The Manifest
AI financing cracks are showing in CoreWeave's debt stress and Anthropic's slowing growth ahead of its IPO, even as Meta, Nvidia and others keep flooding the market with big open-weight models, while US container gateways and European barge routes stay tight despite the usual peak season cooldown.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01CoreWeave CDS distress – 'investors sour on AI debt'The Loadstar
- 02Anthropic's Q3 revenue hits $16.8B but growth slows ahead of planned November IPOThe Decoder
- 03Meta launches an open-weight AI model that runs on MacsTechCrunch AI
- 04No post-peak easing at US container gateways, despite drayage woesThe Loadstar
- 05GlobalFoundries building $21B New York fabs under US-Japan trade dealSupply Chain Dive
Models & Releases
5 storiesMeta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
An 816B open model tuned to run on consumer Mac hardware shrinks the gap between frontier labs and anyone with a capable workstation. Worth testing for document-heavy planning and forecasting tasks that need long context without a cloud bill.
Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop
Open weights plus NVFP4 and a 1M token context are aimed squarely at enterprise agents that need to run continuously. The catch is getting full performance still means buying Nvidia hardware, so budget accordingly.
Black Forest Labs' open-weights image model FLUX 3 Image Klein runs on consumer GPUs with just 8 GB of VRAM
A smaller, faster image model that still beats some proprietary tools means product and marketing teams can run visual generation locally instead of paying per call. A good candidate for catalog and packaging mockups.
Google wants Gemini 4 Argon to be the only AI model robots need
A single model meant to plan, simulate and evaluate robot behavior is a tidy pitch for warehouse automation vendors, but the demo prompts still lean on heavy human guidance. Treat it as early stage, not a drop-in picker.
Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index
Calibrated probabilities in a single forward pass at a low token price is useful for teams building automated approval or exception-routing logic. Whether it earns a spot in production depends on cost per transaction, not the benchmark score.
Supply Chain & Ops
5 storiesNo post-peak easing at US container gateways, despite drayage woes
Houston and other gateways are still running hot even as the broader peak season cools elsewhere, so drayage capacity stays the bottleneck to watch through year end. Keep buffer days in inland moves rather than assume relief is coming soon.
Low Rhine water level driving barge cargo to capacity-strained road and rail
Barges near Duisburg running at two thirds capacity means more freight gets pushed onto already tight road and rail networks around Antwerp and Rotterdam. European shippers should build in mode-shift contingency now rather than after a rate spike.
FedEx digitizes booking process for European cold-chain pharma shipments
A simpler booking flow for temperature-controlled freight is a small operational fix, but it matters for pharma shippers juggling compliance paperwork across markets. Watch for other carriers to follow with their own digitized cold-chain lanes.
Unilever puts business results first in factory digital twin rollout
Rolling digital twins to 40 sites only works if each one proves a measurable gain before scaling further, a discipline most manufacturing AI pilots skip. A useful template for anyone stuck justifying a twin project to finance.
How UNFI tackles higher fuel prices
Hedging diesel costs while layering in AI route optimization is a pragmatic combination rather than a single fix. For distributors, the lesson is that software savings and financial hedges need to work together, not instead of each other.
Deals & Market
5 storiesCoreWeave CDS distress – 'investors sour on AI debt'
Over 51 billion dollars in debt backing one neocloud's infrastructure build-out is now spooking bond and equity investors alike. If financing for AI data centers gets more expensive, expect that cost to eventually show up in compute pricing for everyone downstream.
Anthropic's Q3 revenue hits $16.8B but growth slows ahead of planned November IPO
Revenue growth dropping from 140 percent to 45 percent quarter over quarter is the kind of deceleration that makes IPO investors nervous, even at 16.8 billion in revenue. Watch the November listing as a real-time read on how much more the market will pay for AI growth stories.
First look: J.B. Hunt's Q3 earnings
Beating lowered expectations after a prior earnings warning is a modest win, not a turnaround signal. Intermodal volumes and pricing discipline will tell the real story over the next two quarters.
AP Møller Capital acquires Singapore's Portek International from Mitsui
A feeder terminal operator with footholds across Africa, Southeast Asia and the Baltic gives Maersk's investment arm more port infrastructure leverage outside its core network. Expect more of these smaller terminal bolt-ons as carriers diversify beyond vessel ownership.
Prologis tops Q3 estimates, raises development starts forecast again
Another raised forecast from the largest warehouse owner suggests demand for logistics real estate is not slowing despite rate pressure elsewhere in freight. Worth tracking as a leading indicator for where distribution capacity gets built next.
Research & Frontier
5 storiesGPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece
Solving unpublished theoretical physics problems for under three dollars a run is a genuine capability jump, even if researchers say the model still lacks the judgment to pick which problems matter. A preview of what cheap expert-level reasoning could eventually do for network optimization work.
The World Won't Wait: Evaluating and Teaching Agents Under a Live Clock
A benchmark that penalizes correct-but-late agent actions is directly relevant to anyone deploying agents on live operational systems, where a right answer delivered too slowly is still a miss. Worth reading before trusting any agent with real-time routing or pricing decisions.
Evaluating Model Capabilities After Benchmark Saturation
As top models all score near perfect on standard tests, this paper argues for mining richer signal out of saturated benchmarks rather than retiring them. Useful framing for procurement teams comparing vendors whose marketing all quotes the same maxed-out scores.
AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
A structured way to link agent harness configuration to outcomes and resource use is overdue, given how many production agent failures trace back to the surrounding tooling rather than the model. Teams building internal agent evals should take a look.
Same Answer, New Box: How Much of RLVR's Out-of-Domain Gain Is Answer Format?
Some reported gains from reinforcement learning on verifiable rewards may just be models learning where graders look, not getting smarter. A healthy reminder to stress test any vendor claim about cross-domain improvement before trusting it operationally.
Org & AI Architecture
5 storiesAnthropic is letting METR researchers look under the hood with employee-level access
Giving an outside safety evaluator employee-level access and no veto power over what it publishes is a notable trust move for a company heading into an IPO. Other frontier labs will face pressure to match it or explain why they won't.
Bank of America puts 4,000 people on fixing vulnerabilities uncovered by Anthropic's Mythos and other AI models
AI finding more vulnerabilities than humans can patch is becoming a real staffing problem, not just a security win. Any ops team adopting AI for audits or compliance checks should plan the remediation headcount before celebrating the detection rate.
Report: FTC probes Anthropic, OpenAI after Trump dismissed AI fears as 'hoax'
Regulatory scrutiny on security incident disclosure puts AI vendors in the same bucket as companies that delay data breach reporting. Enterprise buyers should ask vendors now how they handle incident disclosure, before it becomes a contract negotiation after the fact.
Opinion: How politics and policy are rewriting the rules of parcel delivery
Trade policy, national security rules and labor regulation are colliding in e-commerce logistics faster than most parcel networks can redesign around them. Worth a read for anyone setting multi-year network plans that assume today's policy environment holds.
Podcast: How AI will transform freight forwarding beyond the hype
A former Forto technology executive's honest take on where agentic AI actually stands in forwarding today versus the next three to five years. A good gut check against the louder vendor claims making the rounds this week.