AI infrastructure bottlenecks
Query: ai infrastructure bottleneck · data as of 2026-10-12
Screened 40 high-moat US-listed AI infrastructure bottleneck names across seven layers (AI semis/accelerators, semicap & advanced packaging, memory/HBM, networking & optical, power & electrical, cooling & thermal, data-center/cloud), ranked by business model score then market cap, with a themed heatmap and per-name moat commentary.
Screen of 40 high-moat AI infrastructure supply-chain names across seven bottleneck layers, ranked by business model score then market cap. Data as of October 12, 2026.
AI Supply-Chain Bottlenecks — Ranked by Moat Depth
| Bottleneck Layer | Symbol | Company | Moat Score | Growth Score | Market Cap | Rev Growth YoY | Gross Margin | 1M Return | 6M Return | 1Y Return | % vs 150D SMA | Momentum Potential | P/E (TTM) | Forward P/E | Comments |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Semicap & Advanced Packaging | ASML | ASML Holding N.V. - New York Re | 9.5 | 8.5 | 7.437e+11 | 15.58% | 52.83% | 3.06% | 17.07% | 79.67% | 6.96% | 95.4 | 59.55 | 29.4 | Sole supplier of EUV lithography and the only path to leading-edge logic and HBM-adjacent DRAM scaling. High-NA EUV adoption is the next choke point; export-control policy toward China is the main swing risk. |
| AI Semis & Accelerators | NVDA | NVIDIA Corporation | 9 | 9 | 5.465e+12 | 65.47% | 71.07% | 3.68% | 19.82% | 17.81% | 9.05% | 99.9 | 34.66 | 17.63 | Owns the CUDA software ecosystem that makes accelerator switching costs enormous. Main debate is whether hyperscaler in-house silicon and cheaper inference economics compress its pricing power over the next cycle. |
| Data Center & Cloud | GOOGL | Alphabet Inc. | 9 | 8 | 4.314e+12 | 15.09% | 59.65% | 4.20% | 9.91% | 46.42% | 1.70% | 75.1 | 17.7 | 23.88 | Vertically integrated AI player: owns its TPU accelerator line, frontier models, and the largest distribution surface for AI services. In-house silicon insulates it from merchant GPU pricing. |
| Data Center & Cloud | MSFT | Microsoft Corporation | 9 | 7 | 4.018e+12 | 17.79% | 67.94% | 9.18% | 41.36% | 4.44% | 25.05% | 96.4 | 30.13 | 22.96 | Azure plus an in-house Maia roadmap lets it capture AI demand on both the model and infrastructure layers; the capex burden is the counterweight investors are watching. |
| Data Center & Cloud | AMZN | Amazon.com, Inc. | 9 | 9 | 2.799e+12 | 12.38% | 50.29% | 1.07% | 8.19% | 13.96% | 4.63% | 97.1 | 20.88 | 24.95 | AWS is the profit engine funding an aggressive custom-silicon (Trainium/Inferentia) strategy that lowers its own AI unit costs and pressures merchant chip vendors. |
| AI Semis & Accelerators | TSM | Taiwan Semiconductor Manufactur | 9 | 8 | 2.305e+12 | 25.03% | 56.12% | 2.84% | 20.84% | 49.74% | 8.90% | 93.8 | 33.28 | 20.4 | The true single point of failure in the AI supply chain — near-monopoly on leading-edge foundry and advanced packaging (CoWoS), which gates every accelerator and HBM stack. Geographic concentration in Taiwan is the structural risk. |
| AI Semis & Accelerators | AVGO | Broadcom Inc. | 9 | 8 | 1.693e+12 | 23.87% | 67.77% | -1.83% | -6.27% | 3.58% | -6.33% | 97 | 59.12 | 18.16 | Custom XPU/ASIC co-design plus networking silicon makes it the primary beneficiary of hyperscalers diversifying away from merchant GPUs; the software franchise is a recurring cash cow. |
| Semicap & Advanced Packaging | AMAT | Applied Materials, Inc. | 9 | 8 | 3.948e+11 | 4.39% | 48.67% | 8.99% | 26.02% | 127.09% | 5.44% | 56.8 | 43 | 27.13 | Broadest deposition/etch toolset and a key enabler of advanced packaging and HBM capacity expansion. Demand is tied directly to leading-edge wafer starts and memory capex. |
| Networking & Optical | ANET | Arista Networks, Inc. | 9 | 8 | 2.66e+11 | 28.60% | 64.06% | 5.66% | 38.72% | 33.27% | 23.68% | 99.9 | 66.52 | 40.87 | Dominant in high-speed Ethernet switching for AI back-end fabrics, where scale-out cluster bandwidth is a hard constraint. Risk is hyperscaler in-house switch silicon. |
| Networking & Optical | CSCO | Cisco Systems, Inc. | 8 | 6 | 4.547e+11 | 5.30% | 64.94% | 3.26% | 41.13% | 67.92% | 9.65% | 93.3 | 34.64 | 20.81 | Incumbent networking franchise with a large installed base and growing AI/optical portfolio; slower growth but cash-generative and defensive within the layer. |
| Data Center & Cloud | ORCL | Oracle Corporation | 8 | 9 | 4.165e+11 | 17.35% | 65.82% | -8.24% | -11.08% | -53.12% | -12.88% | 35.7 | 23.57 | 12.59 | OCI is winning large AI training contracts, but the buildout requires heavy debt and capex, leaving the equity acutely sensitive to execution on backlog conversion. |
| Semicap & Advanced Packaging | LRCX | Lam Research Corporation | 8 | 6 | 3.874e+11 | 23.68% | 48.71% | 3.91% | 16.01% | 120.56% | 3.36% | 95.7 | 53.65 | 26.79 | Etch and deposition share leader with outsized leverage to NAND/HBM layer-count increases; memory capex cycles drive its earnings volatility. |
| AI Semis & Accelerators | ARM | Arm Holdings plc | 8 | 8 | 2.726e+11 | 22.79% | 97.54% | -3.61% | 61.97% | 49.56% | -0.14% | 89.4 | 257.82 | 83.46 | Licenses the CPU instruction set nearly every mobile and increasingly data-center design depends on — a pure royalty toll on compute proliferation, though valuation leaves little room for error. |
| Semicap & Advanced Packaging | KLAC | KLA Corporation | 8 | 9 | 2.481e+11 | 11.71% | 61.30% | 5.25% | 7.75% | 81.44% | -0.88% | 39 | 51.95 | 28.79 | Process control monopoly-ish position: as leading-edge geometries shrink, inspection and metrology intensity per wafer rises, making it a structural share gainer. |
| Networking & Optical | APH | Amphenol Corporation | 8 | 9 | 2.125e+11 | 51.71% | 36.88% | 2.86% | 19.03% | 37.49% | 14.09% | 99.6 | 21.49 | 13.4 | Interconnect and high-speed cabling content per rack is rising faster than rack count, giving it a quiet but widening role in AI cluster buildouts. |
| Power & Electrical | ETN | Eaton Corporation, PLC | 8 | 9 | 1.657e+11 | 10.33% | 37.59% | 0.31% | 7.11% | 16.31% | 5.50% | 86.1 | 43.41 | 26.61 | Electrical backbone supplier for data centers — switchgear, busway, UPS — with a large backlog tied to multi-year capacity commitments. Power is now the binding constraint on AI buildout. |
| Data Center & Cloud | EQIX | Equinix, Inc. | 8 | 8 | 1.019e+11 | 5.36% | 51.09% | -0.49% | -1.34% | 30.48% | -0.58% | 89 | 66.53 | 55.25 | Largest global interconnection platform; latency-sensitive AI workloads need cross-connect density that is hard to replicate. Power availability and cost of capital are the key constraints. |
| Semicap & Advanced Packaging | SNPS | Synopsys, Inc. | 8 | 8 | 9.63e+10 | 15.12% | 76.98% | 26.46% | 20.28% | 3.74% | 14.40% | 82.4 | 114.99 | 29.11 | Design tools that every custom AI ASIC must pass through, plus IP blocks for chiplets and high-speed interfaces. A royalty-like toll on silicon design complexity. |
| Semicap & Advanced Packaging | CDNS | Cadence Design Systems, Inc. | 8 | 7 | 9.595e+10 | 14.12% | 85.87% | 20.41% | 20.90% | -0.07% | 4.21% | 85.8 | 69.27 | 36.53 | Peers with Synopsys at the EDA duopoly; simulation and verification intensity scales with chip complexity and multi-die packaging adoption. |
| Other AI Infra | MPWR | Monolithic Power Systems, Inc. | 8 | 9 | 6.483e+10 | 26.43% | 55.18% | 7.02% | -3.57% | 35.62% | -3.38% | 76.4 | 80.44 | 37.91 | Power-management silicon content per AI server rising as rack densities climb; shares the power-delivery bottleneck with a strong analog franchise. |
| Networking & Optical | TEL | TE Connectivity plc | 8 | 9 | 6.291e+10 | 8.94% | 35.22% | 2.51% | -6.55% | -1.27% | 4.17% | 81.7 | 21.28 | 16.84 | Connector and sensor content grows with every generation of higher-speed rack architecture; diversified end markets temper the cyclicality of pure AI capex names. |
| Semicap & Advanced Packaging | TER | Teradyne, Inc. | 8 | 7 | 6.001e+10 | 13.13% | 58.22% | 1.09% | 3.79% | 164.89% | 4.52% | 76.6 | 52.73 | 33.23 | Test equipment is a rising bottleneck as advanced packaging and HBM stacks multiply the number of test insertions per device. |
| AI Semis & Accelerators | CRDO | Credo Technology Group Holding | 8 | 9 | 3.775e+10 | 205.68% | 68.04% | 23.25% | 49.48% | 32.87% | 2.41% | 98.6 | 80.02 | 22.02 | Active electrical cables and SerDes IP are critical for scale-up connectivity inside AI racks; high growth but concentrated customer base. |
| Cooling & Thermal | VRT | Vertiv Holdings, LLC | 7.5 | 9 | 9.215e+10 | 27.69% | 36.32% | -6.86% | -20.17% | 41.88% | -16.96% | 71.3 | 54.15 | 26.31 | Liquid cooling is moving from optional to mandatory as rack power density exceeds air-cooling limits; Vertiv's thermal and power distribution share make it a direct bottleneck play, though order timing is lumpy. |
| AI Semis & Accelerators | AMD | Advanced Micro Devices, Inc. | 7 | 8 | 9.701e+11 | 34.34% | 49.52% | 15.13% | 140.75% | 155.16% | 33.89% | 99.6 | 151.59 | 38.42 | The credible merchant alternative to Nvidia in accelerators, with an open software stack (ROCm) as the differentiator. Execution on rack-scale systems is the swing factor. |
| AI Semis & Accelerators | MRVL | Marvell Technology, Inc. | 7 | 9 | 2.389e+11 | 42.09% | 51.02% | 12.63% | 102.58% | 193.87% | 31.02% | 96.1 | 91.05 | 42.54 | Custom AI silicon plus optical DSPs and high-speed interconnect — levered to the same hyperscaler ASIC wave as Broadcom, with a smaller but faster-growing base. |
| Memory / HBM | WDC | Western Digital Corporation | 7 | 8 | 1.448e+11 | 50.70% | 38.78% | -13.47% | 10.57% | 223.86% | -16.00% | 92.3 | 14.36 | 12.19 | Nearline HDD capacity is a genuine bottleneck for AI data lakes, with a duopoly structure and long lead times supporting pricing. HBM exposure comes via its memory spin-off and flash. |
| Networking & Optical | GLW | Corning Incorporated | 7 | 9 | 1.269e+11 | 19.14% | 35.97% | -11.49% | -15.79% | 70.07% | -11.07% | 78.3 | 67.87 | 34.09 | Optical fiber and specialty glass plus a growing role in advanced packaging substrates; a picks-and-shovels play where capacity additions take years. |
| Data Center & Cloud | DLR | Digital Realty Trust, Inc. | 7 | 8 | 6.804e+10 | 10.04% | 56.99% | -3.81% | -4.82% | 5.77% | -2.72% | 59.7 | 50.32 | 62.31 | Wholesale data-center landlord with long-duration leases to hyperscalers; the moat is powered land and interconnection, while development yield and rates drive the equity. |
| Networking & Optical | CIEN | Ciena Corporation | 7 | 8 | 5.743e+10 | 18.79% | 42.04% | 15.87% | -15.61% | 147.55% | -7.57% | 52.2 | 135 | 42.02 | Optical transport and coherent routing for data-center interconnect; a durable installed base but exposed to the timing of carrier and hyperscaler capex waves. |
| AI Semis & Accelerators | ALAB | Astera Labs, Inc. | 7 | 9 | 5.623e+10 | 115.13% | 75.69% | 11.29% | 94.32% | 43.77% | 16.30% | 98.2 | 159.66 | 50.69 | Connectivity silicon for scale-up fabrics (PCIe/CXL retimers, fabric switches) — a relatively new but fast-widening bottleneck inside AI racks. |
| Power & Electrical | EME | EMCOR Group, Inc. | 7 | 9 | 3.437e+10 | 16.62% | 19.33% | -0.19% | -3.97% | 14.67% | -2.39% | 64.6 | 24.23 | 21.2 | Electrical and mechanical contractor that builds the physical data centers; backlog provides visibility but margins depend on labor availability and project mix. |
| Power & Electrical | NVT | nVent Electric plc | 7 | 9 | 2.635e+10 | 29.51% | 37.74% | 0.25% | 22.59% | 67.68% | 6.63% | 88.1 | 45.09 | 25.23 | Liquid-cooling enclosures and electrical protection are becoming standard equipment in high-density AI halls, giving it a direct thermal/power bottleneck role. |
| Power & Electrical | HUBB | Hubbell Inc | 7 | 6 | 2.505e+10 | 3.84% | 35.32% | 2.96% | -11.64% | 14.55% | -2.58% | 66.6 | 28.07 | 20.73 | Grid and utility electrical components with pricing power in a transformer-constrained market; growth is more tied to grid replacement than AI directly. |
| Semicap & Advanced Packaging | ONTO | Onto Innovation Inc. | 7 | 8 | 1.343e+10 | 1.82% | 49.72% | -2.67% | 8.36% | 100.40% | -1.12% | 72.5 | 102.14 | 23.49 | Inspection and metrology for advanced packaging and HBM — a niche with rising content per wafer as chiplet and 3D stacking adoption grows. |
| Semicap & Advanced Packaging | AEIS | Advanced Energy Industries, Inc | 7 | 8 | 1.124e+10 | 21.38% | 37.66% | -2.26% | -25.86% | 56.23% | -11.96% | 68.6 | 51.9 | 18.48 | Precision power supplies and RF power for semiconductor process tools; levered to equipment capex rather than directly to AI demand. |
| Semicap & Advanced Packaging | CAMT | Camtek Ltd. | 7 | 7 | 8.986e+09 | 15.57% | 50.46% | -0.23% | -15.61% | 29.39% | -8.52% | 79.4 | 196.84 | 30.21 | Inspection systems specialized in advanced packaging, where defect control becomes critical for multi-die AI chips. Small-cap with high sensitivity to packaging capex. |
| Power & Electrical | SPXC | SPX Technologies, Inc. | 7 | 9 | 8.437e+09 | 14.17% | 40.51% | -12.18% | -23.87% | -9.42% | -18.87% | 29.5 | 29.55 | 17.22 | Niche HVAC and electrical infrastructure for mission-critical facilities; a smaller, more cyclical way to own the cooling/thermal bottleneck. |
| Power & Electrical | POWL | Powell Industries, Inc. | 7 | 8 | 6.754e+09 | 9.08% | 29.37% | 1.58% | -18.98% | 80.05% | -18.95% | 82.6 | 35.45 | 26.05 | Electrical switchgear and power systems with an order book tied to data-center and utility projects; backlog conversion and input costs drive results. |
| Power & Electrical | BDC | Belden Inc | 7 | 8 | 4.115e+09 | 10.33% | 37.98% | -13.84% | -19.09% | -7.52% | -9.06% | 47 | 17.09 | 10.48 | Cabling and networking infrastructure for industrial and data-center settings; a lower-multiple way to own rising bandwidth and connectivity demand. |
AI Supply-Chain Bottlenecks — Return Heatmap by Layer
[heatmap: AI Supply-Chain Bottlenecks — Return Heatmap by Layer]
Screen Thesis
Framing the AI bottleneck trade
AI capex is not a single demand wave — it cascades through a chain of scarce, hard-to-scale layers, and the deepest moats sit where capacity cannot be added quickly and substitution is hard.
The seven bottleneck layers in this screen
- AI semis & accelerators — the design layer, where software ecosystems (CUDA, ASIC co-design relationships) create switching costs that are far more durable than any single generation of silicon.
- Semicap & advanced packaging — the physical chokepoints. Leading-edge lithography, deposition/etch, process control, and CoWoS-style packaging capacity gate how many accelerators and HBM stacks the world can actually produce. Tool lead times are long and capability is concentrated in a handful of vendors.
- Memory / HBM — HBM bit supply and nearline storage are both tight; capacity additions take multiple years.
- Networking & optical — cluster bandwidth is as binding a constraint as compute. High-speed switching, optical interconnect, and cabling content per rack is rising faster than rack count.
- Power & electrical — power, not chips, is increasingly the gating factor for new data-center capacity. Transformers, switchgear, and grid equipment are on multi-year lead times.
- Cooling & thermal — rack densities have pushed liquid cooling from optional to mandatory for the newest AI halls.
- Data center & cloud — the operators who monetize the tokens and control powered land, interconnection, and in-house silicon roadmaps.
How the screen is built
Names are drawn from a curated AI supply-chain universe and must clear two tests: a high business model / moat score (7+) and positive revenue growth. Ranking is by moat depth first, then market cap — the largest, most entrenched franchises at the top of each layer.
What to watch
The key risk to the whole group is a shift in the AI cost curve: if compute gets cheaper via custom silicon and better inference efficiency faster than demand grows, pricing power migrates from second-derivative suppliers toward platform owners. That argues for favoring names with structural scarcity (tool monopoly, process control, powered land, switchgear lead times) over names whose advantage is purely a function of unit-volume growth.
Run a version of this yourself
Describe the idea in plain English and Relaxfolio returns a researchable shortlist with the charts, fundamentals, and risks attached.
Educational research tool — not investment advice. A screen is a starting point, not a recommendation.