404K Semi-Ai

404K SEMI-AI Evening Briefing 2026-09-01 — Downstream Bottlenecks: GPU Demand Spills Over to Racks, Optical Interconnects, and Data Center Capacity

404K Semi-Ai's avatar
404K Semi-Ai
Sep 01, 2026
∙ Paid

目录

  • Pre-Market Highlights

  • AI & Semiconductor Value Chain

  • GPU / CPU / ASIC

  • AI Cloud & Data Center Operators

  • HBM / DRAM / NAND / SSD / HDD

  • Foundry & Key Components

  • Optical Communications & Photonics

  • Space & Satellite

  • Internet / Platforms

  • Software / SaaS

  • Consumer Electronics / Intelligent Vehicles

Overview

404K SEMI-AI | 2026-09-01

Pre-Market Highlights

AI capex continues to expand, but the critical delivery bottleneck has shifted from raw GPU availability to whether an entire AI factory can be deployed on schedule. Nvidia provided robust growth guidance as Vera Rubin enters multi-cloud rack deployments; meanwhile, optical interconnects, PCBs, liquid cooling, power infrastructure, and regulatory permitting increasingly dictate the timing of revenue recognition.

Enterprise demand is broadening as paid AI adoption accelerates across manufacturing, construction, and healthcare, though switching costs between foundation model providers remain low. Software vendors are rapidly transitioning AI monetization from per-seat add-ons to consumption-based pricing; while this expands total revenue potential, customer retention, gross margins, and underlying usage require close validation.

Near-term risks center on execution rather than underlying demand. A high proportion of flagship data center projects worldwide face delays, and hyperscalers' massive lease commitments remain contingent on power availability, permitting, and physical construction. As prices continue to rise across memory, wafers, IC substrates, and optical components, hardware supply chain revenue visibility is improving, but margins may not necessarily expand in tandem.

AI & Semiconductor Value Chain

GPU / CPU / ASIC

  • AI Factory Platforms: Delivery bottlenecks have shifted from discrete GPU and HBM silicon to full-rack validation. As Vera Rubin and Spectrum-6 (SPECTRUM6) enter multi-cloud deployments, customers must concurrently secure switches, optical interconnects, PCBs, liquid cooling, and power delivery; bare chip shipments no longer equate to recognizable system revenue. Key watchpoints include full-rack yields, joint qualification, and power energization timelines.

  • Nvidia
    1) Management guided next fiscal year sales growth to 70%, well above consensus of 45%; FY27 Q2 revenue reached $96.2 billion, with Data Center revenue at $89.0 billion (92% of total).
    2) Vera Rubin's revenue opportunity per GW expanded to $40 billion, up from $18 billion for Hopper and $25 billion for Blackwell.
    3) Baird maintained an Outperform rating with a $500 price target, projecting that inference architectures will sustain long-term growth.

"Nvidia remains unequivocally supply-constrained. That is the underlying message behind providing guidance not just for the next quarter, but for the entire upcoming fiscal year."

  • Broadcom: OpenAI and Broadcom completed the initial RTL for the Jalapeño inference ASIC in approximately 9 months, optimizing the chip for latency, performance-per-watt, and KV cache memory slicing. While custom ASICs shorten design cycles, software complexity and model lock-in still limit their displacement of general-purpose GPUs. Key watchpoints include mass-production yields, software enablement, and actual procurement volumes.

  • AMD: Next-generation thermal management is extending from cold plates down to the package interface. Reports indicate AMD is evaluating graphene thermal interface materials (TIM) between the package and the cold plate or lid to enhance GPU thermal conductivity; the critical test is whether commercial-grade materials can pass reliability and high-volume manufacturing qualification. Key watchpoints include sample validation, lifetime reliability data, rack-level integration, and supplier qualification progress.

  • Cerebras: Partnering with Compute Nordic Finland to construct a 165 MW AI data center, with initial 50 MW construction underway and plans to expand capacity first to 80 MW and ultimately to 165 MW. Backed by a 7-year contract with an estimated investment of €1.0B to €1.7B, delivery timelines hinge on physical construction and grid energization.

AI Cloud & Data Center Operators

  • Data Center Projects: Supply constraints extend well beyond silicon. Among tracked flagship data centers under construction globally, roughly 40% have been delayed or canceled over the past year, 15% are progressing slowly, and only 30% are tracking on schedule. Given the lack of full methodology disclosure, this dataset serves primarily as an indicator of execution risk rather than a broad industry benchmark, requiring cross-validation against more comprehensive project samples.

"Among the most high-profile data centers under construction globally, roughly 40% have been delayed or canceled over the past year, 15% are progressing relatively slowly, and only 30% are 'on schedule.'"

  • CoreWeave: Vera Rubin racks have entered CoreWeave's deployment roadmap, signaling that neo-cloud providers are shifting from procuring bare GPUs to integrating full-rack systems. Revenue recognition remains contingent on networking, liquid cooling, power infrastructure, and joint customer validation—hardware delivery does not instantly translate into operational compute. Key watchpoints include rack energization rates, utilization levels, and customer acceptance.

  • Nebius: Secured zoning approvals and power supply contracts for the entire 1.2 GW capacity across its Pennsylvania campus, following prior approval of the initial 600 MW. The company reaffirmed its deployment schedule: 260 MW by October 2027, 660 MW by January 2029, and over 1 GW by November 2029; regulatory approval remains distinct from actual grid energization and commercial operation.

  • Lambda & Hut 8: Anthropic signed a $35 billion compute agreement with Lambda, with Hut 8 constructing a 700 MW campus in Texas while Nvidia anchors the underlying lease and supplies GPUs. By chaining silicon, cloud services, and real estate leases into a single structure, execution risk concentrates heavily on construction timelines and counterparty performance.

  • Huaqin Technology: Supernode products completed canary deployments with two major tier-1 clients, with monthly run-rates stabilizing above 100 racks in late July. While mass-production milestones are visible, racks require continued co-validation with domestic GPUs, networking, and thermal solutions. Key watchpoints include monthly output sustainability, qualification conversion into scaled revenue, manufacturing yields, cash collection, and lead times.

  • Honor Electronic: H1 data center power supply revenue reached RMB 1.136 billion (+18.1% YoY), with high-power server power supplies up 41.4% YoY to RMB 936 million (82.4% of segment revenue). Initial small-batch shipments for its first Google project are slated for Q4, while high-voltage direct current (HVDC) solutions remain in the sampling phase and should not be prematurely booked as firm orders. Key watchpoints include design wins, volume ramp, and revenue recognition.

HBM / DRAM / NAND / SSD / HDD

  • Micron: Management highlighted acute memory shortages across all segments, noting limited visibility into when supply will balance demand. DDR4 spot prices carry a 43% premium over contract prices, compared to 16% for DDR5, reflecting severe legacy capacity crowding as HBM and AI servers consume standard DRAM wafer allocations. Key watchpoints include contract pricing trends, customer inventory levels, wafer capacity allocation, and incremental supply additions.

"We have entered an environment of severe memory shortages across all market segments. Despite our best efforts on the supply side, we cannot currently foresee when supply will catch up with demand."

  • Samsung Electronics: HBM5 targets 2x the performance of HBM4E, a 20% gain in performance-per-watt, and a 20% reduction in thermal resistance, with even more aggressive targets for subsequent zHBM. Meanwhile, zNAND-O is slated to sample in 2028 featuring 10x DRAM's bit density and 7x NAND's read bandwidth alongside improved power efficiency; both technologies remain preliminary roadmap targets rather than committed production specs.

  • SK Hynix: Q2 wafer procurement jumped 104.0% QoQ to KRW 670.6 billion against just 10.5% output growth, driving raw material inventory up 61.9% in H1. Procurement significantly outpaced production, likely reflecting advance locking of wafer supply ahead of price hikes and preparation for upside HBM orders; realization remains contingent on downstream conversion. Key watchpoints include inventory turnover and HBM4E qualification.

  • SanDisk: Positioned High-Bandwidth Flash (HBF) as a new memory tier bridging HBM and NAND, with Grade 3 targeting 512GB capacity and 3.072 TB/s bandwidth—providing 8x to 16x the capacity of HBM at comparable cost. While ideally suited for MoE weights and massive KV caches, inadequate software-level data placement remains the primary hurdle. Key watchpoints include controller readiness, runtime software support, and customer validation.

Foundry & Key Components

  • TSMC
    1) Supply chain checks indicate that sold-out AI capacity is driving price hike discussions of 10% to 15% across process nodes, with N2, N3, and N5 potentially facing an additional 10% hike by late 2027; margin realization hinges on broad customer acceptance.
    2) Management projects 2029 AI/HPC system computing power to reach 50x 2024 levels, stressing that HBM, packaging, substrates, testing, and optical interconnects must scale in tandem.

"The critical signal is that TSMC no longer frames silicon photonics scaling merely as a foundry or CoWoS capacity issue."

  • Samsung Electro-Mechanics: Disclosed a 1-year MLCC supply contract for 2027 valued at KRW 1.0722 trillion (9.5% of annual revenue). Combined with two prior high-reliability AI rack capacitor orders totaling KRW 750 billion, rising rack power densities and capacitor content are translating into tangible backlog, though client identities and per-system dollar content remain undisclosed. Key watchpoints include delivery execution and realized margins.

  • Victory Giant Technology: H1 revenue reached RMB 11.6 billion (+28.8% YoY) with net profit attributable to the parent at approximately RMB 2.8 billion (+33.3% YoY); Q2 gross margin fell 6.7 percentage points YoY to 32.1%. Capex rose 228% YoY to approximately RMB 7.1 billion, as concurrent capacity expansion and raw material cost inflation compress the profit elasticity of incremental revenue.

  • Envicool: H1 revenue rose 17.2% YoY to RMB 3.017 billion, while net profit attributable to the parent fell 14.3% YoY to RMB 185 million; data center room-level and cabinet-level thermal management revenues grew 25.8% and 15.3%, respectively. Despite growing liquid cooling demand, domestic price competition, prolonged qualification cycles, and expansion costs have caused profit realization to lag revenue growth. Key watchpoints include gross margin trajectory and operating cash flows.

  • Hengtong Optic-Electric: H1 optical communications revenue surged approximately 130% YoY, with gross profit multiplying 5x from the prior-year period; Q2 optical fiber ASP exceeded RMB 100, with A1 and A2 shipments each accounting for roughly 40%. The 800-metric-ton Phase II optical preform facility in Inner Mongolia is scheduled for commissioning by late 2026; capacity ramp and North American customer qualification represent key growth drivers. Key watchpoints include ASP sustainability and utilization rates.

User's avatar

Continue reading this post for free, courtesy of 404K Semi-Ai.

Or purchase a paid subscription.
© 2026 lihua · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture