美国半导体与硬件：每周投资者脉搏：RAISE峰会要点

Weekly Investor Pulse: RAISE Summit Takeaways CITI'S TAKE We tuned into the RAISE 2026 AI Summit last week and recap keytakeaways. We maintain AMD #1 (agentic CPU demand), CBRS #2 (fastinference pure play), and like NVDA #1/AVGO #2 in value mega capcompute semis. Atif MalikAC+1-415-951-1892atif.malik@citi.com AI Demand:The explosive growth of token-intensive agentic workflows and real-world code generation applications is driving unprecedented demand for AIinference infrastructure.Machine learning compute requirements are currentlyexpanding at roughly 10x year-over-year.Unlike the dot-com era, today’s AIinfrastructure spending is being underwritten by companies generating hundreds ofmillions—and in many cases billions—of dollars in tangible revenue. While thedemand outlook remains exceptionally strong, the increasing reliance on privatecredit and debt financing represents a risk factor for the investment cycle. Asiya Merchant, CFA+1-415-951-1752asiya.merchant@citi.com Papa Sylla+1-212-816-9476papa.sylla@citi.com AI Interns Replacing Traditional Chatbots:Leading AI labs are increasingly movingbeyond conventional chatbots and towards AI-native “interns” equipped withcoding tools and autonomous research functions. Advanced models are nowrecursively accelerating their own development by assisting in engineering, softwarecreation, and research workflows.This self-reinforcing flywheel is compressingmajor model release cycles from the historical six-to-nine-month cadencetoward monthly iterations, significantly increasing the pace of innovation. Hardware Disaggregation:Hyperscalers are rapidly adopting disaggregatedinference architectures to extract greater efficiency from existing Nvidia GPUdeployments. Rather than relying solely on GPU AI clusters, operators are buildingspecialized infrastructure optimized for distinct workloads, including latency-sensitive inference, real-time tool use, and agentic task execution.Thisarchitectural shift is becoming increasingly important as AI workloads diversify. The Resurgence of CPUs:While GPUs remain the dominant platform for modeltraining,CPUs are experiencing a significant resurgence as agentic AI workloadsscale.Complex inference pipelines, orchestration layers, reinforcement learningenvironments, and enterprise AI agents increasingly rely on CPU-intensiveprocessing, expanding the addressable market for general-purpose compute. Memory Squeeze:Data centers are on track to consume approximately 70% ofglobal memory-chips, creating supply constraints that could ripple across the See Appendix A-1 for Analyst Certification, Important Disclosures and Research Analyst Affiliations broader electronics ecosystem. Major hyperscalers are already pursuing multi-year supply agreements to secure productioncapacity.Industry supply-demand dynamics suggest that meaningful relief may not arrive until at least 2028, when newmanufacturing capacity is expected to come online. Hyperscaler Custom Silicon:Traditional CPU and GPU architectures are increasingly constrained by networking, power, andefficiency bottlenecks. In response, hyperscalers are accelerating investment in custom silicon purpose-built for AIworkloads. These chips are optimized around specific models, inference tasks, and agentic workflows running on internalinfrastructure.Google’s latest roadmap illustrates this specialization, with separate silicon platforms dedicated to low-latency inference (8i) and large-scale model training (8t), highlighting the industry’s broader move toward workload-specific compute architectures. The KV Cache Bottleneck:As the cost of multi-turn agentic workflows continues to rise, hyperscalers are increasinglyfocused on optimizing context memory management. A key strategy involves offloading KV cache data from expensive GPUsto lower-cost storage layers, reducing redundant computations and improving system utilization.These techniques candeliver cache-hit rates exceeding 95%, reduce time-to-first-token latency by 20x, and materially lower inference costs.Moreover, because flash is in shortage, hybrid storage architectures including hard disk drives are seeing a resurgence formassive AI datasets. Latency as a Competitive Advantage:Within enterprise environments, inference speed is becoming a critical determinant ofuser adoption and productivity. Faster response times reduce workflow friction, minimize context switching, and encourageusers to tackle increasingly sophisticated coding and analytical tasks.LLM latest models can run at 750 tokens per secondon Cerebras hardware or an order of magnitude faster than standard GPU fleets. Edge AI and Small Language Models:Growing pressure on centralized cloud infrastructure is accelerating the shift towardedge computing and specialized small language models (SLMs). Enterprises are increasingly deploying task-specific modelsthat deliver domain expertise while operating on substantially lighter hardware footprints.Chips like the Nvidia RTX Sparksuper