Core Conclusions
The first landing point for AI compute CapEx is not "data center buildings" but "high-value compute units": capital flows first into GPUs/ASICs/TPUs and their supporting HBM, advanced packaging, and leading-edge wafers, then diffuses outward to rack-scale servers, switch chips, switches, optical modules, liquid cooling, and the power-distribution side. What forms high revenue and profit earliest and most directly is still the accelerator, HBM, foundry, and advanced-packaging links. Microsoft's latest disclosure puts its AI business at an annualized revenue run rate above 37 billion dollars; NVIDIA's FY2026 data center revenue reached 193.7 billion dollars; Broadcom's Q1 FY2026 AI semiconductor revenue hit 8.4 billion dollars with Q2 guidance of 10.7 billion dollars; Micron's FQ2 2026 revenue was 23.86 billion dollars at a 74.4% gross margin. All of this reflects capital accelerating along the "compute chip—memory—network—infrastructure" chain.
The largest revenue elasticity over the next 12–24 months will not necessarily come from server OEMs, but from links that are "supply-scarce with high per-unit value": GPUs/ASICs, HBM, advanced packaging, switch ASICs/AI networking, plus high-power liquid cooling and power-distribution equipment will show higher ASP and profit elasticity than general-purpose server OEMs/ODMs. NVIDIA's full-year FY2026 GAAP gross margin still reached 71.1%, Broadcom's Q1 FY2026 adjusted EBITDA margin was 68%, and Micron's gross margin has climbed to 74.4% in the AI era—evidence that the profit pool clearly concentrates in the high-barrier component layer.
The most direct core beneficiaries remain NVIDIA, TSMC, SK hynix, Broadcom, and Micron: they correspond, respectively, to accelerators, leading-edge process/packaging, HBM, custom ASICs/AI networking, and AI memory. NVIDIA's FY2026 data center revenue was 193.7 billion dollars; TSMC will raise 2026 capital spending toward the high end of 52 billion to 56 billion dollars and expects AI-related demand to be "extremely strong"; SK hynix is widely viewed as the HBM leader, and as early as 2025 customers were racing to lock in 2026 capacity; Samsung also said in Q1 2026 that current customer demand far exceeds supply and that the 2027 supply-demand gap could widen further.
AI server ODMs are major revenue beneficiaries, but not the center of the profit pool: Quanta, Wiwynn, Foxconn, Foxconn Industrial Internet, and Celestica will clearly capture orders and incremental revenue, but their margins are constrained by their systems-integration nature, customer bargaining power, and NVIDIA's growing control over the full system. Citing supply chain information, Tom's Hardware reports that NVIDIA is pushing more complete system integration to a higher level, compressing ODMs' differentiation and margin room.
The most real and most lasting bottleneck in this round of AI expansion is not a single GPU shortage, but a "stacking of multiple bottlenecks": HBM/DRAM supply, CoWoS/2.5D packaging, leading-edge process capacity, high-power supply and distribution, liquid-cooling retrofits, land and grid access, and construction and electrical labor will all jointly limit how fast clusters come online. When Meta raised its 2026 CapEx to 125 billion to 145 billion dollars, it explicitly cited higher component prices and data center costs; Microsoft and the market both point to rising memory/component prices lifting CapEx; Oracle's OpenAI data center project has also been reported as delayed due to material and labor shortages.
"AI CapEx is very high" does not equal "newly added physical compute grows proportionally year over year": part of the 2026 CapEx surge at the hyperscalers stems from sharp price increases in HBM, DRAM, NAND, advanced packaging, and certain key components. Microsoft-related figures suggest roughly 25 billion dollars of CapEx increment is tied to hardware costs; Meta likewise attributes its CapEx increase to higher component prices, especially memory. In other words, within the AI infrastructure chain, both "price" and "volume" are now driving revenue growth simultaneously.
The links with the best margins and strongest bargaining power remain concentrated in companies built on "standard-setting + ecosystem lock-in + capacity scarcity": GPUs and the software ecosystem sit with NVIDIA; leading-edge process with TSMC; HBM in the SK hynix/Samsung/Micron oligopoly; the custom ASIC and switch-chip inflection with Broadcom; EUV equipment with ASML; and advanced-packaging equipment benefiting Applied Materials and others. Applied Materials has guided its 2026 packaging business to grow more than 50%, showing that advanced packaging has moved from a "supporting link" to a "core profit amplifier."
The inference era will not simply weaken hardware demand; it will redistribute the value: training leans more on GPU count, HBM bandwidth, and scale-up networking; inference emphasizes cost-per-watt, memory bandwidth, caching and storage layers, rack-level cooling, power efficiency, and larger-scale scale-out networking. NVIDIA has explicitly positioned Grace Blackwell with NVLink as the "king of inference," stressing lower token cost; AWS is also advancing Trainium and data center architecture upgrades, showing that the inference era is not about "buying less hardware" but "switching to a different hardware mix."
In-house ASICs from cloud providers will erode NVIDIA's "share narrative," but not total semiconductor spending: Google TPU, AWS Trainium, and Meta's in-house inference/training chips essentially shift part of the CapEx from NVIDIA to Broadcom, Marvell, TSMC, advanced packaging, HBM, and networking components. AWS has disclosed that its chip business runs at an annualized revenue run rate above 20 billion dollars; Broadcom states clearly that its AI growth comes from "custom AI accelerators and AI networking." This means ASICs are more about ecosystem redistribution than total-pool compression.
What the market is most likely to overvalue is not "AI fabs" or "AI optical modules" themselves, but companies "with murky disclosure, low AI revenue mix, yet already priced as pure AI plays": typical examples include certain general-purpose server OEMs, general-purpose IDC REITs, early CPO-concept beneficiary chains, and small-cap liquid-cooling/connector companies that have not disclosed their AI revenue mix. By contrast, companies like NVIDIA, Broadcom, Micron, Microsoft, and AWS have at least provided evidence of AI revenue, AI guidance, component price increases, or order visibility.
The most obvious overheating in valuation has already spilled over from "the GPU itself" to "high-elasticity small caps in the supporting chain": when Barron's reported on Applied Optoelectronics, it implied a forward 12-month valuation of about 55.1x; Nebius continued to sharply raise its CapEx guidance to 20 billion to 25 billion dollars amid high growth, while its customer concentration and profit delivery still require verification; some CPO/high-speed optical-interconnect concept stocks are already trading on 2027–2028 penetration rates.
The most important industrial catalysts over the next 12–24 months are not a single new product launch, but five threads validating at once: first, actual shipments and system ramp of NVIDIA's Blackwell/GB300/Rubin; second, HBM3E/HBM4 qualification and capacity lock-in; third, delivery on TSMC's advanced-packaging capacity expansion; fourth, whether hyperscaler CapEx continues to be revised upward; and fifth, whether power, liquid cooling, and gigawatt-scale campuses connect to the grid on time. A problem in any one link would change the revenue-recognition cadence of the entire chain.
The biggest risk is not "AI ending," but "misallocation and deferral under strong demand": this includes a 2027–2028 cyclical swing-back after GPU/HBM/packaging supply over-expands, or, conversely, delayed compute deployment driven by grid, construction, material, and labor constraints. This would keep upstream chips highly prosperous while deferring revenue recognition for mid-stream servers/optical modules/data center infrastructure, creating a scissors gap inside the chain.
Transmission Chain Panorama
AI Compute CapEx Transmission Chain Overview
Chain Position Segment Core Products AI Demand Driver Revenue Recognition Key Customers Supply Bottleneck Margin Profile Representative Companies Public/Private Benefit Strength Investment Elasticity Core Evidence Upstream capital source Cloud providers/model companies/NeoClouds DC CapEx, GPU leasing, data center development Training + inference scale-up, model companies building/leasing compute CapEx deployment, long-term supply contracts, cloud service revenue Microsoft, Meta, Alphabet, Amazon, Oracle, OpenAI, xAI, CoreWeave, Nebius Power, land, construction, component prices End-user ROIC depends on utilization, not hardware margin Microsoft, Meta, Amazon, Alphabet, Oracle, Nebius, CoreWeave Public + private 10 10 Meta 2026 CapEx of 125 billion to 145 billion dollars, Amazon maintaining about 200 billion dollars, Alphabet/Microsoft both around 190 billion dollars per market figures, Nebius raising annual CapEx to 20 billion to 25 billion dollars. Core compute GPU/AI accelerators H100/H200/B200/GB200, MI series, TPU, Trainium, MTIA The most central value in training and high-end inference Chip/module/full-system shipments Hyperscalers, NeoClouds, model companies Wafers, HBM, CoWoS Among the highest gross margins NVIDIA, AMD, Google TPU, AWS Trainium, Meta MTIA Public + non-public/internal projects 10 10 NVIDIA FY2026 data center revenue of 193.7 billion dollars, Q4 data center revenue of 62.3 billion dollars. Memory HBM/DRAM HBM3E/HBM4, DDR5, LPDDR Accelerator bandwidth needs, inference caching, server memory upgrades Memory die/stack shipments NVIDIA, AMD, hyperscalers, ASIC makers Yield, TSV, wafers, packaging Extremely strong margins in tight cycles SK hynix, Micron, Samsung Public 10 9 Micron FQ2 2026 revenue of 23.86 billion dollars at a 74.4% gross margin; Samsung says current supply is far below demand, with a larger 2027 gap. Manufacturing Wafer foundry 3nm/5nm/4N/N3/N2 Full ramp of GPUs/ASICs/switch chips Wafer starts and long-term agreements NVIDIA, Broadcom, Apple, AMD, hyperscalers Leading-edge process capacity High margin, capital-intensive TSMC, Samsung Foundry, Intel Foundry Public 9 8 TSMC raising 2026 capital spending to a high 52 billion to 56 billion dollars, AI demand "extremely strong." Back-end manufacturing Advanced packaging CoWoS, SoIC, 2.5D/3D, interposer Rising GPU+HBM packaging complexity Package-unit shipments NVIDIA, AMD, ASIC makers CoWoS, substrates, test Scarce supply, improving margins TSMC, ASE, Amkor, Intel EMIB, Hanmi Semi, Applied Materials Public 9 9 Reuters points to persistent advanced-packaging tightness; Applied Materials expects 2026 packaging revenue to grow more than 50% year over year. Key materials ABF substrates High-layer-count substrates, FC-BGA Large packages, high-IO chips Shipments + long-term orders NVIDIA, AMD, Broadcom, Taiwanese ODMs Qualification cycles, yield Cyclical + structural AI improvement Ibiden, Shinko, Unimicron, Nan Ya PCB Public 7 8 Larger AI-chip packages and higher IO drive high-end ABF demand, but public financial disclosure by segment is limited and needs further verification. Mid-stream systems AI server ODM/OEM HGX servers, GB200/GB300 racks, rack-scale systems Cluster deployment, rack-level delivery Full-system/rack shipments Microsoft, Meta, AWS, Oracle, CoreWeave, xAI GPU, HBM, power and liquid-cooling adaptation High revenue elasticity, low-to-mid margins Quanta, Wiwynn, Foxconn, Wistron, Celestica, Foxconn Industrial Internet, Dell, SMCI Public 8 9 Celestica Q1 2026 revenue up 53% year over year; Tom's Hardware says NVIDIA is going deeper into system integration, squeezing ODM margin room. Networking Switch ASIC / switches 800G/1.6T switches, AI Fabric, InfiniBand, Ethernet Scale-up / scale-out network upgrades Chip/system shipments Hyperscalers, NeoClouds, server ODMs High-end ASICs, optical interconnect, software stack High gross margin at the chip layer, mid-to-high at the system layer Broadcom, Arista, NVIDIA Networking, Cisco, Celestica, Accton Public 9 8 Broadcom's AI revenue growth is driven by "custom AI accelerators and AI networking"; Arista's 2026 AI networking revenue figures keep being revised upward. Optical communications Optical modules/lasers/silicon photonics 400G/800G/1.6T, EML, DSP, VCSEL, silicon photonics Training-cluster scale-out, cross-rack bandwidth Module/component shipments Hyperscalers, switch makers, OEM/ODM Optical chips, DSP, yield Improving leader margins, but fast price competition Coherent, Lumentum, AAOI, Innolight, Eoptolink, Fabrinet Public 7 8 Coherent Q3 FY2026 revenue up 21% year over year; AAOI won hyperscale-customer orders, but profit delivery is still ramping. Storage Enterprise SSD / AI storage systems NVMe SSD, object storage, parallel file systems Training data lakes, RAG, checkpoints and inference caching SSD/system shipments, subscription maintenance Hyperscalers, enterprises, GPU clouds NAND, controllers, software integration Medium for SSD, higher for software-defined storage Micron, Samsung, Solidigm, Pure Storage, VAST Data Public + private 6 7 NVIDIA has tied its data center storage platform to BlueField-4, showing that AI storage is moving from optional to part of the system architecture. Cooling Liquid cooling Cold plates, CDUs, rack liquid cooling, immersion High power density of GB200/GB300 Equipment shipments + engineering integration Hyperscalers, OEM/ODM, IDC Design validation, on-site retrofit Margins improve fast, but heavily engineering-driven Vertiv, Schneider Electric, Eaton, Delta Electronics, Auras, Asia Vital Components, Submer, LiquidStack Public + private 8 8 AWS is advancing liquid-cooled DC architecture; a single GB300 NVL72 rack's liquid-cooling BOM already approaches 50,000 dollars. Data center power UPS/PDU/busbar/switchgear/transformers UPS, busways, switchgear, power-distribution systems Rising AI-rack power density, megawatt-scale campus buildout Product shipments + EPC/integration Hyperscalers, IDC, GPU clouds Switchgear, transformers, construction labor Mid-to-high margins, strong order visibility Vertiv, Eaton, Schneider Electric, Legrand, nVent, Hubbell Public 8 7 Oracle AI campus construction has been reported as delayed by materials and labor; AWS is also restructuring its power architecture. Downstream compute services GPU cloud/NeoCloud GPU instances, managed compute, leasing Asset-light expansion by model companies, short-term demand spikes Charged by duration/reserved capacity Meta, Microsoft, Anthropic, smaller model companies GPU supply, financing, customer concentration Extremely high revenue elasticity, but high CapEx and leverage CoreWeave, Nebius, Lambda, Crusoe Public + private 7 10 Nebius Q1 revenue up nearly 8x year over year, raising annual CapEx to 20 billion to 25 billion dollars and saying demand significantly exceeds GPU supply. Judging Where the Money Ultimately Flows
If we break down "a cloud provider/model company spending one dollar of AI CapEx" into its most likely ultimate economic destination, I lean toward the following order:
Accelerators and the HBM bound to them
Leading-edge process and advanced packaging
Rack-scale servers, NVLink/NIC/switch ASICs, switches, and optical modules
Liquid cooling and power-side retrofits
Storage systems and SSDs
Data center shells, civil construction, and general facility leasing
This is why many companies that seem "further from the GPU" actually deliver financial results faster than general-purpose IDC operators: in the AI era, the most expensive, most scarce, and earliest-locked items are precisely chips, memory, packaging, networking, and high-power infrastructure—not simple white-box floor space.
CapEx Breakdown and Demand Scenarios
Where Cloud Providers Mainly Spend Their AI CapEx
In FY26 Q3, Microsoft disclosed that its AI business already runs at an annualized revenue run rate above 37 billion dollars, with Azure still growing 40%; Meta raised its 2026 CapEx to 125 billion to 145 billion dollars and explicitly listed higher component prices and data center costs as the main reasons; Amazon maintains about 200 billion dollars of CapEx in 2026, and AWS's AI services already run at an annualized revenue run rate above 15 billion dollars; per market reports, Alphabet raised its full-year AI data center investment plan to about 190 billion dollars, and Google Cloud's order backlog reached 462 billion dollars. Taken together, the bulk of cloud providers' AI CapEx can be summarized as: accelerator systems, memory, full-system servers, networking, optical interconnect, data center power and cooling, and GPU cloud/third-party cloud contracts.
More importantly, the 2026 increase in hyperscaler CapEx does not come entirely from "building more data centers." Public statements from Meta, Microsoft, and Amazon all show that memory and component price increases have become a major driver of rising CapEx, meaning the same one dollar of CapEx increasingly flows to upstream scarce components rather than mid-stream low-margin systems.
A Rough Cost Structure for AI Servers and Training Clusters
Since vendors do not publish complete BOMs, public figures can only support range estimates. Combining NVIDIA's full-system price ranges, liquid-cooling BOMs, and upstream revenue structure, I prefer the following framework:
Object Cost Component Approximate Share Notes Single 8-GPU / HGX-class server GPU+HBM+motherboard/NVSwitch 55%–70% The absolute core of value; for a rack-scale full system, the share may be higher CPU+standard DRAM+SSD 5%–10% Important to a training machine's value, but below the accelerator NIC/DPU/switch connectivity 8%–12% Not negligible in a training cluster, especially high-speed networking Chassis, power supply, cooling/liquid-cooling adaptation 8%–12% Share rises as power density increases ODM/OEM integration, warranty and service 3%–5% Large revenue, but margins typically weaker than upstream chips GB200/GB300/rack-scale systems Accelerators+HBM+NVLink/NVSwitch 60%–70% NVIDIA delivers a deeper bundle, pushing value further up Networking/switching 10%–15% The larger the scale, the higher the share Power+liquid cooling 5%–10% A single rack's liquid-cooling BOM is already substantial Storage and rack integration 5%–10% Rubin-class systems may even carry higher NAND value The public market diverges considerably on NVL72 pricing: citing supply-chain figures, Tom's Hardware says the GB200 NVL72 costs about 2.8 million to 3.4 million dollars, while the GB300 NVL72 inference-oriented system can reach 6 million to 6.5 million dollars; the same source says Rubin NVL72 quotes have risen to a 5 million to 7 million dollar range. Meanwhile, a single GB300 NVL72 rack's liquid-cooling system BOM already approaches 49,860 dollars, and Rubin will push it higher still. Since these figures are not official vendor list prices, they are better suited to relative-share judgments than to precise absolute valuation.
Which Links Have the Highest Value, the Best Margins, and the Lowest Substitutability
The highest value sits with GPUs/AI accelerators, followed by HBM, advanced packaging, switch ASICs, and high-end networking. NVIDIA's FY2026 data center revenue of 193.7 billion dollars already far exceeds the revenue scale of nearly all server OEMs and networking suppliers; Broadcom's AI semiconductor revenue reached 8.4 billion dollars in a single quarter; Micron's gross margin rose to 74.4% in the AI memory cycle. All of this shows the profit pool still sitting up at the "key component layer."
The best margins sit with GPUs/AI accelerators, HBM, switch ASICs/custom ASICs, and certain equipment/software-bound links. Conversely, the easiest to substitute are general-purpose server contract manufacturing, low-end optical components, and general-purpose hardware without a software-hardware platform binding. Tom's Hardware reports that NVIDIA is reducing the ODM's role in the full system, which precisely shows that the mid-stream integration link has weaker long-term bargaining power than upstream.
The most likely to see supply-demand mismatch fall into four categories: First, HBM/DRAM, because demand lock-ins and supply expansion are out of sync; second, CoWoS/advanced packaging, because beyond wafers it is also constrained by interposers, test, and substrates; third, power equipment and grid connection, because data center projects come online far faster than the grid and switchgear can be delivered; fourth, liquid cooling, because large-scale on-site deployment, validation, and maintenance are not purely a capacity question.
Long-Term Compounding Attributes versus Short-Cycle Prosperity Attributes
Tracks with stronger long-term compounding include: GPU/AI accelerator platforms, leading-edge foundry, HBM leaders, switch ASICs/AI networking, and data center power and thermal-management platforms. These links are either locked in by software ecosystems, protected by very high capital barriers, or benefit long-term from "continually rising AI power density."
Tracks more tilted toward short-cycle prosperity include: server ODMs, general-purpose optical modules, ordinary DDR/NAND, ABF substrates, and a portion of GPU cloud companies that depend on specific projects. They can clearly capture the upcycle but are vulnerable to one or two customers switching platforms, inventory adjustments, or price competition.
Demand Transmission and Three Scenarios
Scenario Core Assumption Main Drivers Most-Benefited Links Representative Companies Main Risks Conservative Hyperscalers slow CapEx from 2027, with model efficiency improving faster than token growth Small models, compressed inference, cloud-provider ROI pressure Power/liquid cooling, network optimization, storage-efficiency tools Vertiv, Schneider, Arista, Ciena, Pure Storage Prices fall after GPU/HBM capacity expansion; NeoCloud financing under pressure Base Hyperscalers keep raising AI investment, but price increases absorb part of the physical expansion Training and inference advance together, with inference's share rising The full GPU/HBM/packaging/networking/liquid-cooling chain NVIDIA, TSMC, SK hynix, Broadcom, Micron, Arista, Vertiv Grid and construction slow the pace of coming online Aggressive Agents, multimodal, video generation, and enterprise inference all break out Token explosion, long context, video and agents continuously online GPU/ASIC, HBM, switches, optical modules, power distribution, GPU cloud NVIDIA, Broadcom, SK hynix, Micron, Celestica, Innolight, CoreWeave, Nebius Leverage/debt risk, valuation bubble, policy and export controls I lean toward the base scenario: inference will not weaken total hardware demand but will extend the value from "simply running more cards in training" to "larger-scale networking, power, cooling, caching, and storage." The latest disclosures from Microsoft, AWS, Meta, and Nebius all show AI revenue growth and AI CapEx expansion still happening in sync, with no chain-wide stall yet visible.
Deep Dive into Segments
Master Segment Table
Segment Segment Logic How AI Demand Becomes Revenue Current Supply-Demand Key Customers Price/Margin Trend Capacity Bottleneck Leaders Catalysts Main Risks Investment Appeal GPU Core compute for training and high-end inference Chip/module/system shipments Still tilted toward tight balance Hyperscalers, NeoClouds, model companies High ASP, high margin Wafers, HBM, packaging NVIDIA, AMD Blackwell/Rubin ramp In-house ASIC diversion, export restrictions 10 AI ASIC Cloud providers' in-house control of TCO ASIC design services + volume production Strong demand, share still smaller than GPU Google, AWS, Meta, OpenAI-related chains High margin once mature Leading-edge process, packaging, software ecosystem Broadcom, Marvell, TSMC TPU/Trainium/MTIA volume growth High software-migration cost, customer concentration 8 HBM A hard requirement for AI memory bandwidth Stack/die ASP rising Clearly tight NVIDIA, AMD, ASIC makers Price and margin rising sharply TSV, yield, wafers SK hynix, Micron, Samsung HBM3E/HBM4 qualification Pullback after over-fast expansion 10 DRAM/NAND HBM crowds out general memory supply Die/SSD/module shipments AI tightens the general market Server/PC/phone/SSD Cyclical upswing Competing for capacity with HBM Micron, Samsung, SK hynix Price increases and inventory restocking Large cyclical swings 7 Wafer foundry A required path for all high-end AI chips Wafer starts and long-term contracts Leading-edge process still tight NVIDIA, Broadcom, Apple, AMD High utilization improves margins 3nm/2nm capacity TSMC, Samsung, Intel N2, 3nm volume production Geopolitical risk 9 Advanced packaging The core of GPU+HBM complexity Package volume production Tight balance, leaning tight NVIDIA, AMD, ASIC chains Improving margins CoWoS, substrates, test TSMC, ASE, Amkor, Hanmi CoWoS expansion Downstream switching packaging routes 9 ABF substrates Indispensable for high-end FC-BGA Substrate shipments Structural improvement but still cyclical GPU/ASIC/switch chips Recovering from the bottom Qualification, layer count, yield Ibiden, Shinko, Unimicron, Nan Ya PCB Rising AI mix Strong cyclical attributes 7 AI server ODM Revenue amplifier Full-system/rack delivery Strong demand Microsoft, Meta, AWS, Oracle High revenue, low-to-mid margin GPU, liquid cooling, test Quanta, Wiwynn, Foxconn, Foxconn Industrial Internet, Celestica Volume delivery of GB200/GB300 Weak ODM bargaining power 8 Server power supplies and connectors Rising AI-rack power PSU, high-speed connector shipments Tilted tight ODMs, hyperscalers Improving margins Copper cabling/high-speed materials Delta, Amphenol, TE Rack power upgrades Fast project switching 8 Data center switches Scale-out infrastructure Switch system shipments Strong demand Hyperscalers, NeoClouds Mid-to-high margin ASIC, optical modules Arista, Cisco, Accton, Celestica 800G/1.6T upgrades CapEx pullback 8 Ethernet and InfiniBand The AI networking standards battle NICs + switches + software stack Dual-track in parallel NVIDIA, Arista, hyperscalers Chips better than systems Ecosystem compatibility NVIDIA, Broadcom, Arista Rising Ethernet penetration Compatibility-versus-performance contest 8 DPU/NIC Data movement and security offload NIC/DPU shipments Rising importance in the inference phase Cloud, storage, AI clusters Mid-to-high margin Hardware-software co-design NVIDIA, Broadcom, Marvell AI-native storage architecture Integrated into the SoC 7 Optical modules High-speed interconnect upgrades 800G/1.6T module shipments Volume up, price down Hyperscalers, switch makers Depends on capacity and price war DSP, lasers, yield Innolight, Eoptolink, AAOI, Coherent, Lumentum 800G→1.6T volume production Prices falling too fast 7 Silicon photonics A long-term direction, but penetration still early Chips/modules/platforms Hot theme, slow delivery Hyperscalers, switch makers Not yet fully proven Process and packaging Intel, Cisco/Acacia, Ayar, Lumentum CPO progress Commercialization later than expected 6 Storage systems Training data and inference caching SSD/parallel file systems/object storage Improving demand Hyperscalers, enterprise AI More stable margins from software-hardware combination Controllers, software Micron, Samsung, Pure, VAST Penetration of AI-native storage platforms Built into the cloud 6 Liquid cooling A must-have as AI density rises Cold plates/CDUs/system integration High growth Hyperscalers, ODMs, IDC Improving margins, strong engineering attributes On-site delivery and maintenance Vertiv, Schneider, Eaton, Auras, Asia Vital Components GB300/Rubin coming online Diverging technology routes 8 Data center power equipment The real scarcity is in grid connection and distribution UPS/PDU/switchgear/busbar Tilted tight Hyperscalers, IDC, GPU clouds High order visibility Transformers/switchgear/labor Vertiv, Schneider, Eaton, Legrand Gigawatt campuses Cyclicality and project delays 8 GPU cloud and NeoCloud The vehicle for outsourcing demand Charged by duration/reserved capacity Strong demand but capital-intensive Model companies, enterprises Revenue explosion, margins under pressure GPU supply, financing, customer concentration CoreWeave, Nebius, Lambda, Crusoe Large Meta/OpenAI/Anthropic deals High leverage and concentration 7 Key Judgments at the Segment Level
In this round of AI expansion, the strongest "industrial trend" does not equal the strongest "deliverable profit." For example, the narratives for both GPU cloud and server ODMs are strong, but the former is constrained by financing and concentrated customers, the latter by NVIDIA's system control and thin-margin attributes; whereas power, liquid cooling, and switch ASICs, though their stories are less dazzling than the GPU's, often deliver more solid financial results and are most likely to surprise expectations after 2026.
In addition, CPO/silicon photonics is worth tracking, but it currently looks more like a mid-term technology option than a near-term profit mainline. In its coverage of Ciena's earnings, Barron's noted that Broadcom's management threw cold water on the pace of the in-rack switch from copper to CPO, arguing it may not happen this year or even next; this shows the market's imagination around CPO may be running ahead of revenue delivery in the short term.
Investment Targets and Key Companies
Master Investment Target Table
Note: the table below prioritizes preserving AI benefit path, customer evidence, financial elasticity, and valuation judgment. Where public figures on "AI revenue mix, forward PE/EV/EBITDA, orders/backlog" are insufficient, the item is uniformly marked "needs further verification." This is not a gap but the boundary of public disclosure.
Company Ticker Market Segment AI Benefit Path Key Customers/Evidence AI Revenue Evidence Financial Elasticity Valuation Judgment Tier NVIDIA NVDA US GPU/networking Cloud-provider AI CapEx flows straight to GPUs, NVLink, networking AWS, Google, Azure, Oracle, Meta, CoreWeave, etc. with public deployments/partnerships FY2026 data center revenue of 193.7 billion dollars Extremely high Highest quality but not cheap A Broadcom AVGO US ASIC/switch chips TPU/Trainium/switch ASICs + AI networking Google, Meta and other custom chains Q1 FY2026 AI revenue of 8.4 billion dollars, Q2 guidance of 10.7 billion dollars Extremely high High quality, high expectations A Micron MU US HBM/DRAM/SSD AI memory and SSD AI data center customer base FQ2 2026 revenue of 23.86 billion dollars at a 74.4% gross margin Extremely high Cyclical upswing but already re-rated A Arista ANET US Switches Hyperscaler AI network upgrades Meta, Microsoft are high-exposure customers AI networking revenue figures revised upward continuously High Pricey but strong delivery A Vertiv VRT US Power/liquid cooling High-power and thermal-management retrofits for AI racks Hyperscalers, colos, OEMs Data center demand driving strong orders and CapEx expansion High Expectations already high A/B Applied Materials AMAT US Semiconductor equipment Packaging and front-end capacity expansion TSMC, Samsung, logic/packaging fabs 2026 packaging revenue growth guidance >50% High Relatively more balanced A Coherent COHR US Optical components AI datacom optical interconnect AI networking customer base Q3 FY2026 revenue +21% year over year, AI datacom benefiting Mid-to-high Already prices in plenty of AI expectations B Lumentum LITE US Optical components Optical modules/lasers AI data center customers Strong AI narrative, segment delivery needs ongoing tracking Medium On the high side B/C Applied Optoelectronics AAOI US Optical modules 800G/1.6T and hyperscaler orders Microsoft, Amazon, etc. Fast-growing data center orders, Barron's reports a high forward valuation Extremely high High elasticity, high risk B Oracle ORCL US AI cloud/data center OpenAI/Stargate/OCI AI clusters Large OpenAI, Meta deals Surging backlog contracts and long-term leases High High debt and execution risk B TSMC 2330/TSM Taiwan/ADR Wafer foundry/packaging Core of GPU/ASIC/switch chip manufacturing NVIDIA, AMD, Broadcom, Apple 2026 CapEx revised to a high level, extremely strong AI demand High Scarce, core position A Quanta 2382 TT Taiwan AI server ODM HGX/GB200 rack full systems Hyperscalers, cloud providers High prosperity in AI server business, segment detail needs further verification High revenue elasticity Margins below upstream A/B Wiwynn 6669 TT Taiwan AI server ODM High-end AI servers/racks Hyperscalers High AI mix, strong elasticity High Depends on shipment cadence A/B Hon Hai 2317 TT Taiwan Servers/racks AI server and rack assembly NVIDIA ecosystem, hyperscalers AI servers as a growth focus Mid-to-high Legacy business dilutes valuation B Delta Electronics 2308 TT Taiwan Power supplies/power Server power supplies, data center power ODMs/hyperscalers AI drives rising power density Mid-to-high Relatively balanced valuation A/B Auras 3324 TT Taiwan Liquid cooling/thermal Cold plates and thermal modules AI server chain Clear benefit, public segment detail needs verification High Hot segment B Accton 2345 TT Taiwan Switches White-box switches/AI Ethernet Hyperscalers 800G/1.6T following network upgrades High Needs tracking of customer cadence B Ibiden 4062 JP Japan ABF substrates High-end GPU/ASIC substrates GPU/CPU/ASIC makers Structural benefit, but financial elasticity depends on the cycle Medium Expectation gap in cyclical recovery B/C Shinko Electric 6967 JP Japan ABF/packaging High-end packaging substrates CPU/GPU/ASIC chains Real benefit path Medium Strongly cyclical B/C Fujikura 5803 JP Japan Optical connectivity/cabling High-speed interconnect, optical and thermal management Data center chain Benefiting from AI network buildout Medium Still has an expectation gap B Advantest 6857 JP Japan Test equipment AI chip test demand GPU/logic/memory makers Rising AI test load High High valuation but strong logic B SK hynix 000660 KS South Korea HBM/DRAM HBM bound to AI accelerators NVIDIA, etc. HBM leadership, tight supply Extremely high Heavily re-rated in 2026 A Samsung Electronics 005930 KS South Korea HBM/DRAM/Foundry Benefits at both the memory and logic ends Hyperscalers, AI chip customers Record chip profit in Q1 2026, says supply far below demand High Labor and execution risk B Hanmi Semiconductor 042700 KS South Korea Packaging equipment HBM/advanced-packaging equipment Memory makers Direct benefit from HBM expansion High Needs tracking of order durability B Innolight 300308 CH A-shares Optical modules 800G/1.6T optical modules North American cloud providers are core Real benefit, high AI mix High Already hot but strong delivery A/B Eoptolink 300502 CH A-shares Optical modules 800G/1.6T growth Overseas cloud customers Real benefit High High volatility B Foxconn Industrial Internet 601138 CH A-shares AI servers Server/rack assembly and system integration North American cloud providers, NVIDIA chain AI servers as one main growth point High revenue elasticity Average margins A/B Inspur 000977 CH A-shares AI server Domestic AI cluster servers Internet and government/enterprise customers Clear domestic benefit path, but heavy export-restriction disruption Mid-to-high Higher risk than Taiwanese ODMs B/C Victory Giant/Wus Printed Circuit/Shengyi Technology Multiple tickers A-shares PCB/materials High-frequency high-speed boards and materials Server/switch chains Indirect benefit, limited public AI mix Medium More cyclical + AI C Lenovo 0992 HK Hong Kong Servers/devices AI servers and edge AI Enterprise/cloud customers Real AI business but a large overall base Medium Cheap valuation but limited elasticity C GDS/VNET 9698 HK / VNET Hong Kong/US IDC Providing high-power data centers and colocation Hyperscalers/enterprises Benefit is relatively indirect Medium More of a real-estate + power logic C Schneider Electric SU FP Europe Power/liquid cooling/distribution Data center power and thermal-management platform Global cloud providers and IDC Direct benefit from AI campus construction High High quality but not cheap A Eaton ETN US/strong European business Distribution/liquid cooling UPS, switchgear, power + thermal-management acquisitions Data centers and industry Data center share keeps rising Mid-to-high Still has M&A catalysts A/B Legrand LR FP Europe PDU/busbar/racks Data center infrastructure Global customers Real but not pure benefit Medium Relatively balanced B/C ASML ASML Europe Semiconductor equipment Leading-edge process capacity expansion TSMC, Samsung, Intel AI driving EUV demand High Extremely strong long-term barrier A Deep Analysis of Key Listed Companies
Note: the following focuses on the companies "most worth further research." In the valuation column, financial data directly verifiable from company disclosure is listed where possible; forward valuations lacking a uniform public figure are marked "needs further verification."
NVIDIA
Chain position: GPU, NVLink, InfiniBand/Ethernet networking, AI system platform.
AI exposure: the highest. The primary beneficiary of hyperscaler AI CapEx.
Customers: AWS, Google Cloud, Microsoft Azure, Oracle, Meta, CoreWeave, Anthropic, etc. NVIDIA's FY2026 announcement also disclosed a multi-year, multi-generation strategic partnership with Meta and continued expansion of its AWS/Anthropic/CoreWeave partnerships.
Financials: FY2026 revenue of 215.938 billion dollars, up 65% year over year; data center revenue of 193.7 billion dollars, up 68% year over year; GAAP gross margin of 71.1%; net income of 120.067 billion dollars.
Orders/capacity: no separate backlog disclosure, but dense collaboration around Blackwell, Rubin, and hyperscale customers.
Moat: the CUDA software ecosystem, system-level integration capability, NVLink/NVSwitch, and a platform spanning from chip to factory scale.
Valuation analysis: high quality and high certainty, but market expectations are also extremely high; current detailed PE/EV/EBITDA needs further verification.
Catalysts: Blackwell/Rubin volume production, continued descent of the inference-cost curve.
Risks: slowing AI spending, in-house ASIC diversion, export controls.
Research conclusion: Strong beneficiary / high certainty / not cheap, but still the anchor of the chain.
TSMC
Chain position: core of leading-edge process and advanced packaging.
AI exposure: extremely high, but unlike NVIDIA it does not show up in a single AI metric; instead it shows up broadly through GPUs/ASICs/switch chips.
Customers: NVIDIA, Broadcom, AMD, Apple, hyperscaler in-house ASICs.
Financials and trends: Reuters reports record Q1 2026 profit, with 2026 capital spending positioned at the high end of 52 billion to 56 billion dollars, full-year revenue growth guidance revised upward to more than 30%, and AI demand "extremely strong."
Barrier: leading-edge nodes, packaging synergy, yield, and customer switching cost.
Valuation: a high-quality core asset; detailed forward valuation needs further verification.
Risks: geopolitics, rising overseas-fab costs.
Conclusion: Strong beneficiary / high certainty / an extremely strong long-term moat.
Broadcom
Chain position: custom ASICs, Ethernet switch chips, AI networking.
AI exposure: very high.
Customers: Google, Meta, and other hyperscaler custom chains.
Financials: Q1 FY2026 AI semiconductor revenue of 8.4 billion dollars, up 106% year over year; Q2 guidance of 10.7 billion dollars; Q1 total revenue of 19.311 billion dollars, adjusted EBITDA of 13.128 billion dollars.
Barrier: customer co-design, switch chips and the networking stack, economies of scale.
Valuation: a high-quality AI-substitution beneficiary, but the market has already priced it fairly fully.
Risks: customer concentration, custom ASIC project delays.
Conclusion: Strong beneficiary / high elasticity / not cheap.
SK hynix
Chain position: HBM leader.
AI exposure: extremely high.
Customers: NVIDIA is the most central.
Financials and trends: Reuters and the market broadly see it maintaining a clear lead in the HBM market; in 2025 there was already a state of 2026 supply being locked in ahead of time. The sharp 2026 re-rating of its share price and market cap shows the financial benefit has been confirmed by the market.
Barrier: HBM process first-mover advantage, customer qualification, the yield curve.
Valuation: both prosperity and valuation rising; watch for high-level volatility.
Risks: subsequent capacity expansion, Samsung catching up, customer concentration.
Conclusion: Strong beneficiary / high elasticity / high prosperity but cyclical risk cannot be ignored.
Micron
Chain position: HBM/DRAM/NAND/SSD.
AI exposure: high.
Financials: FQ2 2026 revenue of 23.86 billion dollars, GAAP gross margin of 74.4%, operating margin of 67.6%, capital spending of 5 billion dollars; cloud memory business revenue of 7.749 billion dollars, core data center business of 5.687 billion dollars.
Barrier: forms an oligopoly with SK hynix/Samsung.
Valuation: a dual lift from cycle + AI; if prosperity continues, earnings elasticity is very large.
Risks: a reversal of the memory cycle.
Conclusion: Strong beneficiary / high elasticity / strong cyclical attributes.
Arista Networks
Chain position: AI data center switches and Ethernet.
AI exposure: high.
Customers: Meta, Microsoft are high-exposure.
Evidence: MarketWatch reports it raised its 2026 revenue growth target to 25% and its AI networking revenue target to 3.25 billion dollars; Investopedia also notes that Meta and Microsoft together account for about 40% of company sales.
Barrier: the Ethernet operating system, hyperscale-customer qualification, the AI Ethernet trend.
Risks: customer concentration, white-boxing, ASIC supply.
Conclusion: Strong beneficiary / high certainty / valuation on the high side.
Vertiv
Chain position: UPS, PDU, thermal management, liquid cooling, rack infrastructure.
AI exposure: high, and easily underestimated.
Evidence: Barron's reports it benefits from AI data center infrastructure demand, with about 250 million dollars of 2025 capital spending to expand capacity; the market also broadly views it as a key platform vendor in the "chip-to-grid" chain.
Barrier: global delivery, project integration, service capability.
Risks: high expectations, project delays, intensifying competition.
Conclusion: Strong beneficiary / high certainty / expectations already high but long-term logic remains.
Celestica
Chain position: AI server and network platform manufacturing.
AI exposure: high.
Financials: IBD reports Q1 2026 revenue up 53% year over year to 4.047 billion dollars, with full-year revenue guidance raised to 19 billion dollars; the company also flagged multi-year growth in high-speed Ethernet switching and custom ASIC platforms at its investor day.
Barrier: high-bandwidth Ethernet platforms, deep binding with North American customers.
Risks: customer concentration and the low-margin ODM nature.
Conclusion: Mid-to-high beneficiary / high elasticity / need to track margins and order structure.
Quanta
Chain position: leading Taiwanese AI server ODM.
AI exposure: high.
Logic: benefits from NVIDIA HGX/GB200 and hyperscaler in-house AI clusters.
Public evidence: long-term high prosperity in the company's AI server business is already a market consensus, but English-language public segment disclosure is limited, so this report suggests treating it as a continued deep-dive name rather than giving overly granular financial figures here.
Conclusion: Mid-to-high beneficiary / high revenue elasticity / margins need ongoing verification.
Wiwynn
Chain position: high-end cloud server and AI server ODM.
Logic: deeply bound with hyperscalers, with an AI server revenue mix typically higher than diversified ODMs.
Conclusion: Mid-to-high beneficiary / high elasticity / worth a deep dive.
Foxconn Industrial Internet
Chain position: AI servers and rack systems.
Logic: one of the few manufacturing platforms in the Chinese market that can genuinely take on the North American cloud-provider and NVIDIA system chain.
Conclusion: Mid-to-high beneficiary / strong revenue elasticity / margins are not the biggest highlight.
Innolight
Chain position: high-end optical modules.
AI exposure: high.
Logic: a direct beneficiary of North American cloud providers' 800G/1.6T upgrades.
Risks: price competition.
Conclusion: Mid-to-high beneficiary / high elasticity / need to track ASP and yield.
Delta Electronics
Chain position: server power supplies, data center power distribution and energy management.
Logic: high-power AI racks raise the value of PSUs and power-supply systems.
Conclusion: Mid-to-high beneficiary / relatively strong certainty / relatively acceptable valuation.
Schneider Electric
Chain position: data center power distribution, thermal management, and integration.
Logic: an "infrastructure-platform beneficiary" of AI campus construction.
Conclusion: Strong beneficiary / high certainty / less of a trading stock than semiconductors, but steady financial delivery.
Ibiden
Chain position: ABF substrates.
Logic: high-end GPU/ASIC packaging drives substrate demand.
Conclusion: Medium beneficiary / cyclical recovery + AI expectation gap / need to track the substrate-prosperity inflection.
Applied Materials
Chain position: semiconductor equipment, especially advanced-packaging equipment.
Logic: the more complex AI chips become, the more packaging equipment is needed.
Evidence: the company's 2026 packaging business growth guidance exceeds 50%.
Conclusion: Strong beneficiary / mid-to-high certainty / relatively balanced valuation and quality.
Oracle
Chain position: AI cloud, data centers, and the OpenAI/Stargate hosting platform.
Logic: not an upstream chip company, but converting AI infrastructure CapEx into OCI revenue and long-term capacity contracts.
Risks: debt and construction execution.
Conclusion: Clear beneficiary / high risk / need to keep verifying financing and grid connection.
Further Research List
Starting from both "revenue-growth certainty" and "irreplaceability of industrial position," I think the listed companies most worth a continued deep dive are:
NVIDIA, TSMC, Broadcom, SK hynix, Micron, Arista, Vertiv, Applied Materials, Schneider Electric, Quanta, Wiwynn, Innolight, Delta Electronics, Celestica, Foxconn Industrial Internet. Of these, the first nine lean toward "high certainty," and the latter six toward "high elasticity."
Private Companies and Scoring Tiers
Important Private Companies and Primary-Market Opportunities
Company Country/Region Segment Core Products/Capabilities Customers/Partners Funding and Valuation Relation to Listed Companies Investment Focus Main Risks Crusoe US GPU cloud/data center development AI cloud, campus development, power infrastructure OpenAI/Stargate-related ecosystem, enterprise customers Needs further verification Synergy with NVIDIA, Oracle, data center developers Integrated "compute + energy + campus" Capital-heavy, project execution Lambda US GPU cloud AI training and inference cloud Model companies, enterprises Needs further verification Deeply bound with the NVIDIA ecosystem A key entry point for smaller model companies outsourcing compute GPU supply and financing Groq US Inference chips LPU inference acceleration Enterprise and inference scenarios Needs further verification Partly a substitute/complement to NVIDIA Potential for inference cost advantage Ecosystem and software migration Ayar Labs US CPO/optical interconnect Optical I/O chips Chip and system makers Needs further verification Interaction with the Broadcom, Intel, NVIDIA ecosystem If CPO lands, it will carry extreme leverage Late commercialization cadence Celestial AI US Memory interconnect/optical interconnect Photonic Fabric ASIC/GPU ecosystem Needs further verification Related to advanced packaging and memory-bandwidth routes Solves the "compute constrained by I/O" pain point Uncertain penetration timing LiquidStack Europe/US Liquid cooling Immersion liquid cooling Data center customers Needs further verification Competes and cooperates with Vertiv/Eaton/Schneider Clear benefit if 3,600W-class GPUs become widespread Route competition Submer Europe Liquid cooling Immersion and cooling infrastructure Data center customers Needs further verification Competes and cooperates with listed power/thermal-management companies Large elasticity in high-density AI scenarios Commercialization scale OpenAI infrastructure projects US In-house compute/leasing Stargate, third-party cloud mix Oracle, AWS, Google, CoreWeave OpenAI's valuation and funding figures are huge and need ongoing verification Directly drives upstream procurement Watch shifts in its "build vs. lease" structure Financing, governance, partnership terms xAI infrastructure projects US In-house compute Colossus mega cluster NVIDIA, multiple supply chain parties Funding and valuation need ongoing verification Drives GPUs, power, and cooling at scale Watch power density and infrastructure expansion Environment, grid connection, regulation Company Tiers and Investment Priority
Category Definition Representative Companies Rationale for Categorization Tier A Core direct beneficiaries of AI compute CapEx NVIDIA, TSMC, Broadcom, SK hynix, Micron, Arista, Vertiv, Schneider, Applied Materials Can clearly convert CapEx into revenue/profit and sit in high-barrier links Tier B Clear beneficiaries, but with cyclical or valuation risk Samsung, Celestica, Quanta, Wiwynn, Innolight, Eoptolink, AAOI, Coherent, Oracle, Delta Clear business benefit, but disrupted by price, project landing, customer concentration, or valuation Tier C May benefit long-term, but near-term financial elasticity is weak Equinix, Digital Realty, Legrand, Lenovo, some PCB/materials, some ordinary storage companies The logic exists, but the transmission chain is long or the AI mix is not high Tier D Strong AI narrative, but insufficient actual financial evidence Some pure CPO concepts, some small-cap liquid-cooling/connector concepts, some general-purpose IDC concept stocks No sufficient evidence of AI orders, AI revenue mix, or profit delivery Scoring Model and Total-Score Ranking
I suggest the weights follow the user-given framework:
Direct AI demand exposure: 25%
Chain position and bargaining power: 20%
Financial quality: 15%
Growth elasticity: 15%
Technical barrier: 10%
Valuation reasonableness: 10%
Future catalysts: 5%
Using this framework, I provide a research priority ranking:
Rank Company Total Score Scoring Rationale Summary NVIDIA 90 The highest in direct exposure, bargaining power, technical barrier, and earnings certainty TSMC 88 Irreplaceable in the chain, with the broadest benefit reach SK hynix 86 HBM leader with scarce supply, but cyclicality slightly higher than the top two Broadcom 84 ASIC + AI networking dual drive, concentrated customers but strong delivery Micron 82 Extremely large AI memory elasticity, slightly higher cyclical attributes Arista 81 A high-certainty beneficiary of Ethernet AI networking Vertiv 80 A power/liquid-cooling platform beneficiary, high expectations but solid logic Applied Materials 79 Clear benefit from packaging equipment, relatively balanced valuation Schneider Electric 78 A "shovel seller" for AI campuses, financially steady but a slightly weaker trading stock Quanta 77 Very strong revenue elasticity, but a profit-pool position below upstream Innolight 76 Clear benefit, but price competition must be watched closely Delta Electronics 75 Benefits from AI PSUs and power distribution, relatively steady valuation Celestica 74 High elasticity, but customer concentration and the manufacturing nature cap valuation Samsung Electronics 73 Real benefit, but internal execution and labor variables are pronounced Oracle 70 Strong growth, but high risk too Valuation, Risk, and Final Conclusions
Valuation and Market-Expectation Analysis
Companies already pricing in AI expectations fairly fully: NVIDIA, SK hynix, Arista, Vertiv, AAOI, Nebius, and parts of Coherent/Lumentum valuations—the market no longer prices them on a traditional cyclical framework alone, but trades on the 2027 scale of AI infrastructure. When Barron's reported on AAOI, its forward 12-month valuation of about 55.1x is a typical example.
Companies that still have an expectation gap: I prefer some names among power and thermal-management platforms, advanced-packaging equipment, ABF substrate leaders, and Ethernet AI networking leaders. The reason is not that they haven't risen, but that the market still tends to treat them as "supporting components" rather than "bottleneck assets." Oracle's project delays, the AWS Titus architecture upgrade, and Meta's component price increases all prove that "power distribution + liquid cooling + infrastructure" has moved from optional to required.
Good companies but too expensive: NVIDIA, Arista, Vertiv, some optical-module/optical-component names, Nebius. They may still be good companies, but not necessarily a "good price" at any given moment. By contrast, names like TSMC, Applied Materials, and Schneider are often more balanced between "quality and valuation."
Cyclical bottom reversal + AI demand: Micron, Samsung, ABF substrate-related names, and some high-end PCB/material chains are the more typical "cycle + AI" overlay. If the industry does not over-expand, this kind of combination sometimes beats pure-narrative segments on fundamental delivery.
Largest near-term earnings elasticity: Micron, SK hynix, Broadcom, Celestica, some high-end optical-module companies. Strongest long-term moat: NVIDIA, TSMC, ASML, Broadcom, some HBM leaders. Most asymmetric risk-reward: high-leverage GPU clouds, optical-module/CPO concept stocks already priced for multi-year-out delivery, and companies that depend on large deals but lack ongoing segment disclosure.
Risk Analysis
The most important risk in this theme is not a single "AI bubble," but the following concurrent risks:
Cloud-provider AI CapEx slowdown: current hyperscaler CapEx is enormous; if ROI pressure rises in 2027, it will first hit mid-stream systems and supporting components.
GPU/HBM supply loosening: if 2027–2028 capacity expands too fast, memory and parts of packaging are hit first.
ASIC substituting for GPU: this will erode NVIDIA's share but not necessarily total spending, more likely carving the profit pool over to Broadcom, TSMC, etc.
Model efficiency gains: may lower per-token hardware demand, but may also expand the deployment scope; the net effect on total volume bears watching.
Server and optical-module price competition: mid-stream and some optical components may have "volume without profit."
Data center construction delays: Oracle/OpenAI projects have been reported as delayed by materials and labor.
Power and grid constraints: the real constraint on AI campus construction increasingly comes from grid connection and power supply/distribution, not the GPU itself.
Customer concentration: Arista, AAOI, GPU clouds, and some ODMs all have concentrated large customers.
Geopolitics and export controls: leading-edge process, equipment, and high-end GPUs are all affected.
Valuation-bubble risk: especially in NeoClouds, optical interconnect, and small-cap supporting names.
Final Conclusion
AI compute infrastructure is not a "supporting role" in the AI value chain, but the layer where the real profit pool and the real bottlenecks are most concentrated. If we view AI as an industrial system, then the CapEx of cloud providers and model companies will not scatter evenly across the whole chain, but flow highly concentrated to a few links that hold technical barriers, scarce supply, and customer lock-in: GPUs/ASICs, HBM, leading-edge process and advanced packaging, AI networking, and power and liquid-cooling infrastructure.
I think the five segments most worth watching now are:
GPUs/AI accelerators
HBM and AI memory
Advanced packaging and leading-edge process
AI networking and switch ASICs/switches
Data center power and liquid cooling
The ten listed companies most worth in-depth research are:
NVIDIA
TSMC
Broadcom
SK hynix
Micron
Arista Networks
Vertiv
Applied Materials
Schneider Electric
Quanta
The five private companies/projects most worth ongoing tracking are:
Crusoe
Lambda
Groq
Ayar Labs
OpenAI/Stargate infrastructure projects
The three points most easily misunderstood by the market are:
"AI CapEx growth" does not equal "newly added physical compute growing at the same rate year over year", because much of the 2026 increment comes from memory and component price increases.
"Server revenue growth" does not equal "server companies earning the most profit"; the profit pool still sits in GPU/HBM/ASIC/packaging.
"More efficient inference" does not equal "falling hardware demand"; in many cases it just shifts demand from training to networking, power, cooling, caching, and low-cost compute architectures.
The indicators most worth tracking over the next 6–12 months include:
Hyperscaler quarterly CapEx and cloud backlog
The actual delivery cadence of NVIDIA's Blackwell/GB300/Rubin
HBM3E/HBM4 qualification and order lock-in
Delivery on TSMC's CoWoS/advanced-packaging expansion
Data center grid-connection capacity and campus MW/GW progress
Ethernet 800G/1.6T and liquid-cooling penetration
For the next research list, the directions I suggest prioritizing for secondary deep dives are:
HBM: the strongest profit elasticity, but also the most prone to "misjudging cycle versus prosperity."
Advanced packaging: it determines the true deliverable capacity of GPUs/HBM.
AI server ODM: high revenue elasticity, but the margins and NVIDIA system integration need to be disentangled.
GPU cloud / NeoCloud: financing, contracts, power, and customer concentration are critical.
Liquid cooling and power distribution: the "bottleneck infrastructure" most likely to be undervalued by the market in the future.
If I had to pick a single secondary research direction most worth a continued deep dive, I would prioritize "HBM + advanced packaging": these two links are the core crossover point in the current AI CapEx transmission chain that can simultaneously explain revenue-growth certainty, margin expansion, supply bottlenecks, and valuation divergence.
Open Questions and Limitations
Some companies in Taiwan, Japan, and A-shares disclose AI business segments less transparently than US stocks; this report has tried to avoid fabricating figures, and any unverified AI revenue mix, orders, and forward valuations are marked "needs further verification."
The precise BOM of a single AI server or full rack system is not a publicly disclosed item by vendors; this report uses range estimates, suitable for industrial research but not for precise cost-method valuation.
Funding, contracts, and utilization at NeoClouds and primary-market companies change quickly, suited to high-frequency tracking and should not rely on a single snapshot.
This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.
Full report
Sign in to read the full report
Sign up free to unlock the full text, the Baillie growth scorecard, and full-text search.
Log in / Sign up free