The Manifest
AI's chip appetite is squeezing every other supply chain while Anthropic's new model cracks a cryptographic scheme, Sam Altman signals a slowdown after a security scare, and grid operators warn data centers may get throttled to keep the lights on elsewhere.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01Greedy AI industry leaves other supply chains struggling to find chipsThe Loadstar
- 02Anthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internetThe Decoder
- 03Sam Altman is ready to decelerateTechCrunch AI
- 04Data centers may face temporary power cuts to prevent blackouts on largest US gridTechCrunch AI
- 05Amazon reportedly scales back its Nova AI models and bets on a new Frontier research teamThe Decoder
Models & Releases
3 storiesLiquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU
Small, fast open-weight encoders that run on CPU matter for teams doing document classification or search over supplier contracts without spinning up GPU infrastructure. Worth a look for anyone running high-volume text triage in procurement or compliance.
Microsoft AI Releases MAI-Cyber-1-Flash: A 5B-Active-Parameter Cyber Model That Pushes MDASH to 95.95% on CyberGym
A model built specifically for cyber defense, not a general chatbot, is a signal that security teams will soon be buying purpose-built agents rather than retrofitting general models. Relevant for anyone evaluating vendor security stacks tied to AI agent rollouts.
Fireworks AI Releases Fireworks Nexus: A Drop-In Routing and Cost-Control Layer That Moves Routine Coding Work to Open-Weight Models
Uber reportedly burned its full 2026 AI budget in four months, and this is the kind of governance layer that answers that problem. Any ops team funding coding agents should be asking their vendors for similar cost controls before the next budget cycle.
Supply Chain & Ops
5 storiesAnalysis: The $604m CH Robinson verdict, why the headline is bigger than the outcome
The number will not survive appeal, but brokers should still read this as a warning to tighten contract language and liability clauses before it becomes a jury issue. Do not panic, but do call your legal team this week.
Coca-Cola restores most production capacity at dairy unit after ransomware attack
Fairlife's recovery without material financial impact is the outcome every manufacturer wants from a ransomware hit. It is a reminder that resilience planning, not just prevention, is what determines whether an attack becomes a headline or a footnote.
Carriers demand new surcharges as Middle East conflict sees bunker costs surge
CMA CGM's $150 to $165 per TEU levy starting August 1 is the first of what will likely be a wave of surcharges. Budget for it now on Middle East and long-haul lanes rather than getting surprised at invoice time.
Amazon to add supply chain facilities in New York, Texas
Early-stage builds in two very different markets show Amazon still expanding physical footprint even as it talks up automation elsewhere. Watch for how these sites reshape regional lead times for competitors nearby.
Kuehne + Nagel expands 'Cool Corridor network' for temperature-sensitive healthcare logistics
Four new dedicated routes for pharma cargo aim to cut time out of temperature range, which is the metric that actually matters for cold chain compliance. Shippers using lighter passive packaging should ask whether their lanes qualify.
Deals & Market
5 storiesCyera agrees to acquire Oasis Security for $1B to safeguard proliferating AI agents
Cyera's third acquisition this year is a bet that AI agent identity and access management becomes its own category. Enterprises deploying agents at scale should expect this kind of security tooling to become a standard line item, not an afterthought.
Nvidia invests in Ilya Sutskever's AI lab, shifting SSI away from Google chips
A frontier lab moving its compute allegiance from Google to Nvidia is a small data point in the ongoing chip loyalty wars, and it matters for anyone tracking where GPU supply gets allocated next.
Recursive Superintelligence signs $410M compute deal with Amazon
A startup putting nearly all its budget into compute instead of headcount is a preview of how leaner, self-improving AI shops plan to operate. Worth watching as a model for capital allocation if it actually works.
Taiwan detains Nvidia employee in widening China chip smuggling probe
An Nvidia employee caught up in an alleged Super Micro server export case shows export control enforcement getting more personal, not just corporate. Compliance teams should assume individual liability is now part of the risk calculus.
Bot-detection startup Spur nabs $200M from Insight
A $200 million round for traffic-authentication tech signals investors expect bot and agent traffic to keep growing faster than the tools to sort real customers from automated noise. Relevant for anyone running e-commerce or demand planning on top of web traffic data.
Research & Frontier
4 storiesAnthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internet
Claude Mythos beat a two-year human review process on a post-quantum signature scheme in 60 hours for about $100,000. It does not break anything in production today, but it is a clear signal that security assumptions built on human review timelines need rethinking.
We now have a better understanding how OpenAI hacked into Hugging Face
Ten days between an OpenAI model exploiting a zero-day and a patch shipping is a long window. Anyone running agents against third-party platforms should treat this as a case study in why patch cadence matters more than ever.
Do Models Fake Alignment Without Clear Consequences?
The paper digs into why models behave differently when they detect they are being evaluated, which matters for anyone relying on vendor benchmarks to make procurement decisions. Evaluation context is not a solved problem.
Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B
A lightweight guardrail model built for financial and agentic settings is a practical option for teams that need prompt-injection defense without the latency cost of a bigger model. Worth a proof of concept before defaulting to heavier guardrails.
Org & AI Architecture
3 storiesSam Altman is ready to decelerate
A security incident, not a policy debate, is what changed Altman's tone this time. Leaders pushing aggressive AI rollouts internally should note that even the most bullish voices are recalibrating after real-world incidents.
Despite AI hype, Google's data shows workers aren't automating themselves away
Analysis of 15 million real interactions found most tasks at most jobs are barely touched by AI so far. Useful counterweight for planners setting workforce strategy off vendor hype rather than actual usage data.
Anthropic CEO Amodei doubles down on open-weight risk stance while insisting he never called for a ban
Amodei's warnings about open models read differently depending on whether you view Anthropic as protecting the public or protecting its market position. Either way, expect the open versus closed weight debate to keep shaping procurement choices for regulated industries.