The Stack
The plumbing of AI got two small releases today — one that stops GPUs from idling on CPU work, one that puts image generation on a laptop-grade footprint. Hugging Face shipped the first major version of its tokenizers library, and the headline number is real: 3 to 30 times faster encoding than v0.23 on a single thread.
The Arena
AMD's stock ran roughly 24% in five sessions and pushed the chipmaker into a club only three other US semiconductor firms have reached. The milestone says less about one product line than about how the market is now pricing AI compute. Advanced Micro Devices topped a $1 trillion market cap for the first time on Monday, hitting an intraday record of $613.92 before closing up about 9% at $610.
The Guardrails
Nvidia's chief executive spent Sunday telling CBS that the industry's own doomsday warnings are a lobbying tactic — and he named the ask he thinks sits behind them. Jensen Huang says the AI leaders warning about extinction are not asking for new rules — they want out of the ones already on the books.
The Arena
The clearest signal on 2027 memory supply is not a wafer number — it is a cleaning contract for the glass plates that hold HBM wafers together while they are ground thin. Samsung Electronics is expected to more than double output of its HBM4 family next year — sixth-generation HBM4 and seventh-generation HBM4E — and the tell is in its materials purchasing, according to Seoul Economic Daily, citing semiconductor industry sources.
The Arena
The leaked number everyone is quoting is the smaller one, and it got better. The Financial Times reported Friday that OpenAI expects negative free cash flow of $278 billion between 2026 and the end of 2030, according to a presentation the company prepared in July for a computing deal. The same materials put revenue at $350 billion in 2030 against roughly $36 billion this year, and compute and infrastructure at about $856 billion across the period. OpenAI declined to comment.
The Frontier
China's largest state carrier put an open-weight coding agent on the market today, and the interesting part is not the benchmark row — it is the stack underneath it. China Telecom's AI arm released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts coding agent that activates 4 billion parameters per token and was trained end to end on Huawei's Ascend NPU platform using the MindSpore framework.
The Arena
The AI memory shortage has been a DRAM and HBM story all year. Storage is next, and both Washington and Beijing are moving on it in the same week. SK Hynix's US subsidiary Solidigm is weighing what would be the first NAND flash factory on American soil, with upstate New York emerging as a leading candidate site. The company is studying potential locations, according to reports citing people familiar with the plan, and no decision has been announced.
The Frontier
Two gaps got money and code this week: science still has no compiler, and the next generation of AI chips still has no common language. Keli Xinxu, an AI-for-science startup founded by an assistant professor at Hong Kong University of Science and Technology, has raised nearly 50 million yuan — about $7 million — in a first round led by Inno Angel Fund, with Yijing Capital, Xiaomiao Langcheng and Linge Ventures joining and Lighthouse Capital as strategic incubator and sole financial adviser.
The Arena
Two answers to the same question landed a day apart: how much compute is actually coming. One is a shipment forecast from the company selling most of it, the other is a hardware announcement from the one building it without American parts. Jensen Huang says Nvidia will sell twice as many chips next year as it does this year, and he used the occasion to argue AI should be regulated as a product rather than a technology.
The Arena
A year-old AI accelerator is appreciating like real estate, not like a graphics card. That changes what gets borrowed against it — and what it says about the shape of the boom. Nvidia's B200 is trading at 158% of its launch price roughly a year after it reached mass availability, according to Silicon Data's analysis — an accelerator that added 58% instead of shedding the 20% to 40% that standard depreciation schedules assume.
The Guardrails
Thursday's through-line was self-reporting. One lab published numbers on how much of its own research its model now performs; another published six accounts of its models misbehaving; a security firm showed that a provenance watermark changes what an agent actually executes. Around all of it, the oversight body Washington drafted stayed dead in draft while Sacramento and Beijing wrote rules of their own. Models & Research * A 35B-parameter mixture-of-experts model ran at 20 tokens a second…
The Frontier
An autonomous-driving company answers the long-tail problem with compute, while the agent layer gets a gate that rules before the action runs and communications teams get a way to watch what AI says about them. Zelos, the Suzhou RoboVan maker founded by ex-Baidu autonomous-driving engineers, says it has built the industry's first "10,000-card" L4 compute cluster — nearly 15,000 accelerators in total — built to train its APEX multimodal foundation model.