2026-08-25基础设施OpenAI's first chip has benchmarks now, and OpenAI ran themAt Hot Chips on Tuesday, OpenAI published the first measured results for Jalapeño, the inference accelerator it designed with Broadcom. The claim is more throughput per kilowatt and lower per-user latency at the same time, against an NVIDIA Blackwell system. Volume deployment is a 2027 story.阅读约5分钟→
2026-08-24基础设施NVIDIA ships a decode chip with a competitor's name on itGroq 3 LPX enters full production at Hot Chips, splitting inference across two kinds of silicon. The efficiency figures are NVIDIA's own and still pending review.阅读约6分钟→
2026-08-23其他下周前瞻The week ahead: Hot Chips talks first, NVIDIA reports secondHot Chips 2026 runs Sunday to Tuesday at Stanford with NVIDIA, AMD, Intel, Google, Microsoft, Meta and OpenAI all on the accelerator program. NVIDIA and Salesforce report on Wednesday, Marvell on Thursday. No frontier lab has a dated announcement inside the window.阅读约4分钟→
2026-08-22每日摘要一周回顾The week the money moved and nobody shipped a modelBetween 16 and 22 August, NVIDIA put up to $105 billion of its own credit behind OpenAI's Ohio leases, Etched reached a $21 billion valuation, Alibaba disclosed a $2 billion quarterly loss on its model business, and OpenAI cut a frontier price with an expiry date attached. No lab announced a frontier model.阅读约6分钟→
2026-08-21开发者工具OpenAI finally discounts GPT-5.6 Sol, and puts an expiry date on itSol's standard rate drops to $4 per million input tokens and $20 per million output tokens. OpenAI's changelog calls the price promotional and good at least through 21 November.阅读约4分钟→
2026-08-20基础设施Alibaba's new AI segment shows the model lab losing $2 billion a quarterA reporting change splits the compute business from the one that builds and serves the models. Compute grew 45% and more than doubled its profit. Labs and apps lost RMB13.9 billion on RMB3.3 billion of revenue, which Alibaba attributes partly to Qwen app inference costs.阅读约5分钟→
2026-08-19政策OpenAI previews a way to police API misuse without reading the promptsPrivate Safety Processing looks for abuse patterns across a customer's sessions and emits a signal instead of the text. It arrives while Anthropic is doing the opposite for its frontier models.阅读约5分钟→
2026-08-18安全OpenAI says its largest planned frontier training run is still on holdThe company published its post-breach security overhaul on 18 August: a two-week pause on reinforcement-learning training, rebuilt sandboxes, monitoring that costs about a fifth of the compute it watches, and one run that has not restarted.阅读约5分钟→
2026-08-17基础设施NVIDIA guarantees up to $105 billion of OpenAI's Ohio data center leasesThe chip maker filed an 8-K disclosing residual value guaranties on roughly 4.25 gigawatts of OpenAI's leases at a campus in Pike County, Ohio, with its payment obligations capped at $105 billion. The clause that best explains the deal is the one that cancels it.阅读约6分钟→
2026-08-16其他下周前瞻The week ahead: Imagen 4 switches off, and China's two AI platforms reportBetween 16 and 22 August the confirmed AI calendar holds three things: a Google image-model shutdown on Monday, Baidu's quarter on Tuesday and Alibaba's on Thursday. The heavier week starts the day after it ends.阅读约4分钟→
2026-08-15每日摘要一周回顾The week of the point release, and a $500 billion bet on the compute underneathBetween 9 and 15 August the labs competed on price, latency and how long a model keeps working — not on capability. The largest number of the week was NVIDIA's, and it was a financing announcement.阅读约5分钟→
2026-08-14产品Anthropic explains how Claude's text watermark works, and where it barely registersThe mechanism is keyed word selection. The interesting part is Anthropic's own list of cases where it produces almost no signal, including code.阅读约5分钟→