AI capex keeps climbing

Arista Networks sells Ethernet switch systems and the EOS software stack, primarily into data centers, as one reportable segment; Microsoft and Meta are its largest customers and two customers were cited at 42% of revenue. that is the networking leg of the thesis: scale-out and scale-across switch ports bought with hyperscaler capex. explicitly disclosed AI back-end plus front-end networking is guided at about $3.6b of roughly $12.6b 2026 revenue (~29%), and the broader cloud-titan data-center switching book sits on the same capex budgets, while campus (~$1.25b) and enterprise routing are outside the theme. q2 2026 revenue $3.036b, up 37.7% y/y, with 1.6 tbps AI fabric platforms and linear pluggable optics shipping. a data-center switch system vendor with a disclosed, growing AI cluster line but a meaningful non-AI enterprise and campus remainder → 0.7, between material exposure and pure play.

Amphenol is a diversified connector, cable and sensor maker (second-largest connector share globally) selling into automotive, mobile devices, defense, industrial, broadband and IT datacom. the thesis lives in IT datacom — copper and optical interconnect, backplane and cable assemblies for ai clusters, plus a role in the NVIDIA·Cisco USFP pluggable optics ecosystem. IT datacom was 48% of q3 2026 sales (up 117% y/y, 77% organic) and management says roughly three-quarters of that is ai-related, so ai interconnect is about a third of revenue and the fastest-growing piece; book-to-bill 1.19 on 99% order growth is the backlog signal the brief wants. the other half of the book — automotive 10%, mobile, defense, industrial, broadband — has no link to hyperscaler capex. material exposure for the networking and interconnect angle, not a pure play → 0.6.

Coherent is a photonics manufacturer whose largest business is optical transceivers (800G, 1.6T), lasers and optical components sold into AI data centers and telecom. in fiscal 2026 (june year-end) the Datacenter & Communications segment was $5,275m of $7,118m revenue (74%) and 79% in the june quarter, growing 40% for the year while Industrial (lasers, materials) fell to $1,844m (26%) and is being pruned via divestitures. the datacenter line alone grew 41% on AI transceiver demand, and the communications piece is partly AI-driven data center interconnect and scale-across. that is a direct fit for the networking and interconnect angle: optics are the product, hyperscaler cluster builds are the customer, and backlog leads revenue. the roughly one-quarter industrial and non-AI telecom remainder keeps it short of a pure play → 0.75.

Nvidia designs the GPU accelerators and the NVLink · InfiniBand · Spectrum Ethernet networking that hyperscalers, AI labs and neoclouds buy to build AI data centers. its Q2 FY2027 (quarter ended July 26, 2026) 8-K shows data center revenue of $89.0 billion against total revenue of $96.2 billion, roughly 92% of the book, up 117% year on year; gaming, professional visualization and automotive are the residual. the Q1 FY2027 10-Q puts hyperscalers at about half of data center revenue, with the rest from AI clouds, enterprise and sovereign buyers — the exact capex flow the thesis describes. for the accelerators and AI silicon angle this is the merchant accelerator supplier itself, so the thesis is the business → 1.0 pure play. net sign positive: rising hyperscaler capex converts directly into GPU and rack-system unit volume at 75% gross margin.

TSMC is the dedicated foundry that fabricates and packages nearly every merchant and custom AI accelerator — Nvidia and AMD GPUs, Broadcom and Marvell custom ASICs, hyperscaler in-house silicon — plus the CoWoS advanced packaging that binds HBM to them. it does not design accelerators, so it sits one step behind the angle's named category, but it is the common manufacturing chokepoint for the whole leg. by business mix: Q3 2026 revenue $46.7b, up 41% y/y, with the HPC platform at 68% of revenue driven by AI accelerators; narrowly defined AI accelerator revenue (AI GPU, AI ASIC, HBM controller) was high-teens percent of 2025 revenue and is guided to a mid-to-high 50s CAGR through 2029. smartphone is still 21% and HPC includes non-AI CPUs, PC and networking parts, so the accelerator leg is a large, rising share rather than the whole business → 0.6 material exposure. hyperscaler capex flows to TSMC whether chips are merchant or captive.

Micron is a vertically integrated memory maker — DRAM about 73% of revenue, NAND the rest — and the only US-listed HBM producer. fiscal Q4 2026 (sep 2026) revenue was $54.2b, with the cloud memory unit at $16.3b (30% of revenue, HBM and data-center DRAM) and the core data center unit at $18.0b (data-center DRAM plus SSDs), so data center is roughly 63% of sales; mobile and client (24%) and automotive and embedded (13%) are the non-thesis remainder. HBM revenue is growing faster than the company, most calendar 2027 HBM bit supply is contracted at higher prices, and 26 strategic customer agreements cover over 35% of revenue through 2030 — direct evidence that hyperscaler building programmes convert to committed memory orders. memory is the entire business but roughly a third of it is commodity and client-facing, so 0.8 for the memory angle rather than a full pure play.

AMD designs CPUs and GPUs for data centers, PCs, game consoles and embedded markets. the data center segment — EPYC server CPUs plus Instinct AI GPUs reported together — was $6.7b in Q2 2026, 107% higher year over year and 58% of the $11.5b total; client and gaming was about a third and embedded under a tenth. the accelerator leg of this portfolio is the Instinct line, which AMD does not break out, so the GPU share of total revenue is below half today even though it is the fastest-growing piece, with MI450 and Helios rack shipments to OpenAI and Meta (up to 6 GW each) and Anthropic (up to 2 GW) starting in 2H 2026. EPYC CPUs also sell into hyperscaler capex but are general compute. majority of revenue tied to data center capex, accelerators a rising minority → material exposure, 0.65 rather than pure play.

Credo is a fabless connectivity vendor whose entire product line — ZeroFlap active electrical cables, optical DSPs and ZF optical transceivers, retimers, SerDes chiplets and IP — links accelerators and switches inside hyperscale and neocloud AI clusters. fiscal 2026 revenue was $1.3 billion, up roughly 206% from $436.8 million, with the 10-K attributing the increase primarily to AEC unit volume ramps at hyperscale data-center customers; top-10 customers were about 90% of revenue and two hyperscalers each exceeded 10%. there is no non-data-center segment of any size, and the company's silicon photonics acquisition (DustPhotonics) extends the same interconnect franchise from copper to 1.6T optics. for the networking and interconnect angle this is the thesis itself: cables, DSPs and SerDes whose demand scales with ports per accelerator. pure play, 1.0.

Broadcom is a fabless semiconductor designer plus an infrastructure software vendor (VMware, CA, Symantec). in fiscal Q3 2026 AI semiconductor revenue was $16.7b of $29.6b total revenue (56%), up 221% year over year, with custom XPU accelerators about 73% of that AI revenue (roughly 41% of company revenue) and AI networking silicon the rest (roughly 15%). Q4 guidance of $21.7b AI semiconductor revenue on $34.8b total takes the AI share to about 62%. the remaining book is infrastructure software (30% of revenue) and non-AI semiconductors (about 14%), neither of which expresses the thesis. against the accelerators and AI silicon angle Broadcom is the largest merchant custom-ASIC designer, with six XPU customers and a $161.2b non-cancellable Anthropic commitment, but the software third keeps it short of a pure play → 0.7.

Modine makes thermal hardware. its Data Centers segment sells chillers, dry coolers, CRAC/CRAH units, fan walls, rear-door heat exchangers, coolant distribution units and immersion systems to hyperscale, colocation and neocloud operators — the cooling leg of the thesis. data center products were 35% of fiscal 2026 sales (up from 25%) and the segment grew 90% to $348.6m in the june 2026 quarter, against $261.6m commercial HVAC (itself partly data-center coils) and $277.8m vehicular Performance Technologies. the vehicular business was spun to Gentherm on 1 october 2026, so the remaining company is roughly 55–60% data center cooling by revenue and rising, with data center sales guided up 60–80% in fiscal 2027 and a $4bn+ capacity agreement with one hyperscaler covering 2027–2029. a majority but not whole business → 0.7, above material exposure, short of pure play because commercial HVAC&R remains about two fifths of sales.

Dell is a broad it hardware vendor: ISG (servers, networking, storage) plus CSG (commercial and consumer PCs). the thesis runs through the AI-optimized server line inside ISG — Q2 FY27 AI server revenue $16.4b of $47.0b total (about 35%), ISG $31.8b (68%) of which traditional servers and networking $10.5b and storage $4.9b, CSG $15.0b (32%) with no thesis exposure. FY27 guide of $74b AI servers on $192b revenue puts the line near 39% of sales, and a $95b AI server backlog against $60.9b of quarterly orders shows hyperscaler and AI-lab building programmes converting to committed orders ahead of revenue, exactly the integration leg. AI systems are the largest and fastest-growing product line but still well under half of revenue, with PCs and storage diluting the mix → material exposure, 0.6 rather than pure play.

Vertiv designs, builds and services data-center power and thermal infrastructure — UPS, switchgear, busbar, rack power distribution, air and liquid cooling, cold plates and CDUs, prefabricated modules, racks — plus a global services network on that installed base. FY2025 net sales were $10.2b with data centers the dominant end market (independent commentary puts it near 85%, communication networks ~10%, commercial and industrial ~5%). the thesis's power and cooling leg is literally the product set: kW per rack and liquid-cooling adoption lift content per data hall, and the committed-order channel is visible in backlog rising to $15.0b at end-2025 from $7.2b, organic orders up over 100% and Q3 2026 net sales of $4.3b up 61%. the ~15% telecom and industrial tail and the non-AI share of enterprise and colo data center work keep it just short of 1.0 → 0.9.

ASML sells lithography systems (EUV and DUV) and installed-base service and upgrades to chipmakers; it is the sole EUV supplier and roughly 90% of the lithography market. the whole business sits in the semiconductor equipment leg: Q3 2026 total net sales €11.7b, of which €8.4b systems and €3.3b installed base management, with logic 54% and memory 46% of system sales. management attributes the current order wave to AI-driven capacity expansion for advanced logic and DRAM (memory system sales guided up about 75% in 2026) and is adding 30% to EUV and DUV immersion capacity for 2027 with orders stretching to 2029–2030. but lithography demand is pulled by every semiconductor end market — China mainstream logic and memory, mobile, auto — and the company does not disclose an AI-specific revenue split, so AI capex is the dominant but not sole driver. second-order, lagged exposure across the entire book → 0.6.

Camtek sells optical inspection and metrology tools for semiconductor wafers, with the bulk of demand from advanced packaging — the CoWoS-type 2.5D/3D and HBM stacking capacity that foundries, OSATs and memory makers add when hyperscaler capex rises. management put AI-related products at about 50% of fiscal 2025 revenue with another 20% from other advanced packaging; by Q2 2026 advanced packaging was about 75% of revenue, guided to about 80% by year-end, with more than 20% of the $600m-plus 2026 order book from HBM customers. the remaining quarter is 2D inspection for CMOS image sensors, compound semis and silicon photonics. within the semiconductor equipment and test angle this is a focused advanced-packaging and HBM capacity play rather than a broad toolmaker, so the tier sits between material and pure at 0.75 — the thesis drives most of the business but not all of it.

Celestica is an electronics manufacturing and design services firm whose CCS segment (84% of Q3 2026 revenue of $5.83bn) builds data-center hardware for hyperscalers; ATS (aerospace, industrial, health tech, capital equipment) is the remaining 16%. against the server and rack integration angle, the directly matching line is Compute & Storage — AI servers, liquid-cooled racks and storage — at $1.6bn or 28% of revenue in Q3 2026, up 242% on hyperscaler AI compute program ramps. the larger CCS line is Networking ($3.3bn, 56%), 800G and 1.6T switch systems built under the HPS design-led model, which is data-center system integration but belongs to the networking leg. so the whole of CCS is AI build-out hardware, yet only a bit over a quarter of the company is server and rack integration proper, with that share rising fast → 0.6 material exposure rather than pure play for this angle. three CCS customers are 35%, 15% and 15% of revenue.

Jabil is a contract manufacturer (EMS) with three segments. the Intelligent Infrastructure segment — cloud and data center infrastructure, networking and communications, capital equipment — was 50% of fiscal 2026 net revenue of $36.0b, up from 32% two years earlier; within it, AI-related revenue (server and rack integration, liquid cooling, power and photonics built for hyperscalers) was about $14.4b, roughly 40% of sales, with four customers above $1b each. the other half of the company is Regulated Industries (automotive, healthcare, energy) at 35% and Connected Living and Digital Commerce at 15%, which have nothing to do with the thesis. for the server and rack integration angle this is the single largest line but not the whole business, and management guides AI-related revenue to $22.1b of $44.5b in fiscal 2027, about half → material exposure, 0.6.

Lumentum designs and manufactures optical components and systems — EMLs and CW lasers for transceivers, pump and narrow-linewidth lasers, 800G and 1.6T cloud transceiver modules (via the 2023 Cloud Light purchase) and optical circuit switches — sold to hyperscalers, AI infrastructure providers and network equipment makers. fiscal 2026 revenue was $3,014m, up 83%, with components $649m and systems $357m in the june quarter driven by cloud and AI; in fiscal 2025 the former Cloud & Networking segment was $1,411m of $1,645m (86%) and the industrial tech remainder has since shrunk relative to the AI ramp. the residual is industrial lasers, 3D sensing and legacy telecom, roughly a tenth to a fifth of sales, so the business is the networking and interconnect leg almost in whole → 0.9 rather than 1.0.

Lam Research sells wafer fabrication equipment — etch and deposition tools plus spares, upgrades and service (CSBG, roughly $2.5bn of $8.22bn september-quarter revenue). every dollar it earns is semiconductor equipment, so against this portfolio's semicap angle the business is a clean fit with a broad portfolio. the thesis link is second-order: hyperscaler capex → chipmaker and memory-maker capex → tool orders. management's own model is about $10bn of WFE per $100bn of AI capex, and the AI-tied lines are visible — memory 46% of systems revenue with DRAM spend aimed at HBM, foundry 46% led by leading-edge nodes and advanced packaging (up ~70% in calendar 2026), Aether dry resist ramping in HBM. dilution comes from china at 35% of revenue including mature-node spend, NAND conversion cycles and the installed-base service book, which track the broad WFE cycle rather than AI capex alone. pure semicap, material AI share → 0.7.

Super Micro Computer designs and assembles servers, storage and full rack-scale systems in a single operating segment, sold mainly to cloud, neocloud and enterprise AI buyers. this is the server and rack integration leg almost by definition: AI solutions were about 60% of fiscal Q4 2026 revenue and over 80% the quarter before, management guides AI above 80% going forward on backlog, and the company booked more than $60 billion of new orders in the June quarter with fiscal 2027 net sales guided to $65 to 72 billion versus $39.1 billion in fiscal 2026. the preliminary Q1 fiscal 2027 raise to $18.5 to 19.5 billion was attributed to AI platform demand. the non-AI remainder is CPU servers, storage and edge, a shrinking slice. near pure play for this angle; held just under 1.0 because the AI share swings quarter to quarter with project timing and gross margin sits at 10 to 18%.

Teradyne sells automated test equipment — semiconductor test is 82% of revenue (q3 2026: $1,072m of $1,306m), with product test ($138m, board, optical and high-speed interconnect test) and robotics ($96m, Universal Robots and MiR cobots) making up the rest. the thesis runs through test intensity: ai accelerators and custom asics raise soc test time per device, and hbm and server dram drive record memory test revenue ($301m in q3). management puts ai-related revenue above 60% of total sales, and the brief's own classification of equipment and test is second-order and material rather than direct. non-ai semi test (mobile, auto, industrial), robotics and defense product test are outside the thesis. that mix — majority ai-driven but a tool supplier one step removed from hyperscaler orders and carrying an unrelated robotics line — sits at material exposure, above the 0.6 floor for this angle.