404K SEMI-AI Morning Briefing 2026-07-28 — Return Pressure Mounts: Cloud Capex Raised, Memory and Equipment Orders Remain Strong
目录
After-Hours Summary
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
CSP/Cloud Capex and AI Cloud
GPU/CPU/ASIC and Memory
Foundry, Equipment Testing, and Advanced Packaging
Optical Communications, Power, and Space
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Jensen Huang’s Selected Portfolio
Overview
404K | 2026-07-28
After-Hours Summary
The latest available closing session was July 24. The S&P; 500 rose 0.10%, the Nasdaq 100 fell 1.12%, and the S&P; 500 Equal Weight Index rose 0.78%. The broader market did not decline across the board; pressure was concentrated mainly in heavily weighted technology stocks that had previously rallied sharply, while the equal-weight style relatively outperformed.
Semiconductors were the main drag: SMH fell 3.27%, SOXX fell 4.40%, and the AI & Big Data theme fell 2.28%. Software rose 1.01%, cloud computing rose 1.73%, and cybersecurity rose 0.78%, as capital continued rotating from high-beta hardware segments toward software and cloud services with greater revenue visibility.
On a 20-day basis, the relative-strength readings for cybersecurity, cloud computing, and software were 92.6, 81.5, and 77.8, respectively, versus only 0, 3.7, and 7.4 for equal-weight semiconductors, semiconductors, and AI & Big Data. Fundamentals, however, continue to expand: cloud capex guidance is still being raised, while memory pricing, equipment orders, and optical-communications demand are strengthening. The market debate has shifted from “Is there AI demand?” to “When will cash flow cover the investment?”
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
Model monetization: AI usage is becoming more distributed, while revenue remains highly concentrated. OpenAI, Anthropic, and Google account for only 52% of tokens on Vercel AI Gateway but capture 90% of spending, generating more than 8 times as much revenue per token as other vendors on average. Leading models still command brand and performance premiums, but model routing and lower-cost substitutes will continue to compress inference pricing.
“On average, each token generates more than 8 times as much revenue for the three leaders as for other vendors.”
OpenAI
1) Its European headquarters in Dublin plans to add approximately 250 positions and lease 88,000 square feet of office space. Its existing Ireland team has approximately 100 employees, and the expansion points to growing demand from European customers and developers.
2) The $40 billion bridge loan supporting SoftBank’s investment in OpenAI added 21 lenders, with approximately $7 billion redistributed and the remaining $33 billion still held by the original underwriters and senior lenders. Beyond model demand, financing costs and syndicate absorption capacity are beginning to constrain the pace of expansion.
AI occupational penetration: AI has reached 68% of US occupations, but within each occupation it covers only approximately 1/5 of the tasks in a typical job. At this stage, it looks more like targeted workflow enhancement and does not directly imply complete job displacement. The next questions are whether enterprises can turn fragmented usage into repeatable processes and whether total factor productivity can rise in tandem.
“AI reaches 68% of US occupations.”
Low-cost model routing: Users believe Cursor Router’s Auto Intelligence mode can deliver output comparable to Fable at approximately 60% lower cost. This counterexample suggests that enterprises will automatically route tasks across different models; unless a single model vendor can maintain a clear and widening performance advantage, both revenue per call and customer stickiness may come under pressure.
Returns on AI infrastructure: Morgan Stanley believes compute demand could substantially exceed supply for many years. Its token-economics model indicates that both large models and more efficient models can generate strong returns on the underlying infrastructure. The real validation points are not the absolute amount of capex, but customer contracts, utilization, inference pricing, and free cash flow after depreciation.
“Our highest-conviction view is that compute demand is likely to substantially exceed supply for many years.”
CSP/Cloud Capex and AI Cloud
Google
1) 2026 capex guidance was raised to $195 billion to $205 billion. Second-quarter capital expenditures were $44.9 billion, of which approximately 60% went to servers; operating cash flow was $39.1 billion, and free cash flow was approximately negative $5.9 billion.
2) Google Cloud revenue grew 82% to $24.8 billion, with operating profit of $8.8 billion, a 35.6% margin, and a $514 billion backlog, indicating that the investment is underpinned by real demand.
3) Phillip Securities upgraded the rating from Neutral to Accumulate, while lowering its price target from $450 to $425, with the debate centered on the pace of margin and cash-flow realization.
“Infrastructure expansion has already fallen behind demand.”
Amazon
1) The company said its approximately $200 billion of capex in 2026 is supported by customer commitments, including more than $100 billion committed by OpenAI, and that a large portion of the 2026 investment will be monetized in 2027 to 2028.
2) Near-term free-cash-flow pressure is a deliberate choice, but the risks are also clearer: customer commitments must convert into AWS revenue on schedule, while the chip business still needs to prove that it can become a new profit pillar.
“We are willing to make large-scale capex investments and absorb near-term free-cash-flow headwinds in exchange for substantial medium- to long-term free-cash-flow surpluses.”
Microsoft: The market expects FY2027 capex could grow approximately 60%, with growth dependent on Azure, Microsoft 365, and enterprise AI demand continuing to absorb new capacity. Cantor Fitzgerald maintained its Overweight rating and $502 price target, arguing that enterprise adoption of intelligent, context-driven automation will expand Microsoft’s long-term AI revenue opportunity.
Oracle: AI infrastructure projects are expected to deliver gross margins of 30% to 40%, an after-tax EBITDA conversion rate close to 100%, and a return on invested capital close to 30%. Project economics are attractive, but growth in remaining performance obligations remains constrained by financing and leverage; the key questions are whether new orders bring sufficient upfront payments and whether the capital structure can keep pace with construction.
TeraWulf: The Hawesville, Kentucky data center has been leased to Anthropic, but it has not yet been confirmed which chip supplier will provide credit support. The company’s finance chief said one of Nvidia, Amazon, or Google may provide credit support for chip sales worth tens of billions of dollars. Whether new cloud projects can begin construction is now tied to chip selection, supplier credit, and financing terms.
“One of the chip suppliers with an investment-grade credit rating will provide credit support for the project because it will sell tens of billions of dollars’ worth of chips to the project.”
Nebius: Vertical integration allows new cloud providers to troubleshoot everything from code down to PCBs, cooling units, and physical components inside the data center, while customizing server memory to customer requirements. The advantage is not merely owning GPUs, but bringing hardware, networking, on-site operations, and software optimization under one team; the next test is whether failure rates, utilization, and cost per token can outperform fragmented procurement.
“We can debug everything—not only drilling down to a specific line of code, but also identifying a particular component on a PCB, a cooling unit, or another physical component inside the data center.”
Akamai: Approximately 126 MW of additional capacity provides upside potential for edge-cloud revenue, but high capex may keep free cash flow negative through 2027. Whether expansion translates into returns depends on the contracting pace for new capacity, customer utilization, and the depreciation schedule, rather than data-center scale alone.
GPU/CPU/ASIC and Memory
Nvidia
1) Nvidia invested in Safe Superintelligence and established a long-term partnership. Safe Superintelligence will receive the Vera Rubin platform and GPUs at scale, with its compute capacity expected to increase by an order of magnitude; investment terms were not disclosed, while research transparency and circular-transaction risks remain to be validated.
2) Nvidia joined Microsoft, Salesforce, SAP, Red Hat, Cloudflare, and others to launch an open security AI alliance that plans to jointly develop datasets, evaluation frameworks, attack simulators, and red-team tools.
3) Its long-term plans with Korean companies cover GPUs, HBM, AI cloud, and power infrastructure, but the approximately $950 billion figure represents the combined scale of supply and long-term plans and must not be treated as confirmed orders.
“Attackers have frontier AI. Defenders need a frontier AI ecosystem.”
AMD: Committed orders for MI455 Helios reach up to 14 GW, while the potential total addressable market for 2030 was raised to $2 trillion and the related price target was increased to $625. Order strength supports AI-accelerator share gains, but rack-level delivery, customer acceptance, and supply-chain execution still need to be monitored to avoid treating committed capacity directly as revenue.
Intel
1) Baird maintained its Neutral rating and raised its price target from $75 to $125, arguing that supply tightness is increasing attention on the foundry business, output is ahead of target, and 2026 capex may rise.
2) Intel is working with Lens Technology to develop through-glass-via glass-core substrates, offering a new path to lower signal loss and warpage in larger packages; the project remains in the pilot-line preparation stage, and revenue realization depends on yield and mass-production validation.
Broadcom: Its long-term partnership with Samsung Electronics covers 2 nm and below processes, with a planned scale exceeding $200 billion and extending through 2030. Market analysis estimates that Samsung Electronics could meet 75% to 85% of Broadcom’s HBM requirements and support more than $1 trillion in cumulative AI revenue from 2026 to 2030, but the ultimate outcome still depends on order confirmation, process yields, and HBM deliveries.
Cerebras: The related price target was raised to $310, reflecting market expectations for incremental growth from high-speed inference and new cloud deployments. The competitive question is whether specialized systems can continue differentiating through full-system efficiency, delivery speed, and software compatibility once Google, Amazon, and AMD secure more wafer capacity.
