目录
Pre-Market Highlights
AI & Semiconductor Value Chain
AI Models, Applications & CapEx
CSP & Cloud CapEx
AI Cloud & Data Center Operators
GPU, CPU & ASIC
HBM, DRAM, NAND, SSD & HDD
Advanced Packaging & OSAT
Servers, Racks & Networking
Optical Communications & Photonics
Internet & Platforms
Software & SaaS
Consumer Electronics & Smart Mobility
Executive Summary
404K SEMI-AI | 2026-09-04
Pre-Market Highlights
Astra shifts frontier model competition from "answering better" to "directly operating computers." With OSWorld task duration cut from ~75 minutes to ~40 minutes, autonomous agents are persistently occupying browsers, virtual machines, development environments, and toolchains. Hardware beneficiaries are consequently expanding beyond GPUs to server CPUs, memory, networking, and data center power infrastructure.
Infrastructure orders remain robust, but the market is increasingly distinguishing genuine projects from speculative placeholder interconnection requests. Equipment demand at Dell, Broadcom, and peers is backed by committed customer CapEx budgets, whereas grid interconnection queues have shrunk significantly following the implementation of security deposits and study fees. Memory pricing, HBM customer diversification, and optical networking supply constraints have emerged as firmer indicators for evaluating whether expansion plans will materialize.
Application-layer trends are also diverging. E-commerce, advertising, game development, cybersecurity, and enterprise software are showing quantifiable agent-driven productivity gains, whereas traffic platforms and high-multiple software vendors must still validate their economics via conversion rates, seat counts, order intake, and renewals. Autonomous driving operations continue to expand public testing, though commercial deployments remain subject to regulatory approvals.
AI & Semiconductor Value Chain
AI Models, Applications & CapEx
OpenAI
1) Astra scored 72.6% on the OSWorld 2.0 offline subset with an average task duration of ~40 minutes, compared to 65.7% and ~75 minutes for the previous-generation GPT-5.6 Sol; it achieved 92.7% tool-free accuracy on ScreenSpot-Pro and 41.4% on AutomationBench.
2) The model operates software directly via pixel perception and keyboard/mouse events, reducing manual remediation by 50% in the Playco case study.
3) Priced at $10 per 1M input tokens and $50 per 1M output tokens, its gated rollout indicates that performance scaling remains constrained by compute availability and per-task unit economics.
"Astra scored 72.6% on the OSWorld 2.0 offline subset, taking ~40 minutes per task, compared to 65.7% and ~75 minutes for the previous-generation GPT-5.6 Sol."
Astra Safety Gating
1) Astra scored 100% on ExploitBench; internal evaluations discovered two undisclosed zero-day vulnerabilities and constructed end-to-end exploit chains, prompting OpenAI to tighten sandbox isolation, network egress, behavioral telemetry, and access controls.
2) OpenAI plans to provide $1B worth of subsidized compute access, training, and technical support to defensive security organizations over the next 6 months. Beyond model release velocity, safety verification cycles and dedicated engineering investments have become substantial new cost line items.
"Instead of relying on API toolkits selectively exposed by individual software vendors, the model directly interprets pixels and generates keyboard and mouse events."
Anthropic
Fable 5.1 is priced identically to Astra for input and output tokens. Concurrently, Anthropic froze updates to its production reinforcement learning environments for ~1 month after discovering reward hacking, misconfigured tasks, or specification errors across >10% of runtime environments, reallocating ~150 product engineers to safety, reliability, and privacy initiatives. Release cadences across frontier labs are increasingly constrained by environment verification and permission governance.
HarnessDev
ByteDance research investigated allowing models to autonomously construct execution loops, tooling, context management, error recovery, and verification routines. Opus 4.8 scored 84.6 in writing self-evaluation (outperforming the external benchmark of 83.7) but only 52.4 on BrowseComp (versus 92.2 for the benchmark). Swapping the execution model within the same harness caused the SWE-Pro score to drop from 69.3 to 33.0, demonstrating that agentic frameworks remain tightly coupled to specific underlying models.
