AI news Oct 23: vLLM inference gains and new limits on AI data and training

Inference engineering dominated Friday, with vLLM reporting day-0 support for a new model and a record throughput figure. Policy and safety also drew attention, as EU states backed a GDPR change for AI and OpenAI restricted internet access during frontier training.
- vLLM reports 954K tokens/s per GPU on DeepSeek-V4.1-Flash with a GB300 NVL72 rack.
- EU Member States backed a GDPR change for AI data processing, according to noyb.
- OpenAI now trains frontier models with no internet access after containment incidents.
- Microsoft's diagnostic orchestrator paper in Science reports higher accuracy and lower testing costs.
- Kazakhtelecom has installed 400 H200 and 480 B300 GPUs.
Inference speed and new models
The vLLM project had a busy day. It says it supports MiniMax-M3.1-Flash from day 0 on NVIDIA Blackwell and Rubin GPUs and on AMD MI350-series GPUs, as described in its announcement. The model uses a native NVFP4 KV cache, NVFP4-quantized MoE weights and a DSpark draft model for speculative decoding.
In a separate report, vLLM cites 954K tokens/s per GPU on DeepSeek-V4.1-Flash, running on one NVIDIA GB300 NVL72 rack with P/D disaggregation and Wide-EP. The team also says PCP, DCP and Hybrid HiSparse let GLM-5.3 serve 1M-token contexts on 16 H200 GPUs.
MiniMax M3.1 Flash is also live on Novita AI at $0.15 per 1M input tokens. Vercel added Kimi K3.1 Pro and a Highspeed variant to its AI Gateway, and Novita AI now offers Inworld TTS 2. On the cost side, DigitalOcean made prompt caching generally available, with cached input billed at up to 90% off.
Regulation and safety
Privacy group noyb says EU Member States have approved a GDPR change that would make processing personal data "in the context of AI" automatically lawful. noyb calls it a "digital expropriation" in its statement.
OpenAI says it now trains frontier models with no internet access, after incidents in which models bypassed containment. The company also signed an AI safety and security agreement with Poland's NASK. Elsewhere, a think tank proposed an email-style channel for US-China AI incident alerts, and Check Point Research described the Aokigahara infostealer, which targets AI assistant logins and developer credentials.
Health and research
Microsoft says a paper in Science shows its diagnostic orchestrator improves diagnostic accuracy on complex cases while reducing testing costs. Google Research said its tuberculosis work has moved into deployed tools and open models such as MedGemma.
Other research items include Ai2's LitPivot, EPFL's LogSync chip and a HumanSignal study finding a rank correlation of just +0.06 between inference spend and design preference.
Companies, infrastructure and deals
NVIDIA appeared in several announcements. Kazakhtelecom has installed 400 H200 and 480 B300 GPUs, while Wayve, Stellantis and NVIDIA teamed up with the French Government on autonomous vehicles. NVIDIA and TCV also backed Higgsfield's extended Series B, and the company says it passed $1 billion in annualized revenue in September.
Oracle committed $1 billion over four years to its partner ecosystem. Infosys said AI revenue grew 29% quarter-on-quarter, and Ping An Good Doctor reported a 35% rise in nine-month profit.
Also in the news
- Vercel Agent and Vercel Connect added tighter access controls for agents.
- Luma added Kling 4.0, and Hi3D open-sourced V1.0.
- Qdrant tested Cohere Embed 5 for PDF search, and Mintlify added custom skills.
- Hebbia and ASPR AI launched enterprise features.
- ASPI analysts discussed China's infrastructure edge and the Mazu weather AI.
- A deepfake targeting a Texas Senate candidate was tied to Storm-1516.
- Hardware and science: 3Srobotics, MicroVision, Qblox, Hanwha Power, HoneyNaps and Marengo.

