404K SEMI-AI Evening Brief 2026-07-15 — ASML Raises Guidance, Memory Shortage Extends, AI Applications Enter a Cost War
目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
AI Cloud/Data-Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Semiconductor Equipment/Testing
Optical Communications/Optics Value Chain
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Investment-Bank Target Price Changes Over the Past 12 Hours
ASML’s results and capacity expansion demonstrate that AI capex continues to flow upstream into semiconductor equipment, while memory, testing, and optical interconnects are emerging as the next bottlenecks. At the application layer, winners and losers are increasingly determined by traffic, token costs, data security, and products capable of generating real revenue.
404K SEMI-AI | 2026-07-15
Pre-Market Highlights
ASML provides the strongest fundamental anchor for technology stocks before the market opens. The company not only delivered better-than-expected second-quarter results, but also significantly raised its full-year sales and gross-margin guidance and plans further EUV and DUV capacity expansion in 2027. The equipment segment shows no sign of slowing AI investment, while market attention is gradually shifting from GPUs toward lithography, memory, testing, and interconnects.
The memory supply chain presents greater divergence. HBM is crowding out wafer and packaging capacity, keeping commodity DRAM and NAND supply tight. However, rapidly rising prices are also driving capacity expansion, and new supply in 2027–2028 could bring the upcycle to an end. The most important evidence now is not the grand narrative, but long-term agreements, actual orders, yields, and when incremental capacity begins generating revenue.
The application layer is entering a more practical elimination round. User growth for Codex and ChatGPT Work, alongside traffic gains for Gemini and Claude, indicates that demand remains strong. At the same time, enterprises are beginning to scrutinize token costs, data retention, and the leakage of proprietary knowledge. Traditional software budgets have not disappeared, but they are increasingly flowing toward AI, security, and irreplaceable workflows.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI
1) Weekly active users for Codex and ChatGPT Work increased from 6 million on July 12 to 8 million on July 14, with another report suggesting the figure could soon reach 9 million. High-frequency usage is turning AI coding from a trial tool into an everyday work interface.
2) On the product side, the company continues to expand usage limits and collect feedback. The next questions are whether user growth can translate into stable paid adoption and whether inference costs can preserve attractive value for money.
“Too much choice can also be a problem. But it looks like we may reach 9 million soon. Should we reset the ChatGPT Work and Codex usage limits again, or leave some headroom for now?”
Anthropic
1) One operating datapoint suggests the company has achieved and sustained positive free cash flow while still planning to invest as much as possible, indicating that reinvestment by frontier-model companies may not depend entirely on external financing.
2) Its share of website traffic rose from 1.6% 12 months ago to 9.2% one month ago. However, traffic does not equal paid revenue, and enterprise adoption and unit inference economics still require further validation.
Google Gemini: Its share of generative-AI website traffic rose from 9.1% 12 months ago to 27.8% one month ago, far outpacing ChatGPT’s growth. Google’s established search gateway and in-house TPUs provide dual advantages in distribution and cost. The risk is that website visits cannot substitute for API usage, enterprise contracts, or actual profits; the key question is whether traffic continues to convert into cloud revenue.
Meta Models: An unnamed model scored a perfect 30/30 at the 2026 Asian Physics Olympiad, tying the top three human contestants for first place. Improved capabilities support Meta’s efforts to embed models into advertising, recommendations, and agents, but a single competition result does not demonstrate product retention, monetization, or lower inference costs. Validation metrics include model usage, advertising conversion, and inference costs.
Daum, Upstage, and FuriosaAI: South Korea’s domestic Solar model and 24 RNGD chips have been deployed for search-summary services, processing approximately 500 million tokens per day. Daum plans to increase coverage from approximately 20% to more than 50%. The commercial deployment demonstrates that a small-scale domestic technology stack can operate effectively; the key issues after broader rollout are accuracy, latency, and cost.
“Daum’s AI Overview combines search-engine results with the Solar model to summarize user queries and provide key information. Users no longer need to click through multiple links individually, as the AI reads and summarizes the relevant documents.”
AI Cloud/Data-Center Operators
CoreWeave: The company has signed long-term storage agreements with Micron and SanDisk that include price floors and is exploring put options and other instruments to hedge against future declines in chip prices, though none have yet been executed. Supply-security agreements lock in availability during shortages but could become above-market purchasing obligations once new capacity comes online around 2028. Contract duration and minimum purchase volumes also warrant monitoring.
IREN: The debate centers on contract coverage and financing. Hyperscalers such as Meta can fund their own infrastructure with positive free cash flow, while IREN continues to use at-the-market equity issuance to support capex and has been criticized for failing to secure more long-term agreements during the upcycle. Strong demand does not necessarily translate into shareholder returns. Key indicators include long-term contracts, per-share dilution, and construction returns—particularly when capex begins converting into distributable cash.
“If this trend continues, it could place enormous pressure on the margins of independent data-center operators.”
Firmus and Sharon AI: The two companies are purchasing Nvidia’s technology stack to build AI factories in Australia, supported by offtake guarantees, revenue-sharing arrangements, and credit support. Sharon AI plans to raise US$300–400 million after previously securing US$1.6 billion. The model can accelerate compute deployment, but it also brings demand realization, interest expense, and equity-dilution risks onto the balance sheet earlier.
Nebius: The company is described as continuing to execute and grow, but its shares remain under near-term pressure as the market digests the financing intensity of emerging cloud providers, data-center policy, and hyperscalers’ ability to build infrastructure internally. Without new order, customer, and margin figures, price divergence should be treated only as a monitoring signal, not a substitute for operating validation. Confirmation from formal contracts or financial results is still required.
AI Data-Center Financing: Morgan Stanley estimates that AI- and data-center-related bond issuance has reached US$336 billion since the beginning of 2026, already exceeding the US$217 billion issued in all of 2025. Its aggressive full-year forecast is US$580 billion, up 167% year over year. Financing remains available to support construction, but interest rates and credit risk are becoming part of the cost of compute expansion.
GPU/CPU/ASIC
Nvidia
1) Jensen Huang said Vera Rubin has entered production and that “enormous capacity is coming,” directly addressing supply-chain concerns that custom circuit boards could delay rack deliveries. The true validation point is the pace of full-rack shipments, not whether individual GPUs are in production.
2) Morgan Stanley channel checks indicate that near-term growth remains strong, but memory shortages have forced the company to halve LPDDR5 capacity per rack, showing that system-level bottlenecks are beginning to affect product design.
AMD
1) Helios is configured with 72 MI450 GPUs per rack. Oracle has committed to deploying 50,000 GPUs, while Meta and OpenAI are each planning up to 6 gigawatts of AMD compute capacity. First-quarter data-center revenue reached US$5.8 billion, up 57% year over year.
2) Citi forecasts 2027 AI GPU revenue of US$33 billion and shipments of approximately 1.9 million units. Risks include the product launching three to six months after Vera Rubin and the possibility that planned power capacity does not fully convert into purchases.
“AMD is not trying to defeat Nvidia; it is trying to establish itself firmly as the number-two supplier in every hyperscaler’s AI infrastructure stack.”
Intel
1) Following reported improvements in 18A yields, approximately 80%–90% of Nova Lake dies could be moved back to Intel’s own fabs, which currently have monthly 18A capacity of approximately 30,000 wafers. If realized, this would improve both margins and foundry credibility.
2) The July 27 DAC presentation will showcase 14A, 18A-P, EMIB-T, and Foveros. The market will focus on whether these technology demonstrations can translate into external customers and volume-production revenue.
Google TPU: The tenth-generation 2nm TPU, “Icefish,” is reportedly being jointly designed by Google and MediaTek. The compute die is expected to use TSMC’s 1.4nm process, while the I/O die will use Samsung Electronics’ 2nm process and handle HBM data transfer, with mass production potentially beginning as early as 2028. Outsourcing backend design reflects the shortage of advanced-node engineering talent, but the project’s scope remains unconfirmed.
Samsung Electronics Foundry: Samsung Electronics is exploring outsourced backend design due to rising 2nm demand, with potential partners including ADTechnology, Gaonchips, and Alphachips. Standard backend-design projects are worth tens of billions of Korean won, materially less than full-service ASIC projects. Securing an advanced-node project and winning a high-value mass-production contract should therefore be assessed separately.
HBM/DRAM/NAND/SSD/HDD
Samsung Electronics Memory: The company plans to build a DRAM fab in Giheung with monthly capacity of approximately 100,000 wafers. It will also install another 100,000-wafer-per-month line at Pyeongtaek P4 capable of producing DRAM required for HBM4. The expansion confirms that the current shortage is real, but large-scale construction will ultimately increase supply around 2028. Key indicators include groundbreaking, equipment installation, and yields, with equipment orders providing signals before wafer output begins.
SK Hynix: KB Securities maintains a Buy rating and a KRW4.2 million price target, expecting the memory shortage to persist at least through 2028. Morgan Stanley emphasizes that memory supply is stagnating relative to token demand, making an extended upcycle more likely. A counterpoint is that the ADR has traded at an approximately 50% premium; trading prices should not be conflated with fundamentals, and the target price must be interpreted in the context of the Korean-listed shares.
Micron and SanDisk: Both companies are locking in CoreWeave demand through long-term agreements while planning substantial capacity additions by early 2028. Near-term pricing power comes from undersupply in HBM and high-end memory. The longer-term risk is that contract prices exceed spot prices once new capacity comes online. Buyers have already begun exploring hedges, and contractual price floors will determine losses when the cycle turns.
SLC NAND: High-layer-count 3D NAND is crowding out mature-node capacity, while industrial, automotive, and networking customers are shifting from MLC to SLC. Pricing surveys forecast that SLC contract prices in the second half of 2026 will rise another 120%–170% versus the first half. Suppliers have no near-term plans to add capacity, creating strong pricing leverage, but rising customer costs will also accelerate substitution and inventory adjustments.
“In the near term, major global SLC NAND flash suppliers have no plans to add capacity → Their primary focus is on process-node migration, yield improvement, and optimizing output per wafer.”
Memory Cycle: Monthly memory shipments totaled approximately US$63.3 billion in May 2026, up 285% year over year and approximately 10.7 times the early-2023 trough. Pricing, rather than volume, contributed most of the growth. Evidence that AI is extending the upcycle is strong, but historically the longest period of positive annual growth has been only five years. Supply reversal remains a risk in 2027–2028.
Semiconductor Equipment/Testing
ASML
1) Second-quarter revenue was €9.33 billion, with a 54% gross margin and €2.92 billion in net income. Third-quarter revenue guidance is €11–12 billion, while full-year revenue guidance was raised from €36–40 billion to €43–45 billion.
2) Based on capacity of approximately 65 Low-NA EUV systems and 130 DUV immersion systems, the company plans to expand production capacity for each by 30% in 2027 and is evaluating another 30% increase in 2028.
“AI investment is driving demand for advanced logic and DRAM. Customers are accelerating capacity and capital-spending plans, order inflows remained strong in the first half, and long-term purchase commitments and demand visibility have increased accordingly.”
Applied Materials, Lam Research, and KLA: On July 15, UBS raised its price targets to US$705 from US$570, US$435 from US$375, and US$255 from US$218, respectively. Arcuri believes a wafer-fab-equipment market exceeding US$300 billion by 2029 is becoming credible, based on the combined assumptions of AI-driven capacity expansion, equipment pricing power, and potential Terafab investment.
“Investors continue to underestimate the semiconductor-equipment industry’s growing pricing power, and WFE exceeding US$300 billion by 2029 is becoming an increasingly credible scenario.”
Aehr Test Systems: Fourth-quarter revenue was US$18.84 million, with bookings of US$60.7 million and actual backlog exceeding US$100.6 million. Fiscal 2027 revenue guidance is US$130–150 million, representing growth of 160%–200%, excluding memory revenue. The conversion of test qualification into mass production represents upside, while customer concentration and an approximately 20x after-hours price-to-sales multiple are key risks.
Semiconductor Testing Value Chain: As AI chips become larger, consume more power, and adopt chiplet architectures, the number of tests, testing time, and contact points all increase. ASE Technology, King Yuan Electronics, and Amkor provide outsourced testing, while Advantest, Teradyne, and FormFactor supply testers and probe cards. The true validation point is whether demand spilling over from TSMC translates into orders.
Bosch Silicon Carbide: The Roseville fab has received up to US$225 million in subsidies to support as much as US$2 billion in investment and has begun sample production. It aims to begin commercial mass production of silicon-carbide chips on 200mm wafers in 2026. The project will drive demand for epitaxy, etching, metrology, and high-temperature reliability testing, but the yield ramp will determine whether localized supply materializes.
Optical Communications/Optics Value Chain
Applied Optoelectronics: Construction has begun on two new sites in Pearland, Texas, adding nearly 400,000 square feet dedicated entirely to 800G and 1.6T optical transceivers. Monthly capacity is planned to increase from 100,000 units in early 2026 to more than 650,000 by year-end and more than 930,000 in 2027. First-quarter revenue was US$151 million, up 51% year over year; execution and yields are the largest risks.
“The Pearland project announcement provides tangible evidence that the capacity ramp is actually underway.”
UMC and LIGHTIC Technologies: The two companies advanced a 1.6T silicon-photonics platform to production readiness within 18 months and delivered the first batch of 12-inch production wafers, which have been qualified by a leading global cloud-infrastructure customer. UMC plans to open its proprietary platform in 2027 and develop 400G/lane, CPO, and optical I/O solutions. Conversion of customer qualification into volume orders still requires validation.
Tower Semiconductor: The company plans to invest US$3 billion to expand 300mm silicon-photonics, silicon-germanium, and advanced-packaging capabilities in Japan, supported by approximately US$1 billion in subsidies. The first phase is targeted to become fully operational in the fourth quarter of 2027. The company has also disclosed a US$1.3 billion silicon-photonics customer contract for 2027. The combination of capex and contracts validates demand, while the lengthy construction timeline remains a risk.
Nokia: A commercial AI-RAN platform developed with Nvidia is scheduled for sale in 2027. Demonstrations have achieved more than a 20% improvement in spectral efficiency, with targets of 50% in 2027 and twice as much data transmitted over the same spectrum in 2028. The company also plans to introduce software subscriptions alongside hardware. The key question is whether operator upgrade budgets and 6G compatibility can translate into recurring revenue.
Samsung Electro-Mechanics: Second-quarter revenue is expected to be approximately US$2.55 billion, with operating profit of approximately US$310 million. Its share of the high-end server MLCC market exceeds 40%. AI servers are driving demand for both MLCCs and FC-BGA substrates. Earnings leverage depends on the high-end product mix and execution of capacity expansion and cannot be extrapolated linearly from server volumes alone. Key metrics are MLCC utilization and FC-BGA yields.
Internet/Platforms
Google: Gemini’s web traffic share rose from 9.1% 12 months ago to 27.8%, showing that the search gateway is converting model capabilities into user reach. Meanwhile, the 10th-generation TPU further deepens its custom-chip strategy. The platform advantage is clear, but web share cannot substitute for cloud revenue, API calls, and enterprise renewals; the next focus is traffic monetization quality.
Meta
1) Morgan Stanley expects Meta’s compute capacity to approach 3x its current level by 2028, while total hyperscaler compute capacity roughly doubles over the same period. This directly challenges the “compute glut” narrative.
2) An undisclosed model scored a perfect 30/30 at the Asian Physics Olympiad, but capability upgrades still need validation through advertising, recommendation, and agent revenue.
“According to Morgan Stanley, Meta’s compute capacity will approach three times its current level by 2028.”
Amazon: Debt financing by cloud providers continues to fund investment in data centers, chips, and power, with Amazon, Microsoft, Meta, Google, and Oracle among the major issuers. The positive is that compute buildout has not stopped; the pressure is that financing costs are now part of project return calculations. AWS must demonstrate investment efficiency through cloud revenue and utilization, while debt maturity structure will also affect returns.
Microsoft: Enterprise customers are increasing AI infrastructure investment while reassessing software and mainframe budgets. Microsoft benefits from Azure and Copilot but also faces customer scrutiny over token costs, data retention, and exposure of proprietary knowledge. Whether the platform can grow AI revenue faster than infrastructure depreciation is the next key question, while cash flow coverage of capex is the hard metric.
Apple: Apple is reportedly seeking an injunction against OpenAI hardware that could delay sales of the latter’s smart speaker. Apple must both defend its hardware gateway and accelerate the rollout of Apple Intelligence. The material also points to a 2% sequential decline in hardware spending and potentially slower-than-expected services growth; whether on-device AI can drive an upgrade cycle remains unproven.
OpenAI Platform Gateway: Weekly active users of Codex and ChatGPT Work are growing rapidly, while open prompts, mobile access, SSH workflows, and parallel tasks are strengthening developer engagement. The real commercial questions are usage depth among new users, paid conversion, and unit inference costs. User counts alone can overstate platform value; retention and unit gross margin must also be assessed.
“Users only need to ask questions about the business, and the system can conduct relatively comprehensive research and provide answers, making business information retrieval and problem analysis easy and enjoyable.”
Anthropic Platform Gateway: Claude’s web traffic share rose from 1.6% to 9.2%, establishing it as the strongest third platform behind ChatGPT and Gemini. Indications of positive free cash flow support continued investment, but the company must still prove that its premium-priced models generate sufficient enterprise productivity to offset higher token costs.
Daum: AI Overview has adopted the Solar model and RNGD chips, currently covering approximately 20% of searches with plans to exceed 50%, and will expand into verticals such as shopping and restaurants. Search summaries can reduce click friction but may also reshape content distribution and advertising value. The platform must balance answer quality, the source ecosystem, and monetization.
“Daum’s AI Overview combines search-engine results with the Solar model to summarize user queries and provide key information. Instead of clicking multiple links individually, users can have AI read and summarize the relevant documents.”
PayPal: Market reports indicate that Stripe and Advent have proposed an acquisition at $60.50 per share, potentially backed by approximately $50 billion in financing. The specific bid price raises the prospect of a control premium, but no formal offer, board acceptance, or regulatory progress has been disclosed. With the transaction status unclear, short-term price action should not be treated as definitive value.
X Platform: Mobile posting failures are concentrated in long-form and quoted posts, while the web version remains available, suggesting the issue may lie in the Android publishing pipeline rather than a sitewide outage. Some users also report that an algorithm update has significantly reduced contentious content. The platform must validate the fix, content quality, and engagement rates simultaneously; a single user-experience improvement is insufficient to prove ecosystem recovery.
Similarweb: AI Studio was used to generate charts of generative-AI website shares, providing a real-world showcase for its own product. The data are valuable for tracking shifts in attention, but lack absolute traffic, sample, and paid-conversion metrics. Whether platform revenue can grow alongside usage of AI analytics tools requires confirmation from operating data.
“ChatGPT remains in first place, but its share continues to decline; Gemini continues to expand and holds a firm second place; Claude has rapidly risen from a low base to third place.”
Oracle: Oracle has committed to deploying 50,000 AMD MI450 GPUs and plans to build a publicly accessible AI supercluster entirely powered by AMD chips in the third quarter of 2026. For the platform, this adds a second GPU supplier and reduces single-stack risk. Whether it generates differentiated cloud revenue will depend on timely delivery and customer utilization.
Coupang: Coupang Play continues to strengthen its content gateway through European club competitions and entertainment guests, reflecting the e-commerce platform’s strategy of using sports and streaming to increase user time spent. The available information includes no subscription, advertising, or conversion figures, so these events should be viewed only as user-engagement initiatives rather than direct evidence of profit contribution. The next question is whether viewing time can increase transaction frequency.
Starlink: The V5 terminal makes the satellite-network gateway smaller, lighter, and more power-efficient, with peak speeds above 375 Mbps, average power consumption of 35–50W, and a bundled Wi-Fi 6 router. Improved specifications support broader deployment scenarios, but pricing, launch timing, and shipment volumes have not been disclosed, leaving platform revenue upside unconfirmed.
Platform Traffic Landscape: One month ago, ChatGPT, Gemini, and Claude accounted for 52.7%, 27.8%, and 9.2% of web traffic, respectively, for a combined share of nearly 90%. Competition has shifted from a single leader to multiple strong players, but these statistics exclude substantial API and internal enterprise usage. Platform assessments must therefore incorporate revenue, retention, and costs.
“‘Web traffic share’ is not equivalent to active users, paid revenue, API call volume, enterprise penetration, or model capabilities, nor can it be directly mapped to cloud-computing capex or AI hardware demand.”
