目录
Pre-Market Takeaways
AI and the Full Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
CSPs, Cloud Capital Expenditure, and AI Clouds
GPU, CPU, and ASIC
Advanced Packaging and OSAT
HBM, DRAM, NAND, SSD, and HDD
Optical Communications, Optics, and Passive Components
Internet and Platforms
Software and SaaS
Consumer Electronics and Smart Vehicles
AI demand continues to outpace supply. Cloud vendors are starting to answer concerns over heavy investment with contracts, backlog, and returns on capital. Supply-chain constraints are simultaneously spreading to HBF, DRAM, InP lasers, glass-substrate packaging, and data-center sites, while the market recovery still requires verification through delivery, yield, and financing.
404K | 2026-08-04
Pre-Market Takeaways
The technology-sector recovery is spreading from cloud giants to AI cloud providers, optical communications, memory, and software. QQQ rose 1.8%, with Oracle, CoreWeave, and Nebius showing greater upside beta. Palantir's post-market results moved the question of whether AI can translate into revenue and cash flow from narrative to verifiable financial figures.
Supply constraints are shifting. GPUs remain tight, but incremental bottlenecks are increasingly emerging in HBF and DRAM tiered memory, InP lasers, advanced packaging, power and cooling, and data-center sites with fiber connectivity. Demand spillover broadens the range of second-order beneficiaries, while raising the bar for yield, capacity expansion, and delivery execution.
Cloud vendors' heavy investment still has counterarguments. UBS expects five hyperscalers to issue more than $250 billion of bonds in 2026; Oracle also faces leverage, financing, and project-execution constraints. The more important pre-market question is whether orders convert into revenue and capital expenditure into free cash flow, rather than the scale of spending alone.
AI and the Full Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
Moonshot AI
1) Kimi K3 is inferred to interleave KDA and MLA attention at a 3:1 ratio, balancing compressed memory with global modeling; block-attention residuals improve compute efficiency by 1.25× and reduce pipeline overhead to 4%.
2) With B300, DRAM offloading, and speculative decoding, cost is about $0.1712 per 1 million input tokens, 76.89% below Moonshot AI's blended price of $0.74. The risk is that long-input, short-output workloads remain constrained by prefill and cache-hit rates.
"Kimi K3 is inferred to retain Kimi Linear's overall design, interleaving KDA and full-attention MLA at a 3:1 ratio."
Palantir
1) Q2 revenue reached $1.94 billion, up 93% year over year; U.S. commercial revenue was $764 million, up 150%, and the shares rose 11% after hours.
2) Q3 revenue guidance is $2.160 billion to $2.164 billion, up 83% year over year; full-year free-cash-flow guidance is $4.5 billion to $4.7 billion. Commercial and government demand accelerated together, although gross margin of 86.3% was slightly below the 86.9% consensus.
"Q2 revenue was $1.94 billion, up 93% year over year, above the $1.81 billion consensus, and accelerating from 85% in the prior quarter."
OpenAI
Apple sued OpenAI for allegedly stealing hardware trade secrets. OpenAI then disclosed handover messages from a departing employee, saying Apple employees continued asking that person to explain internal documents several weeks after departure. This does not directly rebut the trade-secret allegations, but it brings hardware-team offboarding, access management, and the integrity of the evidence chain to the fore; the next checkpoints are court filings and remediation of internal controls.
"Of course, I can ask several other people, but you are the best person. Even though you no longer work here."
AI Engineering Roles
Agentic AI has entered semiconductor-verification workflows, where it can operate complex EDA tools and handle language syntax, but experience, judgment, and the ability to address unknown problems still require years of development. If companies cut junior-engineer hiring, the supply of senior talent could break down five to 10 years from now; having a senior engineer manage 10 or more agents may also crowd out creative work.
Model Distribution Layer
Amazon, Microsoft, and Google continue to invest simultaneously in proprietary models, proprietary chips, and third-party hardware, with the operating logic of letting customers switch among models and processors. Model neutrality can reduce exposure to a single technology path, but it also shifts competition toward distribution, cost, and migration convenience. Key metrics are cloud-revenue growth, unit inference cost, and customer renewals.
CSPs, Cloud Capital Expenditure, and AI Clouds
Amazon
1) AWS says Trainium capacity is essentially sold out through the end of 2027, with both large customers and startups adopting the cost advantage of custom processors while retaining the option of NVIDIA GPUs.
2) Amazon modestly raised FY2026 capital expenditure, partly because of higher memory prices. Servers typically last five to six years; if they reach breakeven in two to three years, they can still generate substantial cash returns thereafter.
"Trainium is extremely popular. As we have now mentioned multiple times, our Trainium capacity is essentially sold out through the end of next year."
Microsoft
Microsoft continues to expand the data-center footprint needed for AI and cloud workloads. The truly scarce resource is not ordinary land, but sites that simultaneously offer power, cooling, permits, and fiber connectivity. The company modestly raised FY2026 capital expenditure, with higher memory prices contributing part of the increase. Near-term free cash flow remains under pressure; the next checkpoints are Azure growth, commissioning of capacity under construction, and migration to internal compute.
Google
Google's direct rationale for higher investment is strong demand, and it continues to offer a choice among frontier models, proprietary hardware, and third-party GPUs. Google also joined the HBF alliance led by SK hynix and Sandisk, indicating that inference infrastructure is starting to shift from simply stacking HBM toward tiered memory. The next checkpoints are HBF software support, TPU connectivity, and actual deployment scale.
Oracle
Oracle gained 9% during the technology-sector recovery. The bull case rests on visibility into contracted business, AI-infrastructure gross margins of 30% to 40%, returns on invested capital close to 30%, and cash conversion near 100% for mature projects. Counterarguments include high leverage, dependence on capital markets, potential equity financing, OpenAI demand, and project-permitting execution risk.
IREN
IREN signed multi-year AI-cloud contracts with Microsoft and NVIDIA covering about 85% of its year-end target of more than a $4 billion annualized revenue run rate, with a weighted-average term of four years; customer prepayments cover about 45% of the related GPU capital expenditure. Contracted revenue improves cash-flow visibility, but large-scale capacity commissioning, debt management, and customer migration still require validation.
GPU, CPU, and ASIC
AMD
1) Morgan Stanley expects server CPUs to support quarterly performance and has initial confidence in MI455; sold-out products limit the scope for a near-term upside revision.
2) AMD raised its estimate of the 2030 server total addressable market from $120 billion to about $220 billion, with a 50% compound annual growth rate; Morgan Stanley introduced an assumption of 49% server-business growth in 2027. Expectations are already high, so earnings must deliver simultaneously on revenue, supply, and margins.
Intel
1) Intel's HLFF package targets dimensions as large as 240 mm × 240 mm and integrates compute dies, HBM, and I/O; current validation covers 18 dies and 12 HBM groups, while achieving void-free underfill over a flow distance of more than 40 mm.
2) 18A still requires high-volume manufacturing validation. One interviewed expert estimates the actual delay at about two years; insufficient advanced-node data and packaging yield remain the key constraints on catching TSMC.
"Intel 7, Intel 4, Intel 3, and 18A form a chain, with each node inheriting problems from the preceding node that have not yet been resolved."
TSMC
1) MI455 is considered a possible first chip to ship with an active local silicon-interconnect bridge, with its interposer containing two base dies, 12 HBM4 stacks, and two I/O dies.
2) The Kumamoto fab resumed operations quickly, but a genuine production recovery still requires checking equipment calibration, yield, work-in-process losses, and customer deliveries. Restored power should not be equated with restored commercial capacity.
Advanced Packaging and OSAT
CoPoS Equipment Supply Chain
TSMC and Intel are advancing glass-substrate packaging, with validation planned for 2026 and mass production for 2028. Large rectangular panels could increase material utilization from about 65% to more than 90% and reduce packaging cost by 20% to 30%. Grand Process Technology, All Ring Tech, and Gallant Precision map respectively to wet processing, automated handling, and high-precision bonding.
PSMC
PSMC is the exclusive supplier of silicon capacitors inside Intel's EMIB-T package, shortening high-current power-delivery paths and improving power integrity. Package-level demand is migrating toward silicon capacitors, while board-level high-capacitance demand remains served by MLCCs, allowing both paths to benefit. Key variables are silicon-capacitor expansion, packaging adoption, and power density in 2027 to 2028.
HBM, DRAM, NAND, SSD, and HDD
Micron
Major producers' 2027 DRAM and HBM capacity has already been preallocated, while output from new fabs will mostly arrive in late 2027 to 2028. Micron says the AI-driven shortage will persist well into 2027. The counterargument is that aggressive later-stage expansion could restore balance in 2028 to 2030; near term, long-term-agreement pricing and wafer allocation remain the key indicators.
SK hynix
1) Its Q2 DRAM revenue share was only about one percentage point above Micron's, as HBM3E repricing and HBM4 delays narrowed its lead.
2) The company and Sandisk released the first open HBF standard, covering capacities up to 512 GB and bandwidth of 0.4 TB/s to 3.0 TB/s, and demonstrated 375-layer 4D NAND with 2.5× higher performance per watt than the previous generation.
"HBF is positioned between HBM and SSD, using stacks of eight or 16 NAND dies to deliver capacity of up to 512 GB."
Sandisk
Sandisk and SK hynix released the HBF specification to the industry through the Open Compute Project, covering UCIe connectivity, interfaces, electrical characteristics, reliability, packaging guidelines, and software I/O rules. Google and Tenstorrent have joined the alliance, a positive signal for ecosystem expansion; revenue will ultimately depend on processor compatibility, software support, and clearly defined AI-inference use cases.
Samsung Electronics
Samsung Electronics plans to showcase its zHBM and V-NAND roadmaps at FMS 2026. By vertically stacking above the GPU to shorten transmission distance, zHBM targets four times the speed of HBM4 at about one-quarter of its power consumption; zNAND targets 15 times higher processing performance and one-fifth the power consumption. These are aggressive roadmaps, and mass production, yield, and customer qualification still need verification.
CXMT
HP, ASUS, and Acer have begun using CXMT DRAM, primarily because incumbent suppliers cannot meet demand rather than because of a simple low-price substitution. Supply-chain sources say its pricing is not below that of Samsung Electronics and other vendors, and supply beyond this quarter is also difficult to reserve. A potential new Beijing fab still requires verification of project progress, product qualification, and actual output.
Optical Communications, Optics, and Passive Components
Lumentum
The CEO says the shortage of lasers for AI data centers exceeds 30%, while Lumentum and Coherent have each secured about $2 billion of investment for expansion. Morgan Stanley believes the CPO schedule has not slipped, but the nearer-term focus should be NPO and pluggable optical modules. Continued 800G adoption and the ramp of 1.6T support laser demand; the risk is the eventual degree of CPO adoption.
"As discussed in our deep dive, we have not seen a delay in the CPO timeline; our greater concern is the degree of adoption, as NPO appears to be gaining stronger momentum."
Coherent
Coherent and Lumentum share the same underlying laser and optical-component technology stack for NPO and CPO, so an architectural shift would not immediately weaken demand. The company secured about $2 billion of expansion investment, but views differ on whether InP substrates or internal fabs are the main bottleneck. Near-term indicators are the 800G and 1.6T ramps and pluggable transceivers; the longer-term indicator is CPO adoption.
AXT
AXT returned to profitability as quarterly InP revenue reached a record, while demand still exceeds the pace of expansion, backlog is increasing, and the company has begun turning away orders. Another supply-chain report puts the global InP-wafer shortage at 70% and says effective yield from wafer to finished product has fallen from about 90% to 40%. These are strong claims and should be tested against subsequent earnings, the planned doubling of capacity, and expansion progress in 2027.
Samsung Electro-Mechanics
Each AI-server rack may require as many as 440,000 high-end MLCCs, and suppliers are pursuing price increases of up to 30%. High-capacitance MLCCs remain irreplaceable at the board level, with AI-server usage already more than 10 times that of standard servers. Delivery lead times and the durability of pricing will determine earnings leverage, while a growing package-level share for silicon capacitors is a risk to monitor.
Internet and Platforms
Meta
Meta rose 6%, as investors viewed strength in the core business as evidence that AI investment is beginning to generate near-term returns. Alongside Amazon, Google, Microsoft, and Oracle, the company faces greater financing needs; UBS expects the five companies to issue more than $250 billion of bonds in 2026. Whether advertising and AI-product monetization can outgrow capital expenditure is the key to a free-cash-flow recovery.
"The expansion of data centers, GPU clusters, power, and networks is happening so quickly that these companies are turning to the bond market to finance the next phase of cloud and AI growth."
Amazon
Accelerating AWS growth and sold-out Trainium capacity reinforce the platform's bargaining power: customers can use custom processors to lower costs while combining them with NVIDIA GPUs when needed. The platform advantage is that, if demand weakens, excess compute can be redirected to e-commerce, advertising, and internal services. The counterargument is that higher memory prices raise investment and keep near-term free cash flow under pressure.
"We are starting to see both large customers and startups strongly favor the cost advantage of using Trainium."
Google
Google Cloud demand is driving higher investment, while the platform continues to offer proprietary models, TPUs, and third-party GPUs. Morgan Stanley identifies Google as a beneficiary under both closed-source and hybrid-model scenarios, though its advantage may be less secure if open models prevail. Key indicators are cloud revenue, model calls, HBF-alliance implementation, depreciation cadence, and returns on capital.
"Google and Tenstorrent have joined the alliance, whose members hope to use an open standard to promote HBF adoption in AI inference and tiered-memory architectures."
Microsoft
Microsoft holds a distribution position across closed-source, hybrid, and open-model paths. Copilot and Azure provide the enterprise entry point, while internal and third-party compute can bridge one another. Long-lived sites can support five to six server generations, whereas short-cycle equipment can be scaled down if demand weakens. Platform resilience is high, but capital expenditure, bond supply, and depreciation remain pressure points.
"Microsoft released GPT-4 before OpenAI, did not retreat after the Sydney incident, and rapidly brought Copilot to market as the first professional AI tool."
Oracle
Oracle rose 9%, with the upside beta of cloud-infrastructure transactions reflecting contracted visibility and expectations for high returns on invested capital. The platform challenge is that AI could displace parts of the legacy software business, while the infrastructure business depends on OpenAI compute demand, external financing, and project permits. If contracts do not convert into revenue on schedule, valuation and cash flow will both come under pressure.
"The hyperscalers collectively added $1.5 trillion of backlog, while the expected increase in 2027 capital expenditure is only $365 billion."
eBay
eBay fell 6% amid concerns over dilution from Depop investment and competition from Vinted. Vinted's U.S. users have increased about fourfold since January and now equal about 25% of Depop's monthly active users. If eBay adds $100 million of annual investment, EBIT could fall by about 3%. The platform must prove not merely user growth but a path to profitability by 2027 and efficient customer acquisition.
Depop
Depop remains in an investment-for-growth phase. Rapid U.S. user growth at rival Vinted makes its profitability timetable more difficult. If eBay continues providing roughly $100 million of additional support each year, near-term earnings will be diluted. The next indicators are monthly active users, gross merchandise value, marketing payback, and the 2027 profitability target; any delay would amplify valuation pressure on the parent company.
Vinted
Vinted's U.S. user base has increased about fourfold since January and is now about 25% of Depop's monthly active users. The growth shows that demand for secondhand e-commerce remains intact, but it also shifts competition from category expansion toward customer-acquisition costs and fulfillment experience. For eBay, the faster Vinted approaches scale, the more marketing Depop will require and the later profits will arrive.
Apple
Apple's services growth slowed and its shares fell 2%, making it a source of funding for long positions in cloud giants. At the same time, its trade-secret dispute with OpenAI exposed issues in hardware-talent offboarding and internal access. The consumer platform retains a massive entry point, but must simultaneously contend with services growth, the cadence of AI products, and litigation-execution costs.
Reddit
Reddit strengthened during the internet-platform recovery, showing that improved risk appetite has spread from hyperscalers to advertising and community platforms. Unlike cloud giants, Reddit lacks a hard verification metric such as returns on capital expenditure. The more relevant indicators are whether advertising monetization, data licensing, and user growth can produce durable revenue rather than relying solely on sector rotation.
VeriSign
VeriSign strengthened with the internet-platform group, though its business is more stable domain infrastructure than capital-intensive AI. A market recovery can lift valuation risk appetite, but sustained performance still depends on domain registrations, renewals, and pricing power. If the platform trade rotates back toward high growth, the relative beta of a stable cash-flow asset may recede.
Spotify
Spotify rose ahead of results due the next trading day. The platform needs subscription growth, pricing, and margins to validate the recovery rather than relying on a broad technology rally. If user growth and monetization improve together, capital may continue rotating from hardware toward internet services; if guidance disappoints, the pre-earnings gain will increase downside sensitivity.
OpenAI
OpenAI is both an important source of cloud-infrastructure demand and the center of competition in model distribution. Hyperscalers can use near-term third-party compute to serve multi-year customers and later migrate them to internal capacity, improving returns on investment. This also means changes in OpenAI demand directly affect Oracle and new cloud projects, so contract terms and migration cadence require continued verification.
Anthropic
Anthropic's annualized recurring revenue and GPU rental expense continue to rise, offering supporting evidence that AI demand has not weakened as new compute comes online. For platforms, the implication is that model usage continues to pull through cloud revenue. The risks are lower prices, customer switching, and changes to compute procurement as model competition intensifies; the next question is whether revenue growth can cover inference costs.
Hyperscale Cloud Platforms
UBS expects Amazon, Google, Meta, Microsoft, and Oracle to issue more than $250 billion of bonds in 2026, more than twice the 2025 level and more than 10 times the 2024 level. Their shift from cash machines to major bond issuers shows that AI buildouts now exceed near-term free cash flow. Credit spreads, absorption of issuance, and returns on capital will affect valuations.
"Hyperscale cloud providers are no longer simply cash machines. To keep pace in the AI race, they are becoming major bond issuers."
"Servers typically have useful lives of five to six years. If breakeven can be brought forward to two to three years, they can generate substantial free cash flow."
"Remain model-neutral. No single model can dominate over the long term, because different models will overtake one another at different times."
Software and SaaS
Palantir
Q2 commercial and government revenue both accelerated from high bases, with U.S. commercial revenue up 150% and U.S. government revenue up 90%. Financial results support the company's ability to turn tokens into economic value, while the standard version of Nemotron Ultra beat frontier models on five production tasks within 24 hours. Risks include elevated expectations, gross margin slightly below consensus, and differences in guidance methodology.
"Q2 revenue was $1.94 billion, up 93% year over year, above the $1.81 billion consensus, and accelerating from 85% in the prior quarter."
Rubrik
Loop Capital initiated coverage with a Buy rating and a $100 price target, arguing that agentic AI raises the strategic priority of cyber resilience. Channel checks indicate that high-priority workflows can limit project delays, although hardware supply chains may still constrain near-term demand. The valuation is benchmarked to high-growth peers with more than 20% revenue growth in 2026; sales execution and project conversion are the key indicators.
"Cyber resilience is becoming an important long-term growth opportunity, highlighted further by recent security incidents such as the Mythos scare and the Hugging Face incident."
AppLovin
Survey feedback was broadly positive, but growth slowed after May, meaning elevated expectations for the advertising platform require confirmation in Q3 guidance. The price target was raised, but the new figure was not disclosed, so the action alone cannot establish a fundamental view. The next indicators are self-service advertising penetration, budget expansion, whether growth can reaccelerate, and customer retention.
Unity Software
The full release of D28 drove double-digit month-over-month growth for Unity Software in July, showing that broader product access is starting to translate into usage. Its price target was also raised without disclosure of the new figure. The near-term question is whether D28's contribution persists; longer term, the question is whether AI tools and game-development workflows improve customer retention, monetization, and sales efficiency.
"The cache resides successively in GPU HBM, server DRAM, and SSD, while evaluation should combine prefill time and cache size into a 'KV throughput' metric."
CrowdStrike
Cybersecurity and cloud-consumption shares strengthened together, with CrowdStrike benefiting from the higher identity, endpoint, and data-protection requirements created by enterprise AI workflows. Morgan Stanley retains security software as a beneficiary under closed-source, hybrid, and open-model scenarios, indicating that demand is relatively independent of which model path wins. New contracts, renewals, and margins are the key indicators.
Palo Alto Networks
Palo Alto Networks strengthened with the cybersecurity group and was included among security-software beneficiaries under a closed-source-model scenario. The logic is that agent adoption makes permissions, networks, and data boundaries more complex. The counterargument is that security-budget consolidation could concentrate spending among a few platforms, making revenue growth and remaining performance obligations more reliable indicators.
Snowflake
Snowflake strengthened, indicating improved risk tolerance for cloud consumption. AI applications require data governance, storage, and an access layer, but customers will continue optimizing cloud costs. The investment implication therefore rests on whether consumption growth is sustainable, not on an AI label alone. Key indicators are product revenue, net revenue retention, and whether consumption by large customers recovers with model workloads.
MongoDB
MongoDB strengthened with database and cloud-consumption shares, based on the view that agentic applications increase demand for unstructured data and real-time calls. The risk is that customers can still defer expansion or migrate to lower-cost alternatives. Atlas consumption, new customers, and margins are the next indicators; if consumption recovers more slowly than the share price, the sector-wide rally could reverse.
"The central tension is that MLA reduces decoding compute by absorbing techniques but raises prefill cost. It therefore suits decoding-heavy inference, while offering limited benefit for agentic workloads with long inputs, short outputs, and frequent tool calls."
Datadog
Morgan Stanley includes Datadog among infrastructure-software beneficiaries in a hybrid-model scenario, because multi-model, multi-cloud, and on-premises deployment increases observability complexity. Closed-source or open-source architecture is not itself decisive; workload count and log volume generate the revenue. Cloud consumption, customer expansion, and large-deal conversion are the key indicators.
Zscaler
Zscaler joins CrowdStrike and Palo Alto Networks on the list of security-software beneficiaries in a closed-source-model scenario. Agents expand machine identities and access paths, making zero-trust demand more concrete. An industry list cannot substitute for company execution, however; billings growth, remaining performance obligations, and sales efficiency are still needed to determine whether demand is materializing.
Appian
Appian is included among infrastructure-software beneficiaries in a hybrid-model scenario, as low-code workflows could support enterprise orchestration across multiple models. The opportunity depends on workflows reaching production rather than the number of demonstrations. Subscription growth, cloud mix, and customer expansion should be checked before equating model interest directly with revenue.
vLLM
In the Kimi K3 serving architecture, vLLM checkpoints recurrent state every 32K tokens and at prompt boundaries, with the cache residing successively in GPU HBM, server DRAM, and SSD. This reduces repeated prefill cost, but actual memory use is not constant. Once concurrency exceeds eight, cache-hit rates may fall from a theoretical 95% to below 10%, making deployment results highly dependent on workload shape.
Mooncake Store
Mooncake Store handles KDA coarse-grained state and tiered-cache management, connecting GPU HBM, server DRAM, and SSD into a unified storage layer. The software's value comes from higher prefix-hit rates and lower time to first token. The risk is that long-input, short-output workloads and frequent tool calls amplify prefill cost; actual-session cache-hit rates and throughput are the next indicators.
DSpark
With B300 and vLLM, DSpark speculative decoding achieves input throughput of 4,844.1 tokens per second per GPU at P90 output of 30 tokens per second per user, implying a cost of $0.1712 per 1 million input tokens. The result shows that software optimization can change hardware economics, but the conclusion depends on a three-year rental price, a specific trajectory, and cache configuration, and cannot be generalized to all workloads.
OpenRouter
As of July 30, OpenRouter's floor price was $3 per 1 million input tokens and $15 per 1 million output tokens, providing a market benchmark for Kimi K3 serving costs. Aggregation platforms benefit from model switching and price transparency, while price competition could compress take rates. The next indicators are traffic, model supply, and whether enterprise customers can generate stable revenue.
"With B300, DSpark delivers input throughput of 4,844.1 tokens per second per GPU at P90 output of 30 tokens per second per user, corresponding to a cost of $0.1712 per 1 million input tokens."
Consumer Electronics and Smart Vehicles
Apple
1) Services growth slowed and the shares fell 2%, making Apple a near-term source of funds for rotation into cloud giants.
2) Its trade-secret litigation with OpenAI exposed problems in employee offboarding and access controls. The consumer-hardware entry point remains strong, but AI-hardware cadence, services growth, and legal execution are now concurrent valuation variables. Product and services guidance are the next indicators.
"None of this disproves Apple's broader trade-secret allegations. But Apple must now explain why its own employee-offboarding process failed."
Tesla
First-time electric-vehicle buyers in California can receive a $3,500 discount on eligible new Model 3 and Model Y inventory vehicles. Orders must be placed on or after August 3, 2026 and delivered while funds remain available. The policy covers cash purchases, leases, and financing and should support near-term inventory conversion; the next indicators are incremental deliveries and the net pricing effect.
"To qualify, you must place your order on or after August 3, 2026 and complete delivery while funding remains available."
Sony
Sony and TSMC are advancing joint-venture negotiations for an image-sensor line at the new Kumamoto fab, with the objective of signing a definitive agreement. Sony has reserved ¥10 billion and intends to become the controlling shareholder. The partnership would extend mobile-device sensor capabilities into automotive and robotics; key checkpoints are execution of the agreement, production recovery, customer qualification, mass-production cadence, and capital expenditure.
"Overall demand from AI servers plus residual demand from traditional applications continues to exceed supply. Most output from new fabs sufficient to ease the shortage will not arrive until late 2027 and 2028."
HP
HP has begun using CXMT DRAM, primarily because incumbent suppliers cannot provide enough product rather than because it is cheaper. Diversification by PC vendors shows that standard-DRAM tightness has reached system procurement. Whether shortages ease near term depends on CXMT product qualification and continuity of supply, while the cost side also depends on contract pricing, inventory, and whether procurement volumes continue rising.
ASUS
ASUS has joined HP and Acer in turning to CXMT memory, indicating that the PC supply chain is prioritizing security of supply over low prices. Supply-chain sources say CXMT DRAM is not cheaper than products from Samsung Electronics and other vendors, while supply beyond this quarter is difficult to reserve. For ASUS, the next indicators are model coverage, procurement cost, validation progress, and whether inventory returns to normal.
Acer
Acer's adoption of CXMT DRAM provides a 2nd source for the shortfall from incumbent suppliers. The move reduces single-source risk but does not necessarily improve cost, because CXMT offers no obvious discount. If progress at the new fab, product qualification, or yield disappoints, PC supply constraints could persist into later quarters and affect system configurations.
"We also cannot reserve CXMT supply beyond this quarter because many companies are competing to lock in DRAM supply from the company."
CXMT
Adoption by international PC customers raises CXMT's visibility in the consumer-electronics supply chain. The company is also reported to be considering a 2nd 12-inch memory fab in Beijing. Customer adoption and expansion plans must be validated separately: the former already reflects a real shortage, while the latter still depends on construction, equipment, yield, and product qualification and should not be counted as supply in advance.
Samsung Electronics
PC customers' move to CXMT shows that incumbent DRAM suppliers such as Samsung Electronics cannot fully meet standard-product demand. At the same time, the company is redirecting resources toward HBM, zHBM, and high-performance NAND. The consumer-side risk is that prioritizing AI further squeezes PC DRAM; the opportunity is better pricing and margins for high-end memory. Wafer allocation and contract pricing are the next indicators.
SK hynix
HBM is generally considered to consume three to four times as much wafer capacity as conventional DRAM. Prioritizing AI means allocations to commodity-product and module vendors may fall to about 30% of prior levels. For consumer-electronics customers, this implies continued tightness in standard DRAM. SK hynix must balance high-margin HBM, investment in the HBF ecosystem, and the issue of lower pricing in long-term agreements.
Micron
Micron and SK hynix differed by only about one percentage point in Q2 DRAM revenue share, showing how standard-DRAM pricing and product mix are changing competitive positions. Micron says the shortage will extend beyond 2027, so consumer customers will continue to face higher procurement costs. PC demand elasticity, standard-DRAM allocation, and the release of new capacity are the next indicators.
Lens Technology
Lens Technology is working with Intel on glass-substrate packaging, connecting its consumer-electronics materials capabilities with advanced AI packaging. Glass substrates offer low thermal expansion, high flatness, and strong high-frequency transmission, while the CoPoS path could raise material utilization from about 65% to more than 90%. The opportunity depends on the 2026 validation milestone and the 2028 mass-production milestone.
Intel
Diamond Rapids will enter next-generation architecture discussions at Hot Chips 2026, while the consumer side still depends on execution of Panther Lake, 18A, and Foveros packaging. An interviewed expert says 18A previously lacked both yield and parametric metrics and was delayed by about two years. If packaging or front-end yield remains low, client-product cadence and costs will both come under pressure.
TSMC
TSMC's Kumamoto fab has resumed operations, while negotiations for the Sony image-sensor joint venture are advancing. Consumer-electronics and automotive customers should focus on yield for the individual line, work-in-process losses, and delivery rather than whether the entire fab has power. If production resumes before mid-August and customer deliveries are unchanged, the supply-interruption risk will be shown to be manageable and pressure on customer inventories should ease.
Renesas Electronics
Renesas will close its Takasaki plant, which had previously been planned for SiC-device production. The event shows that power-semiconductor expansion is not a one-way process, as vendors continue to adjust assets according to demand, cost, and production-line efficiency. Automotive and industrial customers should watch capacity transfers, alternative-fab arrangements, and continuity of supply rather than equating a plant closure directly with lower supply.
SpaceX
Starlink Mobile plans to launch thousands of next-generation V2 satellites by 2028 to provide global direct-to-device coverage at 5G speeds. Data service is planned to begin in Q3 2027, followed by voice service several months later. The consumer-connectivity opportunity lies in remote and high-latitude regions; launch cadence, device compatibility, and regulatory approvals are the key checkpoints.
"Data service will launch in Q3 2027, with voice service following several months later, potentially making Starlink the only provider covering high-latitude polar regions."