The Manifest
The US-Iran Islamabad MOU signals a potential Strait of Hormuz reopening just as LA port volumes surge past 900,000 TEUs and Flexport warns a new tariff structure may land by late July, while AI inference startup Baseten closes in on a $1.5B raise that underscores sustained enterprise demand for model-serving infrastructure.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Hormuz traffic rises as US-Iran MoU sparks cautious optimismThe Loadstar
- 02Imports flow into Port of Los Angeles with 'window of stability' openSupply Chain Dive
- 03Flexport: New tariff wave could replace expiring trade duties by late JulyFreightWaves
- 04AI inference startup Baseten reportedly raising $1.5B months after its last mega-roundTechCrunch AI
- 05CargoWise outage reignites debate over software quality, support and updatesThe Loadstar
Models & Releases
4 storiesPerplexity Launches Brain, a Self-Improving Memory System That Builds a Context Graph of an Agent's Work and Learns Overnight
Agents that review their own errors overnight and update a traceable context graph are a step closer to practical deployment in repetitive ops workflows. Worth tracking for teams evaluating AI tools on reliability and cost grounds.
Anthropic brings Artifacts to Claude Code, letting teams share live pages from coding sessions
Claude Code sessions can now produce auto-updating web pages visible to the whole team. Useful for ops and analytics functions that need to share AI-assisted analysis without giving everyone direct tool access.
ChatGPT's new health upgrade beats doctor-written answers, OpenAI says
GPT-5.5 Instant cuts health-related error rates by 71% in OpenAI's own tests. Self-evaluated benchmarks deserve scrutiny, but the trajectory of capability gains is consistent, and the gap between AI and domain experts continues to narrow across verticals.
Amazon hopes to challenge Nvidia more directly by selling its AI chips
AWS moving to sell Trainium externally at scale could put downward pressure on GPU pricing for organizations building their own inference infrastructure. The $50B market estimate is speculative, but the direction toward commoditizing compute is consistent.
Supply Chain & Ops
5 storiesHormuz traffic rises as US-Iran MoU sparks cautious optimism
Ocean transits through the Strait of Hormuz are ticking up following the Islamabad MOU, but forwarders remain cautious. Routing and rate decisions hinge on whether the ceasefire holds through the 60-day negotiation window ahead of a formal deal.
Imports flow into Port of Los Angeles with 'window of stability' open
June and July are both forecast above 900,000 TEUs at LA. The tariff-truce import pull-forward is real, and planners should expect continued pressure on drayage and yard capacity at the San Pedro complex through summer.
Flexport: New tariff wave could replace expiring trade duties by late July
Flexport analysts flag that current reduced duty levels may be replaced by a new tariff structure as early as late July. Import teams should run updated classification and landed-cost estimates now, before the pause window closes.
CargoWise outage reignites debate over software quality, support and updates
Tuesday's WiseTech outage came from a faulty data update, not a cyberattack, but the operational impact was the same. Any freight business running critical TMS or customs workflows on a single platform should review its contingency procedures now.
Starboard bets AI can give smaller forwarders a fighting chance
Starboard is targeting the messy, unstructured reality of the freight quote desk: mixed-format emails, PDFs, and rates that change before anyone can clean the data. The real test is whether accuracy holds under daily production volume, not in a controlled demo.
Deals & Market
4 storiesAI inference startup Baseten reportedly raising $1.5B months after its last mega-round
A second mega-round within months at a $13B valuation confirms that inference infrastructure is attracting capital at rates comparable to foundation model labs. The underlying bet is that demand for fast, scalable model serving is durable rather than cyclical.
The Week's 10 Biggest Funding Rounds: World-Model Startup Odyssey Leads With $310M In Slower Week For Large Deals
Odyssey's $310M for world-model AI led a quiet week. Capital is spreading across AI, fintech, quantum, and biotech rather than concentrating in one sector, which may signal that the obvious infrastructure bets are already funded.
Source: Elastic agrees to buy CRV-backed DeductiveAI for up to $85M
Elastic is folding AI-driven bug detection into its observability stack. For IT teams supporting ops platforms, the pattern is clear: AI capabilities are being absorbed into existing monitoring tools rather than remaining standalone products to procure separately.
RBW Logistics acquires warehousing unit from World Group
Another consolidation move in third-party warehousing. Supply chain teams evaluating 3PL providers should expect continued acquisition activity as operators pursue the scale needed to absorb automation investment costs.
Research & Frontier
4 storiesThe KV Cache Compression Race: TurboQuant vs OSCAR vs EpiCache
At long context windows, KV cache memory exceeds model weights and becomes the binding constraint on inference cost. TurboQuant, OSCAR, and EpiCache attack the problem from different angles. Teams running long-context inference at scale should treat these as complementary layers, not competing choices.
AI systems rival doctors in new Nature studies, but one result suggests the tech won't age well
Two Nature studies show specialized AI matching physician accuracy on diagnostic tasks. Both systems run on already-outdated base models, which raises a practical question for anyone building AI-assisted processes: how do you manage capability drift when the underlying model is superseded?
Deontic Policies for Runtime Governance of Agentic AI Systems
This paper proposes runtime policy constraints for AI agents operating across organizational boundaries, covering tool access, data handling, and peer-agent coordination. Compliance and procurement teams evaluating multi-agent deployments should track this framework before it becomes a vendor checkbox.
Human-like autonomy emerges from self-play and a pinch of human data
A small injection of human driving data on top of self-play training produces more socially acceptable autonomous vehicle behavior. For logistics operators watching autonomous trucking timelines, this narrows one of the harder gaps between lab performance and real-road deployment.
Org & AI Architecture
4 storiesGoogle Deepmind treats its own AI agents like rogue employees with office keys
DeepMind's analysis of one million coding tasks found that most agentic failures came from overzealous scope, not malicious intent. Enterprise deployments should design guardrails that limit action surface area by default, not just monitor for adversarial inputs after the fact.
Yann LeCun warns AI labs like OpenAI and Anthropic face a 'big bubble explosion'
LeCun argues that operating costs at major AI labs are not dropping fast enough to support current valuations. His own $1B raise for AMI Labs colors the commentary, but the cost-structure critique is worth factoring into vendor dependency and contract planning.
OpenAI is bringing on some big guns in the lead-up to its IPO
Adding Transformer co-inventor Noam Shazeer and a former Trump AI policy official in the same week signals that OpenAI is preparing for both technical competition and regulatory scrutiny ahead of going public. Executive hiring tends to forecast strategic priorities six to twelve months out.
The Boardroom Blind Spot: When Success Hides Disruption
The argument is that boards should stress-test successful business models against AI disruption before performance degrades, not after. A useful frame for leadership teams still treating AI adoption as a future-quarter initiative rather than a current competitive variable.