目录
Pre-Market Takeaways
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
CSP/Cloud Capital Expenditure
AI Cloud/Data-Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Advanced Packaging/Semiconductor Equipment/High-Speed Interconnects
Power Semiconductors/MLCC/Data-Center Power
Robotics/Autonomous Driving
Internet / Platforms
Software / SaaS
Consumer Electronics / Smart Vehicles
Overview
404K | 2026-08-27
Pre-Market Takeaways
Nvidia has again raised the ceiling for AI infrastructure demand: FY2028 revenue is expected to grow by approximately 70%, while customer forecasts imply even stronger demand growth. Supply—not demand—is the binding constraint on deliveries. The market’s focus is therefore shifting from whether demand will slow to who can provide sufficient memory, packaging, networking, power, and data-center capacity.
Growth still comes at a cost. Higher memory prices and the Rubin ramp will push Nvidia’s gross margin down before it stabilizes, while emerging-cloud financing structures introduce credit risk into the supply chain. Revenue visibility has improved, but execution depends on supply expansion, project energization, customer utilization, and whether price increases can offset higher costs.
Software and end devices are not showing comparable incremental momentum. Salesforce offered counterevidence through improving seat growth, renewals, and Agent usage, while Apple and Tesla are advancing in advanced packaging and Robotaxi operations, respectively. Still, the clearest flows of capital and orders remain concentrated in the hardware layer of AI data centers. The next checkpoints are whether supply can ramp, projects can be energized, and customers can pay as planned.
Full AI/Semiconductor Value Chain
Demand expansion has evolved from purchasing individual GPUs to securing resources for entire AI factories. Chips remain the core revenue pool, but delivery ceilings are increasingly determined jointly by HBM, substrates, packaging reliability, power, and financing. Nvidia’s revenue guidance and long-term commitments have intensified this transmission chain across the board. The key risks are that higher costs, construction delays, and customer credit issues erode strong demand at the delivery stage. The order of validation for the investment chain has also changed: first, whether memory and packaging capacity can expand on schedule; second, whether data centers can be energized; and finally, whether customers renting compute can sustain utilization and payments.
AI Models/Applications and Capital Expenditure
OpenAI
1) The Jalapeño inference chip uses 216GiB of HBM4, provides 15.4TB/s of bandwidth, and has a rated package power of 700W, with tested consumption of approximately 550W or below. Initial deployment is scheduled to begin by the end of 2026, followed by a ramp in 2027.
2) Nvidia disclosed that OpenAI’s existing and planned demand equates to approximately 12GW of compute capacity, indicating that model developers are securing long-term infrastructure in advance. The risk is that Jalapeño has not yet undergone a complete controlled comparison with Rubin using the same model, operating point, benchmark, and inference mode.
“OpenAI has not presented controlled Jalapeño-versus-Rubin results using the same model, benchmark, inference mode, and operating point.”
Anthropic: Anthropic has reportedly signed a 6-year, $45 billion compute-leasing agreement with Nscale, covering 460MW of power capacity at a West Virginia campus. The deployment is expected to use Vera Rubin, with capacity scheduled to begin coming online by the end of 2027. The order further translates model competition into requirements for power, chips, and timely construction; the launch remains contingent on construction and energization.
CSP/Cloud Capital Expenditure
Amazon: AWS plans to deploy an additional 2 million Nvidia GPUs across its global infrastructure by Q2 FY2029 and introduce Vera CPU. The partnership also covers NVHBM, NVLink Fusion, Nemotron, Bedrock, SageMaker, and warehouse-robotics technology. The companies separately plan to build a secure AI data center equipped with 100,000 GPUs. AWS’s in-house Trainium has not reduced external GPU purchases; the next variables to watch are the actual deployment pace, Vera CPU adoption, and project delivery.
“Nvidia and AWS are expanding their full-stack partnership—across GPUs, CPUs, networking, open-source models, and software—to make Agent and physical AI a reality at a speed and scale only we can achieve together.”
Google: TPU v10 is moving to a multi-vendor architecture. MediaTek will retain responsibility for ASIC integration, SerDes, I/O Die, and HBM packaging, while Marvell Technology and AMD may participate in selected modules. Supply-chain reports indicate that the compute die will adopt a known-good-die model, packaging will move toward 3.5D, and the mask count will increase from approximately 9 to more than 12. MediaTek’s content value is expected to rise to approximately $18,000 per unit, as greater complexity expands the value of design services.
AI Cloud/Data-Center Operators
TeraWulf: Kentucky has approved up to 482MW of power for the Justified campus, removing a regulatory obstacle to a 401MW, approximately $19 billion Anthropic lease. The company must pay for 482MW for 6 years even if it consumes less power. The first campus and data halls are expected to require $4.0 billion to $4.5 billion of investment, equivalent to approximately $10 million to $12 million per critical IT MW. Project visibility has improved, but financing, construction, and energization remain incomplete.
CoreWeave, Nebius, and IREN: Nvidia expects its emerging-cloud partners’ installed capacity to increase from approximately 3GW at the end of 2025 to 8GW at the end of 2026. It is helping finance part of that capacity through take-or-pay commitments, then sharing rental revenue above the guaranteed floor. ACIE—which includes emerging-cloud, enterprise, industrial, and sovereign businesses—generated approximately $40 billion in quarterly revenue, versus approximately $49 billion from hyperscalers. The model can drive both rack sales and recurring revenue, but utilization, credit quality, and the ability to redeploy equipment determine the risk.
GPU/CPU/ASIC
Nvidia
1) Q2 FY2027 revenue was $96.2 billion, up 106% year over year; data-center revenue was $89.0 billion, up 117%. Q3 revenue guidance is $108.0 billion, plus or minus 2%.
2) Vera Rubin has begun shipping and is expected to contribute approximately 20% of Q3 data-center revenue. The revenue opportunity per GW increases from $18 billion for Hopper to $40 billion for Rubin.
3) Supply and capacity commitments increased from $119 billion to $279 billion, with memory accounting for most of the increase. Gross margin is expected to bottom at 71% to 72% in Q4.
4) Melius raised its price target from $400 to $420 and maintained its Buy rating. The debate centers on whether guidance for growth above 70% can absorb pressure from memory costs.
“Unconstrained demand would have been much higher. We achieved 100% year-over-year growth this year. The scale of unconstrained demand is enormous.”
“We expect demand from the AI labs that we intend to support with our own balance sheet to account for approximately one-quarter of our business next year.”
MediaTek: TPU v10’s multi-vendor architecture has not diminished MediaTek’s core integration role. The company uses its proprietary SerDes IP to design the I/O Die and integrates the compute die, I/O Die, and HBM. If EMIB-T remains in use, content value per chip will rise as the I/O Die expands and packaging complexity increases. The key items to validate are mass-production yields, MediaTek’s share in subsequent Google generations, and the actual scope of participation by other design-service providers.
HBM/DRAM/NAND/SSD/HDD
The memory chain should be viewed across two horizons. In the near term, Nvidia is expanding supply commitments, Samsung Electronics and Japanese memory manufacturers are advancing capacity expansion, and customers are beginning to secure scarce resources early. Longer term, HBM bandwidth is still failing to keep pace with compute growth, while hybrid bonding remains constrained by yield, thermal management, and throughput. Higher prices can improve memory manufacturers’ cash flow, but they will initially pressure accelerator gross margins. The key variables are HBM yields, enterprise SSD mix, capacity-ramp execution, and fulfillment of long-term agreements. Any shortfall could erode strong demand at the delivery stage.
HBM Value Chain: The common signal from Hot Chips is that AI accelerator compute performance is increasing approximately 3x every 2 years, while HBM bandwidth is growing by less than 2x. HBM4 doubles the interface from 1024 I/O to 2048 I/O, and Micron demonstrated more than 2.8TB/s per stack, but bandwidth still does not fully keep pace with compute expansion. The long-term direction toward hybrid bonding is becoming clearer, but relaxed package-height specifications could delay mass adoption until HBM5 or 20-high stacks. At 16-high and 20-high, thermal management also shifts from a packaging issue to an architectural one.
