The Intelligence Index v5 reranking says more about what gets measured now than about any sudden capability collapse in Chinese open-source models. 1. Artificial Analysis moved 60% of the index weight to private test sets. It added FrontierCode 2.0 for real-world programming and Terminal-Bench-Science for long-horizon research workflows. General, agents, coding, and science now each carry 25%. 2. US frontier closed models now dominate the top. Kimi K3.1 Pro remains top ten but closed-weight. Among actually open-weight Chinese models, DeepSeek V4.1 Flash 1017 ranks 14th and GLM 5.3 ranks 15th, both behind Anthropic's entry-level Claude Haiku 5.5. 3. The widest gap is Terminal-Bench-Science. Leading closed models complete about 60% of tasks. Most Chinese open-source models manage roughly 10%. Benchmark contamination and distillation shortcuts alone do not explain a gap that wide. 4. Chinese builders spent this year on coding and agents, where validated technical paths, user demand, and commercialization opportunities sat. GLM-5.2 and Kimi-3 became genuinely useful coding substitutes over the summer. Scientific reasoning and research-workflow capability received far less post-training investment. 5. DeepSeek's Liang Wenfeng said it plainly in an investor note: high-quality data labeling is a key constraint, half the research team works on annotation, and closing the gap takes about a year. That covers general data; scientific data is a narrower bottleneck still. 6. Anthropic estimates the global biopharma industry at $2 trillion in its prospectus. Both Anthropic and OpenAI are building toward closed-loop wet-lab workflows where models propose experiments, robots execute them, and results feed back into the next hypothesis. AI accelerating its own research stack is the likely 2027 narrative. 7. Nathan Lambert's caution is worth keeping: one index cannot capture performance across all deployment contexts, and private test sets weaken external auditability. The coding cost-performance advantage Chinese open models hold does not disappear because the yardstick changed. The science gap is real, though, and it is where the next round of frontier competition is being defined.

Silicon Valley's AI giants need a bigger application than coding to justify their trillion-dollar capex. Drug discovery is where they are placing the bet. 1. The market signal is real. Tirzepatide and semaglutide, the two GLP-1 drugs, produced roughly $71 billion in combined 2025 revenue. That is more than double OpenAI and Anthropic's combined end-2025 annualized revenue run rate. 2. October marked a step change. Frontier AI companies announced investments, acquisitions, self-built wet labs, and industrial partnerships nearly every week. The goal has shifted from improving molecular screening to owning the full research loop. 3. Nvidia calls this Lab in the Loop. The model proposes a hypothesis, designs an experiment, the wet lab executes it, and the data feeds back into the next round of model decisions. 4. Kevin Weil's departure to start Warp Sciences is the clearest signal. The former OpenAI science lead is building an AI-automated drug discovery lab in the US. He explicitly aims to pull work back from Chinese contract research organizations like WuXi AppTec. 5. The data is the prize. Automated labs produce compounds, and they also produce proprietary experimental data, including negative results. That data is the feedback mechanism that lets models keep improving. 6. China's lead in CRO and CDMO execution is real, built on years of outsourced research and skilled lab engineers. Lab in the Loop is designed to dilute that advantage by moving both execution and learning back to Silicon Valley. 7. Google's VERID framework is one serious attempt at the missing infrastructure. It acts as a digital twin of a wet lab, tracking equipment layouts, reagent identity, and liquid volumes in real time. That lets AI agents simulate experiments before execution. 8. Closed loops are already running. OpenAI and Ginkgo Bioworks ran six design-test-analyze cycles, testing over 36,000 reaction combinations and cutting cell-free protein synthesis costs by about 40%. Anthropic has built its own Bay Area wet lab and validated a new CRISPR-like enzyme system. 9. The barrier is forming. Anthropic's September protein design competition with Adaptyv explicitly excluded participants from China and other regions. That signals frontier AI and wet-lab infrastructure are becoming linked in ways that may not stay open. For China, the pattern is familiar. Its AI companies have followed Silicon Valley's validated paths before, but arriving a few steps late can mean missing the window when compounding advantages accumulate.

OCP Global Summit this week showed AI infrastructure splitting along lines of labor: CPU for agent execution, GPU plus specialized accelerators for compute, shared memory, and power, cooling, and networking as design starting points. 1. AMD's Helion rack separates agent execution from model inference. The pure-CPU config scales to 32,768 cores for running agent workloads, connected to a separate inference system. Calling models and executing tasks now justify distinct infrastructure. 2. Nvidia contributed Vera Rubin rack and compute-tray designs through its MGX open architecture. That covers liquid-cooled blind-mate connectors, fanless trays, and high-current busbars for racks up to about 250kW. It also opened part of its DSX AI-factory management software. 3. SK Hynix's MARS architecture pools memory across servers instead of tying it to individual CPU/GPU nodes. In multi-turn agent inference, shared KV cache delivered a 3-5x first-token latency improvement versus recompute. Memory vendors now design how a rack organizes compute and data. 4. D-Matrix showed attention-FFN splitting between GPU and its Corsair XPU, across a system connecting up to 1,024 XPUs. Specialized architecture is entering real heterogeneous inference pipelines rather than demanding wholesale GPU replacement. 5. Google and Microsoft are co-designing from chip to grid. The grid-resilience spec calls for data centers to stay connected, respond automatically, and restore near-full load in under two seconds. Microsoft's Mt. Diablo cabinet separates power conversion from compute racks across a roughly 100kW to over 1MW modular range. 6. Lightmatter and Dell are working on how co-packaged optics become installable, testable, and maintainable products. They are defining mechanical, power, signal, thermal, and management requirements for front-panel, near-package, and co-packaged optics.

ByteDance launched Jianying Hub on September 20, bringing AI generation into its video editor. Creators can develop scripts and storyboards, generate material with ByteDance’s Jimeng and Xiaoyunque, then edit it on the timeline. Independent AI video platforms such as Flova have been building around the same need: keeping a project together without moving files between tools. But we doubt that a more convenient workflow alone gives these platforms much protection, because ByteDance also helps creators get their work seen and paid for. Xiaoyunque offers a direct Douyin distribution channel and cash incentives; Jimeng connects creators to commercial assignments through Xingtu. Distribution is already difficult. According to DataEye, Douyin added 221,900 AI short dramas in the first half of 2026, but only 1,055 exceeded 100 million views. Aishi Technology’s PixVerse is looking for customers outside that content pool. On October 9, its Growth Studio connected to Meta Catalog Product Video, turning product catalogs into video ads for merchants. That gives it a specific paid task beyond helping creators produce more short dramas, though it still has to compete with tools such as Marpipe and Creatify. If the ads perform poorly, merchants can cut the budget for producing those videos first. h/t @TMTPostGlobal