Contents
Market Recap
AI and Semiconductor Supply Chain
AI Models, Applications and Capital Spending
Cloud Service Providers and Capital Spending
AI Cloud and Data-Center Operators
GPUs, CPUs and ASICs
HBM, DRAM, NAND, SSDs and HDDs
Foundry and Advanced Packaging
Semiconductor Equipment and Testing
Optical Communications and High-Speed Interconnects
Internet and Platforms
Software and SaaS
Consumer Electronics and Intelligent Vehicles
Huang’s Selected Portfolio
Memory and Storage Index Performance
Share of U.S. Equity Trading Value: Memory and Storage Stocks
AI investors are increasingly focused on how much work the same wafers, bandwidth and power can deliver. Low-stack HBM is entering customer testing, Google is emphasizing the payback efficiency of custom silicon, and optical interconnects are moving toward higher-power light sources. Orders continue to grow, but Oracle’s cash burn underscores the need to deliver both capacity expansion and investment returns.
404K SEMI-AI | September 14, 2026
Market Recap
The latest completed trading session was September 11. ETFs tracking the S&P; 500, Nasdaq 100 and Dow Jones rose 0.85%, 0.87% and 0.97%, respectively, while the equal-weight S&P; 500 ETF gained 0.80%. The rebound was broad, but all four proxies remained down over the past five sessions: the near-term recovery has yet to reverse the preceding drawdown.
Technology performance remained uneven. The semiconductor ETF rose 1.47%, versus just 0.32% for software. Hewlett Packard Enterprise and Dell gained 12.44% and 11.98%, respectively, while Oracle fell 1.74%. Memory and storage stocks also lagged: Micron declined 0.22%, SanDisk 3.50%, Seagate 3.73% and Western Digital 2.98%.
Weekend discussion centered on inference costs and investment payback. Low-stack HBM seeks to preserve bandwidth with less capacity, Google highlighted relatively short AI-server payback periods, and Oracle faces the cash-flow burden of building capacity before collecting revenue. Debate over model-release schedules must translate into orders, capital spending and actual usage before its implications for hardware demand can be assessed.
AI and Semiconductor Supply Chain
AI Models, Applications and Capital Spending
OpenAI
Financing plans and the pace of model development are shifting. Altman reportedly said the company would not go public in 2026 and discussed pausing at new capability thresholds to advance safety and alignment work. OpenAI also said it would follow through on embedded third-party evaluations.
There are further signs of improving inference efficiency. On the ARC-AGI-3 (AGI3) standard test suite, Astra reportedly performed better at its highest reasoning setting than at a low reasoning setting, at 46% lower cost. This finding applies only to that benchmark. Attributing it to a recurrent-depth architecture remains hypothetical, and the cost reduction cannot be generalized across workloads.
Anthropic
Expanding compute commitments make funding execution increasingly important. Signed cloud, chip and data-center agreements reportedly imply potential spending of up to $517 billion, mostly over the next decade. This covers a longer period and a broader range of projects than the previous $180 billion server-leasing estimate through 2029.
Since October 2025, the company has signed agreements for at least 14.8 GW of additional capacity, driven by businesses including Claude Code and Cowork. It has reportedly selected Nasdaq for a potential IPO targeting October, although final offering terms remain unconfirmed.
Amodei has proposed slowing advances in frontier capabilities and expanding third-party evaluations. Evaluation arrangements and compute contracts should be tracked separately; the proposal does not establish that contracted capacity has been canceled.
Cloud Service Providers and Capital Spending
Cloud capital spending: Morgan Stanley expects capex to grow approximately 12% in 2028, with faster growth in AI semiconductor spending and broadly stable spending on infrastructure such as buildings and AI data centers. The mix could therefore shift toward chips and packaging. The bank also forecasts 50% capacity growth in CoWoS, CoPoS and EMIB-T in 2028, alongside modest memory-price increases. The investment question is whether supply additions align with spending priorities; aggregate capex alone is insufficient, and packaging capacity growth should not be equated with equipment revenue growth.
