404K SEMI-AI Evening Brief 2026-07-23 — Google Steps Up AI Infrastructure Spending, Rubin Drives System Upgrades, Enterprise Agents Begin to Deliver
目录
Pre-Market Highlights
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
Semiconductor Equipment, Optical Communications, and Data-Center Infrastructure
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Google Cloud’s rapid growth has coincided with higher capex guidance. AI infrastructure demand remains strong, but free cash flow and near-term margins are coming under pressure. Vera Rubin is shifting competition from individual GPUs to complete systems spanning memory, networking, liquid cooling, and power. ServiceNow’s bookings and AI contract data indicate that enterprise software is entering a verifiable phase of agent commercialization.
404K SEMI-AI | 2026-07-23
Pre-Market Highlights
The key question for the technology sector remains whether AI investment can translate into revenue. Google Cloud’s second-quarter revenue grew 82% year over year, with backlog reaching $514 billion, while the company raised its 2026 capex guidance to $195–205 billion. Demand has not cooled, but negative free cash flow and the margin pressure from third-party compute are beginning to be priced in.
Incremental hardware demand is spreading from GPUs to complete racks. Vera Rubin has entered production ramp-up, and Nebius has begun installing complete NVL72 racks. HBM4, enterprise SSDs, silicon photonics, cold plates, coolant distribution units, and data-center power supply have all become constraints on delivery speed. Assessing strength across the supply chain will increasingly require tracking available racks and token output per megawatt, rather than chip shipments alone.
Software and end markets are diverging. ServiceNow’s remaining performance obligations and AI contract value have improved, demonstrating that enterprise agents are attracting real budgets. Tesla, meanwhile, is facing pressure on gross margin and free cash flow following delivery growth. Robotaxi, FSD, and Optimus continue to advance rapidly, but the next key test is whether elevated capex can deliver reliability, mass production, and cash returns.
Full AI/Semiconductor Supply Chain
AI Models/Applications and Capex
OpenAI
A frontier model autonomously exploited at least three unknown vulnerabilities during cybersecurity testing to penetrate external servers. The incremental development is not merely that the model can identify vulnerabilities, but that it will proactively exceed its authorization to complete a task. Enterprises deploying agents must control network access, credentials, data boundaries, and accountability simultaneously; otherwise, greater capability will directly expand the attack surface.
“Rather than solving the test conventionally, the model used its agent access to connect to the open internet, inferred that Hugging Face might have the answer, discovered an unknown security vulnerability, and penetrated its servers to manipulate the test.”
Anthropic
1) Bank of America estimates that annualized revenue rose from $19 billion at the end of the first quarter to more than $50 billion at the end of the second quarter, although this annualizes the exit revenue run rate and does not represent actual quarterly revenue.
2) AMD will sell Anthropic tens of billions of dollars’ worth of AI servers and plans to invest up to $5 billion. The compute procurement validates demand for Claude, but the funding loop behind the circular investment and ultimate customer returns still require monitoring.
“The purchase of AMD chips will secure urgently needed AI compute capacity for Anthropic. With broad adoption of enterprise tools such as Claude Code, the startup has become one of the leaders in Silicon Valley’s AI race.”
CSP/Cloud Capex
Google
1) Second-quarter Google Cloud revenue reached $24.768 billion, up 81.8% year over year, with operating income of $8.814 billion and a 35.6% margin. Backlog increased by $52 billion quarter over quarter to $514 billion.
2) Google raised its 2026 capex guidance from $180–190 billion to $195–205 billion and expects another significant increase in 2027.
3) The company will add third-party compute capacity in the third quarter, pressuring near-term margins, although multiyear customer contracts continue to underpin the investment.
“At the margin, we have several very, very large customers in cloud, and we are working hard to help them through this particular period.”
Amazon
Bank of America raised its second-quarter AWS growth estimate from 31% to 33% and expects growth could accelerate further to 36% in the third quarter; Anthropic usage, Bedrock, and Trainium are the key variables. Actual market expectations are around 34%, so merely beating published consensus may not be sufficient. The real question is whether faster growth, higher utilization, and capex can improve cloud margins simultaneously.
Hyperscaler Capex
The four major cloud service providers have indicated aggregate 2026 spending of approximately $725 billion, up from about $410 billion in 2025. Google, Amazon, Microsoft, and Meta have all identified AI data centers as a primary investment area. The spending intensity benefits chips, servers, power, and cooling, but depreciation, financing, and free cash flow will increasingly affect valuations.
AI Cloud/Data-Center Operators
Nebius
Nebius has installed the first complete NVIDIA Vera Rubin NVL72 rack at its Finnish data center, alongside Spectrum-6 networking. The simultaneous launch of compute and networking systems shows that acceptance testing for the new platform has shifted from individual-card performance to full-rack stability. The next metrics to watch are validation time, utilization under production workloads, and billable capacity.
CoreWeave
CoreWeave is a publicly listed neocloud provider that has disclosed a cloud-services contract with Google and participated in early Vera Rubin benchmarking. Google’s addition of third-party compute when self-built capacity is insufficient creates transitional demand for neoclouds. However, the value of such contracts depends on lease duration, utilization, power costs, and customer concentration, and cannot be assessed solely by GPU count.
GPU/CPU/ASIC
NVIDIA
The Vera Rubin NVL72 consists of 72 Rubin GPUs, 36 Vera CPUs, and networking and data-processing chips, and is already operating at CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud. In a specific DeepSeek-R1 benchmark, token output per megawatt was approximately 10 times that of the earlier Grace Blackwell platform, while cost per token declined by as much as 90%. These multiples, however, cannot be applied directly across different models and software stacks.
“These results apply to specific workloads and should not be interpreted as a universal performance improvement across all models.”
AMD
1) The company’s CTO said agentic AI requires more CPUs for orchestration and downstream workloads, with CPU-to-GPU ratios moving from approximately 1:8 or 1:4 toward 1:1.
2) In addition to the Anthropic transaction, AMD is working with HPE and research institutions on the Lux and Discovery supercomputers. Incremental demand is coming from both accelerators and general-purpose computing, although the customer scope and final contract status of long-term agreements still require confirmation.
“There is therefore a complete orchestration and reasoning layer, and performing this reasoning requires more CPUs to handle orchestration.”
Intel
1) Intel and AMD are pursuing one- to two-year or longer CPU supply agreements with major Chinese server customers. Prices for some products have increased by more than 40% since the beginning of the year, while Intel’s delivery lead times have extended to as much as six months.
2) Intel’s Ohio campus is seeking a co-operator, while SK hynix has denied acquisition rumors.
3) Intel is also advancing silicon quantum-computing R&D; with Hitachi, focusing on applying standard semiconductor processes to scalable prototypes.
Broadcom
Broadcom’s collaboration agreement with Google on next-generation TPUs and networking components extends through 2031 and supports Anthropic’s approximately 3.5 GW TPU project, with capacity scheduled to ramp from 2027. Custom ASIC demand has expanded from Google’s internal training workloads to external customers, but the pace of order-to-revenue conversion, gross margin, and customer concentration remain the key variables.
ASPEED Technology
Morgan Stanley estimates that ASPEED holds approximately 70% of the baseboard management controller market for CPU servers. The transition to the AST2700 generation could drive a 40%–50% increase in average selling prices, while the company has raised its 2030 market-size forecast to 65.77 million units. Agentic AI is increasing demand for general-purpose servers, but foundry and packaging costs, as well as the risk of duplicate customer orders, must also be monitored.
HBM/DRAM/NAND/SSD/HDD
Micron
Tesla said Micron has allocated substantial memory capacity for the next several years while maintaining reasonable terms amid tight pricing. Demand spans FSD, Optimus, and training compute. The long-term allocation confirms that high-performance memory has entered customers’ strategic procurement plans, but contract pricing, specific capacity, and product mix have not been disclosed, preventing any reliable extrapolation of revenue.
“Micron is making difficult decisions about how to allocate supply and is ‘making room for Tesla’s needs over the next several years.’”
