404K SEMI-AI Evening Brief 2026-07-24 — AI Capex Accelerates, Intel Supply Expands, Optical Interconnect Orders Materialize
目录
Pre-Market Takeaways
Full AI/Semiconductor Supply Chain
AI Models, Applications, and Capex
CSP/Cloud Capex
AI Cloud/Data Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry, Advanced Packaging, and Equipment
Optical Communications, PCBs, and Data Center Power
Robotics and Space Infrastructure
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
The primary constraint in AI infrastructure continues to broaden from individual chips to memory, power, packaging, and optical interconnects. Intel’s above-consensus results and higher capex validate the server cycle; AMD has again raised its CPU and accelerator TAM estimates; and Nokia, Prysmian, and Amkor Technology have provided hard evidence through orders and capacity expansion.
404K SEMI-AI | 2026-07-24
Pre-Market Takeaways
The incremental opportunity across the technology supply chain is shifting from “strong GPU demand” to “what remains in short supply across the full system.” Intel’s server business, AMD’s rack-scale solutions, and NVIDIA’s advance payments for advanced packaging all point to supply expansion. Meanwhile, CXL, optical networking, 800V HVDC, and cooling solutions show that as inference scales, memory capacity, power, and data movement increasingly determine whether projects can enter service on schedule.
Capex continues to accelerate, but cash-flow constraints are coming into view. Google’s quarterly capex rose to $44.9 billion and free cash flow turned negative. OpenAI’s Australian data center has switched to waterless cooling, turning water conservation into higher electricity consumption and equipment investment. The market is beginning to demand cash returns from this spending. Investors must track revenue growth, supply delivery, and free cash flow together rather than focusing solely on construction scale.
At the company level, Intel’s results and guidance were well above expectations, but it still lacks an anchor foundry customer and faces financing pressure. AMD’s TAM increase is substantial, but the real validation points remain MI450 and Helios shipments beginning in September and scaling toward several gigawatts of deployment in 2027. Orders in the optical communications supply chain are more tangible, with Nokia and Prysmian outlining order-conversion and long-term capacity-expansion paths, respectively.
Full AI/Semiconductor Supply Chain
AI Models, Applications, and Capex
OpenAI
1) The S7 data center being developed by OpenAI and NEXTDC in Western Sydney will have 612MW of capacity. Because recycled-water pipelines and permits were not secured on time, the project has switched to waterless cooling.
2) The new design combines liquid circulation around the chips with external fans and dry heat rejection. Direct water consumption will decline, but electricity use, equipment footprint, and operating costs may increase. Whether the project enters service on schedule will depend on local grid capacity and total cooling costs.
“Waterless cooling will reduce water consumption but could increase the burden from electricity use, equipment footprint, and operating costs.”
AI inference memory hierarchy: Running the 70-billion-parameter Llama 3 model in FP16 with standard multi-head attention requires approximately 140GiB for model weights. At batch size 1, the KV cache reaches 160GiB with a 64K context and 320GiB with a 128K context, equivalent to 2.3 times the model-weight footprint. Remote memory over CXL can expand capacity but introduces higher latency. Key validation points are offloading efficiency, software compatibility, and deployment at scale.
Data center capex: Industry estimates indicate that global data center capex will expand from $455 billion in 2024 to more than $1 trillion in 2029. By 2028, incremental electricity demand from semiconductors and AI is expected to exceed 10,000kW. Capital continues to flow into AI infrastructure, but incremental spending will increasingly target power reliability, energy efficiency, HVDC, and solid-state transformers rather than accelerator chips alone.
CSP/Cloud Capex
Google
1) Second-quarter 2026 revenue approached $120 billion, up 24% year over year, while cloud revenue grew 82%. Capex reached $44.9 billion during the quarter, nearly doubling year over year, primarily reflecting investment in data centers, GPUs, networking, and power.
2) Free cash flow was negative $5.9 billion, showing that accelerating cloud revenue and AI infrastructure investment are both scaling rapidly. Management said the cost per AI response fell to its lowest level since launch after stronger search models went live. The next question is whether unit-cost reductions can outpace capex growth.
AI Cloud/Data Center Operators
CoreWeave: In an independent benchmark serving MiniMax M3, CoreWeave delivered 357 output tokens per second, approximately 1.8 times the performance of the runner-up. Time to first token was 6.6 seconds, with an all-in price of $0.22 per million tokens. Combining high speed with the lowest price should support inference customer wins, but a single benchmark cannot substitute for evidence of sustained utilization, customer diversification, and cash returns.
GPU/CPU/ASIC
AMD
1) JPMorgan said AMD raised its estimate for the 2030 AI accelerator TAM to approximately $1.4 trillion and the data center CPU TAM to approximately $220 billion. The report assumes GPUs account for 60% of the accelerator market and AMD captures a 15% share, implying approximately $125 billion in data center GPU revenue.
2) MI450 and Helios shipments are scheduled to begin in late September, ramp rapidly from the fourth quarter of 2026 through the first half of 2027, and reach several gigawatts of deployments in 2027.
3) Following the company’s conference, Jefferies maintained its Buy rating and raised its price target from $515 to $640. JPMorgan maintained its Neutral rating, with the disagreement centered on share capture, R&D; investment, and valuation.
4) AMD Helios: MI455X incorporates 12 HBM4 stacks, providing 432GB of memory and 23.3TB/s of bandwidth. A Helios rack comprises 72 MI455X accelerators, 18 sixth-generation EPYC Venice processors, Pensando networking, and the ROCm software stack. The larger memory capacity supports models and KV caches while reducing model partitioning across GPUs. AMD claims token throughput of up to 34 times that of MI355X, but this still requires validation through power consumption, stability, and customer utilization in production systems.
“Management also raised its estimate for the 2030 data center CPU TAM to approximately $220 billion.”
Intel
1) Second-quarter revenue was $16.13 billion, up 25% year over year and 19% sequentially. Adjusted gross margin was 41.8%, and EPS was $0.42. Data Center and AI revenue reached $6.26 billion, up 59% year over year, while server supply remained tight.
2) The midpoint of third-quarter revenue guidance is $16.3 billion, with gross-margin guidance of 42%. Intel raised 2026 capex to more than $20 billion, with equipment investment increasing approximately 40% year over year, and expects another significant increase in 2027.
3) Intel plans 14A risk production in the second half of 2027 and volume production in 2028, while 18A-P has entered risk production. JPMorgan raised its price target from $45 to $85 but maintained its Underweight rating; Goldman Sachs maintained its Neutral rating. The external foundry business still lacks a named anchor customer, while capex and financing risks remain counterpoints to the thesis.
“DCAI revenue was $6.26 billion, up 59% year over year, with a 39.5% operating margin. Operating profit surpassed the Client Computing business for the first time.”
Cerebras: Its partnership with AMD brings heterogeneous inference partitioning to the forefront. The prefill stage is more constrained by compute throughput, while decoding is more constrained by memory bandwidth. Cerebras will deploy AMD Helios in its data centers, although available information does not clarify whether the systems will be leased or purchased. The investment implication is that inference clusters may assign prefill and decoding to different processors. Commercial validation will depend on the deployment model, system utilization, and cost per token.
HBM/DRAM/NAND/SSD/HDD
Micron: Micron and AMD jointly developed a benchmark designed to reflect real-world agentic workflows. Results showed up to a 3.8-fold increase in AOPS and a 2.9-fold improvement in CPU energy efficiency versus the prior-generation platform. The results indicate that CPU-memory coordination directly affects response speed when AI assistants retrieve answers from large document sets. The benchmark must still be reproduced across more models, datasets, and customer environments before it can inform server procurement decisions.
SK hynix
1) SK hynix is developing 3D-stacked DRAM-on-Logic for on-device AI with a major US customer and is recruiting relevant engineers in San Jose. The technology places bare DRAM dies above a logic chip to shorten interconnects, improve bandwidth and energy efficiency, and conserve space in mobile devices.
2) Channel checks also indicate the risk of order adjustments in general-purpose server DRAM, PC DRAM, and SSDs, with third-quarter DRAM bit-shipment growth potentially slowing from 18% to 8%. This does not imply weaker HBM demand, but fourth-quarter general-purpose DRAM pricing should be monitored separately.
