404K SEMI-AI Tech Evening Brief 2026-07-07 — Memory Profit Realization, TSMC Capacity Expansion, AI Power Bottlenecks
目录
Premarket Takeaways
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
Semiconductor Equipment/Test
Advanced Packaging / PCB / Substrates / Optical Communications
Internet / Platforms
Software / SaaS
Consumer Electronics / Smart Vehicles
Investment Bank Target Price Changes Over the Past 12 Hours
Samsung Electronics posted record second-quarter profit, yet Korean semiconductor stocks fell; capital is starting to test the durability of AI capex. Tonight’s main thread centers on three areas: whether memory price increases can continue, whether TSMC and ASML’s capacity expansion can absorb AI demand, and whether AI data-center power, packaging, optical interconnects, and financing can keep up.
Premarket Takeaways
The AI hardware trade has entered the earnings validation phase. Samsung Electronics reported second-quarter revenue of KRW171 trillion and operating profit of KRW89.4 trillion, with profit up roughly 19x YoY and above multiple consensus estimates. Yet Samsung Electronics and SK hynix fell, and the KOSPI triggered an intraday program-selling halt. The market is not rejecting the memory upcycle; it is asking how long price hikes, capex, and supply discipline can last.
Marginal information across the semiconductor chain remains constructive. J.P. Morgan raised its TSMC price target to NT$3,100 and expects 2026-2028 capex of US$58 billion, US$78 billion, and US$84 billion, respectively. ASML’s 2027 EUV shipment signal is viewed as a potential breakout catalyst for the stock. On the demand side, the cycle is spreading from GPUs to HBM, CoWoS, ABF, CPO, MLCCs, power, and data-center capacity.
The AI narrative in software and internet is more bifurcated. Bank of America reinstated Shopify at Buy, arguing that payments and checkout infrastructure could become the backend for agentic commerce. Amazon Rufus users convert at roughly 2x the rate of non-Rufus users. Morgan Stanley initiated Kingsoft Cloud at Overweight. The question is also more direct: can AI generate revenue, or does it merely push free cash flow into heavier capex?
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
AI application demand
1. Model traffic is still scaling. Deutsche Bank data show aggregate visits to leading models rising from roughly 4.2 billion in December 2024 to nearly 10 billion in May 2026, while ChatGPT’s absolute visits rose from about 3.7 billion to 5.5 billion.
2. This provides a demand floor for compute, memory, and networking, but it also shifts the investment debate toward revenue quality: whether traffic growth can cover training and inference costs is the next validation point.
"U.S.-headquartered models fell to 35% volume share, but still captured 85%+ of spending"
OpenAI/Claude/Google model ecosystem
1. According to reports, OpenAI, Claude, and Google are offering startups hundreds of thousands of dollars’ worth of compute and incentives to attract enterprise customers.
2. This suggests AI platform competition has entered a “compute subsidies for customers” phase. In the near term, this benefits GPUs, cloud, and data centers; over the long term, the question is whether customers stay and generate paid usage.
DeepSeek V4
DeepSeek V4 launched two variants and supports a 1 million-token context window. This number matters for applications because it means the context bottleneck for long documents, codebases, and enterprise knowledge use cases continues to ease. For hardware, it means inference VRAM, bandwidth, and KV cache costs become harder to ignore.
AI financing structure
1. AI compute access is spreading from a small number of giants to startups and enterprise customers. Market views suggest outstanding AI debt financing could exceed US$7 trillion by 2029.
2. This is not simply a “chip shortage”; the pace of supply release is jointly determined by data centers, GPU leasing, long-term offtake, and interest rates.
"Creative structures are needed to get GPUs in the hands of startups + other companies"
CSP/Cloud Capex
Large cloud providers
1. JPMorgan warned that Amazon, Google, and Meta’s combined free cash flow could shift from roughly US$125 billion in 2025 to negative US$80 billion in 2027, driven by rising capex for data centers, GPUs, and AI infrastructure.
2. This explains tonight’s divergence: AI spending remains strong, but cash-flow pressure will make the market more selective.
Google
Google appears simultaneously in cloud AI, model-traffic, and nuclear-fusion investment threads. The harder investment implication lies on the cloud side: TPUs, Gemini, and cloud-customer subsidies form a closed loop, but whether capex can translate into accelerating cloud revenue remains more important than any single financing round or strategic investment.
Amazon
1. Amazon’s AI and retail data provide support at the same time: Rufus-assisted shoppers convert at roughly 2x the rate of non-Rufus users. Another data point puts Amazon e-commerce revenue at US$490 billion and Walmart retail revenue at US$720 billion, while Amazon trades at 29x forward P/E and Walmart at 37x.
2. The market’s underappreciated angle is advertising, cloud, and the AI commerce entry point, not traditional retail revenue scale.
"Amazon AI-assisted Rufus shoppers convert at ~2X that of non-Rufus shoppers"
Microsoft
Microsoft-related signals center on the enterprise AI orchestration layer. DA Davidson argues Copilot is moving from a model narrative to the “orchestration layer” for enterprise AI, routing tasks among models, internal data, and agents. The investment implication is that software revenue depends on workflow attachment, not just model capability.
AI Cloud/Data-Center Operators
IREN
1. IREN’s AI infrastructure partnership aligned with NVIDIA DSX has been repeatedly cited, with a deployment ceiling of up to 5GW, NVIDIA receiving investment rights of up to US$2.1 billion, and Sweetwater expected to become the flagship DSX architecture deployment.
2. The US$11.3 million per MW framework reprices IREN from a miner into an energized AI data-center asset, but financing, customer contracts, and power delivery remain risks.
"Strategic partnership with IREN to deploy up to 5GW of NVIDIA DSX-aligned AI infrastructure."
TeraWulf
Needham reiterated Buy and raised its TeraWulf price target from US$28 to US$33, citing strong demand reflected in 20-year AI infrastructure leases. Compass Point also raised its target from US$28 to US$40. The contract framework is roughly US$19 billion and 401MW of critical IT load, with initial capacity expected online in 2H27.
CoreWeave/Galaxy Helios
Galaxy Helios Phase I will deliver 200MW of gross power and 133MW of critical IT load to CoreWeave, while CoreWeave has committed to 526MW across Phases I-III. Competition in AI cloud is moving upstream from GPU count to power, campuses, offtake contracts, and financing cost.
Nebius
The bull case for Nebius is end-to-end vertical integration: physical data centers, self-developed server racks, software stack, and inference delivery layer. Materials say “all available capacity is sold out every quarter,” but hyperscaler compute subsidies could pressure external AI cloud pricing. The next issues to watch are unit token cost and renewals.
Kingsoft Cloud
Morgan Stanley initiated Kingsoft Cloud at Overweight with a US$15 price target. It expects 2025-2028 revenue CAGR of 35% and adjusted EBITDA CAGR of 79%, with AI revenue contribution rising from 34% in 4Q25 to more than 40% in 2026 and more than 60% in 2028. This is a Greater China AI cloud beta name; the key follow-through will be AI revenue mix and margin realization.
GPU/CPU/ASIC
NVIDIA
1. NVIDIA responded to rumors of Rubin and Kyber rack delays by emphasizing that its chip roadmap remains intact. This directly dampened market concern over rack-level delivery cadence.
2. Another thread is memory cost: materials indicate 35%-40% of total capex will be spent on memory. Higher cloud capex is not entirely about buying more GPUs; it is also absorbing HBM, DRAM, and system costs.
"NVIDIA SAYS AI CHIP ROADMAP REMAINS INTACT"
AMD
Goldman Sachs raised its AMD price target from US$450 to US$640 and reiterated Buy, citing AI hardware demand and supply constraints. The supply chain also cited AMD Venice CPUs as a driver of advanced packaging demand, with total packaging demand in 2027 written in a range of 1 million to 5.5 million units. The next validation points are MI450/Helios and server CPU share, not one-day sentiment.
Broadcom
Broadcom and Apple signed a new multi-year agreement on July 6, extending custom chip supply through 2031. The 8-K states that Broadcom will continue to develop and supply custom ASICs used in multiple generations of Apple products. Market reports say Apple accounts for roughly 20% of Broadcom revenue, and Broadcom has more than 70% share in the custom AI accelerator design-outsourcing market. If Baltra deploys as soon as next year, Apple’s AI server chip roadmap will continue to pull Broadcom.
Intel/AMD and CXL
Samsung Electronics delayed mass production of CXL 3.1 memory modules, reportedly due to delays in Intel and AMD CPU platforms supporting CXL 3.1/PCIe 6.0. For interconnect-chip companies such as Astera Labs in PCIe 6.0 and CXL 3.1, platform delays will freeze part of the product cycle. Near-term capital still prefers commodity DRAM and HBM.
HBM/DRAM/NAND/SSD/HDD
Samsung Electronics
1. Samsung Electronics’ preliminary second-quarter revenue was KRW171 trillion, with operating profit of KRW89.4 trillion, up roughly 1,810%-1,812% YoY and implying an operating margin of about 52%. Profit was above estimates such as KRW84.2 trillion, KRW84.6 trillion, and KRW87.3 trillion, though revenue was slightly below some high-end expectations.
2. The trading reaction was colder: Samsung Electronics fell as much as roughly 10% intraday and closed down about 7%, while SK hynix also declined. The KOSPI closed down 4.91% and triggered a program-selling halt. The market is asking whether memory prices, AI capex, and capacity have already been reflected.
"Revenue: KRW171.0T (+129% YoY)"
"Operating Profit: KRW89.4T (+1,812% YoY)"
SK hynix
1. SK hynix is viewed as the memory trio’s purest HBM exposure, and institutions expressed up to US$7 billion of aggregate purchase interest in an ADS offering.
2. Another framework says SK hynix plans to list ADRs on Nasdaq with a deal size of roughly US$28-29 billion, 57% HBM share, and 2026 supply already sold out. The investment implication is that global capital can more easily trade the HBM leader, but ADR premiums and discounts in the Korean local shares may create structural spreads.
"SK hynix is the best positioned one of the memory trio"
Micron
1. Micron benefits from the DRAM and HBM cycle, with materials citing 1,398% YoY profit growth. Hiroshima fab expansion is also cited at US$9.3 billion.
2. The debate is that commodity DRAM supply could catch up with demand as early as 2028. Whether pricing power can be sustained long term depends on HBM attachment, LTA pricing, and AI server shipments.
SanDisk/Kioxia/Western Digital/Seagate Technology
NAND and enterprise storage signals continue to heat up. SanDisk profit was US$3.62 billion, while ADATA said third-quarter DRAM contract prices will rise 20%-30% and NAND will rise 35%-40%. TrendForce is more conservative, expecting conventional DRAM to rise 13%-18% QoQ in 3Q and NAND to rise 10%-15%. The direction of price increases is consistent; disagreement over slope is the core trading issue.
Memory industry pricing
UBS raised HBM demand to 33.1 billion Gb in 2026, up 90% YoY, and 58.7 billion Gb in 2027, up 77% YoY. It expects DDR contract prices to rise 32% QoQ in 3Q26 and 18% QoQ in 4Q26. Bernstein says 35% of total bit shipments in FY27-29 have been locked in by long-term agreements, with prices only at a 10% discount to June 2026 ASP.
Foundry
TSMC
1. J.P. Morgan maintained TSMC at Overweight and raised its price target from NT$2,500 to NT$3,100. It expects 2026/2027/2028 revenue of NT$5.34/7.19/8.81 trillion and capex rising to US$58/78/84 billion.
2. The key assumptions are that AI data-center revenue CAGR is raised to 69%, N3 capacity reaches 240k wfpm by end-2028, N2 reaches 170k wfpm by end-2028, and N3/N2 prices rise 8%-10% in 2027.
"Raise Datacenter AI CAGR to 69%"
Samsung Electronics Foundry
Supply-chain reports say Samsung Electronics’ foundry order backlog has risen to KRW50 trillion, or about US$32.7 billion, with customers such as Meta interested in the 2nm process. The point here is not Samsung Electronics’ memory profit, but whether AI ASICs can provide a second profit curve for Samsung Electronics Foundry. Near term, yield, customer switching, and advanced packaging support still need validation.
Semiconductor Equipment/Test
ASML
J.P. Morgan maintained ASML at Overweight with a €1,900 price target and a US$2,200 ADR target. It expects second-quarter revenue of €8.6996 billion and gross margin of 51.7%. The report argues near-term results are not the focus; what the market really wants is the 2027 EUV shipment signal: 90 tools would be mildly positive, while 90-100 tools would be very positive.
"ship 90 EUV tools in '27 would be mildly positive...90-100 would be very positive"
Applied Materials/Lam Research/KLA
Kenanga estimates that each additional US$100 billion of AI data-center investment translates into roughly US$8 billion of WFE spending. Ex-China WFE spending is expected to grow 36% in 2026 and 33% in 2027. Memory WFE is expected to grow 48% in 2026 and another 26% in 2027, with equipment beneficiaries expanding from logic to memory, etch, and deposition.
PSK Holdings/PSK
KB Securities assigned PSK Holdings a KRW220,000 target price and Buy rating, saying its Fluxless Reflow has more than 90% share in the CoWoS process, with revenue mix of 62% Reflow and 33% Descum. PSK’s target price is KRW280,000, with expected three-year sales and operating profit CAGRs of 27% and 34%, respectively. This is equipment beta to CoWoS and HBM TSV expansion.
ASE Technology Holding
Investment-bank checks indicate ASE Technology Holding’s target price has been raised to NT$805, citing LEAP advanced packaging and broad-based price increases in 2H26. LEAP revenue is expected to reach US$3.6bn/US$6.4bn in 2026/2027, driven by AMD Venice, Nvidia Vera, and TSMC oS/CP. The company fits better under advanced packaging than under AMD.
Advanced Packaging / PCB / Substrates / Optical Communications
ABF Substrates
Morgan Stanley raised target prices for Unimicron, Nan Ya PCB, and Zhen Ding, citing an earlier and stronger ABF price-upcycle. Nan Ya PCB raised BT substrate prices by about 20%-30% in 2Q, and ABF by about 10%. The ABF supply-demand model shows the shortage widening to 25% by 2030, with prices expected to rise 20%-25% in 2026, more than 25% in 2027, and 25%-40% in 2028.
"The substrate pricing cycle is proving stronger and earlier than expected"
CPO / Silicon Photonics
Morgan Stanley expects CPO switch shipments to rise from 23,000 units in 2026 to 59,000 in 2027 and 200,000 in 2030, implying a 144% CAGR from 2024 to 2030. TSMC’s PIC capacity plan is to expand from 10,000 wafers/month in 2026 to 15,000 wafers/month in 4Q, and at least 25,000 wafers/month in 2028. The bottleneck is testing and packaging, with wafer-level testing targeted to fall from 6 hours to 3-4 hours.
