The Manifest
Trump's 50% Canada tariffs kick in and Ottawa vows to match them dollar for dollar just as a DRAM shortage pushes Nvidia AI server prices up 15%, while new research shows agentic AI usage exploding even as frontier labs still can't say how they'd contain a rogue model.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Trump's 50% tariff on many Canada imports in effect as talks stallSupply Chain Dive
- 02Memory shortage reportedly drives Nvidia AI server prices up about 15 percentThe Decoder
- 03Inherent, founded by DeepMind alumni, says its AI 'teammate' just outperformed Anthropic and OpenAI at replicating researchTechCrunch AI
- 04Frontier AI labs still won't say how they'd contain a rogue modelTechCrunch AI
- 05AI is becoming AI's biggest customer as agentic token usage jumps 14x on OpenRouterThe Decoder
Models & Releases
4 storiesMeet FreeToken: An Edge-Native MoE Serving Engine that Runs 753B GLM-5.2 on a Single Workstation GPU
Splitting cache misses between PCIe and CPU makes it possible to run a 753B model without a data center behind it. Worth a look if your team needs to pilot frontier-scale models without committing to cloud spend first.
RayNeo's new AI glasses skip the camera, focus on text overlays
No camera and no speakers means fewer privacy objections for warehouse or field deployments, but also fewer use cases. Watch this as a narrower, cheaper alternative to full AR headsets for simple pick-and-check tasks.
Netflix tests language model as alternative to hand-built recommendation logic
Swapping thousands of hand-crafted features for plain-text behavior signals is a template that applies well beyond streaming, think demand forecasting or assortment planning built on messy historical data.
Vercel Introduces 'Is Agentic', a Free Agent-Readiness Scoring Tool That Audits Public Websites Using Ora's 100+ Checks
As buyers increasingly route through agents rather than browsers, this gives procurement and marketing teams a concrete checklist for whether vendor or supplier sites are even legible to those agents.
Supply Chain & Ops
3 storiesTrump's 50% tariff on many Canada imports in effect as talks stall
The Section 338 levies are live and Canada is promising dollar-for-dollar retaliation, so anyone sourcing across the border needs updated landed cost models now, not after the next round of talks.
U.S. Border Patrol finds 20 undocumented immigrants inside tractor-trailer at Texas truck stop
A reminder that carrier vetting and load security still matter more than any AI dashboard. Fleets should double check who is actually behind the wheel on brokered loads through border regions.
GM targets development costs with commercial vehicle ECM
GM is trying to commoditize the engine control work that adds no differentiation to a truck program. If it holds up, expect faster and cheaper commercial vehicle development cycles for fleet buyers.
Deals & Market
3 storiesMemory shortage reportedly drives Nvidia AI server prices up about 15 percent
DRAM shortages from Samsung, SK Hynix and Micron are hitting hyperscalers just as they lock in more AI capex. Expect that cost to eventually show up in the price of every AI service built on top of this infrastructure.
AI is becoming AI's biggest customer as agentic token usage jumps 14x on OpenRouter
Agents have out-consumed humans on tokens since early last year, though cheap cached prompts keep the actual cost curve flatter than the growth number suggests. Budget for agent workloads separately from chat usage.
How China's gray market sells Claude tokens at a fraction of the price
Geoblocking and selfie verification are being routed around through transfer stations selling access at a tenth of list price. Anyone relying on export controls or region locks for competitive or compliance reasons should assume they are leakier than advertised.
Research & Frontier
5 storiesInherent, founded by DeepMind alumni, says its AI 'teammate' just outperformed Anthropic and OpenAI at replicating research
If an agent can reliably replicate published findings, that is a useful sanity check tool for any R&D-heavy operation before committing capital to a new process or material.
AI could make scientists do more work less well, not less work better, study argues
Time saved by AI tends to get reinvested into starting more projects rather than polishing existing ones. Worth remembering before assuming AI automatically improves the quality of internal analysis or reporting.
Study explains why AI agents benefit from 'skills' and when they fail
Skills help mainly by imposing structured workflows, not by adding knowledge, but large skill libraries make agents worse at finding the right instructions. Keep skill sets small and curated if you are building internal agent tooling.
World models that ignore human beliefs predict the wrong actions, new research shows
Physics-only simulators miss what people actually want or intend, which matters for any planning tool that models human decisions inside a supply chain, not just object motion.
Psychological methods reveal major weaknesses in AI security testing
Popular safety benchmarks can be gamed by blanket blocking requests, which inflates scores without making a model more useful or actually safer. Treat vendor safety claims as marketing until independently verified.
Org & AI Architecture
4 storiesFrontier AI labs still won't say how they'd contain a rogue model
A new study finds essentially no public containment plans at leading labs despite increasingly odd model behavior. Enterprises betting critical processes on these systems should be asking vendors this question directly.
OpenAI says California should strengthen its AI safety bill
A notable reversal from a company that previously opposed SB 53. Regulatory posture from major labs is shifting fast enough that compliance teams should track state-level bills, not just federal signals.
Harvard's $699 startup bootcamp offers AI avatars of its instructors
A cheap, scalable way to deliver feedback at volume, and a preview of what corporate training and upskilling programs will look like as AI avatars replace some live coaching.
Decoding AI's Open-Source Course Maps Three Ways to Run an Agent Loop and the Provider Economics Behind Each
Changing only the harness moved a coding agent from 30th place to top 5 with the same model. Teams obsessing over which model to buy may be optimizing the wrong variable entirely.