The Chokepoints

Sole manufacturer of EUV lithography systems. Intel's own 10-K calls ASML the 'sole supplier' of EUV tools deployed in every leading-edge node. There is no second source — and ASML's own optics come from a single supplier.

ASML makes lithography systems, along with software, metrology, inspection, and installed-base services. Its EUV and DUV tools pattern the most advanced logic and memory chips, so AI accelerators, HBM, and AI-era CPUs depend on its systems. The company describes itself as a holistic lithography solution provider and has one reportable segment.

The only foundry capable of fabricating leading-edge AI accelerators at volume. Multiple chip designers file TSMC as their sole wafer source, and the chairman says switching takes five to seven years.

TSMC is a dedicated semiconductor foundry: it manufactures wafers for customers from their proprietary chip designs rather than selling chips under its own brand. Its process portfolio spans CMOS logic, mixed-signal, radio frequency, embedded memory, and BiCMOS, plus 3DFabric advanced packaging and testing services. In the AI buildout, it fabricates the leading-edge logic that GPUs, AI ASICs, XPUs, data-center CPUs, and networking chips rely on, and it packages many of those chips with CoWoS, SoIC, and CoPoS.

The AI training GPU standard. Multiple customers file NVIDIA as a single-source supplier, CUDA's software ecosystem locks in the installed base, and management says supply is constrained through the entire VeraRubin lifecycle.

NVIDIA designs the processors, networking, and software used to train and run AI. Its GPUs and rack-scale systems, sold as GB300 and NVL72, are the compute engine for frontier models, while its Spectrum-X Ethernet and InfiniBand connect hundreds of thousands of chips into single clusters. The CUDA software platform is the layer developers build against, and Data Center revenue is now 91.5% of the company.

Seven times larger than its nearest competitor in process control, and the gap has widened 360 basis points since 2021. Even rival equipment makers — Applied Materials, ASML, Lam Research — are KLA customers.

KLA makes process-control and process-enabling equipment and services for semiconductor and electronics manufacturing. Its dominant segment, Semiconductor Process Control, sells inspection, metrology, and software that help fabs manage yield from R&D through high-volume manufacturing. In FY2025, wafer inspection was $6.20B, patterning $2.20B, and services $2.68B.

Designs the custom AI silicon for Google, Meta, and other top hyperscalers. Arista's own 10-K names Broadcom as a sole source for switching chips, and the only 100-terabit Ethernet switch in the industry is theirs.

Broadcom reports in two segments. Semiconductor Solutions designs and supplies custom AI accelerators (the company calls them XPUs or TPUs), Ethernet switches and routers, SerDes, optical components, storage connectivity, RF filters, wireless connectivity, broadband and industrial chips. Infrastructure Software is built around VMware Cloud Foundation, mainframe software, cybersecurity and enterprise application products. Six AI customers anchor the silicon business — Google, Anthropic, Meta, OpenAI and two unnamed customers — each committed to multi-year, multi-gigawatt deployments supported by Broadcom's custom chip designs and its networking fabric.

Half of the EDA duopoly. Thirty-five companies depend on its tools and IP to design chips, and management says hyperscalers' 'entire custom-silicon strategy is built on the availability of Synopsys IP.'

Synopsys describes itself as a leader in engineering solutions from silicon to systems. It sells EDA software that accelerates and automates chip design; hardware-assisted verification systems, including ZeBu, HAPS and Xebu, used to emulate and prototype designs before fabrication; Design IP — pre-configured, proven circuits such as PCIe, USB, LPDDR and die-to-die blocks that designers drop into a chip; and Ansys simulation and analysis software that predicts how a product performs in the real world. The company reports two segments: Design Automation, which includes EDA, Ansys and Other, and Design IP. The AI build-out reaches Synopsys through chip design: AI accelerators, networking chips, HBM memory systems and multi-die packages all require EDA tools, verification and interface IP before they can be built, and their thermal, structural and fluid behaviour feeds Ansys simulation from chip to grid.

The other half of the EDA duopoly, supplying tools to 47 companies including every major hyperscaler custom-silicon program. No third player covers the full design flow.

Cadence makes computational software, accelerated hardware, and silicon IP used across semiconductor and electronics design. Its product categories are Core EDA at 71% of Q1 2026 revenue, semiconductor IP at 14%, and System Design & Analysis at 15%. It sells place-and-route, sign-off, verification, and analog tools; Palladium and Protium hardware; interface and memory IP such as PCIe, UCIe, HBM, and LPDDR6; and multiphysics tools including Clarity, Celsius, and Integrity 3D-IC. This spans the full design flow, so when AI chip and system builders add transistors, chiplets, and advanced packaging, they consume more Cadence design work.

One of exactly three companies capable of making HBM memory. The CEO says Micron can fulfill only 50 to 66 percent of key customer demand, with multi-year lead times for new capacity.

Micron Technology designs and manufactures DRAM, NAND flash, and NOR memory chips, along with solid-state drives and multi-chip packages, under the Micron and Crucial brands. Its memory products are foundational to every tier of the AI stack, from high-bandwidth memory (HBM) for GPU accelerators to high-capacity DDR5 for cloud servers and LPDDR for AI-enabled edge devices.

The instruction-set architecture behind roughly 100 percent of DPU and SmartNIC silicon and about half of hyperscaler CPUs. Twenty-two million developers and 250 billion cumulative chip shipments create lock-in that is architectural, not contractual.

Arm architects, develops, and licenses high-performance, energy-efficient compute platforms. The 20-F describes CPUs that address performance, power, and use-case requirements, plus GPU and NPU accelerators and system IP such as interconnects. As of March 2026, Arm added Arm-designed silicon with the Arm AGI CPU, its first merchant data-center product. The architecture sits across smartphones, cloud server CPUs, DPUs and SmartNICs, automotive electronics, and edge devices.

One of three companies in the world that can build HA-class gas turbines for AI power, with roughly three-year lead times and slots sold out through 2030. A major competitor has stopped taking orders beyond 2030.

GE Vernova is a global electric power equipment and services company. Its products generate, transfer, orchestrate, convert, and store electricity, and its installed base generates approximately 25% of the world's electricity. The company operates three reportable segments: Power, Wind, and Electrification.

One of three companies that supply ultra-high-purity industrial gases to leading-edge fabs, locked in by 15-to-20-year take-or-pay on-site contracts. Once a plant is piped into a fab, switching is effectively impossible.

Linde plc is the largest industrial gas company worldwide. Its primary products are atmospheric gases—oxygen, nitrogen, argon, and rare gases—and process gases—hydrogen, helium, carbon dioxide, carbon monoxide, electronic gases, specialty gases, and acetylene. It also designs and builds gas-production equipment through its Engineering segment, which constructs turnkey process plants for air separation, hydrogen, synthesis, olefin, and natural gas applications. The AI buildout depends on Linde's ultra-high-purity gases for advanced semiconductor fabs, supplied from on-site plants under long-term contracts.

The only U.S. company with proven technology to produce HALEU nuclear fuel — confirmed by competitor filings and an NNSA sole-source procurement. No domestic alternative arrives before approximately 2029.

Centrus makes the fuel that keeps nuclear reactors running and is rebuilding U.S. commercial uranium enrichment. Its LEU segment supplies separative work units (SWU), natural uranium hexafluoride, uranium concentrates, uranium conversion, and enriched uranium product; its Technical Solutions segment produces HALEU, the fuel required by many next-generation reactor designs. Management describes Centrus as the only company with a proven American technology for commercial LEU and HALEU demand — a management claim, not independently verified.

Switchgear and electrical distribution with data center orders up 240 percent year-over-year and a 228 GW backlog — roughly twelve years at current build rates. Spending over a billion dollars on 24 factory expansions.

Eaton is an intelligent power management company with FY2025 revenue of $27.4 billion, serving customers in 180 countries. It supplies electrical and cooling infrastructure from the utility grid to the chip: switchgear, UPS, power distribution, power quality, and liquid cooling. Its data center / distributed IT end market is the main channel through which AI demand reaches Eaton.

The switching fabric that stitches thousands of GPUs into a single cluster. Number-one market share in high-speed switching, $8.9 billion in purchase commitments, and 52-week component lead times.

Arista Networks is a data-driven, client-to-cloud networking company anchored by its Extensible Operating System (EOS) and Network Data Lake (NetDL). It sells Ethernet switches, routing, optics, and software that connect AI accelerators and data centers; it does not sell AI compute. The company serves cloud and AI titans, AI and specialty providers, and enterprise customers across AI centers, data centers, campus centers, and WAN centers.

Half of the nearline HDD duopoly with Western Digital. Both companies refuse to add unit capacity, both are fully allocated for years, there is no third manufacturer, and flash storage remains far too expensive per terabyte to substitute.

Seagate makes hard disk drives, SSDs, storage subsystems, and the Lyve edge-to-cloud platform. The 10-K describes the company as a leading provider of data storage technology and infrastructure solutions. Its AI-infrastructure connection runs through high-capacity nearline HDDs, including HAMR-based Mozaic drives, which supply the low-cost mass-capacity storage tier beneath compute-intensive AI workloads.