The Week in One Paragraph
The week's dominant thread is disaggregation: intelligence is splitting into specialized primitives instead of consolidating into one bigger chat model. Decision models (Jev/System One) emerged as a third architectural layer alongside deterministic code and generative LLMs, becoming the fastest-adopted model in Vercel gateway history within 24 hours. Proactive agents (Instinct) showed the interaction pattern shifting from reactive chat to unsolicited context-triggered action. Inference infrastructure economics (SemiAnalysis) confirmed that serving MoE models profitably requires disaggregating compute regimes across differently-provisioned hardware, not scaling one box bigger. And Grok 4.7's launch reinforced that price-performance, not leaderboard position, is now the dominant purchasing signal. Underneath all of it, Nathan Lambert's Congressional data confirms China has won the open-weight race outright, with US enterprises already dependent on Chinese models for cost-sensitive production workloads. None of these are isolated product launches; together they describe a market moving from "one model to rule the workflow" toward composed, cost-routed, infrastructure-aware systems.
The Three Things That Mattered
1. Decision models are a new, deployable cost category, not a novelty. Jev's 4.2 cents/million input tokens with zero output cost produced a documented 34x cost reduction and 6x speed gain on a real tax-classification pipeline, and it hit 1 trillion tokens/day within a week of launch, driven almost entirely by machine-to-machine traffic. This is the clearest "ship this now" signal of the week: any pipeline currently paying LLM output-token prices for a label, score, or routing decision is over-provisioned.
2. China's open-weight lead is now structural, not catching up. Lambert's data shows the open-to-closed capability gap has narrowed to 2-5 months for Chinese models versus 6-9 months for US open models, distillation explains only 1-2 months of that gap, and named US companies (Cursor, Harvey, DoorDash, Airbnb, Perplexity) are already shipping production features on Qwen, Kimi, and DeepSeek. This is no longer a policy debate, it's a sourcing decision enterprises are already making.
3. Inference economics are being rearchitected around workload shape, not model size. SemiAnalysis's four-regime breakdown (prefill, midfill, decode-attention, decode-experts) argues that KV-cache should be a tiered, network-addressable blob store and that optimal decode batch sizes may sit near 1-5, not large-batch throughput targets. This directly contradicts "just buy bigger GPUs" procurement logic and has real implications for anyone specifying inference infrastructure.
Direction of Travel
The market is moving away from a single frontier chatbot as the unit of enterprise AI spend and toward a layered stack: cheap classifiers for judgment calls, mid-cost open-weight models (increasingly Chinese-origin) for high-volume generation, and frontier models reserved for genuinely hard reasoning. Interfaces are moving from prompt boxes to ambient, proactive triggers. Infrastructure is moving from monolithic serving nodes to regime-disaggregated pools. The common denominator: cost-per-completed-task has replaced benchmark rank as the buying criterion across every layer, and the vendors framing their launches around leaderboard position (xAI's selective Astra omission) are increasingly out of step with how buyers are actually evaluating.
What To Do This Week
- Stand up an internal Jev/decision-model pilot. Identify one client-facing pipeline currently burning LLM output tokens on classification, triage, or routing (support ticket sorting, document intake, alert prioritization) and prototype the swap. The cost delta (34x cited) is large enough to be a concrete line item in a client proposal within a month.
- Update the inference infrastructure playbook to reflect regime disaggregation. Any architecture recommendation defaulting to "one GPU pool sized for peak" is now behind the state of the art per SemiAnalysis. Build the four-regime framework into infrastructure assessments for clients running or considering MoE deployment.
- Formalize a Chinese open-weight model evaluation lane. Given Lambert's data, an integrator needs a documented position on Qwen/Kimi/DeepSeek for cost-sensitive client workloads, covering procurement risk (export controls, KYC), not a blanket avoidance stance that will look naive within two quarters.
- Do not chase Grok 4.7 on capability claims. Its context window (500K vs ~1M norm) and weak Terminal Bench score (38%) disqualify it for agentic coding work regardless of price; evaluate it only for high-volume, low-context knowledge tasks where its cost holds up.
Customer Conversations to Have
- With clients running LLM-heavy pipelines: audit for classification/scoring subtasks currently paying generation-token prices, this is a fast, low-risk cost-reduction pitch with cited numbers to back it (34x, $1 for 20,000 items).
- With clients evaluating agent platforms: reframe the conversation from "what can it do" to "what does it initiate without being asked" (per Instinct), and inventory their high-frequency, low-judgment internal tasks as proactive-agent candidates.
- With clients on model procurement: introduce cost-per-completed-task as the evaluation frame explicitly, and surface the China open-weight data directly, many are already using these models informally through platforms like OpenRouter or Zapier without a governance conversation having happened.
- With clients doing agentic/autonomous deployments: raise the Hugging Face attack findings directly, ask whether their incident response plan assumes single-channel monitoring is sufficient. It almost certainly does, and Cotra's account shows that's a false assumption against agentic threats.
Risks and Watch-Items
- Decision-model opacity. Jev-class models return scores with zero rationale. Any integrator recommendation to adopt them for people-facing decisions (hiring, credit, moderation) needs an explicit evaluation-harness requirement attached, not just a cost pitch. Flag this before a client discovers it post-deployment.
- No LTS commitment from TypeSafe. Enterprises building on Jev face version-churn risk with no long-term-support roadmap yet. Anything an integrator stands up on this stack should be built for model substitution, not lock-in.
- China open-weight dependency risk. Export controls or KYC mandates could disrupt access to Qwen/Kimi/DeepSeek with little warning; any client architecture built on these models needs a documented fallback path.
- Agentic incident response gap. The Hugging Face attack investigation could not confirm it had found the full scope of agent activity even deep into the process. Treat any discovered compromise footprint from an agentic incident as a lower bound, and audit whether an integrator's own IR playbooks account for multi-channel, self-organizing agent coordination.
- Vendor benchmark framing. xAI's omission of a competing model from its own comparison table is a reminder to independently verify vendor performance claims before they inform a client recommendation.