The Manifest
Anthropic had a big day in court and in the lab, while OpenAI's messy split with Cursor and another debt-fueled chip buy underline how fragile and capital-hungry the AI stack still is, even as most logistics firms admit they still can't make AI pay.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Anthropic wants to do for physical hardware what its Model Context Protocol did for softwareThe Decoder
- 02Anthropic gets its first court win over the Pentagon's supply-chain risk labelTechCrunch AI
- 03OpenAI cuts off Cursor after SpaceX acquisition, citing Musk's history of breaking contractsThe Decoder
- 04Neocloud Lambda secures $1B in debt to buy more chipsTechCrunch AI
- 05Most in logistics still can't make AI pay – the few that can think smallerThe Loadstar
Models & Releases
5 storiesLAION drops massive open video dataset with 10 million hours of footage for AI research
Ten million hours of open video is now available to train models outside the proprietary datasets big labs keep locked up. Expect cheaper, faster vision models to show up in inspection and quality-control tools sooner than planned.
Anthropic wants to do for physical hardware what its Model Context Protocol did for software
This is basically MCP for robotic arms and lab instruments, cutting integration time from weeks to hours in early tests. It still needs a human supervisor, since Claude sometimes misjudges basic physical cause and effect.
Google Deepmind's AI Co-Scientist now plans experiments, runs lab equipment, and writes scientific papers
A Gemini-based multi-agent system delivered validated results across three science domains, including running its own lab equipment. Worth watching as an early template for how R&D and QA labs might delegate routine bench work.
GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs Independently Converge on the Same Model Architecture
Two competing labs landed on nearly the same architecture independently, a sign the open-weight playbook is standardizing fast. That usually means cheaper, faster models pressuring proprietary vendor pricing within a few quarters.
Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning
A $399 bipedal robot with an open training stack puts real sim-to-real robotics prototyping within reach of small teams. Useful for anyone testing warehouse or last-mile robot concepts before committing to enterprise vendor pricing.
Supply Chain & Ops
5 storiesMost in logistics still can't make AI pay – the few that can think smaller
BCG found only 13% of logistics firms see real returns from AI, and the winners are automating boring, narrow tasks rather than chasing grand visions. If your AI roadmap reads like a moonshot deck, cut it down to one workflow first.
Port of Los Angeles locks in ONE terminal for 30 more years
A 30-year lease extension clears $200 million in zero-emission cargo equipment investment at one of the busiest US terminals. Commitments on this timeline signal ports are betting on automation and electrification regardless of near-term volume swings.
DSV-Schenker integration at full speed – but who will be left standing?
More than a year in, DSV is integrating DB Schenker faster than expected, testing its rapid-synergy playbook against a much bigger, messier deal. Watch which legacy Schenker systems and customer contracts survive the cut.
Canada Post to launch weekend parcel delivery by year's end
Weekend delivery is coming after a new labor contract steadied a carrier that still lost nearly $200 million last quarter. A reminder that operational trust, not just new capacity, is what wins shippers back.
Motive and Highway Restore ELD Data Access After Integration Dispute
An API dispute over compliance data access got resolved, but it exposed how much carrier oversight now depends on data-sharing terms between vendors. Worth reviewing your own ELD and compliance-data contracts for similar leverage points.
Deals & Market
5 storiesNeocloud Lambda secures $1B in debt to buy more chips
Another neocloud is borrowing $1B to buy Nvidia chips and lease them to Microsoft, showing how much of the AI buildout runs on debt rather than equity. Watch these financing terms, a default here would ripple through chip supply and cloud pricing.
Open-weight AI companies are the Valley's hottest acquisition targets
Capital is chasing companies that give models away for free, betting on distribution and services rather than licensing revenue. If you're evaluating vendors, expect more open-weight players to get acquired or repriced within the year.
OpenAI cuts off Cursor after SpaceX acquisition, citing Musk's history of breaking contracts
OpenAI is walking away from Cursor over Musk's SpaceX buyout, even though its models are a small slice of Cursor's traffic. A reminder that AI tool access can get pulled for reasons that have nothing to do with your usage or contract terms.
Meta executive leaves for OpenAI as the social media giant faces growing scrutiny in India
A Meta APAC executive jumping to lead OpenAI's Southeast Asia and Australia operations shows executive talent tilting toward frontier labs over platform incumbents.
Anthropic gets its first court win over the Pentagon's supply-chain risk label
A federal judge ruled the Pentagon's supply-chain risk blacklist against Anthropic was unlawful, though the label stays in place pending a separate case in DC. This is a live test of how government vendor-risk labels for AI companies hold up in court.
Research & Frontier
4 storiesAI benchmarks have a trust problem and Google wants to fix it
DeepMind is piloting double-blind, cryptographically sealed benchmark testing so neither the model maker nor the evaluator can see the other's side. If this catches on, expect procurement teams to start demanding tamper-proof scores before trusting vendor benchmark claims.
An Anthropic researcher just gave us a peek at self-improving AI
Automated systems improved performance on all 10 misalignment benchmarks tested without degrading other capabilities, a small but real step toward models that patch their own weaknesses. Worth tracking for anyone relying on vendor safety claims that assume static model behavior.
CIFQA: A Deterministic Tool-Grounded Multi-Agent LLM Framework for Financial Query Answering
A tool-grounded multi-agent setup aims to stop LLMs from producing confident but numerically wrong financial answers. Relevant to anyone building AI into pricing, invoicing, or contract math, where a plausible wrong number is worse than an obvious error.
The Accuracy-Efficiency Paradox Quantifying Net Energy Loss in on-Device Energy Forecasting
Researchers found more accurate on-device energy forecasting models can trigger more compute and battery drain than they save, undercutting their own point. A caution for anyone deploying edge AI in warehouses or fleets, more accurate is not automatically more efficient.
Org & AI Architecture
4 storiesMeta makes AI glasses slightly less creepy with limit on nonconsensual recording
Meta added a limit so glasses stop recording when the safety light is covered, closing one specific loophole while leaving broader surveillance concerns untouched. A small fix, not a fix for the trust problem itself.
Beatport blocks fully AI-generated music from its DJ marketplace
A DJ marketplace banning fully AI-generated tracks is a content-authenticity line drawn ahead of regulation. Expect more vertical marketplaces to adopt similar bans as a trust signal to buyers.
Bot Auto commits to U.S.-based remote assistance operators
Staffing robotaxi remote-assistance roles exclusively with US personnel gets ahead of a Senate investigation into where oversight actually sits. A template for how autonomous-vehicle operators answer questions about who is really in the loop.
Always-on and self-starting AI agents might be OpenAI's next big play
OpenAI is testing a Codex mode that stays active indefinitely and generates its own follow-up tasks, and it already deleted user data once in testing. Anyone piloting autonomous coding or ops agents should treat always-on modes as a governance question, not a productivity feature.