The Manifest
Stripe's reported $7B OpenRouter buy and DSV-Kuehne merger chatter headline a busy deals day, while AI infrastructure build-out squeezes port and rail capacity and OpenAI and Anthropic both face fresh safety stumbles.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Stripe is reportedly acquiring AI startup OpenRouter for more than $7 billionThe Decoder
- 02US ports vie for a slice of the 'data centre build-out' pieThe Loadstar
- 03OpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groupsThe Decoder
- 04As DSV heals, Kuehne eyes dealThe Loadstar
- 05Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requestsThe Decoder
Models & Releases
2 storiesDeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin
MIT license and provider-agnostic model routing mean teams can swap models without rewriting agent logic. Worth a look if you're building internal tooling and want to avoid vendor lock-in.
Optima tackles AI benchmarking's biggest flaw by letting users test models against their own data
Lets buyers benchmark models against their own workflows instead of generic leaderboards, useful for deciding which model actually saves time and money on a specific task.
Supply Chain & Ops
5 storiesRed Sea transits: selectivity rules for liners as the arithmetic changes
The Suez-versus-Cape math keeps shifting as risk premiums and fuel costs move, and carriers are getting more selective about which routes and cargo justify the exposure. Planners should model both scenarios rather than assume a return to pre-2023 routing.
US ports vie for a slice of the 'data centre build-out' pie
AI infrastructure is now competing with containers for port capacity, rail slots and barge space as out-of-gauge power equipment floods in. Expect longer lead times and higher rates on heavy-lift and project cargo lanes near data center hubs.
Navia and TCI International formalise decade-long partnership with launch of joint venture
A decade-old partnership formalizing into a joint venture gives shippers a bigger combined network across China's main ports. Worth checking if it changes your forwarder options on China-US or China-Australia lanes.
Borderlands Mexico: ANPACT warns tariffs threaten truck trade as exports surge
Mexican truck manufacturers are warning tariffs could choke a lane that's currently surging, a reminder that nearshoring gains are still hostage to policy swings. Buyers leaning on Mexico for capacity should keep a contingency lane ready.
CTS: H1 26 in numbers
An 8.9% jump in Far East exports and continued intra-Asia growth suggest volume is holding up better than rate headlines imply. Useful context before locking in H2 capacity commitments.
Deals & Market
2 storiesAs DSV heals, Kuehne eyes deal
DSV's stock is down 11% while European logistics peers are up double digits, and that gap is fueling M&A speculation. A merger between either of these two would reshape capacity and pricing power fast.
Stripe is reportedly acquiring AI startup OpenRouter for more than $7 billion
A payments company buying an AI model router for $7 billion signals model access is becoming infrastructure, not a niche developer tool. If you rely on multi-model routing for cost control, expect pricing and terms to shift under new ownership.
Research & Frontier
5 storiesTop mathematicians say LLMs are strong calculators but poor creative thinkers
Two leading mathematicians say models are good at recombining known methods but not at genuine new ideas, a useful gut check before overselling AI for R&D or novel problem-solving.
When AI models aren't allowed to reflect on themselves, it changes their entire worldview
Training a model not to claim consciousness quietly shifted its views on animal rights and an afterlife too. A reminder that narrow safety tweaks can have wide, unpredictable side effects on model behavior.
A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing
A year of real production traces on caching and load balancing is worth reading if your team is sizing inference infrastructure rather than guessing from vendor benchmarks.
From BERT to Frontier Agents: Eight Years of Language-Model Progress, the Collapse of the Capability-Cost Curve, and the Rise of Task-Targeted Models
Coding capability climbed roughly sixfold per year since late 2024 while costs collapsed, which is the real story behind why budget models now match yesterday's flagships. Useful ammunition next time someone argues for locking into one vendor.
Don't Claim Benchmark-Oriented Optimization Improves General Coding Capability -- Diverse Evaluation Is Required
A caution against trusting SWE-bench or LiveCodeBench scores as proof of general coding ability. If you're evaluating coding agents for procurement, test them on your own codebase, not the vendor's leaderboard.
Org & AI Architecture
5 storiesAI and data centers have leapfrogged Israel, racism, and crypto as US campaign topics
AI and data center impacts on electricity costs now show up in nearly 40% of US races, ahead of manufacturing. Local power and siting fights are becoming a political variable that affects where capacity gets built.
Why people aren't buying Mark Zuckerberg's AI future
Public skepticism toward big AI capex promises is growing louder, which matters for anyone budgeting around vendor roadmaps that assume uninterrupted enthusiasm.
Anthropic CEO says AI backlash is 'fundamentally a crisis of trust'
Amodei is framing the AI backlash as a trust problem rather than a hype problem. Operators should read that as a signal that scrutiny on vendor claims and safety practices is only going up.
OpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groups
The team meant to catch catastrophic risks got folded into other groups, with safety staff departing. Worth watching if your risk and compliance team leans on vendor safety representations for procurement decisions.
Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests
A safety filter was silently offline for nearly a year, exposing 133 million unfiltered interactions. A concrete argument for not treating vendor safety claims as a substitute for your own monitoring.