Contents
Premarket Takeaways
AI and Semiconductor Value Chain
AI Models, Applications and Capital Spending
Cloud Providers and Infrastructure Spending
GPUs, CPUs and ASICs
HBM, DRAM, NAND, SSDs and HDDs
Foundries
Semiconductor Equipment and Testing
Optical Communications and Components
Servers, PCBs and Substrates
Internet and Platforms
Software and SaaS
Consumer Electronics and Smart Vehicles
Always-on agents are bringing server CPUs back into focus, while AI hardware demand extends into system memory, enterprise SSDs and optical interconnects. Supply remains constrained, but revenue conversion requires separate scrutiny: Micron’s earnings will test pricing and margins, cloud providers’ custom chips will test packaging capacity, and software providers must show that usage can translate into recurring revenue.
404K SEMI-AI | September 28, 2026
Premarket Takeaways
AI demand is spreading across the hardware stack. Muse’s always-on architecture has renewed attention on server CPUs, with lead times reaching 25–30 weeks versus 16–20 weeks in a balanced market. Memory, advanced packaging and optical components also remain constrained. Assessing demand requires tracking both continued customer investment and suppliers’ ability to deliver equipment and bring systems online.
The next checkpoints are product launches and earnings. OpenAI’s September 29 developer conference will focus attention on models, pricing and agent capabilities; Micron’s September 30 results will test memory pricing, customer agreements and HBM4 progress. Applications face a countervailing pressure: enterprises are seeking cheaper models, while personal agents help users cancel subscriptions. Rising usage does not guarantee revenue growth for every platform.
AI and Semiconductor Value Chain
AI Models, Applications and Capital Spending
OpenAI: The developer conference will test the next stage of agent competition. Key items for September 29 include new models, pricing and API changes, and the ability to complete multistep tasks autonomously. Presence previously targeted customer support, sales and billing—workflows that require sustained execution and tool use. For the supply chain, the key questions are whether workloads shift from brief exchanges to continuous operation, and whether pricing covers the resulting resource consumption. Product previews still await confirmation at launch.
Anthropic:
General-purpose model upgrades are changing application development trade-offs. The Claude Code team believes that the next model generation could match the roughly 10%–20% performance gains delivered by application frameworks. Specialized models and fine-tuning still have a role where specific use cases require them.
Opus 5.5 has reportedly been used to check nonrecurring items in financial models, cross-referencing filings, earnings calls and forecast spreadsheets. Such examples suggest that models are beginning to handle detailed professional work, but the path from individual use cases to enterprise adoption at scale remains unproven.
Cloud Providers and Infrastructure Spending
Google:
TPU expansion is increasingly constrained by packaging choices. UBS estimates that Intel’s EMIB-T supply could support more than 2 million TPU v9 chips in 2028. Substrates could become a bottleneck if volumes reach 4 million or more in 2029.
UBS believes MediaTek may be exploring additional CoWoS options with TSMC, although discussions remain preliminary. Supply flexibility depends separately on chip demand, substrate availability and qualification of alternative packaging. Exploratory discussions should not be treated as confirmed volume-production orders.
“UBS therefore believes MediaTek may be exploring additional CoWoS options with TSMC. Discussions remain at an early stage, and CoWoS could eventually be introduced alongside EMIB-T.”
Meta:
Muse’s adoption is accompanied by more concrete infrastructure expansion. Meta says it is expanding compute infrastructure on a 5 GW scale; Muse attracted millions of users within two weeks of launch. Its use of always-on cloud virtual machines also brings CPU demand into focus.
Meta plans to integrate Muse into its Ray-Ban smart glasses and let users customize the wake word. The direction of integration across hardware, models and personal assistants is clear, but the incremental benefits for CPUs and GPUs will depend on the tasks users actually run.
Cloud infrastructure supply and demand: Capital spending continues to rise, while delivery faces multiple constraints. Goldman Sachs expects major US hyperscalers’ AI infrastructure spending to reach $1.2 trillion in 2027, up more than 50%. HSBC sees shortages across power, data center capacity, substrates, foundry capacity, advanced packaging and optical components, with little meaningful relief through at least 2027. Investment supports supply-chain demand, but delivery schedules will constrain realized revenue growth.
