The Lyceum: AI Daily — Jul 29, 2026
Photo: lyceumnews.com
Wednesday, July 29, 2026
The Big Picture
Tuesday’s consequential AI news happened outside the model. Washington restricted foreign robots, PJM Interconnection put future data centers on notice, and the U.S. General Services Administration made agent software easier for federal agencies to buy. The common thread is operational authority: who can connect to the grid, enter the country, access government systems or drive through London. This 24-hour edition does not recycle the older Reuters reports about DeepSeek’s chip project, President Donald Trump’s planned AI oversight order or Anthropic’s foreign-access restrictions; the broader infrastructure-spending story appears here through Meta Platforms’ fresh El Paso financing deal.
Today's Stories
Washington Draws a Security Perimeter Around Physical AI
Washington is drawing its AI security boundary around machines, not just models. The Trump administration has blocked approvals for new Chinese humanoid and quadruped robots, along with networked power inverters used by renewable-energy systems, batteries and data centers, Reuters reported Tuesday. The restrictions took effect immediately for unreleased products, and the Federal Communications Commission can revisit equipment already approved for sale.
If the restrictions hold, suppliers outside China gain an opening in the U.S. robot market, while American data-center builders face a narrower pool of approved power equipment. The policy also establishes a template: Washington can treat any connected machine capable of seeing, moving or interrupting electricity as part of the national-security perimeter around AI.
Failure will look quieter. Broad exemptions, few revoked approvals or continued access through non-Chinese subsidiaries would turn the ban into paperwork rather than a durable market barrier. Watch what the Federal Communications Commission actually removes from sale—not merely what Washington says it can remove.
The AI Boom Meets PJM’s Circuit Breaker
PJM Interconnection is putting large data centers on notice: during power shortages, some may have to go dark. Beginning in June 2027, PJM plans to make some large data centers interruptible, TechCrunch reported Tuesday. PJM operates the largest U.S. wholesale electricity market, serving 67 million people across a region stretching from Virginia to Illinois; the rule is expected to cover facilities drawing at least 50 megawatts.
If the plan works, flexible computing becomes an infrastructure product. Data centers could shift non-urgent workloads, install batteries or pay for on-site generation instead of assuming every server receives uninterrupted electricity. Grid access would cease to be a utility detail and become part of the economics of training and operating AI. (The AI boom finally meets the circuit breaker)
Non-adoption would appear as generous exemptions, delays or data centers relocating beyond PJM’s territory. The decisive signal is whether other grid operators copy the rule. One regional program is a constraint; several would create a new market for interruptible AI computing.
Washington Adds an Agent Workforce to the Federal Catalog
The federal government has added an agent workforce to its shopping catalog. The U.S. General Services Administration added CORAS’s GARY agent-orchestration platform to OneGov on Tuesday. Federal agencies can now procure software that assigns specialized AI agents to reports, analysis and workflows while retaining human approval and an audit trail. (Washington puts an agent workforce on the federal menu)
The offer includes discounts of up to 80% through September 30, 2027. CORAS says its App Factory can produce mission-specific applications in days rather than the six to 12 months associated with traditional development; that is CORAS’s claim, not a measured federal deployment result. (Washington puts an agent workforce on the federal menu)
If agencies adopt GARY, federal AI could move beyond chatbots and into software that performs bounded pieces of administrative work. If they do not, OneGov will become another well-stocked shelf with few buyers. The useful signal will be a named agency publishing mundane but measurable results: hours removed, cases processed or backlogs reduced. (Washington puts an agent workforce on the federal menu)
Baidu’s Robotaxi Enters London Traffic
Baidu’s robotaxi has entered London traffic. Baidu’s Apollo Go and Freenow by Lyft began testing RT6 robotaxis in London on Tuesday. Dozens of vehicles with human safety operators will initially operate in Brent, with public rides planned for 2027 subject to regulatory approval.
Success would give Baidu something harder to export than a model: a physical AI system trusted to operate under Western traffic law. It would also intensify London’s emerging contest among Baidu, Waymo and Wayve, where fleet operations and public confidence may matter as much as driving software. (Baidu’s robotaxi leaves China for London traffic)
Failure will be easy to recognize. Testing remains confined to safety-driver trials, public service slips beyond 2027 or regulators restrict the operating area. The next meaningful milestone is not another demonstration—it is Transport for London allowing passengers into driverless vehicles. (Baidu’s robotaxi leaves China for London traffic)
Recursive Converts Its War Chest Into a $410 Million Compute Contract
Recursive Superintelligence has turned its war chest into a major cloud commitment. The company signed a multiyear, $410 million computing agreement with Amazon Web Services, TechCrunch reported Tuesday. Recursive emerged from stealth in May with $650 million and is pursuing systems intended to improve their own software and products.
If that approach produces useful systems, AI-lab economics may tilt further away from headcount and toward enormous cloud commitments. Amazon Web Services gains durable demand, while smaller labs face a higher financial threshold for competing in compute-intensive research.
The alternative is an expensive capacity reservation attached to a thesis that never becomes a product. Recursive founder Richard Socher told TechCrunch that users should see the company’s first tangible products around October. Shipped software and external users will matter more than the contract’s size.
⚡ What Most People Missed
- Meta’s data center became a finance project: Meta Platforms and BlackRock formed a venture for an approximately $14 billion, one-gigawatt data-center campus in El Paso, Reuters reported. BlackRock-managed funds will own 80%, suggesting AI infrastructure is increasingly financed like an airport or power plant—not purchased like ordinary cloud equipment.
- Agent security attracted a $1 billion letter of intent: Cyera agreed in principle to acquire Oasis Security, which manages the software identities used by services and AI agents. The transaction is not closed, but the proposed price shows how valuable permissioning becomes once agents receive real access to corporate systems.
- Robot perception is becoming a packaged subsystem: LG Innotek and TDK announced plans to develop integrated vision and tactile sensor modules for humanoid robots. No finished module or performance measurement accompanied the agreement, but the supply chain is beginning to package sight, touch, sound and motion below the robot-maker layer.
- The agent-plumbing fight reached court: Runlayer sued Rippling, alleging that Rippling used an extended trial to copy its gateway for connecting agents securely to corporate tools and data. Rippling acknowledges building a competing product but denies misusing Runlayer’s intellectual property.
- Voice cloning still has a consent problem: Fish Audio raised a $52 million seed round after reporting 8 million users and $21 million in annual recurring revenue. Fish Audio says automated takedowns now require less than three minutes, but faster removal does not answer why an unauthorized voice can be uploaded first.
📅 What to Watch
- If the Federal Communications Commission revokes approvals for Chinese robots already sold in the United States, the physical-AI ban will become an installed-base problem rather than merely a border rule.
- If another U.S. grid operator adopts PJM-style curtailment, batteries and workload scheduling will become competitive advantages for AI companies.
- If Microsoft’s July 29 results connect Copilot adoption with improving AI margins, enterprise agents will look less like a usage story and more like a way to repay infrastructure spending.
- If Meta Platforms raises capital spending again on its July 29 call without proportional AI-driven revenue, outside financing structures like the BlackRock venture may spread quickly.
- If Amazon’s July 30 results show faster Amazon Web Services growth without margin compression, contracts like Recursive’s will look like durable cloud demand rather than speculative reservations.
- If a federal agency publishes audited productivity results from GARY, OneGov could become a distribution channel for agent software across government rather than another pilot program.
The Closer
A robot stopped at the border. A data center waiting beside a giant circuit breaker. A London cab carrying a safety driver training for unemployment.
Meanwhile, AI’s plumbing has become valuable enough for a lawsuit—which is how software knows it has finally achieved product-market fit.
Keep one hand near the off switch.
Forward this to the person who still thinks AI runs entirely in “the cloud.”