Compute Labs logo

ComputeLabs Research

AI & Compute Infrastructure — September 19, 2026

Edition of · 15 stories

OpenAI projects $856 billion in compute and infrastructure spending through 2030, alongside $278 billion in negative free cash flow during 2026–2030.

The figures are company projections reported from a presentation, not completed spending or secured financing. Financial_Express, citing presentation materials reviewed by the Financial Times, reports that OpenAI expects approximately $856 billion in compute and infrastructure expenditure through the end of 2030. The presentation identifies this as the company’s largest individual spending category.

Read the full story · 1 source →

Guangxi (广西) published a financial-development plan targeting an AI investment-fund cluster totaling at least RMB 50 billion.

The fund target forms part of Guangxi’s broader financial-sector development policy. The regional government’s General Office issued the Guangxi Financial Industry Development 15th Five-Year Plan (《广西金融业发展“十五五”规划》), with broader financial-development objectives extending to 2030. Its AI provisions call for accelerating the creation of a fund cluster totaling at least RMB 50 billion, rather than reporting that this amount has already been raised or invested.

Read the full story · 4 sources →

California Governor Gavin Newsom signed an executive order commissioning AI-safety recommendations within two months; audits and shutdown mechanisms remain proposals.

The executive order establishes a process for developing safety recommendations. The supplied report says Newsom signed the order on Friday and directed an expert group to produce guidance within two months on strengthening California’s AI-safety laws. Newsom described the action as accelerating safety measures and argued that federal oversight was insufficient.

Read the full story · 1 source →

OpenAI, Anthropic, and Google face a federal antitrust lawsuit alleging collusion over calls to coordinate slower AI development.

The reported case is a civil lawsuit, not a finding of antitrust liability. Telegram reports citing Politico say the complaint was filed on Friday in the U.S. District Court for the Northern District of California, on behalf of four plaintiffs. In addition to the three companies in the digest, the reports name SpaceXAI, using that designation for the defendant associated with Elon Musk.

Read the full story · 3 sources →

University of Michigan researchers introduced Fengshui, a chiplet co-design framework jointly optimizing chiplet pools and application-specific AI accelerator designs.

Fengshui addresses two linked levels of accelerator design. The University of Michigan paper introduces a framework that jointly optimizes the composition of a chiplet pool and the design of bespoke application-specific integrated circuits, abbreviated BASIC in the paper. The framework therefore covers both the available chiplet ecosystem and the accelerator designs built from it.

Read the full story · 1 source →

Researchers introduced IC-ThermBench, an open thermal-model benchmark combining 3D integrated-circuit tasks with a 50,000-sample 2.5D chiplet dataset.

IC-ThermBench is a research benchmark for thermal learning across integrated-circuit configurations. The paper comes from researchers at the University of Technology Sydney, ShanghaiTech University, and Technical University of Munich. Its title describes an open, progressive benchmark for generalizable 2.5D/3D integrated-circuit (IC) thermal learning.

Read the full story · 1 source →

U.S. Energy Secretary Chris Wright ordered continued operation of two Indiana coal-fired power plants, directing utilities and the grid operator to act.

The reported order concerns continued operation of existing coal generation. Financial_Express states that Energy Secretary Chris Wright ordered two Indiana coal-fired power plants to remain operational. The messages identify NIPSCO, CenterPoint Energy, and the Midcontinent Independent System Operator (MISO) as parties directed to act.

Read the full story · 3 sources →

Jintan (金坛) Phase II Unit 2 completed whole-unit startup; the project plans two 350-megawatt compressed-air energy-storage units.

The milestone applies to Unit 2 of the Phase II project. The salt-cavern compressed-air energy-storage project is located in Jintan District, Changzhou, Jiangsu (江苏省常州市金坛区), and POWERCHINA (中国电建) participated in its construction. The reported achievement is completion of whole-unit startup, rather than an explicit announcement that the entire project has entered commercial operation.

Read the full story · 1 source →

Envision’s Xinghe AI data center (远景星河基地) in Ulanqab (乌兰察布) began operations in August, spanning 120,000 square meters.

The reported operating milestone occurred in August 2026. A Financial_Express message attributed to CCTV Finance (央视财经) describes Envision’s Xinghe facility in Ulanqab as occupying 120,000 square meters of building area, approximately 20 standard football fields in the report’s comparison. The area figure is not a measure of electrical or computing capacity.

Read the full story · 1 source →

More than 40 language and multimodal models completed pretraining on Huawei’s Ascend (昇腾) compute platform, which supports PyTorch.

The pretraining count comes from a Huawei executive’s statement. On September 19, Zhu Zhaosheng (朱照生), head of Huawei’s computing strategy and business development department, said more than 40 large language and multimodal models had completed pretraining using Ascend. The figure counts models, not accelerators, servers, or clusters.

Read the full story · 2 sources →

Qwen (千问) launched Qwen3.8-LiveTranslate, a simultaneous-translation model using an Interleave architecture, reducing average per-character latency from 2.8 to 2.3 seconds.

The launch targets real-time simultaneous translation. A Telegram message attributed to the Qwen model team’s WeChat account reports the release of Qwen3.8-LiveTranslate on September 19. The announcement identifies Interleave as the architecture used to restructure its real-time translation approach.

Read the full story · 1 source →

Claude Code 2.1.277 now reads AGENTS.md automatically when a project folder lacks CLAUDE.md, adding support for alternative project instructions.

The update introduces a conditional fallback for project instructions. Beginning with Claude Code version 2.1.277, Claude automatically looks for and reads AGENTS.md when the project folder does not contain CLAUDE.md. The supplied report does not describe this as replacing CLAUDE.md or merging both files when both exist.

Read the full story · 1 source →

Purdue University researchers introduced DRC-Aid, an agentic framework automating local chip-layout design-rule repairs through verification-in-the-loop search while preserving layout equivalence.

DRC-Aid applies an agentic approach to local design-rule correction. Purdue University researchers describe it as a closed-loop framework that formulates design-rule checking (DRC) repair as verification-in-the-loop search. The paper is titled “DRC-Aid: Design-Rule Correction via Agentic Framework utilizing Inference-Time Large Language Models.”

Read the full story · 1 source →

All ComputeLabs Research editions