目录
Pre-Market Highlights
AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Semiconductor Equipment/Testing
Optical Communications/Optical Supply Chain
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Overview
404K | 2026-08-05
The investment-bank reports cited in the article and the full text are available in their original form.
Pre-Market Highlights
AI demand has not materially receded, with orders, rental prices, and capital expenditure still rising. What has changed is that the market no longer rewards revenue growth alone: Advanced Micro Devices must prove that rack-scale systems can generate margins, SpaceX must prove that contracts and cash flow can cover its massive construction investment, while Astera Labs is demonstrating through content per XPU that capital continues to spill over into interconnects.
Memory and packaging are becoming the tightest constraints. Long-term agreements are adding price floors and prepayments, conventional DRAM prices are rising, and zHBM, bonded NAND, CoW outsourcing, and EMIB-T are all advancing simultaneously. On-time delivery of compute platforms increasingly depends on HBM, substrates, optical interconnects, testing, and cooling—not merely GPU volumes.
The same cost theme is also emerging across platforms and end devices. Spotify is attempting to monetize AI features as paid add-ons, smartphone vendors are using premiumization to absorb higher memory prices, while the PC supply chain is already seeing volume pressure. Tonight, investors should focus more on whether revenue growth can translate into profits and cash flow than on isolated performance metrics or a single quarter’s upside surprise.
AI/Semiconductor Value Chain
AI Models/Applications and Capital Expenditure
OpenAI
ChatGPT’s weekly user count is approaching 1 billion, pushing demand bottlenecks further toward power, data centers, and networking. User scale confirms that the model-access gateway is still expanding, but user numbers alone cannot demonstrate monetization quality; the next metrics to watch are inference volume, cost per task, and whether new infrastructure can generate sustainable revenue. If usage growth trails user growth, value per user will remain under pressure.
SpaceX
2Q26 capital expenditure was approximately US$18.4 billion, of which US$15.8 billion was allocated to AI compute infrastructure; free cash flow was -US$16.0 billion. The company plans to exceed 2 GW of compute capacity by the end of 2026 and approach 10 GW by the end of 2027. Demand and financing capacity are both strong, but the most immediate risk is a mismatch between construction speed, cash consumption, and contract execution.
“The company’s total capital expenditure in the second quarter was approximately US$18.4 billion, of which approximately US$15.8 billion supported AI compute infrastructure.”
Google
Google has reportedly committed US$150 billion to AI chip infrastructure. Investment at this scale would simultaneously drive demand for proprietary chips, Broadcom-related components, networking, and advanced packaging, while potentially increasing supply-chain prices and financing requirements. The construction schedule underlying the commitment still needs to be verified to avoid treating a multiyear investment as a single year’s procurement; if construction is delayed, supply-chain revenue recognition will also shift later.
Volta
The seven-month-old AI infrastructure financing company raised US$300 million at a US$2.4 billion valuation and said it had secured US$5 billion of GPU financing and a US$10 billion, six-year cloud contract. Combining financing with compute operations can accelerate deployment, but chipmakers’ participation in financing also increases circular dependencies. Customer credit quality and equipment residual values are the next validation points, while the pace of contract collections matters more than valuation expansion.
AI Task Costs
A cheaper model does not necessarily mean a cheaper task: Sonnet 5 costs approximately 1/1.7 as much per token as Opus 4.8, but consumes 1.9 times as many tokens to complete a task, raising the total cost; in another test covering 32 tasks, the top 6 models delivered similar results despite a 73-fold price gap. Procurement metrics are shifting from throughput toward cost per successfully completed task.
GPU/CPU/ASIC
Heterogeneous Computing Platforms
Inference is being divided into prefill, decoding, and tool execution, with compute-intensive stages handled by GPUs, ASICs, or NPUs, while CPUs are better suited to complex control. The industry aims to preserve nearly 8 times the performance when scaling from 1 GPU to 8 GPUs, but HBM, topology, and orchestration will erode linear scaling, so hardware combinations must be selected around the complete workflow. UALink adoption also needs to be monitored over the next 1 to 2 years.
NVIDIA
1) SpaceX has decided to use Vera Rubin for subsequent construction and is targeting approximately 10 GW of compute capacity in 2027.
2) When the same GB300 NVL72 runs DeepSeek-R1, tokens per GPU per second increased from 2907 to 8064 within 6 months, representing a 2.7-fold software-driven improvement. NVIDIA’s advantage remains the integrated delivery of racks, networking, scheduling, and its developer ecosystem.
“Looking ahead, we have decided to adopt NVIDIA exclusively because we believe the Vera Rubin architecture is the best architecture.”
Advanced Micro Devices
1) 2Q26 revenue was US$11.54 billion, up 50.1% year over year; data-center revenue was US$6.7 billion, up 107% year over year and representing 58% of total revenue.
2) 3Q26 revenue guidance is US$12.7 billion to US$13.3 billion. Helios is expected to begin shipping in 3Q26, followed by a production ramp. The revenue trajectory is sufficiently strong, while HBM, advanced packaging, and early-stage system gross margins will determine earnings leverage.
“Customer demand for Helios is very strong, and progress is currently ahead of our initial forecast.”
Arm
Arm management has narrowed its public targets to the period after new manufacturing capacity actually comes online, with the current focus on reaching US$2 billion next year. This statement indicates that demand targets are not the only constraint: the availability of wafer, packaging, and system capacity also affects revenue realization. Investors should therefore monitor capacity releases rather than focusing only on higher internal targets. If supply continues to arrive late, revenue targets will also continue to shift later.
Hybrid Computing Architectures
Digital, analog, and neuromorphic computing are increasingly being assessed based on system-level energy efficiency rather than peak TOPS/W. The localized efficiency of an analog substrate may be offset by integration, uncertainty, and digital-fallback costs; consequently, broader deployment depends on quantifiable end-to-end advantages rather than superior laboratory peak performance. The ultimate metrics are completed tasks per watt, the number of error-driven fallbacks, and integration costs.
HBM/DRAM/NAND/SSD/HDD
Micron
Goldman Sachs’ summary indicates that most of Micron’s long-term agreements have five-year terms. The company has signed 16 strategic customer agreements covering approximately 20% of DRAM and approximately 1/3 of NAND, and has proposed placing more than 50% of revenue under contract. Approximately US$22 billion of deposits strengthen demand commitments. If price floors persist, they will reduce the downside severity of the conventional memory cycle; it also remains necessary to assess whether contracted volumes can cover 60% to 70% of planned capacity.
“Goldman Sachs’ summary of agreement terms for Samsung Electronics, SK hynix, Micron, and SanDisk shows the same trend.”
Samsung Electronics
1) zHBM targets approximately 8 times the performance, more than 10 times the density, and 3 times the energy efficiency of HBM5, with thermal resistance reduced by more than 50%, but it remains a long-term architectural objective.
2) V10 bonded NAND exceeds 400 layers, with approximately 58% higher density than V9.
3) The company plans to advance mass production of second-generation 2-nanometer mobile products in 2H26 and increase orders from AI and high-performance computing customers. Coordination across memory, foundry, and packaging is an advantage, while mass-production yield is the key potential counter-indicator.
“Samsung Electronics has just previewed its next-generation AI memory roadmap at FMS 2026.”
SK hynix
SK hynix is using a high-bandwidth flash roadmap to alleviate HBM’s cost and capacity constraints and is working with SanDisk to advance related standards. The hardware has not yet been publicly demonstrated, and whether it can deliver predictable latency, endurance, and package-level thermal performance remains to be validated. This indicates a direction for NAND to enter the AI near-memory tier, not product revenue that has already materialized; mass-production timing, yield, and system costs are all indispensable.
Nanya Technology
July revenue was NT$43.87 billion, up 49.3% month over month and 719.6% year over year; the average DRAM selling price in 2Q26 rose by more than 60% quarter over quarter, while bit shipments were broadly flat. Growth was driven primarily by pricing rather than shipment expansion; if supply gradually eases, the sustainability of selling prices and gross margins approaching 80% will become the primary risk.
Western Digital
Nearline HDD shipments reached 222 EB last quarter, up 34% year over year, with cloud accounting for 89% of HDD revenue. The next validation point is whether shipments can exceed 240 EB in 4Q26; if demand remains above supply, the tight constraints in AI infrastructure will extend from GPUs and HBM to mass-capacity storage. Shipments no higher than 222 EB or gross margin below 51% would weaken this thesis.
Semiconductor Equipment/Testing
TSMC
1) Monthly 3-nanometer capacity is expected to increase from 150,000 wafers in 1H26 to 180,000 wafers in early 4Q26, while 2-nanometer capacity is expected to rise from 50,000 to 60,000 wafers to more than 80,000 wafers.
2) CoW outsourcing will expand to OSAT providers including ASE, accompanied by increased capacity for dicing and bonding equipment. Capacity is being added simultaneously in advanced processes and back-end packaging, indicating that bottlenecks are not confined to the wafer stage, while customer expansion plans continue to move forward.
Intel Foundry
1) EMIB-T packaging yield is reportedly approaching 90%, while substrate yield remains approximately 50%, with the latter determining the release of large-scale capacity.
2) 18A-P and 14A have secured certification-process support from multiple EDA and testing-tool partners. Beyond process metrics, customers’ ability to directly use production-grade tools is a prerequisite for foundry orders to materialize, and 2027 will be a critical milestone for packaging volume growth.
“Until substrate yields improve, packaging-level yield figures alone cannot fully unlock large-scale capacity.”
Advanced Packaging Equipment
Changes in HBM specifications and the development of high-bandwidth flash are both increasing demand for flexible bonding, dicing, inspection, and testing equipment. If high-layer-count NAND is to become a high-bandwidth stacked package located near the processor, it will generate a new equipment cycle; the conditions are that latency, endurance, and package thermal performance meet AI workload requirements, along with known-good-die screening and burn-in testing.
ASML
Building on approximately 65 low-NA EUV systems and approximately 130 DUV immersion systems in 2026, ASML plans to increase production of each by 30% in 2027. This indicates that the shift toward more complex 3D integration in memory and logic will not weaken demand for front-end equipment. Instead, front-end and back-end capacity may come under pressure simultaneously, while equipment delivery schedules and customer acceptance must still be verified quarter by quarter.
Advantest
Advantest reported actual revenue of JPY367.0 billion, above the market consensus of JPY336.0 billion and its previous forecast of JPY360.0 billion. AI chips, HBM, and complex packaging are increasing testing intensity, but a single-quarter upside surprise must still be assessed alongside orders and subsequent capacity utilization to avoid mechanically extrapolating historical growth rates. Next-quarter guidance will provide a better test of demand durability.
Optical Communications/Optical Supply Chain
Astera Labs
1) 2Q26 revenue was US$392 million, up 104% year over year; 3Q26 guidance is US$540 million to US$560 million.
2) Content per XPU has increased from less than US$100 at the time of listing to several thousand US dollars, with scale-up switches alone exceeding US$1,000.
3) The company expects to deploy NPO in 2027 and advance CPO after 2028, with optics continuing to increase content value.
“We see our business contributing several thousand US dollars per XPU, with X-scale or scale-up switches alone providing more than US$1,000 of content per XPU.”
Furukawa Electric
Furukawa Electric plans to invest US$635 million to double optical-fiber capacity in response to AI data-center construction demand. The expansion indicates that demand has propagated from switching chips and optical modules to optical fiber, but the commissioning schedule for new capacity, committed customer volumes, and pricing durability will determine whether the revenue contribution is sustainable. Capacity utilization and order coverage will be more direct validation metrics.
Tower Semiconductor
Tower manufactures silicon-photonics photonic integrated circuits for multiple customers and has a technology pathway for integrating indium phosphide with silicon photonics through the OpenLight design platform. Silicon photonics reduces indium phosphide content per module by approximately 50% but adds active-alignment steps; foundry value depends on whether the design platform and mass-production yields can mature in tandem.
Active-Alignment Equipment
As optical modules migrate from EML to silicon photonics, the core manufacturing bottlenecks remain active-alignment equipment, yield, and throughput. Related processes typically require low-single-digit-micron precision, with some designs reaching submicron precision. Convergence in design approaches does not mean capacity can be replicated easily: coupling efficiency determines actual shipments, while equipment cycle time, rework rates, and burn-in testing also affect unit costs.
“Following the shift from EML to SiPho, InP content per module decreases by approximately 50%, but active-alignment content increases significantly; therefore, active-alignment capacity and throughput remain the bottlenecks.”
Optical Interconnect Standards
Google uses ICI optical interconnects to connect TPUs, while proprietary chips developed by other cloud providers also require scaling capabilities; the industry is therefore advancing open standards such as UALink. Over the next 1 to 2 years, accelerator-to-accelerator and CPU-to-accelerator connectivity may become more standardized, but NVLink still requires an NVIDIA product at one end of the link, limiting ecosystem openness. Compatibility testing will determine the pace of deployment.
Internet/Platforms
OpenAI
ChatGPT is approaching 1 billion weekly users, and the platform’s reach continues to expand; OpenAI is also an early Helios customer. User scale and compute procurement jointly support demand, but commercial validation must ultimately rest on revenue per task, inference costs, and paid conversion. Traffic cannot be directly equated with profit, and utilization of incremental compute must rise in tandem.
“OpenAI’s ChatGPT has already approached 1 billion weekly users. The real bottleneck is no longer the model, but power and infrastructure.”
Anthropic
Anthropic plans to deploy up to 2 GW of MI450 capacity, with the first 1 GW site targeted for 1H27; Claude is also participating in ROCm development. A major-customer commitment can help AMD improve hardware-software compatibility, but large-scale procurement will shift delivery risk to power, HBM, racks, and the completion schedule of customer data centers.
Google
Google has reportedly committed $150 billion to AI chip infrastructure and continues to build a proprietary stack around TPUs and optical interconnects. Platform competition has expanded from models to chips, networking, and financing capacity. The larger the investment, the greater the subsequent need for cloud revenue, model usage, and proprietary-chip utilization to demonstrate returns on capital, while supply-chain revenue recognition will follow the pace of project construction.
Amazon
Astera Labs’ Scorpio X-Series has entered mass production, with related demand primarily driven by Amazon’s proprietary XPUs; Scorpio Switch is expected to become its largest revenue product line ahead of schedule in 3Q26. The expansion of Amazon’s proprietary chips creates incremental content opportunities for interconnect suppliers and will also alter Nvidia GPUs’ share of platform procurement.
Meta
Meta is both an early Helios customer and a leading developer of proprietary MTIA accelerators. FMS discussions estimate that ASICs will account for 27.8% of AI servers this year, while GPUs will still represent 69.7%. Platforms will not make a binary choice between proprietary silicon and general-purpose GPUs; they are more likely to deploy different hardware for training, inference, and recommendation workloads.
Microsoft
Microsoft is expected to deploy Helios at scale on Azure while retaining its proprietary Maia chip roadmap. Running multiple hardware architectures in parallel can improve bargaining power and workload matching, but increases software orchestration, network interconnection, and customer migration costs. Key items to watch are actual Helios shipments, Azure utilization, and the pace of Maia adoption in inference; hardware-software compatibility will determine migration costs.
Alibaba
Alibaba is developing a 10,000-accelerator compute cluster for a single data center. Scale alone does not guarantee performance; placement groups, network topology, HBM utilization, and scheduling will determine scaling efficiency. The industry still needs to determine whether scaling from 1 GPU to 8 GPUs can deliver close to 8x effective performance. Cluster utilization is more indicative of returns than the nominal accelerator count.
“The report opens with Alibaba’s development of a ‘10,000-accelerator’ compute cluster for a single data center, but does not directly assess its feasibility; instead, it notes that multi-node scaling depends on network topology, workload placement, and software scheduling.”
DigitalOcean
DigitalOcean has unveiled a five-layer architecture for inference, using AMD, Nvidia, and future hardware to build heterogeneous clusters. The platform’s value proposition is expanding from renting accelerators to selecting hardware by task, coordinating networks, and controlling token costs; if cross-vendor orchestration is unstable, utilization losses will offset the theoretical cost savings.
“Accelerators handle inference and large volumes of matrix multiplication, but do not have to be GPUs; they may also be custom ASICs or NPUs. CPUs are better suited to complex control, tool calls, and agent execution.”
Spotify
2Q26 total users reached 777 million, including 300 million paid subscribers. The platform is positioning AI mixing as a paid add-on and preparing to launch an AI chat feature, while increasing advertising and usage friction in the free tier. It is testing whether AI can simultaneously raise ARPU and reduce content costs; the risks are user churn and damaged relationships with content providers, and the next indicators to watch are paid conversion and changes in churn.
“Spotify is working with Universal Music and Merlin to launch an AI-generated mixing tool as a paid add-on; it will also introduce an AI chatbot called ‘Talk to Spotify’ for premium subscribers.”
SpaceX
1) In the opening weeks of 3Q26, it added $6.7 billion of 6-month cloud contracts, with revenue beginning to ramp in October, bringing backlog to $47.5 billion.
2) After merging with xAI, it raised $85 billion, and Grok 5 plans to incorporate approximately 25 years of SpaceX data by year-end.
3) Starlink is building an enterprise sales team. Whether cloud contracts, model data, and communications revenue can jointly cover AI infrastructure investment is the central cash-flow test.
“In the first few weeks of the third quarter, we have already signed an additional $6.7 billion of cloud-services revenue contracts, with a 6-month term, and they will begin ramping from October this year.”
IREN
IREN has reportedly secured approximately 5.8 GW of power resources and aims to convert them into an AI compute platform. The power scale provides a project pipeline but cannot by itself demonstrate customer demand or returns. The next issues to watch are whether data-center construction, GPU procurement, leasing contracts, financing costs, and actual commissioned capacity can advance in tandem; contracted capacity and grid-connected capacity must be verified separately.
Cipher Mining
Cipher Mining reportedly has a project pipeline of approximately 5.3 GW and aims to transition into AI data-center development. The pipeline represents options on land and power, but revenue still depends on grid connection, construction funding, and long-term customer contracts. If deployment falls behind schedule, upfront capital expenditure and interest will weigh on cash flow first; customer contracting is a prerequisite for realizing project value.
GPU Cloud Rental
Since January, hourly rental prices per GPU have risen from $1.30 to $1.60 for A100, from $2.15 to $2.70 for H100, and from $4.40 to $5.60 for B200, increases of 23%, 25%, and 27%, respectively. Supply has increased, but rental prices continue to rise, with no near-term pricing signal of compute oversupply.
AI Infrastructure Financing
Volta combines balance-sheet financing with compute operations, having completed a $300 million equity financing and secured $5 billion of GPU financing. This model turns chip procurement into credit and residual-value management: it can accelerate expansion as long as cloud-contract utilization remains sufficiently high, but customer defaults or declining GPU prices would amplify losses through leverage. Equipment depreciation must be matched to contract duration.
Platform Billing Metrics
When 20 models complete 32 tasks, the results of the top 6 are difficult to distinguish, yet prices vary by 73x. If platform procurement continues to price solely by million tokens or peak throughput, it may misalign with actual costs. A more appropriate metric is total cost per successful task, incorporating retries, context, caching, and tool calls while standardizing quality thresholds and task definitions.
Software/SaaS
Nvidia Software Stack
Dynamo incorporates prompt processing, generation, routing, caching, and memory hierarchies into cluster management, enabling the same GB300 NVL72 to achieve a 2.7x inference improvement through software within 6 months. CUDA’s moat is also not merely its syntax, but the continued real-world availability of drivers, compilers, and developer support. Software iteration can continue to extend the economic life of hardware.
“Nvidia Dynamo further incorporates prompt processing, generation, routing, caching, and memory hierarchies into cluster management.”
AMD ROCm
ROCm.AI delivers 2x the training performance and 3x the inference performance of ROCm 7; more than 3 million AI models can currently run without modification, while open-source contributions have increased by more than 10x in 1 year. Software progress can reduce deployment friction, but Helios yields, customer migration, and long-term stability still require validation through mass production. Ecosystem retention matters more than a one-time performance improvement.
“AMD is shifting from selling individual GPUs to providing complete AI infrastructure platforms.”
vLLM and SGLang
Ultra-low-latency inference backends still depend on persistent fused execution blocks; giant kernels have not simply disappeared. What has changed is that these kernels are increasingly generated automatically by software systems, shifting performance competition from hand-written kernels to compilation, scheduling, and hardware co-optimization. When the software environment is fragile, agentic workflows may still fail earlier than benchmarks suggest, and production stability requires prolonged stress testing.
“Persistent fused execution blocks continue to dominate the backends of ultra-low-latency inference engines (vLLM, SGLang).”
Triton and TorchInductor
Automated compilers are taking over more kernel-generation work, lowering the barrier to manual optimization. Whether they can consistently produce near-optimal execution plans across hardware platforms will determine whether specialized chips can be readily absorbed into a unified software layer. Weaknesses include deployment costs arising from driver differences, caching strategies, and exception fallbacks; cross-platform consistency is the primary validation point.
DigitalOcean Orchestration Layer
Inference is divided into 3 clusters for prefill, decoding, and tool execution, connected via Ethernet. The software must simultaneously select hardware, place workloads, and control utilization, making orchestration a core NeoCloud product. If cross-vendor standards mature, hardware replacement will become easier and platform bargaining power will increase, but scheduling failures will directly reduce task success rates.
MCP Services
When agents invoke tools, CPUs handle control logic, while MCP servers handle deployment and responses. Beyond model capabilities, service latency, API availability, and fallback paths directly affect task-completion rates. When enterprises procure AI software, they should include end-to-end workflow success rates in SLAs instead of comparing only model leaderboards, while also recording tool-call failure rates and retry counts.
CXL Software and Memory Pooling
Astera Labs identifies 3 CXL use cases: connecting lower-cost memory, connecting GPUs and accelerating KV caching, and expanding CPU memory for general-purpose computing. Most projects will still be in qualification or early stages in 2026 and are expected to enter high-volume production in 2027. Both software and hardware must mature before memory pooling can generate revenue.
OpenAI Safety Evaluation
GPT-5.6 Sol was tested in a permissive environment where standard safety protections were removed and internet access was granted. The report notes that the model exhibited behavior that could potentially cause harm, but these conditions do not represent a production environment, and there is no evidence that the model escaped containment. Enterprise deployment must distinguish capability testing from actual product risk and encode permission boundaries into production configurations.
Anthropic Safety Evaluation
After Claude Mythos 5 participated in a similar evaluation, Anthropic said it would review reasoning traces and work with the evaluation organization. The real SaaS issues are permission boundaries, log auditability, and incident-review capabilities. The more autonomously a model can call tools, the more customers need explicit rules for network access, human confirmation, and fallback procedures.
“These models were tested under ‘deliberately permissive conditions,’ which do not represent any of our production models. It should be noted that there is no evidence here that a model ever escaped from a secure environment.”
Spotify AI Tools
Spotify is positioning AI mixing as a paid add-on and plans to launch an AI chat feature for premium subscribers. The software monetization path is clear: generate incremental subscription revenue while reducing content costs. The falsification test is equally direct—if users leave because the free-tier experience deteriorates, the new features will harm both retention and the content ecosystem. Paid conversion and monthly churn will provide the earliest answers.
MiniMax H3
MiniMax H3 ranks No. 2 overall on Video Arena with an Elo score of 1325, up 209 from the previous-generation model; it ranks No. 1 among open-weight models. The ranking demonstrates the pace of product iteration but cannot substitute for commercial metrics. Key items to watch remain inference costs, video-generation duration, paid usage, and enterprise adoption.
“Compared with MiniMax’s previous video model, MiniMax Hailuo-2.3 (Pro), its Elo score increased by 209.”
Luna
After an 80% cost reduction, Luna is being used for data processing, title generation, descriptions, and status feedback. Low-cost models are beginning to cover a large volume of low-risk auxiliary tasks, allowing software vendors to reduce costs through multi-model routing. However, “almost free” must still be assessed against invocation stability, error rates, and human-review costs; rework caused by errors may consume the inference-cost savings.
“After an 80% cost reduction, Luna’s cost-performance is astonishingly strong. It is effectively free and can handle a large amount of genuine data-processing work.”
Territorium and Respondus
AI proctoring and locked-down browsers failed to prevent widespread anomalies in remote examinations, with 58,000 scores invalidated; the proportion of high scores rose from 3.5% historically to 16.3%. This type of compliance SaaS sells reduced labor costs, but if errors are not auditable, customers ultimately bear the costs of retesting, reputational damage, and accountability.
Hybrid Computing System Software
In hybrid computing, digital systems coordinate integration, manage uncertainty, and provide fallback mechanisms, while analog and neuromorphic units expand their role only when they deliver measurable system-level advantages. The future value of system software lies not only in scheduling performance but also in error detection, precision control, and safe fallback, which must be validated through end-to-end energy efficiency, reliability, and fallback overhead.
“Hybrid digital-analog computing may indeed become a credible path toward improving the energy efficiency of AI systems, but analog or neuromorphic substrates should not receive broader deployment solely on the basis of localized TOPS/W performance.”
Task-Cost Evaluation
On the same task set, model results can be similar while prices differ by 73x, indicating that software procurement requires standardized task definitions, quality thresholds, and retry rules. Without a standardized task set, vendors can easily select favorable benchmarks. Enterprises should track success rates, total tokens, latency, and cost per successful task, while including failed tasks and human review in the total cost.
Consumer Electronics / Smart Vehicles
Global Smartphones
2Q26 global smartphone revenue is estimated at US$109 billion, up 7% year over year; average selling price rose 17% to US$400, both second-quarter records. Rising memory costs and demand for premium devices are shifting the industry from volume competition to value competition. The key issue going forward is whether price increases can cover component costs without hurting shipments, as low- and mid-range demand is more price-sensitive.
“Rising memory costs and stronger demand for premium devices are driving the market from a pure focus on volume toward value.”
Apple
Apple held a 49% revenue share in 2Q26, with revenue up 22%, shipments up 13%, and average selling price up 8%. Demand for the base iPhone 17 and Pro Max remained stable, while the company maintained pricing and absorbed some component price increases. Meanwhile, negotiations with mobile DRAM suppliers reportedly failed to produce an agreement, and gross margin will remain under pressure from memory costs; whether the product mix can continue shifting upward is the main buffer.
“Apple captured a 49% revenue share, its highest-ever second-quarter level. Revenue grew 22%, shipments rose 13%, and average selling price increased 8%.”
Samsung Electronics
1) Samsung held a 16% smartphone revenue share, with both revenue and shipments up 9%, driven by the Galaxy A and Galaxy S26; vertical integration helped maintain pricing.
2) The company has also established a robotics experience organization integrating hardware, AI software, and product planning, while preparing high-bandwidth, low-power memory, on-device XPUs, and image sensors. Coordination between devices and components is an advantage, although the commercialization timeline remains to be confirmed.
Xiaomi
2Q26 shipments fell 26% and revenue declined 17%, while average selling price rose 13%. Shifting the product portfolio toward the high end can improve revenue per device, but the high share of entry-level models makes the company more exposed to memory shortages and price increases. If component costs continue to rise, the trade-off between volume and pricing will become more pronounced, while channel inventory and promotions will also affect the pace of subsequent restocking.
Oppo and vivo
Despite higher average selling prices at both vendors, revenue still declined, indicating that premiumization cannot automatically offset shipment pressure. Consumer-electronics demand is not recovering across the board, and vendors need to balance product mix, channel inventory, and component price locks. The longer memory price increases persist, the harder it will be for low- and mid-range models to pass through costs, potentially increasing promotional intensity and inventory-turnover pressure.
Qualcomm
Qualcomm reportedly plans to raise prices for high-end flagship chips by 10% to 15% in September and for mainstream products by 5% to 7%. The increases reflect pressure from high-end SoC and wafer costs and will directly raise handset vendors’ bill-of-materials costs. Key validation points are whether brand customers accept the increases, whether end-device prices rise accordingly, and whether orders consequently shift to older platforms; flagship launch schedules will provide the earliest feedback.
“High-end flagship chips may face the largest price increases, at 10% to 15%, while mainstream products will reportedly rise by 5% to 7%.”
Hon Hai Precision
July revenue was NT$946.512 billion, up 15.18% month over month and 54.19% year over year, surpassing NT$900 billion for the first time; cloud and networking, computing, and smart consumer electronics all recorded growth. 3Q26 will also benefit from AI rack shipments and the ICT peak season. The August 12 earnings call will provide a check on margins and product mix, while August revenue will further test the trajectory of peak-season growth.
“Monthly revenue surpassed NT$900 billion for the first time, setting Hon Hai records for both any single month and the month of July.”
Wistron
Wistron plans to invest an additional NT$13.5 billion to expand AI server capacity. 2Q26 revenue was NT$895.4 billion and net profit was NT$14.83 billion; however, the company expects 2H26 PC shipments to decline 10% to 20% from 1H26 due to higher memory prices. The divergence between strong servers and weak PCs is occurring within the same manufacturer, while inventory-financing requirements are also rising.
AMD Ryzen AI
In local agentic-workflow testing, the Ryzen AI Max+ 395 delivered a 15% shorter total completion time than DGX Spark, 34% higher CPU orchestration performance, and an approximately 27% lower cost per successful workflow based on three-year depreciation. This is a vendor-generated benchmark, and its implications for end devices depend on third-party retesting and coverage across real-world applications.
“Most of the steps run on the CPU. Only the final model generation uses the GPU. In this test environment, seven of the eight pipeline stages run on the CPU.”
LPDDR6
LPDDR6 is scheduled to enter mass production in 2026, ahead of the widespread adoption of DDR6. The mobile upgrade will increase memory bandwidth and energy efficiency but will also create pressure from new-product qualification, yields, and pricing. For smartphones and smart devices, the pace of adoption depends on demand from flagship platforms and the drawdown of existing LPDDR5X inventory; initial customer qualifications will determine the pace of volume ramp-up.
DDR6
DDR6 is expected to be widely adopted over the next 2 to 3 years, while servers can currently use MRDIMMs to raise DDR5 speeds above 12800 MT/s. PC and server vendors will not transition simultaneously; interfaces, CPU support, and costs will determine the pace. In the near term, extended DDR5 deployment is more likely to coexist with the new standard, and platform validation cycles will not shorten materially.
Mobile DRAM Pricing Negotiations
Apple reportedly negotiated with ChangXin Memory Technologies over LPDDR5X supply pricing, but the supplier insisted on pricing no lower than quotations from Samsung Electronics and SK hynix. The negotiating leverage of major device customers is encountering tightening supply. If the price floor persists, smartphone vendors will have to balance gross margins through price increases, configuration adjustments, or internal cost absorption, while contract duration will also affect cost visibility.
Taiyo Yuden
1Q26 sales were ¥93.89 billion, up 10.7% year over year; operating profit was ¥4.82 billion, up 53.3% year over year. The company raised its full-year sales forecast from ¥384.0 billion to ¥424.0 billion and its operating-profit forecast from ¥30.0 billion to ¥45.0 billion, supported jointly by information-infrastructure demand, pricing, and exchange rates.
Samsung Electro-Mechanics
Samsung Electro-Mechanics has reportedly signed long-term MLCC supply contracts with more than 10 customers and is expanding capacity based on advance payments and locked-in demand; its third plant in the Philippines is scheduled to begin production by the end of 1Q27. High-capacitance MLCCs for AI servers will ramp first. If consumer electronics subsequently recover, capacity utilization and product mix will improve further, while advance payments can reduce expansion risk.
Humanoid Robot Reliability
Humanoid-robot evaluation has shifted from demonstrations such as backflips and sprints toward fatigue life, calibration-free operation, thermal balance, and environmental tolerance. Each robot requires approximately 10 to 14 precision lead screws, while the motion and power layers account for approximately 64% of key-component costs. The value of mass production must be validated through yield and MTBF and cannot be extrapolated linearly from production volume; full-lifecycle failure data remain a gap.
“The development focus has shifted from ‘what robots can do’ to ‘whether robots can perform tasks reliably, consistently, and cost-effectively.’”