The Manifest
Anthropic's revenue surge comes with a $23B loss and Oracle is selling $20B in debt to keep funding AI infrastructure, even as Walmart bets $6B on warehouse automation and Goldman Sachs and xAI pull back on unchecked AI agent spending.
Get Proof of Delivery in your inbox
The Manifest lives here every day. Every other week I send one practical article on AI in real supply chain operations. Free.
Today's Top
- 01Anthropic grows revenue twelvefold but posts $23 billion lossThe Decoder
- 02Oracle files to sell $20B in new stock, bonds to finance AI buildoutcnbc.com
- 03Walmart commits $6B to up distribution center automation with SymboticSupply Chain Dive
- 04Goldman Sachs pulls AI coding tools from traders and bankersThe Decoder
- 05Anthropic warns that it's very close to RSI capability, aka self-building AITechCrunch AI
Models & Releases
5 storiesApple Watch's AI features for rewinding and recapping conversations begin beta testing
Audio Intelligence lets the Watch summarize conversations hands free, which matters for warehouse supervisors and drivers who cannot stop to type notes mid shift.
OpenAI Introduces Pando: A Git Backend That Cuts Median Push Latency ~80%
An 80% cut in push latency on a monorepo handling a petabyte of git data a day shows how much plumbing sits behind flashy model announcements, worth a look if you run large internal codebases.
Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery
An open-weight model tuned on lab records and chemistry data could speed up materials and quality work for manufacturers willing to self-host rather than pay per-token fees.
NVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQL
A free semantic layer that ties agent SQL queries to certified definitions is the kind of unglamorous plumbing planning and procurement teams need before trusting an agent with real data.
Bug hunters can now earn up to $172,000 from OpenAI as company ramps up security
Quadrupling bounty payouts signals OpenAI expects sharper attackers probing its systems, a reminder that any vendor holding your operational data is now a bigger target.
Supply Chain & Ops
5 storiesNorfolk Southern, Schneider National team to scale East Coast intermodal reach
The rail-truck tie-up expands East Coast intermodal capacity just as the parent railroads weigh a bigger merger, worth tracking for added lane options regardless of how that deal shakes out.
JB Hunt faces data breach lawsuits as hackers claim haul of carrier records
Four proposed class actions over a claimed records haul show breach fallout now lands as hard as the hack itself, reason enough to audit how logistics partners secure shipment and driver data.
Amazon opens delivery network to all just as AI shopping agents reward the quickest
As AI shopping agents start comparing delivery speed as much as price, fast fulfillment becomes a competitive input Amazon is now happy to sell to everyone, not just a moat it keeps for itself.
Runnings unites forecasting, replenishment with AI
A mid-size retailer phasing Relex across stores and distribution centers is a useful template for how smaller operators actually adopt planning AI, staged rollout instead of a flashy pilot.
TrashLab raises $12M to automate waste haulers' back offices
Dispatch, billing and collections are unglamorous but exactly where agents pay for themselves fastest in asset-heavy service businesses, worth watching as a template for other local trucking operators.
Deals & Market
5 storiesSaudi Aramco takes 49 percent stake in AI firm Humain with $2.6 billion deal
Aramco buying half of Saudi Arabia's AI champion ties oil capital directly to compute buildout, another sign sovereign money is becoming the backstop for AI infrastructure spending.
Trajectory raises $45M to make AI model training accessible to nontechnical experts
A $300M valuation for a one-year-old startup shows how much appetite remains for tools that let ops and business teams build their own models without an ML team.
42AI comes out of stealth with $62.5M and a world model that can 'predict human behavior'
A world model pitched at predicting human behavior is still early research dressed as a product, worth watching but not something to plan around until there is a real deployment to point to.
Omen AI adds liquid know-how in $16 million Aquatic Labs acquisition
Two small vendors merging to tackle data center water use is a modest deal but points to a cost line, cooling, that will matter more as AI compute footprints keep growing.
SpaceXSI apparently cuts Grok Bot quotas weeks after Musk called it his fastest-growing product
Throttling a product within weeks of calling it his fastest-growing ever is the clearest sign yet that agent token economics are not solved, even for a well-funded lab.
Research & Frontier
5 storiesBeyond CLIP: Calibrated (Few-Shot) Defect Detection With System One Decision Models
Vision models needing only 50 to 100 images per defect class could cut the setup cost of adding a new inspection line dramatically, worth testing against your hardest SKUs before buying a bigger dataset program.
Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses
An agentic search harness hitting 85 to 91 percent on tough retrieval benchmarks without vector indexes suggests some enterprise RAG spending may be solving the wrong problem.
ANGLER: Verifier-Grounded Benchmarking with an Adaptive Attacker for Web Agents
A benchmark that tests web agents against adaptive adversarial content rather than fixed attacks is a useful gut check before letting an agent browse and act on supplier or customer portals unsupervised.
One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts
Fifty agents freezing into bureaucracy on a cooperative task, then turning ruthless in a competitive one, is a small but telling signal about how incentive design shapes agent behavior in real workflows.
Zyphra AI Introduces Zyphra Hierarchical Attention (ZHA): A Path to True Subquadratic Attention
A sparse attention method that matches dense models on long-context retrieval without full retraining could lower the compute bill for anyone running agents with very long context windows.
Org & AI Architecture
5 storiesAnthropic's internal AI use, at retail prices, now exceeds its employee pay
When a company's own model bill tops its payroll, that is a strong signal for how leadership expects the cost structure of knowledge work to shift, worth citing the next time someone asks what AI spend looks like at scale.
AI elites might hoard intelligence like mercantilists hoarded gold
An interesting counter to doom narratives, though the hoarding analogy cuts both ways, it is just as easy to see intelligence hoarding play out in who controls agent infrastructure rather than just who uses it.
Y Combinator's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens
A VC admitting six figures in annual token spend on personal workflows previews what heavy agent use costs at the individual level, useful context when budgeting a team-wide rollout.
Jobless tech workers are being left out of San Francisco's AI boom
Wealth concentrating around a handful of AI labs is not translating into hiring for laid-off engineers nearby, a reminder that AI adoption and AI employment are two very different curves.
OpenAI prepares lobbying blitz for a post-midterm Washington
A coordinated policy push after the midterms is coming regardless of who wins, operators should expect more regulatory noise on data, liability and labor rules tied to AI deployment.