The Manifest
Nvidia, Anthropic and Amazon all struck massive compute and infrastructure deals while Meta quietly scrapped its AI-driven layoff plan after agents failed to deliver, and ocean freight capacity keeps tightening from typhoons, Gulf tensions, and a newbuild ordering wave.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Nvidia snaps up Hugging Face for $12.9 billion as closed AI labs pull awayThe Decoder
- 02Anthropic locks in 45-billion-dollar compute deal with Nscale ahead of IPOThe Decoder
- 03Amazon just tripled its order of Nvidia chips over 'surging demand'TechCrunch AI
- 04Employee revolt and failing agents forced Meta to scrap its AI layoff planThe Decoder
- 05War and weather continue to limit intra-Asia capacity and keep rates highThe Loadstar
Models & Releases
5 storiesGLM-5.3-Flash matches top models at a fraction of the cost, and runs without Nvidia
Z.ai's open model lands near flagship performance running entirely on Chinese chips at a seventh of the cost. Worth watching if you're trying to reduce dependence on any single chip vendor for inference workloads.
Alibaba releases Qwen3.8-Flash-Next, targeting 'ultimate cost efficiency'
A 125B model that only activates 6B parameters per token, beating larger rivals on coding and office tasks at a ninth of the training cost. Another reason to keep pressure-testing vendor pricing before renewing contracts.
Google announces Gemini 3.5 Transcribe for AI-powered speech-to-text
85 languages, filler-word stripping, and lower latency matters for dispatch logs, call center notes, and voice-driven warehouse picking. Worth a pilot if transcription errors are costing you rework today.
IBM's new Granite 4.2 models ride the wave of interest in local LLMs
Open, Apache-licensed models with agentic RL for terminal and code tasks, built for predictable enterprise deployment. Good option for teams that don't want to send operational data to a closed API.
Claude Cowork now runs its own browser inside the desktop app
Anthropic is building autonomous web action directly into its desktop agent. Anyone piloting research or procurement agents should watch how this changes what tasks can run unattended.
Supply Chain & Ops
5 storiesWar and weather continue to limit intra-Asia capacity and keep rates high
Typhoons in China and Middle East unrest are hitting the same trades at once, pushing the intra-Asia container index up a third straight week. Anyone sourcing from the region should plan for continued rate volatility into peak season.
Yard shutdown threat at Mundra follows ban on using off-dock container yards
Adani Ports' embargo on independent depots is a single-port policy fight that could squeeze empty container availability across India's busiest gateway. BCOs and carriers moving through Mundra should have a backup plan.
Ocean rate concerns as orders for new container ships near 40% of global fleet
That much newbuild ordering sets up an oversupply risk a couple of years out, even as rates spike now. Shippers locking in multi-year ocean contracts should factor that swing into their negotiating position.
'Secure capacity' says RXO as it warns of further haulage rate rises
Spot truckload rates just hit a four-year high and RXO expects more upward pressure heading into peak season. Buyers still riding month-to-month capacity should lock in now rather than chase the market later.
Advance Auto Parts rebids carrier contracts, expects to save tens of millions
A concrete example of a retailer squeezing carrier costs through rebids plus shipment accuracy and DC process fixes. Worth benchmarking if your own transportation spend hasn't been revisited recently.
Deals & Market
5 storiesNvidia snaps up Hugging Face for $12.9 billion as closed AI labs pull away
Nvidia is paying roughly 80x revenue to anchor the open-model ecosystem just as closed labs diversify off its chips. A defensive move worth watching for anyone dependent on Hugging Face's model hosting.
Anthropic continues compute-gobbling streak in $45B deal with Nscale
Compute contracts are becoming the dominant line item ahead of an Anthropic IPO. Worth tracking if you're assessing vendor financial stability before signing a long-term enterprise agreement.
Amazon just tripled its order of Nvidia chips over 'surging demand'
Two million more GPUs over two years signals hyperscaler capacity building hasn't slowed. That should keep enterprise cloud AI pricing competitive a while longer.
Viral AI startup Instinct has raised $350M at a $2.5B valuation
A one-year-old startup drawing that much capital and privacy scrutiny is a reminder to vet data practices carefully before adopting any flashy new consumer AI tool into a business workflow.
Chinese Moonshot AI negotiates hosting deals with Microsoft, Amazon, and Google
A Chinese model landing on major US clouds for the first time would reshape vendor options. Enterprises will need to weigh potential cost savings against data sovereignty and compliance questions.
Research & Frontier
5 storiesRENDER: Controlling Reader-Facing Evidence in LLM Memory Evaluation
Useful for anyone building RAG or memory systems: results shift heavily based on how history is packaged for the model, not just what's retrieved. Check your own evaluation setup for this blind spot.
ESQ-Bench: A Multi-Tier Enterprise Oracle Benchmark for Evaluating NL2SQL Dialect Generalization and Silent Semantic Divergence
Shows NL2SQL accuracy claims above 89% fall apart against real enterprise Oracle schemas. Worth reviewing before trusting a vendor's natural-language query pitch on your own database.
Demystifying Reinforcement Learning Post-Training of Language Models
A technical breakdown of how RL post-training actually shapes reasoning behavior. Good background reading before betting a process on a specific model's stated capabilities.
Google Research Introduces GlucoFM: A 0.72M-Parameter Dual-Stream Foundation Model for Continuous Glucose Monitoring
A tiny specialized model beating much larger general ones on niche time-series tasks. The same pattern is likely to show up in demand forecasting models built for narrow supply chain signals.
Robot brain builders are pushing out of their GPT-2 era
Robotics foundation models are still early relative to language models. Expect physical automation in warehouses and factories to lag well behind the pace of chatbot progress for now.
Org & AI Architecture
5 storiesEmployee revolt and failing agents forced Meta to scrap its AI layoff plan
Meta wanted to replace far more of its workforce with agents than previously known, and the plan collapsed when the agents couldn't deliver. A cautionary tale for anyone planning aggressive AI-driven headcount cuts before agents are proven.
Sam Altman says OpenAI will have AGI by the end of 2026 if you accept his definition
The definition matters more than the headline here. Don't let a vendor's own timeline claims drive your automation roadmap decisions.
Bill Gates wants to see a robot tax and 'Human Reserved' jobs to mitigate harms from AI
A prominent voice adding weight to the policy debate on offsetting AI job displacement. Relevant as automation touches more planning, procurement, and back-office roles beyond the factory floor.
What Would Have to Be True for Agentic Coding to Replace Junior Engineers
Lays out falsifiable conditions instead of hype, tested against actual METR, OpenAI and DORA evidence. A good gut check before assuming coding agents let you cut junior hiring.
How do we explain OpenAI's executive exodus?
Leadership churn at a top lab is a signal to diversify vendor relationships rather than betting critical operations on any single provider's roadmap stability.