目录
Pre-Market Highlights
AI & Semiconductor Value Chain
AI Models, Applications & Capex
CSP & Cloud Capex
AI Cloud & Data Center Operators
GPUs, CPUs & ASICs
HBM, DRAM, NAND, SSD & HDD
Advanced Packaging, Interconnects, Thermal & Power
Robotics & Autonomous Driving
Internet & Platforms
Software & SaaS
Consumer Electronics / Smart Mobility
Overview
404K SEMI-AI | 2026-09-03
Pre-Market Highlights
Tonight's most significant marginal shift comes from custom silicon. Broadcom raised its FY2026 AI revenue guidance to $58 billion as six XPU customers accelerated in parallel. Gigawatt-scale deployments across OpenAI, Anthropic, Google, and Meta are transitioning from roadmap planning to generational product rollouts and delivery schedules, shifting the center of AI capex from standalone GPU procurement to full-rack systems.
Demand shows no evident signs of cooling, but execution friction is mounting. GPU utilization is increasingly dictated by KV cache capacity, memory bandwidth, and data movement, elevating the critical role of HBM, CXL, advanced packaging, and high-speed interconnects. Meanwhile, land, power availability, permitting approvals, fiber alignment, and memory manufacturing yields now determine whether backlogs translate into recognized revenue on schedule.
Platform and software ecosystems reveal a different divergence: while Azure continues to accelerate and Google Search and YouTube advertising sustain growth, low-quality AI-generated code, ungoverned production permissions, and inconsistent compute pricing metrics are driving up operational costs. Raw compute capacity is no longer sufficient to evaluate capital deployment quality—the key metrics going forward are throughput per watt, memory availability, deployment execution, and ROIC.
AI & Semiconductor Value Chain
AI Models, Applications & Capex
OpenAI
1) Reports indicate OpenAI entered early 2025 with ~2 GW of compute capacity and is projected to surpass 5 GW by year-end. Together, OpenAI and Anthropic account for ~30% of total incremental compute additions in 2025 and could absorb 40%–50% of global incremental capacity by 2027.
2) Broadcom confirmed that Jalapeño has begun shipping. OpenAI plans to deploy 1.3 GW of Jalapeño in 2027 and over 5 GW in 2028. The custom XPU delivers an estimated ~50% cost reduction versus standard GPU solutions; critical milestones ahead center on land acquisition, power access, and delivery timelines.
"So we're seeing this happen. It's not speculative. It is happening, and we are in the middle of helping make this happen."
Anthropic
1) Following its 1 GW Ironwood deployment in 2026, Anthropic is projected to deploy an additional 5 GW of TPU V8I in 2027, potentially becoming Broadcom's largest XPU customer. Its blended contract rate across Lambda, Hut 8, and Nscale sits at ~$16 billion/GW.
2) Internal projections are even more aggressive: value generated by frontier labs is estimated to grow 4x annually (doubling every 6 months), with top-tier scientist-level models expected in key domains within 12 months—tightly linking compute demand to committed model capabilities.
"The 2019 scaling laws paper convinced him that what was once a 50-year timeline had compressed to under 10 years. That trend spanned 8 orders of magnitude at the time, and has now crossed 15 orders of magnitude."
Moonshot AI
Moonshot AI has reportedly filed confidentially for a Hong Kong IPO and is raising capital around a $50 billion pre-money valuation. Surging demand following the K3 launch has strained capacity, pushing June ARR to ~$300 million. In July, the company closed a $3.5 billion funding round at an estimated ~$35 billion valuation. Sustaining this upward valuation trajectory will require confirmation from subsequent revenue run-rates and compute cost management; the formal prospectus remains unavailable.
Inference Workloads
System bottlenecks in AI are shifting from pure compute capacity to data movement. Long context windows, chain-of-thought reasoning, and agentic workflows continuously expand KV cache demands—exceeding ~140 GB in certain environments, on par with the weight footprint of a 70B-parameter model. In a single query, over 50% of GPU cycles are consumed by data transfer, making utilization entirely contingent on timely data delivery.
"In the AI era, compute itself is not the core bottleneck. The core bottleneck is where data resides and how it moves through the system."
Software Optimization
While FlashAttention, PagedAttention, and speculative decoding continue to curb VRAM footprints, software-level efficiency gains are nearing their theoretical limits. Google's TurboQuant (presented at ICLR 2026) compresses 16-bit KV cache down to ~3–4 bits, reducing memory overhead to ~1/5; however, algorithmic compression cannot eliminate the physical interconnect latency between processors and remote memory.
