目录
Post-Market Summary
AI/Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
AI Cloud and Data-Center Operators
GPUs, CPUs, ASICs, Foundry, and Advanced Packaging
HBM/DRAM/NAND/SSD/HDD
Optical Communications, High-Speed Interconnects, and Data-Center Infrastructure
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Jensen Huang’s Selected Portfolio
Memory Index Performance
Memory Stocks’ Share of US Equity Turnover
Overview
404K SEMI-AI | 2026-08-12
Post-Market Summary
The latest completed U.S. trading session, on August 10, extended the divergence between strong software and weak semiconductors. The S&P; 500 ETF edged down 0.03% and the Nasdaq 100 ETF fell 0.30%; cloud computing, software, and cybersecurity thematic ETFs rose 2.95%, 2.26%, and 2.81%, respectively, while semiconductor ETFs declined 2.28% to 2.55%.
Capital is shifting its focus from individual chips to entire AI factories. Nvidia has disclosed third-party financing platforms exceeding $500 billion, while neoclouds such as CoreWeave and Nebius are competing on orders, power, and capital spending to accelerate expansion. The critical constraints now extend beyond chip availability to energization, construction, acceptance testing, and customer utilization.
Execution is also diverging more sharply at the company level: CoreWeave and Lumentum continue to deliver strong revenue growth, while TSMC’s advanced-packaging yields keep improving. Memory vendors are being repriced based on long-term contract coverage, expansion cadence, and free cash flow. Consumer adoption remains gradual, making smartphone financing, on-device model memory requirements, and automotive design wins more important to track than conceptual product announcements.
AI/Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
Nvidia
1) The company has established third-party financing platforms exceeding $500 billion with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR, and others to fund AI-factory construction. Separating the financing vehicles from chip sales should help ease concerns over circular transactions.
2) Nemotron 3.5 uses a 30-billion-parameter mixture-of-experts architecture but activates only 3 billion parameters per query. On Jetson Thor, it generates 115 tokens per second, versus 89 on Orin.
“Using primarily third-party capital to fund AI-factory investment should help alleviate concerns over circularity, support investment by sovereign nations and neocloud providers, and create annuity-like recurring revenue streams.”
AI inference pricing: Inference prices have fallen to roughly half their May peak, while Goldman Sachs expects the number of tokens processed to rise to 24 times the current level by 2030. Lower prices support higher application usage but require providers to keep improving utilization. Focusing only on token growth while ignoring unit pricing and inference costs risks overstating the revenue trajectory.
Economics of 1GW of compute: Stress testing indicates that a 1GW GB300 cluster would comprise roughly 7,000 racks and 500,000 GPUs, producing 1.46 billion tokens per second. At 60% utilization and $1.73 per million tokens, theoretical annual revenue would be approximately $48 billion, or roughly $35 billion after a 72% discount. The model shows that converting compute into cash flow ultimately depends on utilization, pricing, and the residual value of older GPUs.
Billable capacity: The near-term bottleneck in AI infrastructure is shifting from chip availability to whether facilities can be energized on schedule and pass acceptance testing. Grid connections, construction labor, rack commissioning, and customer certification jointly determine when capacity becomes billable. Even after accelerators arrive, capital expenditure cannot generate revenue until the supporting infrastructure is complete. This is a prerequisite for revenue recognition.
“The key metric has shifted from chip shipments and broad growth rates to billable capacity that can actually be installed, energized, tested, deployed for customers, and converted into revenue.”
AI accelerator market: JPMorgan expects global accelerator shipments to increase from 3.3 million units in 2023 to 23.3 million in 2027, with Nvidia rising from 1.7 million to 9.9 million. Custom ASICs and XPUs are projected to reach a 53% share in 2027, versus 47% for GPUs. Competition is broadening from standalone chip performance to interconnects, software, and complete-system delivery.
AI Cloud and Data-Center Operators
CoreWeave: Q2 revenue was approximately $2.58 billion, up 112% year over year. Adjusted EBITDA was roughly $1.51 billion, with a 59% margin; the net-loss amount was $626 million. Remaining performance obligations reached $104 billion, contracted power totaled 3.7GW, and active capacity stood at 1.5GW. Order visibility is high; the key variables are the pace at which new capacity comes online and financing costs.
“Customer demand is accelerating as enterprise adoption broadens and we continue to deepen our technology platform.”
Nebius: Q1 revenue reached $399 million, up 684% year over year, including $390 million from its core AI-cloud business, which accounted for 98% of group revenue. Annualized recurring revenue rose 54% sequentially to $1.92 billion. The company expects 2026 revenue of $3.0 billion to $3.4 billion and year-end annualized recurring revenue of $7 billion to $9 billion, while guiding to capital expenditure of $20 billion to $25 billion. Delivering this expansion will require both financing and the ability to energize capacity.
IBM and Together AI: The companies signed a $240 million multiyear compute agreement to deploy infrastructure based on Nvidia HGX B300 and Spectrum-X on IBM Cloud beginning in Q1 2027. The system is intended to process 400 trillion tokens per month and increase output capacity 30-fold. The contract shows that enterprise AI-model providers are using long-term commitments to secure training and inference capacity, although delivery still depends on the full network and system coming online.
“IBM and Together AI have signed a $240 million multiyear agreement to deploy clusters purpose-built for large-scale AI inference on IBM Cloud.”
QumulusAI: The company is building a dedicated B300 cluster for DRW under a 1-year base contract with 3 annual renewal options. Contracts signed since June now exceed $246 million, and DRW is its second financial-sector customer. Financial institutions are willing to pay for dedicated compute, but project value must still be validated through renewals, utilization, and power delivery.
“Quantitative investment firms are among the most demanding compute customers and are rapidly increasing their investment in AI infrastructure. We will give the DRW team rapid, dedicated access to the latest Blackwell compute resources whenever required.”
EROC: Backlog has increased to $1.7 billion, approximately 10 times its previous level. This includes 470MW for Anthropic, scheduled for delivery through 2028, and a 366MW Meta project. The company’s 2026 revenue guidance is $435 million to $465 million, with the midpoint approximately 2.5 times the prior level. Power orders have been secured, but revenue recognition still depends on construction, energization, and customer acceptance.
GPUs, CPUs, ASICs, Foundry, and Advanced Packaging
AMD: Following its acquisition of ZT Systems’ manufacturing operations, Sanmina has become AMD’s preferred new-product-introduction partner. Helios racks have entered production, with shipments beginning this quarter, acceleration planned for Q4, and capacity expansion continuing through 2027. AMD data-center revenue reached $6.7 billion, more than doubling year over year. However, the supply chain has not indicated that all Helios rack orders will be handled by a single contract manufacturer, so order allocation must be verified quarter by quarter.
Intel: Supply-chain reports indicate that Google’s ninth-generation TPU will begin using Intel’s EMIB-T packaging in late 2027. Intel expects related monthly capacity to reach approximately 25,000 CoWoS-equivalent wafers in 2027 and 100,000 in 2028; TSMC is expected to have approximately 170,000 wafers of capacity over the same period, with industry capacity totaling around 370,000 wafers. Customer volume production and yields will determine whether the foundry business generates meaningful incremental revenue.
TSMC: CoWoS at 5.5 times reticle size has entered high-volume production, with yields consistently above 98% for multiple AI customers and reaching 99% for some. The company also plans to advance a 14-times-reticle-size solution in 2029 and build 10 advanced-packaging facilities. Larger package areas can support more complex systems, but capacity expansion must simultaneously address substrate, equipment, and thermal-management constraints.
“CoWoS at 5.5 times reticle size has now entered high-volume production. Yields for multiple AI customers remain consistently above 98%, reaching as high as 99% for some.”
CPU ratios: The ratio of CPUs to accelerators in server systems is not fixed. Current configurations commonly approach 1:1, while some agentic workloads may reach 2:1. CPU demand will rise if inference workflows incorporate more planning, retrieval, and scheduling steps. A higher ratio could also indicate longer accelerator wait times, and assumptions of 12% to 22% idle capacity must be tested against actual cluster utilization.
Advanced-packaging bottlenecks: Large packages integrate chips, memory, and interconnects into a single system, but improving yields does not mean delivery is unconstrained. ABF substrates, testing, cooling, and power systems may still determine rack shipments. Investment analysis should track packaging capacity, customer yields, and system acceptance together rather than extrapolating revenue solely from accelerator-wafer volumes. These factors affect both delivery schedules and system costs.
HBM/DRAM/NAND/SSD/HDD
Nanya Technology: July revenue rose 55% sequentially, indicating that higher spot prices are already flowing through to the company. JPMorgan expects the memory shortage to persist for another 9 to 12 months. Nanya plans to increase monthly capacity from 60,000 wafers to more than 100,000 within 3 years, with an initial addition of 30,000 wafers per month coming online in mid-2027. Capital expenditure through 2029 will be approximately $15 billion, and returns on this expansion depend on whether the high-price cycle lasts long enough to offset higher depreciation.
“The report expects the shortage to persist for another 9 to 12 months. Increasing discussions around long-term agreements should also improve pricing visibility and support a floor under prices.”
Kioxia: The company aims to raise long-term contract coverage to 50% by 2028, reducing the impact of NAND pricing cycles on cash flow. Enterprise SSD total addressable capacity is expected to exceed 820EB next year, representing growth in approximately the mid-50% range. The GP1 product is scheduled to sample with selected customers in late 2026 and deliver 10 million IOPS. The ¥800 billion share repurchase places capital returns and expansion discipline on the same balance sheet.
SK Hynix: The company has restarted the second phase of its Dalian NAND expansion, with equipment installation planned for 2026 and production scheduled for the first half of 2027. Its NAND capacity in China will increase by approximately 50%. The expansion targets enterprise SSD demand from AI servers, but equipment installation, yield ramp-up, and the pricing cycle may not align. Investors should watch whether long-term customer orders absorb the incremental bit supply.
