The silicon industry is the backbone infrastructure of the modern economy. From ASML's monopoly in extreme ultraviolet lithography to NVIDIA's compute supremacy in data centres and TSMC's manufacturing at nodes below 3nm, this sector combines irreplicable economies of scale with multi-billion barriers to entry.
Massive AI data-centre investment, chip sovereignty (CHIPS Act) and industrial electrification.
Very high R&D costs (defensive capex), exclusive proprietary patents and critical switching costs.
16 assets with verified 5D Vaultflake radar and intrinsic valuation models.
NVIDIA designs GPUs and systems that have become the default accelerators for training and inference in large-scale AI, as well as remaining a supplier to gaming and professional visualisation. The public report has to hold two ideas at once: the current cycle is extraordinary in revenue and margin, and cycles in semiconductors have historically mean-reverted when customers digest capacity. CUDA software, networking (including Mellanox heritage) and a full rack-scale story make NVIDIA more than a chip vendor. Customers buy a platform. That platform still sits inside a supply chain of TSMC, CoWoS packaging, HBM memory and a handful of hyperscale buyers. Concentration of customers and of manufacturing is a first-order risk, not a footnote. Valuation in a boom often discounts many years of data-centre spend. The radar is useful precisely because it will look 'expensive' when the story is hottest and 'cheaper' when the cycle breaks — neither reading is a timing signal by itself.
"The near-term moat is a combination of performance leadership in AI GPUs and switching costs in CUDA: models, libraries and developer muscle memory are written to NVIDIA's stack. Competitors can ship silicon; they cannot instantly clone the software ecosystem. Packaging and systems integration add another layer. The moat is not a legal monopoly and it is not guaranteed if a new architecture, custom ASICs at hyperscalers, or export controls cut demand. Treat 'unassailable' as marketing; treat CUDA as a real, but contestable, advantage."
ASML builds photolithography systems, including the EUV tools without which leading-edge logic and memory fabs currently cannot print the smallest nodes. It is a Dutch equipment company whose customers are a short list of chipmakers — TSMC, Samsung, Intel and a few others — which is both pricing power and customer-concentration risk. The installed base generates high-margin service and upgrade revenue. Lead times are long. Geopolitics sits on every shipment: export licences to China, allied industrial policy, and the fact that ASML's own suppliers (optics, light sources) are themselves specialised. This is not a consumer brand. It is a bottleneck in the semiconductor capital-equipment cycle. When foundries over-build, ASML's orders pause with a lag; when they under-build, ASML cannot instantly double output. The radar should be read as a cycle plus a structural monopoly-like position, not as a SaaS subscription.
"The moat is a de-facto monopoly in high-NA and EUV lithography, protected by decades of process knowledge, patents, and an ecosystem of exclusive suppliers. The switching cost for a foundry is not 'another vendor's tool' — there is no equivalent EUV competitor at the frontier. That is as close to a structural barrier as listed technology gets. It can still be damaged by export bans that shrink the addressable market, by a prolonged capex winter, or by a future patterning technology that bypasses EUV. Wide is not eternal."
TSMC manufactures chips for other companies. It does not design the phone or the GPU. It prints the silicon that Apple, NVIDIA, AMD and a short list of fabless customers order. The TSM ticker is the US listing of that Taiwanese company. A holder carries Taiwan risk and the listing wrapper, not only the semiconductor cycle. The advantage sits in the leading-edge node and in the packaging that stacks those chips for artificial intelligence. ASML sells the lithography without which that node cannot be printed. TSMC is not ASML. It is the factory. When customers digest capacity, tool orders pause with a lag, and the factory still has to fill lines that are already built. Geopolitics is first order. A blockade, an export ban or a new fab in another country changes the value even if the process is still the best. This report does not state a share of world output and does not set a target price.
"The moat is the yield of manufacturing at the hardest node, built over decades of process knowledge, single-source suppliers and customers who already design against that fab. Switching foundry at the frontier is not a one-quarter vendor change. It is not a consumer brand. It narrows if a large customer designs for another fab, if packaging stops being the bottleneck, or if politics cuts the market the company is allowed to sell into."
AMD designs processors and accelerators. It does not own a fab. TSMC prints most of the silicon. That couples the share to the foundry. A good design that cannot get a wafer does not ship, and a wafer that is late arrives after NVIDIA's. In servers it competes with Intel for the CPU. In artificial-intelligence accelerators it competes with NVIDIA, which also carries the software the models are already written for. Winning a design at one cloud customer is not winning the market. This report does not state a share. The cycle is data centres and PCs. The two can move in opposite directions in the same year. Cash goes back as buybacks and a dividend that is small next to design reinvestment. Read AMD next to TSM and NVIDIA, not as if it ran the furnaces.
"The advantage is the design and the relationship with the server buyer, plus being one of the few credible alternatives to Intel in CPUs and to NVIDIA in accelerators. It is not NVIDIA's software ecosystem and it is not TSMC's factory. The moat narrows if a hyperscaler builds its own chip and stops needing AMD's, or if the leading-edge wafer does not arrive."
Broadcom is a fabless semiconductor and infrastructure-software company. The semiconductor side sells custom accelerators (often called XPUs) to a handful of hyperscalers and the Ethernet switching silicon that ties those clusters together. The software side is the VMware estate plus older infrastructure franchises. A public report that treats Broadcom as 'another AI chip' misses the mix: custom silicon is a design-win business, networking is a merchant franchise, and software is a renewal business with a different margin and a different customer. Broadcom does not own a leading-edge fab. Manufacturing sits at foundries, so capacity, packaging and export rules are shared constraints with the rest of the AI supply chain. The economic question is whether design wins and software renewals keep returns high after the debt taken on to buy VMware, not whether the firm can print a GPU that looks like NVIDIA's. Shareholder returns have included a dividend and buybacks, but leverage from the software deal means cash has a creditor in front of the equity. This page describes that structure. It is not a price target.
"The semiconductor moat is switching cost: a custom accelerator or a switch ASIC that is designed into a hyperscale rack is painful to rip out mid-generation. The software moat is the installed VMware base, where migration cost, not a patent, keeps the renewal. Neither is a foundry monopoly and neither is CUDA. A few customers can represent a large share of semiconductor revenue, which is concentration, not a wider moat. Treat 'AI leader' as a mix description and check design-win disclosures in the filing."
Texas Instruments sells analog and embedded chips for cars and industry. Analog Devices sells in the same trade. It is not a counted catalog and it is not a capital-allocation benchmark by decree. This report does not state the wafer size. The design already inside a car or a machine sticks until the next model. Changing it is a project. If industry builds less, it buys fewer chips. The company's own factory runs half empty, and that cost stays. Cash separates cars from industry, which do not move together. The dividend has to fit. A year of heavy autos is not normal earnings.
"The advantage is the design already inside the car or the machine. The next model can change it. The moat narrows if the factory runs half empty, or if the customer designs in another chip."
Intel sells processors for computers and servers. AMD sells the processor that competes, and Arm is another architecture. It also wants to manufacture chips for others. TSMC already does that trade. It is not the giant by decree and the shift is not already done. Factory spending is not earnings. A subsidy is not the cash from the sale. This report states neither the amount nor how the chip is made. If the server customer stays with another processor, or if the foundry does not fill, the year is soft. Cash is the processor sale, minus that spending. A year of many computers is not normal earnings. The dividend, if any, has to fit.
"The advantage is the computer already using that processor and the factory already standing. The customer can order the server from someone else. The moat narrows if the new factory does not fill, or if the spending does not fit inside the sale."
Arm designs an architecture and processor blocks, and collects a licence and a royalty on the chips other companies build with that design. It does not own a fab and it does not sell the phone. The historical volume is the handset. The data centre and the PC are newer layers, where Nvidia and the hyperscalers use designs based on that architecture. The royalty depends on how many chips ship and on the price that was agreed, not on a fixed toll on the handset. A weak phone year lowers the royalty even if the architecture is still in the pocket. This report does not state a share of the smartphone installed base. RISC-V is the long-run alternative that does not pay that royalty. Today it is more of a threat at the margin than a replacement for the installed base. The share is intellectual property with a unit cycle underneath, not a foundry.
"The moat is the instruction set and the software already written against it. Changing architecture means rebuilding an ecosystem, not swapping a silicon vendor. It is not a monopoly on every processor: Intel and others are still in the server on their own. The moat narrows if an open standard matures enough that the phone or the server no longer needs Arm's licence."
Lam Research builds etch and deposition equipment. Etch is the step that sculpts the wafer. The more layers a stacked memory has, the more times it has to be etched. That is why memory pulls Lam in a way it does not pull a company that only sells lithography. The customer is again the short list: memory fabs and logic fabs. Service on tools already installed is the part that does not switch off when a new order slips. Lam does not replace Applied or ASML. It sits at another step on the same line. The memory cycle is sharper than logic. A year of depressed DRAM prices can cut capex even though etch is still needed over the long run. Export controls do the same job as in the rest of the equipment industry: they take customers out of the addressable market.
"The moat is etch chemistry already qualified in the fab and the base of tools that must be maintained. Replacing it in the middle of a stacked-memory process means rewriting the recipe. It is not a monopoly on every tool on the wafer. Applied and Tokyo Electron occupy neighbouring steps. The moat narrows if memory stops stacking layers, or if a fab qualifies someone else on the same etch."
Applied Materials sells the machines other companies use to make chips, and the service on the machines already installed. The catalog is wide: deposition, implant, polish, and packaging tools. It is not extreme-ultraviolet lithography. That is ASML. Applied sits in many of the other steps on the wafer. The customer is a short list of fabs: TSMC, Samsung, Intel, and the memory makers. When those fabs pause capital spending, the tool order pauses with a lag. Service on the installed base holds up better than new equipment. A weak order year does not erase the business, and a euphoric year is not normal demand. Export bans toward China shrink the market a tool can be shipped into. That is a political variable, as it is for ASML. The share is an equipment cycle plus a wide catalog, not a software subscription.
"The advantage is the catalog and the installed base. A fab does not change process chemistry because someone else has a brochure: the recipe is already qualified on that tool. Applied does not have an exclusive on frontier lithography. Tokyo Electron and Lam compete in parts of the same step. The moat is wide across the set of steps, and narrower in any one step on its own."
KLA sells the systems that look at the wafer and say whether the process has gone wrong. Inspection and metrology: find the defect before a fab fills a whole lot with failures. It does not print the chip. It does not design it. It sits beside the Applied, Lam and ASML tools, at the control step. The customer is the same short list of fabs. As the node gets harder, more measurement is needed, not less. When fabs pause new tools, process control does not switch off entirely: the installed base keeps measuring what is already being made. A new-tool order still follows the capex cycle. This report does not state a share of the inspection market or a margin. The equity is process equipment, with a cycle, plus a base of algorithms that fab has already calibrated.
"The advantage is the tool already qualified and the defect library that fab has built with it. Switching inspector in the middle of a node means recalibrating the line. It is not lithography and it is not etch. Others work in metrology. The moat is wide in reference optical inspection and narrower in any single technique. It narrows if a new fab is born already qualified on someone else."
Synopsys sells the software used to design a chip before it is sent to the fab. The tools cover design, verification and intellectual-property blocks an engineer drops in instead of drawing them from scratch. The model is licence and subscription. There is no furnace. Cadence is the other name this page should be read with. The two of them cover the flow. Neither is the only program in the world, and a large team does not switch flows because the other cut price for a quarter. The chip is manufactured by TSMC or another foundry. Synopsys does not print the wafer. Growth follows the number of designs and how hard it is to verify a chip with billions of transistors. A weak phone year does not switch off software already sitting in an accelerator team's flow. This report does not state a recurring-revenue percentage.
"The moat is the design flow already learned. Engineers, scripts and intellectual-property blocks are written against those tools. Leaving means rebuilding the method, not uninstalling a program. Cadence does the same job. The moat narrows if a large customer takes the flow in-house, or if a new design no longer needs the block Synopsys was paid for."
Cadence is the other large supplier of chip-design software. It covers digital circuits, analog circuits and simulation of the system the chip will live in. It is paid by licence. It does not manufacture the wafer and it does not sell the phone. Read it beside Synopsys, not instead of Synopsys. A design flow often mixes tools from both. The useful comparison is which one weighs more in one customer's flow, not a one-point P/E gap between the two shares. The chip designed here ends up at a foundry. Cadence is not TSMC. Verification and simulation get harder as the design gets larger and as packaging gets harder. That supports spending even when the handset cycle is weak. This report does not state a gross margin or a recurring-revenue percentage.
"The advantage is the same shape as Synopsys, with a different accent: the engineer already has the design, the simulation and the habit written in Cadence. Switching tools in the middle of a chip costs months. There is no legal exclusive on silicon design. The moat narrows if the customer moves the flow to the other vendor or builds it inside."
Micron manufactures memory: DRAM, NAND and the high-bandwidth memory that sits next to the accelerator. It is one of the three large producers, with Samsung and SK Hynix. It is not a logic foundry. It does not design the GPU. It sells the part that GPU needs if it is not to run out of data. The memory price is a brutal cycle. At the top, the margin looks like an extraordinary business. At the bottom, the same fab loses money. Capital spending does not switch off at the bottom, because the next density has to be built first. Lam sells the etch for those layers. Micron is the customer, not the tool maker. High-bandwidth memory ties part of revenue to the accelerator cycle. It does not turn Micron into NVIDIA. This report does not say Micron is the only producer of that memory and it does not state a share.
"The advantage is knowing how to manufacture density at a cost that still fits inside the cycle price, inside an oligopoly of three. The scale of the fabs is the moat, and also the trap: they are not switched off. It is not a designer's software ecosystem and it is not ASML's lithography. The moat narrows if a fourth player manufactures the same density, or if the accelerator stops asking for so much bandwidth."
Marvell sells connectivity chips for data centers and, apart, chips designed for one customer. They are not the same book. Broadcom sells parts of the same trade. It is not the leader of that connection by decree and it is not a replacement already won for another chip. This report does not describe how the chip is made or how the link works. A design win is not a shipment. The customer can give the next one to someone else. A year of heavy data-center building is not normal earnings. Connectivity does not move with the custom chip. Cash is the chip shipped. If something is paid out, it has to fit. It is not a captive customer.
"The advantage is the customer already using that chip in that design. They can award the next one elsewhere. The moat narrows if the design never ships, or if building slows and the other book does not make it up."
Infineon designs and manufactures power and automotive semiconductors: the parts that switch current in a car, in a factory and in the power supply of a data centre. It is not the GPU foundry. The power transistor is a different trade, with different fabs and different customers. The large cycle is automotive and industrial. A weak car year shows up even though a power chip is still needed in every vehicle that does get built. Silicon carbide and gallium nitride are materials of that trade, not an exclusive. This report does not state a world share. The listing is German, in euros. For a Spanish investor the currency matters less than it does for Nestlé, and the car cycle matters more. A dividend exists and should be read against the cycle's cash, not as a coupon. Compare it with logic names only so the product is not confused.
"The advantage is a power process already qualified in the car and in the industrial inverter, plus a supply relationship the manufacturer does not break in the middle of a platform. It is not a monopoly on the transistor. Others make power devices. The moat narrows if the customer qualifies someone else on the next platform, or if car volumes stay down for years."