vLLM v0.31.0: DeepSeek optimization and large-scale serving breakthroughs

Chinese frontier labs published safety results for just 3.6% of model releases; Beijing's Framework 3.0 tightens applications while leaving frontier development unrestricted

SemiAnalysis conducted original analysis of 857 model releases from nine leading Chinese AI developers (ByteDance, Alibaba, Tencent, Baidu, DeepSeek, MoonShot, Zhipu, MiniMax, StepFun) from 2021–2026, finding only 31 releases (3.6%) accompanied by published safety results, with just 9 (1.1%) available at launch. Reasoning models are 93% without published results; no Chinese frontier text model shipped with dangerous-capability evaluation. The remaining 813 releases (94.9%) have no safety disclosure at all. China's stated safety posture contradicts its actual practice. While Beijing's AI Safety Governance Framework 3.0 (released September 14, 2026) introduces new risk categories—agents, embodied AI, models deceiving evaluators—and triggered seven regulatory texts in five days, all govern deployment and content, not frontier model development. The comprehensive AI Law promised in 2023–2024 was shelved in 2025. Instead, China pursues rapid AI+ penetration targets (70% by 2027, 90% by 2030) with safety duties only at the application layer. The result: a regime stricter than the US on outputs but looser on frontier capabilities—no compute thresholds, no pre-release safety gates, and no duties triggered by training scale or model capability.

OpenAI's September-end ARR disclosed at ~$50B, well below $70B media reports; gap stems from cloud partnership revenue accounting

The Financial Times reported that OpenAI told investors its annualized revenue run rate at September end was approximately $50B, significantly below widely circulated media estimates of $70B. The discrepancy appears to stem from how partnership revenue sharing with major cloud providers is counted. The disclosure sparked a 5%+ drop in Oracle shares and broader AI stock selloff, underscoring how dependent the sector is on accurate OpenAI metrics and the opacity of private company financial reporting.

Cache pricing emerges as real LLM retention lever; mid-tier competitive tier solidifies at $2–$10 input range

Jamin Ball's analysis in Clouded Judgement reveals that while headline input/output pricing drives acquisition—Astra and Fable both at $10/$50 per 1M tokens—cache read rates determine actual agent workload costs. Providers range from $1/cached token (Astra, chat-optimized) to $0.01 (Luna, loop-optimized). In a realistic 40-turn agent session with 85% cache hit, Fable comes in ~16% under Astra despite identical vanilla list pricing, while Grok 4.7 is cheapest on vanilla but more expensive than Sol and Terra when cache costs are factored in. A distinct mid-tier has emerged below frontier models, with Sol and Terra pricing $2–$10 per 1M input tokens versus Astra/Fable's $10, creating market segmentation by use case and workload optimization. Ball argues this mirrors how egress pricing created cloud lock-in: as agents run longer and perpetually, cache hit rates will drive switching costs. Swapping models mid-session means rebuilding cache from scratch, creating real retention leverage independent of headline pricing.

Periodic Labs scales autonomous lab infrastructure for materials discovery; treats RL environments as noisy physical experiments

Liam Fedus and Ekin Dogus Cubuk's Periodic Labs, founded September 2025, has built industrial-scale autonomous lab infrastructure and is now deploying high-throughput experimentation for materials discovery. The team combines LLM experts, solid-state physicists/chemists, and hardware engineers in a Bell Labs-style model to ground AI in real experimental data rather than internet text. Periodic's "synthesis superintelligence" approach treats reinforcement learning environments as physical experiments with inherent noise and uncertainty. Rather than optimizing against known answers, the lab trains AI systems on noisy X-ray diffraction data, furnace degradation, measurement errors, and incomplete telemetry—requiring decision-making under uncertainty fundamentally different from digital RL. This allows discovery of genuinely novel materials by avoiding memorization of existing scientific literature. Fedus and Cubuk argue that intelligence alone is insufficient for scientific discovery; frontier models will still need to physically experiment, and failed experiments may be some of the most valuable training data.

vLLM v0.31.0: DeepSeek-V4.1-Flash Optimization, Fast Restart, and Large-Scale Serving Improvements

vLLM Contributors in vLLM Project (GitHub) published a release titled "vLLM v0.31.0: DeepSeek-V4.1-Flash Optimization, Fast Restart, and Large-Scale Serving Improvements". Main takeaways:

• DeepSeek-V4.1-Flash now uses FlashMLA mega attention with NVFP4 compressed KV cache as SM100 default, plus fused TP all-reduce and MoE finalize, achieving significant throughput gains
• New `vllm preload` CLI launches weight-cache daemon keeping post-quantized weights resident in GPU memory across engine restarts, reducing startup overhead
• Large-scale serving improvements include MoonEP balanced EP all2all backend, prefill context parallelism with data parallelism, and DeepEPv2 with sequence parallelism for multi-cluster deployments
• Model Runner V2 now supports draft-model speculative decoding and custom logits processors, with LiLiCorr drafter and async scheduling for DFlash
• Scheduling controls added: `--max-num-active-seqs` caps RUNNING admission independently, `--long-prefill-token-threshold` adapts to waiting prefills, improving request queuing under load

Why Can't the U.S. Build Ships?

Brian Potter (Construction Physics), guest post on Noah Smith's Substack in Noah Smith's Substack (noahpinion.blog) published an essay titled "Why Can't the U.S. Build Ships?". Main takeaways:

• U.S. built only 5 oceangoing commercial ships in 2022 vs. China's 1,794 and South Korea's 734; China's shipbuilding capacity is 232x larger than U.S.
• U.S. ships cost roughly 2x more to build than foreign equivalents, a cost gap that has persisted for 100+ years, not a recent decline
• U.S. competitive shipbuilding ended in the 1890s when transition from wooden sail to steam-powered metal ships required skills American builders lacked and Britain dominated
• Protectionist cabotage laws (Jones Act, 1920) preserved domestic shipbuilding by requiring U.S.-built vessels for domestic trade, but prevented competitive development
• WW1 and WW2 demonstrated massive U.S. shipbuilding capacity (5,000+ ships in WW2, 60% of world tonnage post-war) but failed to create sustainable commercial industry afterward