The Manifest
China unveiled a rival global AI governance body and Moonshot's Kimi K3 renewed doubts about the West's compute edge, while ocean and air freight signal the peak season is already turning even as logistics providers keep layering agentic AI into warehouses and pricing desks.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01China's new World Artificial Intelligence Cooperation Organization is President Xi's clearest play yet for a parallel AI orderThe Decoder
- 02Just like Deepseek, China's Kimi K3 is forcing Western AI labs to question their compute advantageThe Decoder
- 03How Apple's big lawsuit could disrupt OpenAI's IPO plansTechCrunch AI
- 04Databricks hits $188B valuation, extending its run as AI's favorite second actTechCrunch AI
- 05Tariffs, Iran war prompt ocean shippers to scrap tradition: Los Angeles port headSupply Chain Dive
Models & Releases
5 storiesJust like Deepseek, China's Kimi K3 is forcing Western AI labs to question their compute advantage
A 300-person team shipped a model early testers rank near Opus 4.8, which undercuts the argument that scale of spend is the only moat. Buyers evaluating vendor lock-in should watch how fast open-weight options close the gap on the specific tasks they actually run.
Anthropic slashes Claude Fable 5 limits in Max and Team Premium and pushes Pro users toward API pricing
Anthropic reversed course on pulling Fable from subscriptions but is quietly cutting usage caps in half, a sign that unit economics on premium models are getting harder to hide. Teams on Pro plans should model out API costs now rather than after the July 20 switch.
GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did
Full Access Mode let the model overwrite a temp directory variable and delete entire home directories without asking first. Anyone piloting agentic file or system access should keep destructive actions behind an explicit confirmation step regardless of what the vendor's safeguards promise.
NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB
A top-ranked open embedding model with an NVFP4 variant that holds accuracy while roughly doubling throughput on Blackwell is a real option for teams running retrieval at scale. Worth benchmarking against whatever embedding stack currently powers internal search or RAG.
Google Cloud's Always-On Memory Agent Replaces RAG and Embeddings With Continuous LLM Consolidation on Gemini 3.1 Flash-Lite
Ditching vector databases for a running ingest-consolidate-query loop over SQLite is a genuinely different pattern for agent memory, not just another RAG wrapper. Ops teams building long-lived assistants for planning or procurement should watch how this performs at real data volumes before betting on it.
Supply Chain & Ops
5 storiesTariffs, Iran war prompt ocean shippers to scrap tradition: Los Angeles port head
LA logged over a million TEUs in June, but Gene Seroka is flagging a murky second half as tariffs and Middle East risk push carriers away from long-standing routing habits. Planners should treat H2 ocean capacity assumptions as provisional and keep contingency lanes live.
Ocean capacity is rising, demand is cooling, and 'the market is starting to turn'
The first weekly drop in container spot rates since April, with Shanghai-Rotterdam down 1% and Shanghai-Genoa off 3%, is an early signal that peak season pricing is topping out. Buyers who held off on contract negotiations may finally get leverage back.
AI cargo props up air cargo market amid warnings of an early peak
AI infrastructure and semiconductor shipments are keeping air freight rates up 33% year on year even as ecommerce volumes soften, per Flexport data. Shippers dependent on air capacity should expect data center buildouts, not holiday retail, to be the swing factor this cycle.
Kuehne+Nagel rolls out cloud-native contract logistics solution
KN SwiftLOG's move to a cloud-native, agentic version of its Blue Yonder-based WMS across 1,000-plus sites is a concrete data point on how fast agentic tooling is reaching warehouse floors, not just pilots. Worth asking your 3PL what their equivalent rollout timeline looks like.
Rampant Prologis – the two sides of the data centre story
Prologis signed a record 67 million square feet of leases this quarter and is repositioning itself as more than a warehouse landlord as data center demand reshapes its portfolio. Anyone leasing industrial space should watch whether AI infrastructure crowds out traditional distribution capacity in key markets.
Deals & Market
5 storiesDatabricks hits $188B valuation, extending its run as AI's favorite second act
Databricks keeps proving that a data platform vendor can rebrand as an AI company and get rewarded for it, backed by its own research on open-weight coding model savings. Enterprises evaluating vendors should look past the valuation headline to whether the cost claims hold up in their own workloads.
How Apple's big lawsuit could disrupt OpenAI's IPO plans
A trade secrets suit alleging misconduct up to OpenAI's chief hardware officer, with 400-plus ex-Apple staff now on payroll, lands at the worst possible moment for IPO timing. Enterprise buyers with long OpenAI contracts should factor in some governance uncertainty over the next two quarters.
Zuckerberg's plan to sell excess AI compute could finds its first big customer in Anthropic
Meta renting spare data center capacity to Anthropic would be a notable crack in the assumption that every hyperscaler needs to hoard its own compute. Watch whether this becomes a template for other labs to buy capacity rather than build it.
Why the first GPU financiers are turning to inference chips in a $400 million deal
Lenders that started by financing training GPUs are now underwriting inference silicon, which tells you where the next capex wave is expected to land. Ops leaders planning multi-year AI infrastructure budgets should note financing terms are shifting toward inference-heavy assumptions.
AI-driven memory crunch jolts India's smartphone market
Memory chip demand from AI data centers is squeezing supply for consumer electronics makers, pushing up smartphone prices in India. Any procurement team sourcing DRAM or NAND for non-AI products should expect this competition for memory capacity to persist into next year.
Research & Frontier
5 storiesOpen-weight models now match frontier cyber performance from just four months ago at a fraction of the cost
The UK's AI Security Institute finds open models like GLM-5.2 and DeepSeek V4-Pro now trail closed frontier cyber capability by only four to seven months, down from six to ten a year ago, and that safety fine-tuning on open weights is easy to strip out. Security teams should assume attacker tooling parity is closer than defender readiness.
Sakana AI's Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation
A biologically plausible training rule that avoids weight transport is early stage, with modest CIFAR-10 accuracy, but it's a genuine alternative research direction rather than another backprop variant. Not production relevant yet, worth a bookmark for the neuromorphic and edge-training crowd.
Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning
This reworks the standard RLHF-style rollout topology for agents that operate inside executable sandboxes, which matters for anyone training agents to run real tools rather than just chat. Early stage but relevant to teams building agentic RL pipelines for internal automation.
ToolAnchor: Anchoring Counterfactual Context to Boost Agentic Tool-use Capability
Addresses a real operational headache: agents trained on a fixed toolset get stuck once you add a new tool without retraining from scratch. Useful reading for anyone maintaining agent fleets that need new integrations added incrementally.
Intelligent Three Level Learning Architecture for Autonomous UAV Swarms in Search and Rescue
A hierarchical learning setup for drone swarms doing search and rescue is a preview of where warehouse and yard inspection drones could head, layering different learning mechanisms by role. Still academic, but worth tracking for anyone piloting drone fleets in logistics facilities.
Org & AI Architecture
5 storiesChina's new World Artificial Intelligence Cooperation Organization is President Xi's clearest play yet for a parallel AI order
Xi's 5,000 training slots for Global South countries and a new cooperation body with ASEAN, the African Union and BRICS is a deliberate bid to build AI governance infrastructure outside Western institutions. Multinational operators should watch which standards bodies their overseas partners start aligning with.
The Pentagon's new AI playbook treats slow adoption as a bigger risk than 'imperfect alignment'
The Navy's new strategy explicitly ranks the risk of moving too slowly above the risk of imperfect model alignment, a striking reversal of the usual caution-first posture in defense. Any org still debating whether to slow-walk AI adoption for safety reasons should read this as a signal of where institutional risk tolerance is heading.
Linus Torvalds tells AI critics in the Linux kernel community to fork off
Torvalds' blunt endorsement of AI-assisted code review tooling in one of the most conservative open source communities is a useful data point for any org still fighting internal battles over AI adoption. When the skeptics lose this argument, it's a sign the debate is mostly over.
Netflix's 300 AI productions show how fast the technology is spreading through entertainment
Netflix says AI-assisted footage on one docuseries ran twice as fast at half the cost, and the savings will likely go toward more content rather than a smaller budget. A clean example of AI adoption funding output growth instead of headcount cuts, worth citing in any internal ROI conversation.
Neil Rimer thinks the AI money is coming back out
The Index Ventures co-founder argues the concentrated wealth AI is generating in Silicon Valley will eventually have to redistribute, one way or another. Worth a read as a counterweight to the usual capex-only framing of AI's economic impact.