404K SEMI-AI Morning Brief 2026-07-31 — Returns Begin to Diverge: Improving Cloud Revenue, With Power and Storage Constraints Determining the Pace of Monetization
目录
After-Hours Summary
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditures
CSP/Cloud Capital Expenditures
AI Cloud/Data-Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundries and Advanced Packaging
Semiconductor Equipment/Testing
Optical Communications/High-Speed Interconnects
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Jensen Huang’s Selected Portfolio
Overview
404K | 2026-07-31
After-Hours Summary
The latest available closing session was July 29. The S&P; 500 fell 1.54%, the Nasdaq 100 fell 2.04%, the equal-weighted S&P; 500 fell 0.90%, and small caps fell 1.64%. Pressure was greater on heavyweight technology stocks, indicating that investors are still reducing exposure to high valuations and heavy capital spending; the market has not yet returned to a broad-based rally.
Sector divergence was clearer than the index moves: the semiconductor ETF fell 4.79%, extending its 20-day decline to 23.12%, while the technology sector fell 2.64%. Software and cloud computing rose 0.64% and 0.48%, respectively, with their 20-day relative-strength readings climbing to 63.0 and 74.1. The market is beginning to split “AI investment” into two categories: companies that can translate it relatively quickly into cloud revenue and software monetization are proving more resilient, while the hardware supply chain remains under pressure from capacity, power-supply, and capital-spending cycles.
The marginal change is coming from fundamental validation. AWS is improving in both growth and margins, while Meta is already seeing returns from AI in advertising. However, grid interconnection, memory supply, advanced packaging, and testing capacity still determine when revenue can be realized. The most important metrics ahead are no longer just capital expenditures, but billable compute capacity, backlog conversion, wafer starts, memory pricing, and actual deliveries.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditures
OpenAI
1) Annualized revenue in July has already exceeded revenue for the entire second quarter, as model demand continues to translate into cloud revenue and inference consumption.
2) Elastic expanded its partnership, integrating reasoning models, cybersecurity models, and Codex into its enterprise search, security, observability, and data-governance platforms. The next point of validation is whether model companies can continue converting usage into enterprise contracts.
Anthropic
CoreWeave is issuing a new data-center term loan to support capital expenditures related to Anthropic, Jane Street, and Hudson River Trading. The loan is expected to be priced at 96 to 97, with a spread of 550 basis points over the benchmark rate. Financing costs have risen materially from the previous loan, showing that growth in compute contracts and balance-sheet pressure can coexist.
European AI Gigafactories
The EU plans to select up to 7 projects, provide as much as €10 billion in public funding, and attract at least €20 billion in private capital. Each facility will be equipped with at least 100,000 advanced AI chips, and aggregate capacity is expected to increase by more than 100% relative to the existing network of 19 AI factories. The tender will create new demand for model training, fine-tuning, and inference, but winning bids will not be announced until July 2027.
Enterprise AI Adoption
AI project returns exhibit a pronounced J-curve: process redesign, data preparation, and organizational costs occur first, while benefits are often delayed. Under a model assuming an initial success probability of only 5%, a system builder running 50 projects annually for 5 consecutive years may not diverge from an organization that merely accumulates projects until years 4 to 5, and may not break even until year 8. Heavy near-term investment does not automatically prove that long-term capabilities have already been established.
JPMorgan Chase
The company estimates that AI has already created $1.0 billion to $1.5 billion in value, making it one of the few corporate cases to quantify returns. Disclosures with real substance should cover shared data pipelines, deployment tools, postmortems of failed projects, and multi-use-case cost-benefit analysis, rather than merely reporting partners or project counts. Investors can use these metrics to distinguish capability accumulation from budget consumption.
CSP/Cloud Capital Expenditures
Microsoft
Improving Azure growth is sending a more positive signal on investment returns, placing the cloud platform at the forefront of AI-demand monetization. Microsoft’s cloud backlog is reportedly approximately $627 billion. The investment implications of this figure depend on revenue conversion over the next 12 to 24 months, renewal pricing, and capital-expenditure growth; the total contract value alone is insufficient, and Azure gross margins also need to be monitored.
Amazon
1) AWS grew 37% year over year, its fastest growth in 18 quarters, with an operating margin of approximately 39%. Annualized AI revenue reportedly increased from $15 billion to $25 billion.
2) The company generated total sales of $200 billion and EBIT of $27.5 billion; fiscal 2026 capital expenditures remain approximately $200 billion. Revenue and margins are already improving, but subsequent cash flows must still demonstrate returns on the heavy investment.
"AWS growth, margins, and AI revenue are all improving, but the ROI on high capital expenditures and third-quarter guidance still need to be validated."
Meta
1) Second-quarter revenue was $60.8 billion, constant-currency advertising revenue grew 27%, and third-quarter guidance was $61 billion to $64 billion.
2) Fiscal 2026 capital-expenditure guidance was raised to $130 billion to $145 billion, while the expense range increased to $165 billion to $169 billion. Advertising-efficiency gains have already been realized, but non-advertising AI monetization and long-term returns on compute remain points of valuation disagreement.
Google
Second-quarter capital expenditures reportedly approached $45 billion, while the company did not repurchase shares for a second consecutive quarter, indicating that capital allocation is clearly shifting toward AI infrastructure. Separately, Gemini has entered Oracle AI Agent Studio, Fusion Cloud Applications, and NetSuite, giving the model access to additional enterprise workflows. The next question is whether cloud-revenue growth can offset depreciation and cash-flow pressure.
Oracle
Gemini has been integrated into AI Agent Studio, Fusion Cloud Applications, and NetSuite, strengthening Oracle’s multimodel AI and workflow-automation offerings. The company also disclosed 1.2 GW of contracted capacity and a 2.8 GW upper limit under agreements, indicating that demand remains strong. However, converting contracted capacity into live revenue still depends on power supply, equipment, data-center facilities, and customer acceptance.
Modular Data-Center Construction
Modular solutions can reportedly shorten the overall construction cycle by approximately 36%, to 7 to 9 months, reduce capital expenditures per megawatt by approximately 8%, cut on-site labor hours by approximately 63%, and reduce demand for licensed electricians by approximately 85%. For a 50 MW data hall, earlier commissioning would generate approximately $200 million in undiscounted benefits. Speed has become part of the return equation for cloud capital expenditures.
AI Cloud/Data-Center Operators
CoreWeave
The $3.1 billion data-center loan completed in May was priced at 99 with a spread of 450 basis points over the benchmark rate and attracted $19 billion in orders. The new loan is expected to be priced at 96 to 97, with the spread widening to 550 basis points. The contracts are associated with customers including OpenAI, Cohere, and Anthropic, yet credit conditions have weakened; financing pricing is a key metric for assessing the quality of neocloud expansion.
Nebius
The company disclosed an average power usage effectiveness of 1.25 in 2025 and estimated that its renewable-energy and waste-heat-recovery projects avoided 102 GWh of electricity consumption. For AI cloud operators, low power usage effectiveness directly reduces non-compute electricity consumption, but the investment implications still depend on commissioned capacity, customer utilization, and revenue per unit of electricity.
Cipher Mining
Morgan Stanley believes the company’s approximately 2 GW of interconnection applications may be classified as ERCOT base load. The overall process may place 65 GW of projects into the base-load category, above the previous expectation of 35 GW. If the determination is finalized, existing interconnection rights could accelerate data-center projects, but studied load does not guarantee that the entire requested capacity will be obtained.
Galaxy Digital
Morgan Stanley lists the company among covered names that may receive favorable interconnection determinations relatively early. Approximately 140 GW of studied load may still remain in the system-upgrade-study phase for an extended period, and final capacity and timing will depend on study results expected in April 2027. Project-classification notices are therefore only the first hurdle; leasing and construction progress will determine revenue.
TeraWulf
Existing interconnection rights and convertible sites make mining-facility conversion a relatively fast route to securing power. Morgan Stanley estimates that the potential US data-center power shortfall from 2026 to 2028 is approximately 38 GW, while probability-weighted fast-power solutions could provide 27 to 46 GW. The company’s value depends on signed leases, actual operating MW, and per-watt economics.
Bloom Energy
Second-quarter revenue was $1.065 billion, up 165.5% year over year; GAAP gross margin was 33.4%, and operating cash flow was $226.4 million. Full-year revenue guidance was raised to $3.9 billion to $4.2 billion. The company says it has received validation or approval from major US hyperscalers and more than 10 AI customers, but it has not yet provided a phased disclosure of MW shipped, accepted, and in paid use.
"The central issue is that ‘power equipment in place’ does not equal ‘billable compute capacity online.’ A genuine project must sequentially clear six hurdles: site permitting, reliable fuel, generators, electrical engineering, data-center facilities, and customer acceptance."
Vertiv
1) Second-quarter revenue was $3.27 billion, up 24% year over year; adjusted operating profit was $738 million, with a margin of 22.6%, and free cash flow was $925 million.
2) Full-year revenue guidance is approximately $14 billion, implying more than $8 billion of revenue in the second half. UBS maintained its Buy rating and $370 price target; delivery timing and supply-chain congestion remain near-term risks.
GPU/CPU/ASIC
Changes Across GPU/CPU/ASIC Platforms
AI-chip competition continues to shift from peak performance at the individual-chip level toward system economics. Demand for CPUs, ASICs, and GPUs is simultaneously driving wafers, memory, substrates, packaging, and testing. What ultimately determines the revenue trajectory is whether customers can obtain complete systems and put them into operation. A delay in any part of the supply chain will push order recognition further out.
NVIDIA
1) Jetson AGX Thor provides up to 2070 FP4 TFLOPS of AI compute and 128 GB of memory, targeting real-time inference for humanoid robots and autonomous systems.
2) The company is evaluating the use of Intel EMIB in a future processor composed of 4 GPUs, indicating that package size and capacity have become constraints on platform scaling.
Arm
1) First-quarter revenue was $1.29 billion, operating margin was 41.2%, and earnings per share were $0.45; the midpoint of second-quarter revenue guidance was $1.38 billion.
2) The demand pipeline for AGI CPUs exceeds $2 billion in fiscal 2027 to fiscal 2028, and capacity supporting the initial $1 billion of revenue has been secured. Goldman Sachs maintained its Sell rating and raised its price target from $150 to $160; Morgan Stanley maintained its Neutral rating and raised its price target from $202 to $212.
"The upgrade to licensing-growth expectations offsets slowing mobile royalties, but the business mix is shifting from more durable royalties toward more volatile licensing revenue."
Qualcomm
1) Third-quarter revenue was $9.947 billion, while automotive revenue was $1.588 billion, up 61.4% year over year. Fourth-quarter earnings-per-share guidance was $2.15, below the market expectation of $2.36.
2) Two custom-chip projects for cloud providers have received purchase orders and begun wafer production, with a fiscal 2027 data-center revenue target of $5 billion. Baird raised its price target from $300 to $400 and maintained its Outperform rating.
HBM/DRAM/NAND/SSD/HDD
Memory Market
Supply-demand conditions begin to diverge in 2027: AI servers and HBM continue tightening DRAM supply, while NAND is trending toward oversupply because of capacity expansion and weak consumer demand. The hardware supply chain can no longer be characterized by a single “memory cycle.” DRAM contract prices, NAND spot prices, wafer starts, and capital expenditures should be tracked separately, as the two inventory cycles may diverge further.
