The Manifest
New US tariffs on Canada and a looming sanctions bill are rattling global sourcing just as Nvidia's AI chip dominance faces its first real competitive test and a ransomware hit forces Coca-Cola's Fairlife unit offline.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01US slaps 50% tariffs on many Canada importsSupply Chain Dive
- 02Anthropic's landmark $1.5B copyright settlement is approvedTechCrunch AI
- 03Nvidia's grip on AI chips weakens as Microsoft turns to AMD and Anthropic may followThe Decoder
- 04Ransomware attack forces Coca-Cola to suspend US production at dairy unitSupply Chain Dive
- 05Graham's sanctions bill set to detonate global sourcing? Model it!The Loadstar
Models & Releases
5 storiesNVIDIA Releases Cosmos 3 Edge: A 4B-Parameter Open World Model That Reasons and Generates Robot Actions On-Device
A 4-billion-parameter world model built to run on-device, meaning robots and vision systems can reason and act without a cloud round trip. Worth tracking for warehouse and yard robotics vendors that need low-latency perception.
Meta Open-Sources Astryx: An Agent-Ready React Design System With 150+ Accessible Components, Seven Themes, and a CLI
Meta open-sourced its internal React design system after eight years and 13,000 apps, giving builders a mature, agent-ready component library for free. Useful if your team is standing up internal tools or dashboards and wants a shortcut past custom UI work.
Alibaba's Tongyi Lab Releases Qwen-Audio-3.0-TTS, a Hosted Text-to-Speech Model in Flash and Plus Tiers Across 16 Languages
A hosted TTS model in 16 languages aimed at production use cases like IVR and customer support bots. Not open weights, so it competes on reliability and language coverage rather than customization.
Xiaomi-Robotics-1 shows that more data beats bigger models when training robots to move
Xiaomi trained its robot model on over 100,000 hours of handheld-gripper motion data instead of scaling model size, and it worked better. A reminder that for physical automation, data collection strategy often beats bigger compute budgets.
Google's 'Frozen v2' chip reportedly bakes Gemini's architecture directly into silicon for efficiency gains
Google is reportedly baking Gemini's architecture directly into a custom chip due in 2028, targeting 6 to 10x efficiency gains over current TPUs. If it lands, expect Google to undercut OpenAI and Anthropic on inference pricing.
Supply Chain & Ops
5 storiesEC parcel tax bid for 'level playing field' creates bumpy ride for air cargo flows
The EU's new €3 duty on low-value ecommerce imports is landing unevenly across member states, and air cargo networks are scrambling to adjust routing and customs handling. Shippers moving small parcels into Europe should expect inconsistent enforcement for months.
Fuel surcharges wallop FedEx, UPS shippers as Amazon looms
Elevated fuel surcharges are pushing shippers to seek discounts just as Amazon's delivery network expands its footprint. Long term, that competitive pressure could reset parcel pricing across the market.
WD-40 hit with 100% spike in some sourcing costs from Iran war
A concrete data point on how the Iran conflict is hitting input costs for ordinary consumer goods makers, not just energy and shipping. Worth checking your own sourcing exposure to the same raw material chains.
Trump offers to cut aluminum tariffs in half for onshoring promises
Companies that commit to expanded domestic aluminum production by January 2029 get relief from the 50% levy under a new federal program. It's a capital commitment play more than a tariff rollback, buyers should read the fine print before counting on savings.
A new era for shippers, with blanked sailings now a fact of life
Carriers are now routinely pulling 10 to 14% of scheduled capacity via blanked sailings across the main east-west trades, per Sea-Intelligence. Treat capacity uncertainty as structural, not a temporary blip, when building ocean contracts.
Deals & Market
5 storiesAnthropic's landmark $1.5B copyright settlement is approved
The court approved Anthropic's landmark payout, but it settles one case and leaves the bigger question of training-data copyright unresolved. Expect more suits to use this settlement amount as a benchmark.
Nvidia's grip on AI chips weakens as Microsoft turns to AMD and Anthropic may follow
Microsoft is standing up AMD's Helios platform at scale for Azure, and signs point to Anthropic testing AMD hardware too. If buyers diversify away from Nvidia, GPU pricing power starts to shift, which matters for anyone budgeting AI infrastructure costs.
Gritt exits stealth with $34 million for robots to build solar plants, then, everything else
Gritt raised $34M to automate the hardest physical tasks on construction sites, starting with solar and expanding from there. One more sign investors are betting on robotics for asset-heavy, labor-constrained industries beyond warehouses.
Prologis 'disappointed' that talks with Segro 'did not provide meaningful clarity'
Prologis's bid for Segro is stalling after a follow-up meeting failed to clarify a path to a deal. Watch this one, a combination would reshape European logistics real estate and warehouse availability for occupiers.
Moonshot pauses new Kimi K3 subscriptions after GPU demand maxes out in 48 hours
Moonshot had to stop selling subscriptions because demand outran its GPU capacity within two days. A useful reminder that compute constraints, not model quality alone, are still capping how fast AI vendors can scale.
Research & Frontier
5 storiesRater State Bias in RLHF Preference Data: An Audit Framework
Researchers show that human raters' stress and fatigue during labeling can quietly bias the preference data used to train models. Anyone relying on RLHF-tuned models for decision support should ask vendors how they manage annotator conditions.
PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection
A single prompt injection into a planner agent can cascade corrupted instructions through every downstream sub-task. Relevant for anyone piloting multi-agent systems in procurement or logistics workflows where a compromised plan could quietly propagate bad decisions.
Deterministic Replay for AI Agent Systems
Agent runs are notoriously hard to reproduce because of sampling variance and external API state, which makes debugging production failures painful. This work on deterministic replay is a step toward the kind of auditability ops teams will need before trusting agents with real transactions.
RouteCost: A Production-Inspired Multi-Stage Framework for Pre-Order Shipping Cost Estimation in E-Commerce
A practical framework for pre-order shipping cost estimation that accounts for destination demand mix, dimensional pricing, and consolidation effects, not just distance. Directly useful for ecommerce teams trying to tighten margin planning before checkout.
Hugging Face says an AI agent hacked its infrastructure, and it used AI to fight back
An autonomous agent reportedly carried out a multi-thousand-action attack on Hugging Face's production systems, and commercial AI models' own safety guardrails got in the way of the defenders' analysis. A concrete early case of agentic attacks outpacing agentic defense tooling.
Org & AI Architecture
5 storiesTrump's latest AI czar has already resigned
The director role for the Center for AI Standards and Innovation has turned into a revolving door since David Sacks departed. Policy uncertainty at the top makes it harder for companies to plan around federal AI rules.
OpenAI is scared of open-weight models. Should the US be?
Talk of banning Chinese open-weight models exposes how hard it still is to build a durable business model around AI. Operators evaluating open-weight options should watch this debate before locking into any single vendor's roadmap.
How supply chain leaders can avoid common AI pitfalls
The advice is simple but often ignored, get the data right first, pilot before scaling, and be willing to kill projects that do not pay off. Worth reading if your organization is under pressure to show AI wins fast.
Trump administration reportedly builds a slow-motion ban on Chinese AI models through sanctions and soft pressure
Rather than an outright ban, Washington appears to favor sanctions and liability pressure to steer companies away from Chinese models. Procurement and IT teams evaluating cheaper open-weight options should factor in this regulatory drift.
YouTube clarifies policies around AI slop and upsetting videos
YouTube tightened definitions of what AI-generated content can't monetize, an early sign platforms are drawing harder lines on synthetic media quality. Relevant if your marketing team is leaning on AI video for content production.