and a sharp new analysis argues nobody has actually priced the collateral underneath it.
- AMD Signs $5B Anthropic Investment Tied to a 2-Gigawatt Chip Supply Deal. AMD will invest up to $5 billion in Anthropic while Anthropic commits to as much as 2 gigawatts of compute built on AMD’s Instinct MI450 chips, making the chipmaker a stakeholder in its own biggest customer.
- OpenAI’s 2030 Infrastructure Budget Jumps 25 Percent to $750 Billion. OpenAI reportedly raised its planned infrastructure spend through 2030 by a quarter, and the first bill on the list is a $20 billion Georgia campus with no financing plan disclosed.
- TSMC Adds $100 Billion for Arizona, Pushing Its Total Pipeline to $265 Billion. TSMC’s CFO is betting a full-year capital expenditure hike on what he calls a multi-year AI demand megatrend.
- GPU Collateral Has Three Prices, and Lenders Are Only Quoting One. Face value, liquidation value, and going-concern value diverge sharply on GPU-backed loans, and an analysis argues the unpriced gap is the risk sitting under the entire AI financing stack.
Washington Turns Up the Pressure on Chinese AI: A Sanctions Threat Meets a Government-Built Model
Treasury is threatening sanctions over an alleged case of model theft even as the Department of Energy backs an open government model built the opposite way.
- Bessent Threatens China Sanctions Over Alleged Model Distillation. Treasury cites Moonshot’s Kimi K3 as a possible case of intellectual property theft, even though the timeline undercuts the White House’s own version of events.
- DOE and Arcee AI Open GS1, a Public Open-Weight Science Model. Genesis-Science-1 bets on government-owned weights and community-contributed research data, with a first submission deadline of August 6.
The New Layer Above the Models: Routing, Permissions, and Evals for AI Agents
OpenAI, Cursor, LangChain, and Anthropic all shipped new plumbing this week for controlling, routing, and evaluating AI agents rather than just building bigger models.
- OpenAI’s Presence Puts Guardrails Around Enterprise Agents. OpenAI is now selling the permissions, policy, and escalation layer that sits around its agents, not just the model underneath them.
- Cursor Ships a Router That Picks the Model for You. Cursor Router promises frontier quality at 60 percent lower cost, though the number comes from Cursor’s own early-access customers rather than an independent benchmark.
- LangChain’s New Skill Turns Agent Traces Into Real Evals. The released Eval Engineering Skill reads a codebase and its production logs, then compiles them into runnable Harbor test suites.
- Anthropic Turns Its Labor Data Into a Claude Chat Feature. A new connector lets Claude answer questions about jobs and automation by querying Anthropic’s own Economic Index directly inside chat.
Today’s Quick Hits
- A Popular AI Benchmark Stayed Clean, and That’s the Real Finding. Dylan Castillo tested 1,008 SVGs across seven models to check whether AI labs train on Simon Willison’s pelican-bicycle benchmark, and found little evidence they do.
- Willison Says Model Gatekeeping Left Hugging Face’s Defenders Exposed. Simon Willison argues the OpenAI-Hugging Face breach shows gated frontier access, shaped partly by export-control politics, left defenders under-armed against attackers facing no such limits.
- Physical Intelligence Denies It Is Being Acquired by Anthropic. A weekend rumor that Anthropic would acquire robotics startup Physical Intelligence spread across X despite a denial from its CEO, and one confirmed detail keeps the story alive.
- Amazon Cuts Jobs Inside Its AGI Unit. Amazon confirmed layoffs in the group pursuing artificial general intelligence, even as it pledges continued AI spending.
- Anthropic Tests Managed Projects That Act Without Prompting. Internal builds show a project mode where Claude organizes tasks and runs scheduled work on its own, with no launch date set.