404K SEMI-AI Tech Evening Brief 2026-07-08 — AI Infrastructure Broadens, Memory Prices Rise, Optical-Electrical Chain Fills In
目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
CSP/Cloud Capex
AI Cloud/Data Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry and Advanced Packaging
Semiconductor Equipment/Testing
Optical Communications/Optical Chain
High-Speed Interconnect/Connectors/Thermals and Power
MLCC/Passive Components
Robotics/Autonomous Driving and Space/Satellites
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Investment Bank Target Price Changes in the Past 12 Hours
The AI infrastructure trade continues to broaden from a single GPU theme into memory, servers, optical interconnects, substrates, power supplies, and electricity. The most important pre-market items to watch: cloud capex estimates are still being revised higher, memory price increases are starting to flow through to earnings and financing, and supply bottlenecks in optical communications and power are being repriced by the market.
Pre-Market Highlights
Tonight’s technology signals are concentrated: AI demand has not slowed, but capital is shifting from “who has GPUs” to “who can deliver complete AI factories.” Dell, Quanta Computer, Penguin Solutions, AAOI, TSMC, and SK Hynix are sending the same message: orders are pressing into DRAM/NAND, HDDs, optical components, ABF, CPO, PICs, power supplies, and assembly capacity.
Memory is the hardest pre-market variable. Prices and long-term agreements for DRAM, NAND, HBM, and enterprise SSDs are all moving higher. Samsung Electronics and Penguin have already validated this on the profit side, while SK Hynix’s planned US ADR issuance repackages the AI memory bottleneck as a dollar-denominated asset. The risks are also clearer: deleveraging in Korean tech stocks, ADR dilution, oil prices, and long-end rates will amplify volatility.
There is also incremental progress at the application layer. Anthropic’s Claude Code, API consumption, and pre-IPO financial signals continue to support the narrative that compute investment can be profitable. Meta’s image model, Tencent’s Hunyuan and WeChat Agent, and Amazon’s conservative capex guidance all show that AI applications are still transmitting demand back into cloud and hardware supply chains.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
Anthropic/Claude
1. Research reports say Anthropic secretly filed for an IPO on 2026-06-01, with GAAP EBIT reaching US$1bn in 3Q26, a 6% margin, and ARR rising from US$9bn at end-2025 to US$30bn in 1Q26.
2. Claude Code and API consumption are the key variables. API accounts for more than 75% of ARR, net dollar retention is 500%, and ARR per MW is expected to reach US$60mn within the year. The investment implication is direct: if model revenue can cover inference costs, compute procurement by cloud and neocloud companies is not merely cash burn.
ARR rose from US$9bn at end-2025 to US$30bn one quarter later
OpenAI/Frontier Labs
OpenAI and Anthropic’s combined compute demand is estimated to exceed 100GW by 2030, while currently available compute is just above 6GW. Additions of 2.5GW in 2025 and 5GW in 2026 are not in the same order of magnitude as the future 90GW gap. This is the underlying assumption for every hardware chain tonight: if it is wrong, valuations compress; if it is right, supply remains under pressure.
Meta AI
Meta is launching the Muse Image generation model while multiple data points suggest it is still looking for more power and compute capacity. The market previously worried that Meta would cut AI capex, but management messaging looks more like monetizing old cards while continuing to build out new ones. For Nvidia, memory, servers, and neoclouds, the key question is not whether Meta changes the narrative, but whether 2026-2028 contracts and power can land.
"Despite those efforts, Meta is still hungry for even more computing power"
China Open Models
The market is beginning to discuss the possibility that China may restrict overseas access to top AI models, involving open-weight models such as Qwen, Doubao, and GLM. If low-cost models face overseas restrictions, Western companies will have stronger incentives to build complete AI stacks in-house. If they are not restricted, cheap tokens will continue to expand application usage, in turn pulling demand for compute, memory, networking, and power infrastructure.
Tencent Holdings
Citi expects Tencent’s 2Q26 revenue to grow 9.3% YoY to RMB201.7bn, with Non-GAAP net profit up 5.1% YoY to RMB66.25bn. On the AI side, daily token consumption increased 20x after the Hy3 preview, and WorkBuddy reached 8.85mn MAU in March. Tencent’s investment implication is not simply model launches; WeChat, cloud, enterprise productivity, and buybacks all support valuation.
CSP/Cloud Capex
Cloud Capex
Jefferies estimates that the four major US hyperscalers will spend about US$700bn on capex this year and more than US$800bn next year. Including Oracle, Anthropic, OpenAI, and neoclouds, the total could exceed US$1tn next year. The pressure point in this number is cash flow: the four major cloud companies’ capex/operating cash flow ratio is expected to rise from 41% in 2023 to 92% in 2026.
"capex is estimated to surpass US$1tn next year"
Google/Meta/Amazon
BofA raised capex estimates for three CSPs over the next three years, saying the combined total could reach US$705bn in 2027 and US$816bn in 2028. Google’s power usage grew 37% YoY in 2025 to more than 43TWh, with PUE maintained at 1.09. Amazon’s actual capex has exceeded guidance for three consecutive years, by an average of 7.2% above the midpoint. Cloud providers are still locking in supply through debt, power, and data centers.
Amazon
Amazon’s capex guidance has been conservative over the past three years: FY23 guidance was US$50bn versus actual US$52.7bn; FY24 guidance was US$75bn versus actual US$83bn; FY25 guidance was US$125bn versus actual US$131.8bn. If the market looks only at guidance, it may underestimate AWS’s pull on servers, ASICs, networking, power, and data center construction.
Microsoft
Microsoft has two layers of signals: first, prior calendar-2026 capex of about US$190bn, making it still a major source of storage and server demand; second, some Excel and Outlook functions have already replaced external models with the in-house MAI model, processing tens of thousands of prompts per week. This creates pricing pressure for model companies, while for the hardware chain, even proprietary models still consume inference compute.
AI Cloud/Data Center Operators
Dell Technologies
Evercore raised Dell’s target price to US$500 and maintained Outperform, citing AI infrastructure demand that is materially above supply, with the 2027 supply-demand imbalance potentially worse than in 2026. Dell already has 5,000+ enterprise AI factories in production, up more than 50% over the past six months. Constraints are broadening from GPUs to DRAM/NAND, CPUs, HDDs, and optical components.
"demand remains materially ahead of supply"
Penguin Solutions
Penguin’s FY26 Q3 revenue was US$478.7mn, above consensus of US$421.4mn; adjusted EPS was US$0.84, above consensus of US$0.56. AI-driven businesses accounted for 74% of revenue and grew 104% YoY. Integrated Memory revenue was US$275mn, up 111% YoY. This company validates the storage content in system integration, not merely upstream price increases.
CoreWeave/Nebius
Neocloud trades are under short-term pressure, but fundamental divergence matters more. CoreWeave has backlog of US$99.4bn, Q1 2026 revenue of US$2.1bn, up 112% YoY, and adjusted EBITDA margin of 56%. Nebius’s Q1 AI cloud revenue grew 841% YoY and it guided to year-end 2026 ARR of US$7bn-9bn. Next, watch token output per card, financing costs, and customer concentration.
TeraWulf
TeraWulf’s 401MW Kentucky data center lease payments may receive investment-grade credit support. Its wholly owned Lake Mariner site has construction cost of US$2.2mn per MW, below Abernathy’s US$5mn. For AI data center operators, the battleground is shifting from “how many MW” to “who has lower power, contract, capital, and delivery costs.”
GPU/CPU/ASIC
Nvidia
Nvidia’s marginal change is still Rubin. BofA expects HBM value per rack to rise from about US$150,000-300,000 for Blackwell to about US$400,000 for Vera Rubin, with Vera Rubin servers entering mass production at Quanta this quarter. The supply chain also says Nvidia denied Kyber rack delays, and the CCL upgrade trend remains unchanged. This supports HBM, ABF, CCL, and server ODMs.
AMD
Morgan Stanley expects AMD GPU shipments of 2.4mn units this year and 2.7mn next year. AMD remains the core AI accelerator substitution narrative, but tonight’s stronger conclusion comes from the supply chain: a 20% increase in cloud-provider GPU prices is being read as rising demand, not merely higher memory cost. AMD still needs shipments, software stack progress, and customer concentration to validate the story.
Broadcom/MediaTek
The ASIC debate centers on margin. Reports say MediaTek’s TPU margin is about 35%, while Broadcom’s is about 65%. This means that as Google TPU, custom ASICs, and cloud providers’ in-house chips heat up, the market will increasingly ask whether Broadcom’s high margins can be sustained over the long term. Apple’s multi-year agreement with Broadcom exceeding US$30bn also reinforces long-term order signals for wireless connectivity and custom chip manufacturing.
Intel
Intel is being placed in two chains: server CPUs and advanced packaging. Morgan Stanley believes Intel’s EMIB competition may push TSMC to bring forward CoPoS adoption to 2028 for Feynman. At the same time, the market is enthusiastic about Intel’s SOTP narrative, but without order, capacity, and profit validation, valuation elasticity also brings higher volatility.
HBM/DRAM/NAND/SSD/HDD
Micron
BofA reiterated a US$1,550 target price on Micron, saying memory accounts for 35%-40% of cloud AI capex, 2-3x historical levels, while memory stocks still trade below 10x forward P/E. Bernstein’s target price is US$1,300, noting that 2QCY26 traditional DRAM contract prices rose about 74% from 1QCY26. Micron’s risk is that the near-term rate of change peaks, but AI inference and enterprise SSDs are still consuming supply.
"Memory now represents 35-40% of cloud AI capex"
Samsung Electronics
Samsung Electronics’ preliminary 2Q26 revenue was KRW171tn, with operating profit of KRW89.4tn, up 28% and 56% QoQ respectively. Multiple institutions estimate DRAM operating margin above 80% and NAND at about 69.5%. Mirae Asset maintains Buy with a KRW550,000 target price. The market is worried about a cycle peak, but current profits are mainly from AI memory pricing and capacity reallocation.
SK Hynix
SK Hynix plans to raise about US$28bn through a US ADR issuance, with 10 ADRs representing one common share, equivalent to about 2.5% of market cap. Proceeds will go to Korean chip fabs and EUV equipment, precisely the constrained parts of HBM. Near term, this is dilution and deleveraging pressure; medium term, it dollarizes the AI memory bottleneck and may bring an accessibility premium.
SanDisk/Western Digital/Seagate Technology
The NAND and HDD trade is rotating from cyclicals back into AI infrastructure. Bernstein rates SanDisk Outperform with a US$3,000 target price. Enterprise SSD demand and DDR4 8Gb shortages could drive 3Q26 contract prices up 50% QoQ. But SanDisk, Western Digital, and Seagate Technology have already pulled back sharply in the short term, and the market is asking how long the price-increase curve can continue.
Kioxia
Bernstein maintains Underperform on Kioxia with a JPY40,000 target price. The debate is that NAND is more exposed to Chinese competition and supply recovery. If enterprise SSDs and high-capacity NAND remain tight, Kioxia has pricing leverage. If LTAs and new capacity begin to pressure prices in 2027-2028, Kioxia’s valuation recovery will be weaker than that of DRAM/HBM leaders.
Foundry and Advanced Packaging
TSMC
TSMC’s 2Q26 revenue is expected to rise 11% QoQ, with gross margin around 68%; 3Q26 revenue is expected to rise 13% QoQ. 2026 capex is expected at US$56bn and 2027 at US$73bn, below consensus of US$80bn-85bn, with the constraint possibly coming from EUV supply. The target price was raised to NT$2,900, based on N2/N3/CoWoS expansion accelerating for AI demand.
"2Q26 GM: ~68%"
Rapidus
Rapidus said its foundry pricing will be lower than TSMC’s, which is its entry strategy as Japan’s second source for advanced process capacity. Low pricing can win trials, but the real barriers in advanced foundry remain yield, delivery time, ecosystem, and long-term stable mass production. Before customer and capacity figures are available, this can only be treated as a supply-chain diversification signal, not evidence that TSMC’s share is being materially rewritten.
ABF/Substrates
Morgan Stanley says the ABF substrate supply gap will widen to 25% by 2030. Servers, AI GPUs, AI ASICs, and networking applications will account for more than 80% of ABF value share by 2030, up from about 60% in 2025. CY26/CY27/CY28 ABF prices are expected to rise 20%-25%, 25%+, and up to 25%-40% YoY. Substrates have moved from supporting components to constraints on the pace of AI shipments.
Semiconductor Equipment/Testing
Applied Materials/Lam Research/KLA/Teradyne
BofA believes semiconductor momentum will broaden from memory and compute into equipment in 2H26, naming Applied Materials, Lam Research, KLA, and Teradyne. The logic is not one quarter of orders, but memory expansion, EUV, advanced packaging, and longer test times jointly raising WFE and testing demand. The risk is short-term deleveraging in semiconductor stocks, with equipment names likely sold first as high-beta exposure.
ASML
TSMC’s 2027 capex may be constrained by EUV supply, which instead highlights ASML’s bottleneck role in the AI expansion cycle. If EUV deliveries limit advanced process capacity, the AI-related expansion pace at TSMC, Intel, and Samsung Electronics will all be constrained by equipment supply. But ASML is also pulling back with the SOX in the short term, and its valuation is sensitive to rates and capex delays.
SCHMID/PCB Equipment
SCHMID received repeat orders of more than EUR37mn from Chinese customers for HDI-ML and mSAP capacity expansion used in AI server boards and optical module applications. 2026 order revenue rose from EUR44.3mn to EUR81.7mn. This signal is small, but the direction is clear: AI server boards and optical module expansion have entered the equipment procurement phase.
Optical Communications/Optical Chain
AAOI
AAOI’s Q1 2026 revenue was US$151.1mn, up 51% YoY. Q2 guidance is US$180mn-198mn, and full-year revenue guidance is above US$1bn. 800G demand exceeds capacity, and the combined 800G and 1.6T capacity target is 500,000 units per month by end-2026. Whether the long-term CPO narrative can be realized depends first on the 800G/1.6T capacity ramp.
"800G demand is currently outpacing production capacity"
CPO/PIC
Morgan Stanley/Yole expect CPO switch deployments to rise from 5,000 units in 2025 to 200,000 units in 2030, a 144% CAGR. TSMC’s PIC monthly capacity is said to have increased from 500 wpm to 10,000 wpm, with 15,000 wpm expected in Q4. Optical interconnect is no longer just a long-dated concept; it is entering wafer-capacity and switch-deployment metrics.
InP substrates
The supply-demand gap for optical-communications InP substrates is large: effective global capacity this year is about 600,000-750,000 wafers, versus demand of 2.6 million-3.0 million wafers. Prices for 2-inch optical-communications InP substrates have risen from about US$800/wafer in 2025 to US$2,300-2,500/wafer in April 2026. If CPO and 1.6T continue to scale, InP may become the next pricing anchor in the optical chain.
Marvell Technology/Credo/Coherent
High-speed interconnect is moving from the board edge into the chip core, and SerDes, optical links, AEC, and CXL are all being repriced. UBS expects the CXL-related ASIC attach market to reach US$7 billion-10 billion by 2030. Names discussed by the market as beneficiaries include Marvell Technology, Credo, Astera Labs, and Coherent. The next step is to watch actual customers and design-ins.
High-Speed Interconnect/Connectors/Thermals and Power
Quanta Computer
Quanta Computer’s June revenue was NT$385.19 billion, up 102.9% YoY; Q2 revenue was NT$1.04 trillion, up 105.6% YoY. Vera Rubin servers enter mass production this quarter, visibility from top-tier CSP demand extends to 2028, and the company also plans to open three more AI server assembly plants in the US. On the ODM side, both orders and capacity are already expanding in parallel.
