目录
Post-Close Summary
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
CSPs and AI Cloud
GPU/CPU/ASIC and High-Speed Interconnects
HBM/DRAM/NAND/SSD/HDD
Foundry, Equipment, and Optical Communications
Space/Satellites
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Jensen Huang's Selected Portfolio
Overview
404K | 2026-08-07
Post-Close Summary
As of the latest close on August 5, the S&P; 500 fell 0.20%, the Nasdaq 100 declined 0.90%, and the Dow gained 0.44%. The technology sector fell 0.53%, communication services declined 1.04%, and the semiconductor ETF lost 1.04%. The broader market showed no consistent direction, while performance among heavyweight technology stocks remained sharply divided.
Medium-term relative strength diverged from single-day performance. The cloud computing and software ETFs gained 10.54% and 9.55%, respectively, over 20 days, with RPS20 ranking among the leading themes; the semiconductor ETF fell -3.93% over 20 days but rebounded 12.99% over 5 days. Capital remains within the AI theme but increasingly favors companies that can deliver orders, cash flow, and margins.
The latest developments continue to reinforce demand for AI capital expenditure, GPU leasing, memory, and optical interconnects, but post-earnings price reactions have become more demanding. Even if software companies accelerate growth, they must demonstrate guidance delivery and paid conversion; even if memory companies secure long-term agreements, they must address whether pricing trajectories, inventories, and gross margins can hold.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI
1) GPT-5.6 Sol has become the default model supporting instant responses and deep reasoning for Plus and Pro users, while Free and Go users can use GPT-5.6 Luna text chat without limits. Product tiering is expanding free access to highly capable models.
2) Southern Company said OpenAI’s contract in Georgia has created demand approximately 1 GW above recently approved capacity. Model competition is translating directly into pressure on power supply and data-center construction.
"The CXL technology stack has moved from the standards and demonstration phase into customer validation, with larger-scale deployment more likely to emerge in 2027 and continue into 2028."
Microsoft: Following the launch of its Hyderabad data center, Microsoft now operates 4 cloud regions in India, with Adani Group and HDFC Bank among the first customers. The company has committed approximately US$20.5 billion to expand its India operations. The value of the additional local capacity must be validated by Azure customer growth and AI-workload utilization, while the pace of customer migration to the new region remains another factor to monitor.
Google
1) The company plans to raise up to US$25 billion through the US bond market, across as many as 10 maturity tranches ranging from 2 to 40 years; investor demand totaled approximately US$115 billion, about 4.6 times the maximum proposed issuance.
2) The financing follows another increase in the company’s 2026 capital-expenditure outlook. Bond demand was strong, but pressure on free cash flow has also made returns on invested capital more important.
Meta: A new multi-tier sequential learning architecture has been deployed in the user-representation layer of advertising recommendations, seeking both scaling behavior similar to large language models and the low latency required for ad ranking. The validation criteria are straightforward: whether model investment can improve advertising conversion and revenue per unit of compute while preserving response speed and inference costs; the pace of production migration will determine the speed of scaling.
CSPs and AI Cloud
Nscale: The company plans to list in the US in September and has disclosed US$51 billion of contracted revenue; second-quarter revenue exceeded US$100 million, versus US$37 million in the first quarter and US$33 million for full-year 2025. It has approximately 25,000 active GPUs and 289,000 GPUs either active or under contract, and plans to add 10 GW of power capacity. Execution risks center on converting contracts into revenue, GPU delivery, and bringing power capacity online.
Oracle: The company continues to expand its Port Washington AI campus, with physical construction and cloud-capacity investment still progressing. UBS maintained its Buy rating while cutting its price target from US$285 to US$245, arguing that concerns over delays to some OCI projects may be overstated; the key issues to track are whether project launches and lease commitments convert into revenue in tandem, and whether GPU utilization can ramp.
Nebius: The 300 MW expansion of the Vineland data center continues to face community concerns over water and power consumption, noise, and air quality, and a final decision on the project has not yet been made. The shares fell 2.99% over 5 days but were still up 47.75% over that period; behind this high sensitivity, approval progress, financing commitments, and the construction timetable are the clearest potential disconfirmation points.
IREN: The Sweetwater 1 site was energized in May 2026, and reports indicate that 1,400 MW of capacity is ready, with the company potentially expanding its partnership with NVIDIA. Power is a scarce asset, but MW capacity will convert into cash flow only if customer contracts, equipment delivery, and utilization all materialize. Formal contract confirmation is still pending at this stage.
CoreWeave: New cloud providers generate approximately US$9.4 million to US$10.4 million of annualized revenue per energized MW, above the approximately US$3.5 million to US$4.4 million generated by traditional colocation data centers. The premium comes from GPU leasing rather than simply renting data-center space. The next questions are whether lease pricing, utilization, and hardware supply can be sustained, as well as the risk that additional supply depresses rental rates.
"The market model is based on a 2030 CPU market of US$200 billion to US$225 billion, with an estimated midpoint of US$215 billion"
GPU/CPU/ASIC and High-Speed Interconnects
NVIDIA
1) The company is reportedly testing a Rubin Ultra GPU with a reduced HBM configuration, planned for launch in 2027, to lower costs and alleviate potential HBM4E supply constraints.
2) Market-research data indicate that NVIDIA holds a 92.4% share of the sovereign AI-chip market, although it still needs to be verified whether the figure is calculated by revenue or unit shipments. Whether the lower-memory configuration can be offset by faster networking and server optimization is central to next-generation system efficiency.
Astera Labs
1) Second-quarter revenue was US$392.4 million, up 104% year over year and 27% quarter over quarter; PCIe 6 products accounted for more than 50% of revenue for the first time, and third-quarter revenue guidance was US$540 million to US$560 million.
2) Scorpio X contributes approximately US$1,000 of value per XPU, with 10 customers entering pilot production and qualification. Growth is shifting from retimers toward rack-scale switching, but customer qualification, the UALink roadmap, and the Trainium 4 architecture will continue to determine sustainability.
"Astera Labs’ most important near-term growth initiative is expanding from within-tray connectivity into the rack-scale market."
CXL Ecosystem: Larger-scale deployment is more likely to emerge in 2027 and continue into 2028. The base case corresponds to 7.5 million CXL-enabled AI CPUs, approximately 9.4 million endpoints, and a core-chip and connectivity market of approximately US$4.5 billion; the first focus is single-host memory expansion and DDR4 reuse, followed by whether rack-scale shared pools can generate repeat orders.
"DDR4 reuse is the strongest driver of early CXL adoption by hyperscale cloud providers"
Credo: The company operates at the CXL physical-link and high-speed interconnect layer, rather than the memory-controller layer. Rack-scale accelerators currently account for only approximately 9% to 11% of the installed AI-accelerator base, so its opportunity depends on standardized rack-scale deployment and connectivity value per endpoint, rather than solely on growth in total server volumes. Repeat customer orders would be a stronger signal.
"The key tension is that near-term performance depends on copper interconnects and PCIe, while longer-term rack networking may shift toward UALink, NVLink Fusion, and optical interconnects."
HBM/DRAM/NAND/SSD/HDD
SanDisk
1) Data-center revenue increased 15-fold over 5 quarters, with QLC NAND becoming the capacity tier for AI inference; approximately US$94 billion of contracted revenue provides more than 4 years of visibility.
2) Morgan Stanley said long-term agreements cover approximately 50% of 2027 supply and approximately 2/3 of 2028 supply, with a gross-margin floor of approximately 80%.
3) Jefferies maintained its Buy rating but cut its price target from US$3,000 to US$1,750, citing slowing price increases, lower gross-margin guidance, and inventory accumulation.
"The report’s central thesis progresses from industry-scale expansion to networked, recurring demand, before mapping the opportunity across an eight-layer industry framework."
Micron: Management previously said the company could meet only 50% to 2/3 of key-customer demand. Goldman Sachs’ estimated 6% supply shortfall is not measured on the same basis as the gap in satisfying customer demand: the former compares bits produced with bits consumed, while the latter reflects volumes customers want to purchase but cannot obtain. Pricing power will still depend on HBM yields, the pace of capacity expansion, and changes in customer specifications, rather than the headline shortage figure alone.
Samsung Electronics: SUSS temporary bonding and debonding tools are being used in Samsung Electronics’ HBM stacking process, indicating that back-end assembly equipment remains an important constraint on HBM capacity expansion. Samsung Electronics’ foldable-device preorders increased 30%, also providing the memory and consumer-electronics businesses with demand from different cycles. The next factors to monitor are HBM qualification and initial-sales conversion for foldable devices; the two business lines should not be treated as part of the same demand cycle.
SK hynix: Discussion of HBM supply and demand continues to point toward specification upgrades and insufficient supply. If thermal requirements, packaging, and stack height increase further, HBM content per GPU may have additional upside; conversely, NVIDIA’s testing of a lower-memory version also indicates that customers are seeking alternative configurations amid cost and supply constraints. Supply and demand remain in mutual adjustment, and specification changes will affect unit value.
Seagate Technology: Its 44 TB HAMR product has entered volume deployment at 2 hyperscale cloud providers, approximately 2 to 3 quarters ahead of Western Digital. Whether this lead translates into higher enterprise-drive shipments and faster declines in cost per TB is the near-term dividing line in HDD competition; customer capacity expansion, product yields, average capacity, and the cost curve still require ongoing validation, rather than relying solely on technology-launch timing.
Western Digital: The company plans to advance its corresponding HAMR product by the first half of 2027, and its current lag has increased market attention on technology-transition risk. The shares fell 5.36% over 5 days and 5.66% over 20 days; with data-center demand still strong, product qualification, mass-production timing, and cost per TB matter more than aggregate industry demand, while changes in customer share must also be monitored.
