Pick-and-Shovel Investment Analysis · 2026
A full-stack analysis of $600–650B in hyperscaler infrastructure spending and the public companies positioned to capture it.
Step 01 — The Spenders
Across Microsoft, Amazon, Alphabet, and Meta, 2026 capex guidance has entered "industrial policy" territory. Over $600–650B annually, with ~70–75% AI-specific infrastructure: chips, servers, data centers, networking, and power.
MSFT
Microsoft
Vast majority to AI chips and AI-optimized data centers. FY26 run rate from external synthesis.
GOOGL
Alphabet / Google
~60% servers, ~40% data centers/networking. DeepMind, Google Cloud, and Gemini infrastructure.
AMZN
Amazon / AWS
Predominant use: AWS and AI infrastructure — data centers, custom chips (Trainium/Inferentia), satellites.
META
Meta Platforms
Step-function increase explicitly tied to AI compute and "Meta Compute" / "personal superintelligence."
Step 02 — Value Chain Analysis
Every dollar of AI capex flows through a multi-layer value chain. We map 18 public companies across five layers: compute, memory, networking, power/thermal, and energy.
Nvidia
AI accelerators + networking "system" vendor
>90% share of the AI accelerator market by revenue; GPUs remain "a generation ahead" of Google's TPUs per Nvidia's own Blackwell launch commentary.
Q4 2025 data center revenue ~$51B out of $57B total quarterly revenue, +66% YoY, with "cloud GPUs sold out."
~90% of customers now buy Nvidia's switches, NICs, and interconnects alongside GPUs — making it a rack-scale AI platform, not merely a chip vendor.
Mega-deals: Oracle spending $40B for 400,000 GB200 GPUs; a planned $500B U.S. manufacturing/infrastructure program reinforces its centrality.
Competitive Moat
CUDA software ecosystem, full-stack platform (GPU + networking + libraries), and privileged access to TSMC CoWoS capacity.
Advanced Micro Devices
Second-source AI GPU provider
MI300/MI325/MI355 line is the de facto second source for hyperscalers: deployments at Microsoft Azure, Meta (Llama inference), Oracle, and others.
Azure deploys tens of thousands of MI300X GPUs for OpenAI; Meta has "broad deployment" for Llama. Revenue ramped to >$1B/quarter in 2024.
Competitive Moat
Higher on-package HBM capacity (192–288 GB), improving ROCm software, and hyperscalers' explicit desire not to be single-sourced to Nvidia.
TSMC
Foundry + advanced packaging
Near-monopoly provider of 3 nm and 2 nm wafers plus CoWoS advanced packaging used by Nvidia, AMD, Broadcom, Marvell, and hyperscaler in-house chips (TPU, Trainium, Maia, MTIA).
CoWoS capacity sold out through 2025 and into 2026, scaling from 35–40k wafers/month (2024) → 90–110k (2026). Nvidia alone consumes ~60% of CoWoS output.
Competitive Moat
Process leadership at 3 nm and 2 nm, plus advanced packaging (CoWoS, SoIC) as a gating resource for all AI GPUs and custom ASICs.
"AI training and inference are now HBM-limited. HBM TAM is projected to grow from $35B (2025) to $100B by 2028 — a ~40% CAGR."
HBM Supercycle AnalysisSK Hynix
HBM kingpin
~57–62% share of the HBM market in 2025. Roughly 90% of Nvidia's HBM supply comes from SK Hynix, making it the critical memory bottleneck for H100/H200 and early Blackwell systems.
Briefly overtook Samsung in overall DRAM profits, explicitly driven by AI HBM demand. HBM is now its highest-margin segment.
Competitive Moat
Packaging and yield leadership in HBM3/HBM3E; entrenched position as Nvidia's primary HBM supplier.
Micron Technology
U.S. HBM supplier
~20–21% HBM share (Q2 2025), second only to SK Hynix and ahead of Samsung in some quarters. Sole U.S.-domiciled HBM supplier.
Beneficiary of HBM capacity sold out through 2026, with hyperscalers seeking geographical diversification from Korean supply. Also pushing high-capacity DDR5 and data-center SSDs.
Competitive Moat
U.S. domicile (CHIPS Act tailwind), growing HBM3E competency, and broad DRAM/NAND portfolio for AI server shipments.
Samsung Electronics
DRAM + recovering HBM share
HBM share recovered from mid-teens to ~22–35% by late 2025 with strong HBM3E uptake; analysts expect significant HBM4 share for Nvidia's Rubin platform.
One of three global HBM suppliers. Any successful qualification for Nvidia/AMD directly translates into meaningful AI revenue leverage.
Competitive Moat
DRAM scale, HBM4 roadmap, and vertical integration with Samsung Foundry.
Broadcom
Custom accelerators + switch ASICs
AI semiconductor revenue reached ~$20B in FY25, up 65% YoY; total company ~$64B, making AI chips ~30%+ of sales. AI-related backlog exceeds $73B.
Major wins: multi-billion-dollar Google TPU rack orders; a $100B+ lifetime value OpenAI accelerator/networking deal (2026–2029); additional hyperscaler XPUs.
Supplies leading Ethernet switch ASICs (Tomahawk/Trident) and custom AI accelerators, giving Broadcom deep lock-in at both compute and fabric layers.
Competitive Moat
Co-designed ASICs, end-to-end system model (XPUs + networking + software), and hyperscaler concentration.
Arista Networks
AI data center Ethernet
AI center revenue $1.5B (2025) → $2.75B (2026), within a ~$10.65B total revenue outlook (~26% AI mix). 20%+ revenue growth and 60%+ gross margins.
Core supplier of 400G/800G switches for back-end training networks and front-end inference fabrics at hyperscalers. Analysts model revenue nearly doubling by 2030 off AI networking.
Competitive Moat
Deep incumbency with cloud titans, high-end Ethernet switching portfolio, and software-defined networking stack tuned for leaf-spine AI fabrics.
Marvell Technology
Custom AI ASICs + silicon photonics
Custom silicon for Amazon (Trainium), Microsoft (Maia), and other hyperscalers: XPUs, CXL controllers, optical DSPs, and switch silicon. Engaged with 3 of the top 4 hyperscalers.
2 nm data infrastructure platform integrates compute dies, HBM4 interfaces, and 1.6T optical I/O on a single package. Co-packaged optics (CPO) ports projected to rise from <50k today to >18M by 2029.
Competitive Moat
Custom ASIC platform, silicon photonics IP, and multi-hyperscaler engagement (3 of the top 4).
Coherent
Datacenter optics and CPO
Accelerating demand for 800G and 1.6T datacenter optics; datacenter bookings book-to-bill >4x with visibility extending into 2027.
Ramping 6-inch InP capacity; won a "landmark" CPO order from a leading AI datacenter customer, with revenue starting late 2026.
Competitive Moat
Vertical integration in InP lasers/EMLs and leadership positions across 800G/1.6T modules and emerging CPO.
Lumentum
AI optics enabler
Management states "virtually every AI network is powered by Lumentum technology" either directly or via OEMs, reflecting a deep role in high-speed datacenter transceivers and optical circuit switches.
Components segment revenue grew 68% YoY with strong AI/cloud demand; inventory building specifically to support forecast AI network growth.
Competitive Moat
Entrenched position in 400G/800G optics, OCS, and scale-out/scale-up optical fabrics.
Vertiv Holdings
Thermal and power backbone for AI data centers
2024 net sales ~$8.0B (+17% YoY); Q1 2025 up +24% YoY; backlog of $7.9B. Management explicitly cites AI and cloud as primary demand drivers.
Core products: UPS, power distribution, and increasingly direct-to-chip and immersion liquid cooling — CDUs, cold plates, and immersion platforms for high-TDP AI racks (100–150+ kW).
Competitive Moat
Long relationships with hyperscalers, broad global service footprint, and portfolio spanning both power and cooling.
Schneider Electric
Integrated power + cooling + PPAs
Data center and networking growing at double-digit rates; data centers expected to account for >24% of group revenue. U.S. datacenter contracts totaling ~$2.3B.
Motivair acquisition made Schneider a major data center liquid cooling player; offers prefabricated power modules, chillers, and DCIM software; also arranges PPAs for hyperscalers.
Competitive Moat
Full-stack data center infrastructure (power, cooling, racks, DCIM), global scale, and deep channel/OEM partnerships.
Super Micro Computer
AI servers + "AI factory" integrator
Q2 FY26 revenue of $12.7B, +123% YoY, driven almost entirely by AI GPU systems and rack-scale "Data Center Building Block Solutions" (DCBBS) for Nvidia GB300 and AMD MI3xx.
>90% of revenue from AI GPU platforms; DCBBS includes cooling, power shelves, battery racks, and networking — SMCI acts as a prime contractor for AI factories, not just an OEM.
Competitive Moat
Ultra-fast design cycles with Nvidia/AMD, pre-integrated rack solutions, and deep operational leverage to hyperscaler cluster roll-outs.
Eaton
800 VDC and "grid-to-chip" power
Launching 800 VDC reference architectures aligned with Nvidia's "AI factory" standard, covering power conversion and protection for high-density AI racks.
Marketed "grid-to-chip" strategy provides power distribution, backup, and digital control across the AI data center power chain; strong in switchgear, busways, PDUs, and UPS.
Competitive Moat
Broad installed base in electrical infrastructure, partnership with Nvidia, and IP around AI-specific power quality management.
Constellation Energy
Merchant nuclear "AI grid" leader
Largest U.S. nuclear operator, largely merchant — can directly monetize AI data center power demand rather than being capped by regulators.
Signed a 20-year PPA with Microsoft to restart Three Mile Island Unit 1 (Crane Clean Energy Center) for AI data centers; plus a 20-year 1.1 GW PPA with Meta (Clinton plant).
Competitive Moat
Existing nuclear fleet, merchant status, and multi-decade hyperscaler PPAs.
Vistra
Nuclear + flexible generation for AI
Meta signed 20-year agreements for ~2.6 GW of nuclear capacity (Perry, Davis-Besse, Beaver Valley), including uprates, as part of Meta's plan to secure up to 6.6 GW of nuclear by 2035.
Competitive Moat
Merchant generator with nuclear + gas, direct hyperscaler PPAs, and regional positioning in high-growth data center markets (PJM, ERCOT).
NextEra Energy
Renewable + gas developer for data center hubs
Plans to build 15–30 GW of new power generation specifically for data center hubs by 2035. Multi-GW partnerships with Google and Meta, including 2.5 GW of projects for Meta (2026–2028).
Competitive Moat
Project development pipeline, scale in renewables, and explicit data center-hub strategy.
Step 03 — Synthesis & Ranking
Qualitative "Exposure Rating" reflects how central AI/hyperscaler capex is to each company's revenue and growth trajectory. High = core business and/or majority of incremental growth.
| Ticker | Company | Sector | Primary AI Catalyst | Exposure |
|---|---|---|---|---|
| NVDA | Nvidia | Semis – GPUs & Networking | >90% share of AI accelerators; 66% YoY DC growth; 90%+ networking attach rate | High |
| AMD | AMD | Semis – GPUs | MI300/MI325/MI355 at MSFT, Meta, Oracle; second-source AI GPU ramping to multi-billion revenue | High |
| AVGO | Broadcom | Semis – Custom ASICs & Switches | AI revenue ~$20B; $73B AI backlog; XPUs & switch ASICs for Google, Meta, OpenAI | High |
| ANET | Arista Networks | Networking – DC Ethernet | AI revenue $1.5B (2025) → $2.75B (2026); core 400/800G switch supplier | High |
| MRVL | Marvell | Semis – Custom ASICs & Optics | Custom XPUs for Amazon Trainium, Microsoft Maia; 2 nm HBM4 + optics platform | High |
| 000660.KS | SK Hynix | Semis – Memory (HBM) | ~57–62% HBM share; ~90% of Nvidia's HBM supply; biggest direct HBM beneficiary | High |
| VRT | Vertiv | DC Infrastructure – Power & Cooling | $8B revenue (+17%); $7.9B backlog; leading liquid cooling and power for high-density AI racks | High |
| SMCI | Super Micro | Servers & DC Integration | >90% revenue from AI GPU platforms; $12.7B Q2 FY26 (+123%); rack-scale AI factory integrator | High |
| MU | Micron | Semis – Memory (HBM/DRAM) | ~20–21% HBM share; U.S. HBM supplier; AI memory supercycle beneficiary | Medium–High |
| COHR | Coherent | Optics – Datacenter | Rapid ramp in 800G/1.6T optics; >4x book-to-bill; landmark CPO order | Medium–High |
| LITE | Lumentum | Optics – Components | "Virtually every AI network" uses Lumentum; components +68% YoY | Medium–High |
| SU FP | Schneider Electric | DC Infrastructure – Power & Cooling | DC >24% of revenue; ~$2.3B U.S. contracts; Motivair liquid cooling acquisition | Medium–High |
| CEG | Constellation Energy | Utilities – Merchant Nuclear | 20-year PPAs with Microsoft (TMI restart) and Meta (1.1 GW Clinton) | Medium–High |
| VST | Vistra | Utilities – Nuclear & Gas | 20-year PPAs with Meta for ~2.6 GW nuclear; part of 6.6 GW nuclear strategy | Medium–High |
| TSM | TSMC | Foundry & Packaging | Sole leading node + CoWoS provider for most AI GPUs; capacity sold out into 2026 | Medium |
| 005930.KS | Samsung | Semis – Memory & Foundry | HBM3E recovering to 22–35%; expected major HBM4 supplier for Nvidia Rubin | Medium |
| ETN | Eaton | Power – Grid-to-Chip | 800 VDC reference design with Nvidia; grid-to-chip AI power architecture | Medium |
| NEE | NextEra Energy | Utilities – Renewables & Gas | 15–30 GW new generation for DC hubs by 2035; multi-GW Google and Meta partnerships | Medium |
Step 04 — Key Risk Factors
The AI capex supercycle is real, but it is not without risk. Three structural threats deserve close attention before constructing any portfolio around this theme.
01
Hyperscaler capex is rising ~36–70% YoY into 2026. If AI workloads prove less monetizable than expected, or if model efficiency dramatically reduces compute needs — e.g., better algorithms, more efficient inference — hyperscalers could cut or flatten capex beyond 2027. Suppliers with long lead times (HBM, CoWoS, optics, liquid cooling) would face order cancellations and inventory corrections reminiscent of past semi cycles.
02
Custom ASICs (TPU, Trainium, Maia, MTIA) and in-house network fabrics are on a steep ramp. By mid-/late-decade, a larger share of AI workloads may shift from merchant GPUs to first-party silicon, shrinking TAM for Nvidia/AMD even if total flops grow. CPO and on-package optics could also reallocate value from stand-alone optics vendors to whoever controls the XPU package. Foundry concentration at TSMC means any geopolitical or export-control shock could bottleneck the entire chain.
03
AI data center power demand may increase U.S. data-center load 30x by 2035, to >120 GW, with global DC electricity use potentially more than doubling by 2030. Grid, transmission, and permitting constraints are already causing delays. FERC and local regulators are imposing new tariff classes and planning mandates for data centers. Pushback from communities (land use, water, noise, reliability, national-security concerns) could lengthen project timelines or cap expansion in key regions like Northern Virginia, indirectly impacting all suppliers. Additional idiosyncratic risks: export controls on high-end AI chips to China, antitrust/competition policy targeting Nvidia and Broadcom, and FX/geo risk for Korea/Taiwan-centric names.
Step 05 — Portfolio Synthesis
For portfolio construction, the practical next step is to map each name's revenue mix and incremental growth to hyperscaler AI capex, then size positions according to where you want to sit on the AI capex convexity vs. diversification spectrum.
Highest Beta — AI Capex Convexity
Closest to AI GPU, HBM, networking, and cooling bottlenecks explicitly called out in hyperscaler guidance.
Diversified Infrastructure Layer
Strong AI tailwinds buffered by other end-markets; lower volatility, broader exposure.
Long-Duration Power Thesis
Structural beneficiaries of multi-decade PPAs supplying firm or renewable power to data center hubs.
"Think of 2026 as a ~$600–650B AI infrastructure year for the four hyperscalers, with 75%+ of spend directly flowing into hardware and power ecosystems that public suppliers monetize."
Report synthesis