404K SEMI-AI Tech Evening Brief 2026-07-06 — Memory Price Hikes, AI Rack Delays, Cloud Capex Revised Up Again
目录
Pre-Market Key Points
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
Wafer Foundry
Semiconductor Equipment/Test
Optical Communications/Optical Chain
PCB/Connectors/Thermal and Power
MLCC/Passive Components
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Investment Bank Target Price Changes Over the Past 12 Hours
The AI infrastructure trade entered a more differentiated phase today: memory earnings continue to be revised up, Nvidia’s rack-level roadmap is being slowed by manufacturing bottlenecks, and cloud vendors are still locking in power, memory, and external compute capacity. In the near term, capital is still buying AI supply shortages; over the medium term, whether pricing power can hold will depend on the slope of DRAM/NAND price hikes, CPO mass-production yields, and data-center free cash flow.
Pre-Market Key Points
AI infrastructure is not cooling down. The main theme has become “where the real bottlenecks are.” Meta, Anthropic, AWS, and Microsoft continue to add compute and external capacity, while memory, MLCCs, InP, optical modules, power supplies, PCBs, and testing are all being tightened at the same time. The market is starting to distinguish between two things: demand has not disappeared, but the engineering cadence for Kyber, CPO, 800VDC, and high-layer-count PCBs has not kept up.
Memory is today’s firmest pricing line. Samsung Electronics, SK Hynix, Micron, SanDisk, and Western Digital face the same issue: AI servers are consuming HBM, server DRAM, NAND, and enterprise SSD capacity ahead of schedule. The disagreement is also here: Meritz and UBS continue to revise up earnings and target prices, while BofA and TrendForce caution that DRAM/NAND ASP sequential growth may slow after 2Q26.
For software and internet names, the AI evidence is more about revenue quality. Meta’s ad recommendation system is viewed as a case where AI capex has already generated commercial returns; Datadog was downgraded by Bernstein while its target price was raised, indicating strong AI Lab demand but potentially slower core cloud monitoring demand after Q3. Consumer electronics and smart vehicles retain only marginal changes backed by hard information, avoiding the framing of ordinary new-product and sales news as an AI mainline.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
OpenAI/Anthropic
1) Model-layer revenue continues to exceed expectations. Anthropic’s ARR reached US$45bn in May, versus only US$1bn about 15 months earlier, while OpenAI is at roughly US$25bn.
2) Another compute-demand metric shows that in December 2025, the two companies expected combined ARR by end-2026 to be US$61bn, but halfway through 2026 it had already exceeded US$100bn. Model revenue is running ahead of procurement plans, which is more favorable in the short term for AI clouds such as CoreWeave and Nebius. The risk is that once customer revenue slows, contract rates will come under pressure first.
"capacity procurements have fallen short"
Model-Layer Concentration
The AI stack is concentrating toward the middle model layer within the three layers of “chip-model-application.” Materials indicate that other model companies have almost no meaningful revenue, which will make application companies, enterprise customers, and chip companies unwilling to accept procurement control by a small number of model buyers. The investment implication is that the hardware chain wants the model layer to remain competitive; otherwise, the bargaining counterparties for GPU, ASIC, and memory suppliers will become increasingly concentrated.
"If you are a chip company, you do not want a world where only two companies buy your silicon"
AI Application Costs
Heavy users consumed US$135 of Fable credits in about two hours. Although this is only an application-usage sample, it shows that the cost pressure of generative applications will quickly feed through to subscription quotas, inference resources, and cloud bills. For application companies, revenue growth must cover model calls, storage, inference, and distribution costs at the same time. For upstream suppliers, the faster credits are consumed, the easier it is for that to translate into short-term inference compute demand.
Google TurboQuant/DeepSeek MLA
Software optimization is beginning to feed back into hardware TAM. Google TurboQuant can reduce KV cache memory demand to one-sixth of the original level and accelerate GPU inference by up to 8x; DeepSeek MLA claims to reduce KV cache usage by 90%-95%. This will reduce memory demand per task, but it may also lower inference prices and expand total call volume. For the storage chain, the key is tracking which is faster: “lower unit demand” or “higher total usage.”
"TurboQuant compresses KV cache, cutting memory needs by up to 6x"
CSP/Cloud Capex
Meta
1) Mizuho says that in addition to Meta’s US$125bn-US$145bn 2026 capex budget, the company has signed multiple external compute contracts: a five-year US$27bn agreement with Nebius in March, a US$21bn agreement with CoreWeave in April extending to 2032, and an approximately 1.6GW capacity agreement with Crusoe on June 18.
2) Other materials define Meta as AI’s first killer application, because GPU capex has already translated into advertising share recovery through ad recommendation, content ranking, and signal completion. The investment implication is that Meta’s AI spending has not appeared as a pure cost item; it is binding ad revenue and compute procurement together.
"Meta has signed multiple external compute contracts on top of its US$125bn-US$145bn 2026 capex budget"
Amazon/AWS
AWS raised its 3Q26 ASIC server shipment plan by 20%-30% versus the original plan. Trainium 3 began small-volume shipments in May, and L11 rack-level orders enter mass production in July; Trainium 2 is sold out, Trainium 3 is almost fully booked, and customers are already booking Trainium 4. This signal directly benefits ASICs, server motherboards, thermal solutions, chassis, assembly, MLCCs, server DRAM, and FC-BGA. The pressure is that the ramp of in-house chips will squeeze part of the general-purpose GPU demand narrative.
"AWS has raised its ASIC server shipment forecast 20%-30%"
Microsoft
Wolfe believes Microsoft’s disclosed timing and scale may support its status as one of Micron’s early SCA customers. Wolfe estimates Microsoft CY26 capex will increase by US$15bn, of which roughly US$10bn comes from higher memory prices, assuming memory previously accounted for about 30% of short-life asset capex and prices rise by about 40%. Microsoft’s issue is not whether demand exists, but that memory lock-ins and component inflation may push FY27 free cash flow into negative territory.
Google
Ahead of TSMC’s July 16 earnings, the market is watching whether full-year US dollar revenue growth will be revised up from “above 30%” to 34%-36%, with a bullish scenario challenging 40%. Google TPU v9 and Intel EMIB-T are being placed in the same supply-chain watchlist: if TPU v9 can ship at scale as planned, carry HBM4-class memory, and achieve acceptable yields, the advanced packaging bottleneck will not be priced solely by TSMC.
Five Major Cloud Vendors
Tech fund inflows and cloud capex are reinforcing each other. AI capex from Google, Amazon, Meta, Microsoft, and Oracle is expected to exceed US$800bn in 2026 and rise further to US$1.1tn in 2027; another measure says combined 2026 capex from the four major US tech giants will rise about 80% YoY to roughly US$700bn. Spending at this level will continue to support GPUs, ASICs, HBM, networking, optical communications, liquid cooling, power supplies, and MLCCs, but it will also increase depreciation, leasing, and cash-flow pressure.
AI Cloud/Data Center Operators
CoreWeave/Nebius
AI cloud’s near-term pricing power comes from model-company revenue running fast. Materials indicate that combined ARR from OpenAI and Anthropic has exceeded US$100bn and that original procurement plans are insufficient, so they are willing to pay premiums for additional compute. CoreWeave has nearly US$100bn of backlog and revenue growth above 100% YoY, but heavy debt, losses, and hyperscaler self-build remain valuation caps. Nebius benefits from the Meta contract, but delivery, financing, and power milestones still need to be verified quarter by quarter.
Anthropic Australia Tender
Anthropic has launched a 1.4GW AI data-center tender in Australia, with a construction scale of about US$15bn. Around 1GW is planned to come online before the end of 2027, with a decision expected within six weeks, and the project may be split into 4-5 contracts. IREN has 800MW of power access at Bundey in South Australia, but energization starts in 2028, later than the end-2027 milestone. This shows that the key threshold for AI cloud investment is not just GPUs, but the “power year” and “commissioning year.”
"Real power, wrong year."
Crusoe/Meta Capacity
Meta’s approximately 1.6GW capacity agreement with Crusoe reinforces one judgment: leading platforms are still locking in long-term external compute while building their own capacity. If compute were truly oversupplied, platforms would not continue signing multi-year contracts beyond annual capex budgets. For data-center operators, contract duration, grid-connection timing, and power costs are more important than simple cabinet scale.
South Korea Data Center Power
South Korea’s “three major super projects” target 8.4GW by 2029 and an additional 10GW by 2035, totaling 18.4GW of AI data centers. GNC Energy has about 70%-80% domestic market share, announced an approximately KRW30bn order on July 2, and has secured confirmation letters of about KRW100bn. The benefit to power-security companies is clear, but whether orders can convert into high profit depends on project starts, grid connection, and the backup power configuration ratio.
GPU/CPU/ASIC
Nvidia
1) Nvidia’s Kyber NVL144 rack has been cited by multiple materials as delayed by more than 12 months to 2028 due to manufacturability issues. The NVL72x2 back-to-back rack was canceled after opposition from major cloud vendors, and Rubin Ultra may be reduced from 4 dies to 2 dies.
2) Jefferies judges that no Kyber in 2027 has become highly likely, and that Rubin Ultra will continue using Oberon/NVL72, which is negative for the overall PCB supply chain and positive for copper-cable manufacturers.
3) This is not an AI demand collapse, but a rack-level manufacturing issue involving PCB midplanes, CPO, 800VDC, cooling, and yields. The “on-time cross-generation” premium in Nvidia’s valuation needs to be discounted, while AMD’s MI series and Google TPU gain a substitution window.
4) Nvidia supply-chain investment: market rumors suggest Nvidia may lock in supply through additional investment and seek more favorable ASPs. The materials do not provide investment amounts or contract terms, so this can only be treated as a signal of supply-chain capital binding. If true, the strategic value of optical communications, high-speed interconnects, and key materials suppliers would rise; the risk is that the rumors are unverified and cannot be written directly as signed orders.
"Kyber NVL144 rack architecture has been delayed more than 12 months to 2028 due to manufacturability issues"
Intel
1) Intel confirmed price increases for some consumer and server CPUs, with some Xeon processors rising by more than US$1,000. Price hikes can ease cost pressure, but they will raise server procurement costs for cloud vendors and enterprises.
2) Intel EMIB-T is viewed as a path to attack TSMC’s advanced packaging bottleneck. The key observation point is whether Google TPU v9 can pair with HBM4-class memory and ship at scale. If validated, Intel’s value will lie in advanced packaging and heterogeneous integration, not only traditional CPUs.
AMD/Google TPU
Nvidia’s Kyber delay creates a window for AMD’s MI series and Google TPU, but current materials do not provide specific order or share elasticity. The investment judgment should be framed as “more optional supply,” rather than “Nvidia’s share must decline.” The real items to track are hyperscaler procurement lists, MI platform delivery timing, TPU v9 yields, and HBM4-class memory support.
ARM/Qualcomm
High-signal materials do not provide new target prices, orders, or financial changes for ARM or Qualcomm. Qualcomm remains pulled between weak handset terminal demand and data-center catalysts. Its current positioning remains as a GPU/CPU/ASIC watch item; ordinary investor-day expectations should not be written as formal incremental information.
Cerebras
Cerebras remains an AI dedicated accelerator watch item, but the materials do not add new orders, revenue, target prices, or customer data. If OpenAI or enterprise inference customer progress is later confirmed, it can be included as a company-specific section; for now, it should not be added merely to fill out the chain.
HBM/DRAM/NAND/SSD/HDD
Samsung Electronics
1) Meritz raised its target price for Samsung Electronics from KRW420,000 to KRW500,000 and maintained Buy, forecasting 2Q26 operating profit of KRW90.1tn, revenue of KRW182.1tn, and an operating margin of 49.5%, above consensus of KRW75tn-KRW84tn.
2) DRAM BG/ASP are +12%/+50%, respectively, while NAND BG/ASP are +3%/+63%; DS pre-bonus operating profit is KRW109.5tn, and memory pre-bonus operating profit is KRW112tn, but LSI/foundry losses exceed KRW2tn.
3) The risk is that memory price increases will raise component costs for end markets such as MX and home appliances, so a semiconductor profit surge and pressure on terminal demand may appear at the same time.
"2Q26E operating profit KRW90.1tn"
SK Hynix
UBS expects SK Hynix operating profit in 2026/2027/2028 to be KRW327tn/KRW623tn/KRW667tn, respectively, equivalent to about US$213.5bn/US$406.8bn/US$435.5bn. The company is also reported to plan a sale of 177.9mn ADS, representing 17.79mn ordinary shares, potentially raising about US$28bn. If the issuance proceeds, the funds will be used for new chip plants and equipment procurement in South Korea. The risk is that financing expands supply expectations, but actual capacity still takes time to come online.
Micron
1) Micron’s Hiroshima project involves investment of about US$9.3bn to produce 1γ DRAM and HBM, with shipments expected around summer 2028. The Japanese government will provide support of up to about US$3.1bn, with total support of about US$4.8bn.
2) Its 2026 HBM is described as fully presold, with new customer orders delayed to late 2027 and key customers receiving only 50%-67% of requested volumes.
3) Wolfe also links Microsoft’s capex upgrade to Micron’s SCA customer relationship, indicating that the debate around Micron has shifted from “whether demand exists” to “whether pricing can hold and when expansion can ease shortages.”
"Memory is no longer just about PCs and smartphones"
SanDisk/Western Digital
SanDisk and Western Digital continue to be driven by NAND and enterprise SSD pricing. TrendForce data show enterprise SSD revenue up 86.1% to $18.46bn, with contract prices up roughly 80% in the same quarter; BofA estimates 512Gb NAND wafer contract prices at about $25, roughly 10x the February 2025 trough of $2.50. The debate is that TrendForce also cut its 2Q26 NAND ASP QoQ forecast from 70%-75% to 55%-60%, suggesting the price-increase slope may shift from extremely strong to strong but slower.
DRAM/NAND Industry
J.P. Morgan’s May WSTS data show global semiconductor sales of $131.9bn, +16.1% MoM and +118.8% YoY; DRAM sales +27.7% MoM, Flash sales +39.8% MoM; ASP +16.7% MoM and +85.0% YoY. The strength is driven almost entirely by memory pricing, not synchronized volume growth across all end markets. The investment implication is that memory has the highest near-term earnings leverage, but industry revenue quality is highly dependent on ASP.
"May industry sales growth accelerated to +119% Y/Y"
Kioxia/YMTC
Kioxia began sample shipments of its 10th-generation BiCS FLASH in July, increasing bit density by 60%, lifting interface speed to 4.8Gb/s, and reducing power consumption. YMTC entered the global SSD configuration for Lenovo’s ThinkBook 14 G9; the signal is not volume, but global PC BOM validation. Both lines show NAND competition will not be determined only by price, but also by power consumption, interface speed, customer qualification, and supply stability.
Wafer Foundry
TSMC
The key variable ahead of TSMC’s earnings is whether full-year US-dollar revenue guidance can be raised. Multiple materials place Nvidia, AMD, Apple, AWS ASIC, and high-performance computing demand on the same order trajectory, with the optimistic scenario challenging +40% US-dollar revenue growth. The risk is not demand, but whether advanced packaging, CoWoS, materials, and the domestic supply chain can keep up.
Samsung Electronics Foundry
Samsung Electronics Foundry achieved its first monthly profit since 2023 in June 2026, driven by HBM4 base die ramping on 4nm, yield of around 80%, and higher fab utilization. External orders include Tesla AI6 and Groq inference chips, while discussions are underway with Meta and Anthropic. This signal shows HBM will bind advanced nodes and foundry profitability together, but Samsung Electronics’ 1Q26 foundry share was still only about 6.5%, versus TSMC at about 72.3%, so the industry structure has not reversed.
"Samsung Electronics Foundry just turned profitable in June 2026"
Intel Foundry/Advanced Packaging
Intel’s EMIB-T opportunity is in advanced packaging, not simply catching up on leading-edge nodes. If Google TPU v9 can ship at scale with HBM4-class memory, Intel can gain validation in the tightest segment of TSMC’s advanced packaging capacity. If shipments, yields, or customer revenue do not appear in parallel, this line remains a technology narrative.
Semiconductor Equipment/Test
Applied Materials/KLA/ASML/Lam Research
In Goldman Sachs’ preview of 2Q semiconductor earnings, the firm said expectations are still being revised up, but stock selection matters more after the sector’s sharp rally. It is positive on Applied Materials for strong DRAM, pricing tailwinds, and visibility through 2028, while more cautious on KLA because a WFE mix with high DRAM exposure may be unfavorable for end-market chains such as ARM. Both J.P. Morgan and Citi indicate WFE is concentrating toward AI, HBM, NAND upgrades, and advanced packaging; for the equipment chain, order visibility matters more than single-quarter valuation expansion.
Japanese Semiconductor Equipment
The Semiconductor Equipment Association of Japan raised its FY2026 sales forecast for Japan-made semiconductor equipment, expecting sales to exceed about $40bn for the first time and set a new high. Incremental demand comes from AI server advanced logic, memory investment, and equipment upgrades. This data point has limited direct attribution for ASML, Applied Materials, and Lam Research, but confirms the cycle for Japan-linked names such as Tokyo Electron, Advantest, DISCO, and Lasertec.
