GearFlow Lab

How Semiconductors are Made: Following the Multibillion-Dollar Journey From
Silicon Wafers to Big Tech CAPEX

How Semiconductors are Made:
Following the Multibillion-Dollar Journey
From Silicon Wafers to Big Tech CAPEX

A single logic chip never comes from one factory. It comes from a relay race spanning six continents, hundreds of specialized firms, and roughly a trillion dollars in annual revenue. One company draws the blueprint. Another rents that blueprint and stamps it into silicon. A third builds the machines that do the stamping. A fourth ships the gas that keeps those machines running.

A fifth wraps the finished die in protective packaging and tests it. Every one of these companies survives on a different business model, carries a different margin structure, and trades on a different set of catalysts. Before you can underwrite a position in $NVDA, $TSM, $MU, $INTC, $AMD, or $ASML, you need the org chart of the empire they all operate inside.

Think of the semiconductor value chain as a functioning empire. Empires need law and design — that’s IP and EDA. They need agriculture and industry — that’s fabrication. They need quartermasters supplying the front — that’s materials and gas.

They need checkpoints and customs — that’s packaging and test. And they need a treasury, because the whole apparatus runs on where tax revenue collects and where it flows next. Map the empire first. Then the individual stock stories click into place.

The Six Districts of the Semiconductor Value Chain

The chip manufacturing process runs in one direction — design, then fabrication, then packaging, then sale. Six distinct districts perform that work, and each one has its own scoreboard.

Every one of those wafers eventually lands with a front-end industry — data centers, consumer lectronics, automotive, industrial, computing, telecom, and defense and space. That’s the destination. Now walk the map district by district.

District One: The Architects — IP Holders and EDA

Nothing gets fabricated until someone draws it, and nobody draws a modern chip from scratch. ARM licenses the instruction-set architecture running in the overwhelming majority of smartphones and an accelerating share of data-center CPUs.

Synopsys and Cadence sell the Electronic Design Automation software that turns a chip concept into a manufacturable layout — without their tools, no fabless company can tape out a design a foundry will accept.

This district behaves like a toll booth: it collects a small royalty or license fee on nearly every chip built anywhere, regardless of who wins the fabless wars downstream. The business model carries software-like margins because the marginal cost of licensing another design is close to zero, and the switching costs are brutal — once an engineering team standardizes on a design flow, ripping it out mid-project is not realistic.

District Two: The Armory — Equipment Makers

Every fab, no matter who owns it, buys its tools from the same short list of vendors, and that concentration is the entire investment thesis. SEMI’s year-end 2025 forecast put global emiconductor manufacturing equipment sales at a record $133 billion for the year, a 13.7% jump, with $145 billion projected for 2026 and $156 billion for 2027 — three straight years of growth pushed almost entirely by AI-driven logic and memory capacity builds.

ASML sits at the top of the pyramid: it holds effectively 100% of the extreme ultraviolet (EUV) lithography market, the only technology capable of printing transistors below 7 nanometers. Full-year 2025 net sales came in at €32.7 billion, and the order backlog already stretches into 2027 — SK hynix alone committed roughly $8 billion for about 30 EUV systems, and Samsung committed close to $4 billion for 20 more.

That backlog explains why equipment stocks often hold up or rally even when chip prices are falling. A foundry doesn’t cancel a lithography order because memory prices dipped this quarter; it’s locking in capacity for a node transition three years out. Equipment demand is driven by the node roadmap, not the spot cycle, which decouples these stocks from short-term chip pricing in a way most investors underappreciate.

District Three: The Quartermasters — Materials Suppliers

Behind every equipment purchase sits a consumable supply chain that rarely makes headlines until it breaks. Shin-Etsu and SUMCO dominate silicon wafer production. Dow and Entegris supply the specialty chemicals, slurries, and ultra-high-purity gases that etching and deposition tools consume by the ton. Hoya makes the photomasks.

Thisdistrict generated roughly $68 billion in the most recent full-year measurement, and itsdefining trait is invisibility paired with irreplaceability: when Russia’s 2022 invasion of Ukraine disrupted neon gas exports — Ukraine had supplied roughly half the world’s semiconductor-grade neon, used in the lasers inside lithography tools — the entire industry felt the squeeze within weeks.

Materials companies rarely carry the multiple of a fabless darling, but they carry something more durable: nobody etches a wafer without them.

District Four:
Fabless Moats in the Semiconductor Value Chain

Here’s the question every retail investor eventually asks: how does a company with zero factories post a 50%+ operating margin? The fabless model answers it. $NVDA, $AMD, Broadcom, Qualcomm, and MediaTek design chips and outsource every step of physical manufacturing to a foundry.

That structure strips out the single largest cost center in the entire semiconductor value chain — the fab itself, which now costs upward of $20 billion to build at the leading edge — and replaces it with a variable cost: a purchase order to $TSM or Samsung Foundry.

The fabless district generated roughly $250 billion in the most recent annual measurement, with independent market trackers now sizing the broader design-house category closer to $270–280 billion as AI accelerator demand accelerates.

The trade-off is dependency. A fabless company’s entire roadmap lives or dies on securing enough leading-edge wafer capacity and advanced packaging slots from a foundry partner it doesn’t control. That’s precisely why $NVDA locks in multi-year capacity commitments with $TSM years in advance — the fabless margin structure only works if the supply chain behind it holds.

District Five: The Industrial Heartland — Foundry and the CAPEX Arms Race

If fabless companies are capital-light, foundries are the opposite: they are among the most capital-intensive businesses on Earth, and that’s the second question worth answering — why does a foundry have to spend tens of billions of dollars every single year just to stay in the game? Moore’s Law forces node transitions roughly every two years, and each new node — the industry is now racing toward high-volume 2-nanometer gate-all-around (GAA) production — requires new tool sets, new fab shells, and new process R&D that obsoletes the prior generation’s equipment. Fall one node behind, and a foundry’s largest customers walk to a competitor within a single product cycle. There’s no coasting in this district.

$TSM has effectively won this arms race for now. Full-year 2025 revenue reached $122.54 billion, up 36.1% year-over-year, and its global foundry market share climbed to 69.9%, up from 64.4% in 2024, according to TrendForce. Samsung Foundry held a distant second at roughly 7% share, trailed by China’s SMIC, Taiwan’s UMC, and GlobalFoundries.

$INTC’s foundry ambitions remain the wild card — heavy capital commitments, uneven execution, and a market watching closely for signs of a genuine turnaround. Foundry revenue itself came in at roughly $158 billion in the baseline measurement of the value chain, and the district’s defining metric for any investor is simple: track capital expenditure guidance quarter over quarter, because a foundry’s CAPEX today is its market share three years from now.

District Six: The Checkpoint — OSAT and the Packaging Renaissance

For two decades, back-end packaging was the least glamorous stop in the chip manufacturing process — commoditized, low-margin, easy to ignore. AI broke that pattern. High-bandwidth memory (HBM) and chiplet-based GPU designs require advanced packaging techniques like chip-on-wafer-on-substrate (CoWoS) that stack multiple dies with microscopic precision, and that packaging step has become as much of a bottleneck as the wafer fab itself. $TSM is expanding its CoWoS capacity to roughly 680,000 wafers in 2025, a 106% jump, and it still can’t fully satisfy demand — spillover orders are flowing to independent OSAT providers.

That’s why ASE, the largest pure-play OSAT with roughly 44.6% global share and $18.54 billion in 2024 revenue, has effectively become an AI infrastructure stock rather than a back-end afterthought. Amkor and China’s JCET round out the top three, together commanding well over 60% of OSAT revenue in a district sized at roughly $42 billion in the baseline measurement, though industry trackers now peg the broader OSAT and advanced-packaging category considerably higher as 2025 volumes ramped. The lesson for investors: when a bottleneck moves, the market capitalization follows it. Packaging didn’t get more glamorous — it got scarcer, and scarcity is what re-rates a sector.

The Standing Army — IDMs and the Memory Supercycle

Integrated Device Manufacturers design and fabricate under one roof, and the IDM district generated roughly $793 billion in the most recent full-year measurement across Samsung, SK hynix, $INTC, $MU, and Kioxia — with the memory-specific slice, DRAM and NAND combined, accounting for roughly $235 billion of that. Memory is where the current cycle is most dramatic.

SK hynix is on pace to post $49.6 billion in 2025 DRAM revenue against Samsung’s $46.4 billion, according to Visible Alpha consensus tracked by S&P Global — the first time in the company’s history it has out-earned Samsung in the category Samsung invented.

The reason is HBM: SK hynix controlled somewhere between 53% and 62% of the high-bandwidth memory market through 2025 depending on the quarter measured, versus Samsung in the mid-to-high 30s and $MU trailing near 20%. In December 2025, $MU announced it would exit consumer memory and storage entirely to redirect capacity toward AI data-center customers — a direct read on where the margin now sits.

Bank of America has labeled the setup a memory “supercycle” comparable to the 1990s boom, forecasting global DRAM revenue growth above 50% and NAND above 45% for 2026, with average selling prices climbing sharply on both.

That’s the mechanism behind Microsoft’s disclosure that roughly $25 billion of its swollen 2026 capital budget traces directly to rising memory and component costs — memory pricing power has become a line item large enough to move a hyperscaler’s own guidance.

The Imperial Treasury — End Markets and the Big Tech CAPEX Engine

Every district above exists to serve a front-line industry: data centers, consumer electronics, automotive, industrial systems, computing, telecom, and defense and space. Right now, one of those front lines dwarfs the rest. Microsoft, Amazon, Alphabet, and Meta are on track to spend a combined $725 billion in capital expenditures in 2026, up roughly 77% from about $410 billion in 2025 — the largest coordinated infrastructure buildout in corporate history, according to multiple sell-side trackers compiling company guidance.

Amazon alone is guiding to roughly $200 billion, Microsoft to about $190 billion, Alphabet to $175–185 billion, and Meta to $115–145 billion after repeated upward revisions. The World Semiconductor Trade Statistics organization now projects the entire global chip market to hit $975 billion in 2026, up more than 25% from an already-record $772 billion in 2025.

This is the treasury that funds everything upstream of it. Every dollar a hyperscaler commits to a new data center eventually splits into a GPU order, a wafer order, a lithography order, and a specialty-gas order — flowing backward through every district this article just mapped.

The Law of Inverted Money Flow

Here is the single mental model worth carrying out of this chapter. Physical product in the semiconductor value chain moves left to right: design becomes a wafer, a wafer becomes a packaged chip, a packaged chip becomes a product on a shelf or a server rack. Money moves in the opposite direction. It starts with the end customer — a hyperscaler opening its capital budget, an automaker funding an ADAS platform, a consumer buying a phone — and flows backward through every district in reverse order.

Trace one concrete chain. Google decides to expand Gemini training capacity and commits capital. That dollar reaches $NVDA, which designs the accelerator and books the revenue as a fabless company. $NVDA’s purchase order flows to $TSM, which fabricates the die and, in turn, writes checks to ASML for EUV capacity, to Applied Materials and Lam Research for deposition and etch tools, and to Entegris and Shin-Etsu for the gases and wafers those tools consume.

Global Top 30 Semiconductor Design Companies (Fabless, IP & EDA) –(As of February 17, 2026)

Country	Company	Market Cap ($M)	Key Products
USA	NVIDIA	4,442,283	GPUs/AI Accelerators, CPUs, Networking, Edge & Automotive SoCs
USA	Broadcom	1,541,719	Data Center ASICs, Custom Accelerators, Storage Interconnects
USA	AMD	338,017	x86 CPUs, GPUs, FPGAs, DPUs, Embedded Processors
USA	Intel	233,716	CPUs, GPUs/Accelerators, AI Accelerators, Networking/Connectivity
USA	Texas Instruments	205,252	Analog ICs, Power Management (PMICs), MCUs
USA	Analog Devices	164,717	High-Performance Analog/Mixed-Signal, RF, Sensors
USA	Qualcomm	150,127	Mobile SoCs, Cellular Modems, RF Front-End, Automotive/IoT AI
UK	ARM Holdings	133,047	Processor IP (Cortex), Compute Subsystem (CSS) Platforms
Taiwan	MediaTek	94,312	Smartphone SoCs (Dimensity), Connectivity, Smart TV/Automotive SoCs
USA	Synopsys	83,623	EDA Tools, Semiconductor Interface/Processor IP, System Analysis
USA	Cadence	81,513	EDA Tools, Design IP, System Analysis & Emulation
Germany	Infineon	68,727	Power Semiconductors (SiC/GaN), Automotive MCUs, Sensors
China	Cambricon	68,633	AI Accelerators (Train/Inference), Edge AI Chips
USA	Marvell	66,603	Data Center Networking, Custom ASICs, Optical Interconnects (PAM4 DSP)
Netherlands	NXP Semiconductors	61,517	Automotive MCUs/Processors, In-Vehicle Networking, Radar/Sensors
USA	Monolithic Power	56,122	Power Management ICs (PMICs), Power Modules
USA	Microchip	42,512	MCUs/MPUs, FPGAs, Timing Devices, Analog/Power Systems
China	GigaDevice	31,676	SPI NOR Flash, MCUs, Embedded Memory Solutions
Japan	Renesas	30,162	Automotive/Industrial MCUs, Automotive SoCs, Analog/PMICs
Italy	STMicroelectronics	29,871	MCUs (STM32), Sensors (MEMS), Power Semiconductors (SiC)
China	Montage Technology	29,458	Server Memory Interface Chips (DDR5 RCD/DB), Interconnects
USA	On Semiconductor	28,452	Power Semiconductors (SiC), Image Sensors, Electrification Systems
USA	Credo Technology	21,936	High-Speed SerDes/PHY, Retimers, Optical/Electrical DSPs
USA	Astera Labs	21,843	PCIe/CXL Smart Connectivity, CXL Memory Controllers
China	Will Semiconductor	21,045	CMOS Image Sensors (CIS), Analog ICs, PMICs
China	VeriSilicon	21,043	ASIC/SoC Design Services, Custom Silicon, IP Integration
USA	MACOM	18,315	RF/Microwave (Defense/Telecom), Optoelectronic ICs, TIAs
USA	Lattice	13,354	Low-Power FPGAs, Hardware Security, Edge AI Acceleration
China	Rockchip	11,272	ARM-based Application Processors, AIoT SoCs, Multimedia SoCs
USA	Rambus	10,975	Memory Interface IP (DDR/HBM), Security IP, SerDes IP Architecture

The finished die then moves to an OSAT like ASE, or to TSMC’s own captive CoWoS lines, for packaging — another leg of spend running in the same backward direction. The wafer traveled left to right. The capital that made it possible traveled right to left, starting at Google’s balance sheet and terminating at a gas supplier in Japan.

This is why sophisticated allocators watch big tech CAPEX guidance as a leading indicator for the entire value chain, sometimes eighteen to twenty-four months before the revenue shows up in a foundry’s or an equipment maker’s earnings print. When a hyperscaler raises its capital budget, that signal is already flowing backward through the empire before most of the downstream companies have said a word about it publicly.

Reading the Districts: What the Rest of This Series Covers

This map is the prerequisite, not the destination. Every district here compresses an entire investable universe: why fabless operating margins compress or expand with foundry pricing power, why a foundry’s CAPEX-to-depreciation ratio predicts its market share years in advance, why equipment order backlogs decouple stock performance from the memory spot cycle, and why a packaging bottleneck can re-rate an entire OSAT sector inside of eighteen months.

Each of those deserves its own chapter, and each business model rewards a different kind of diligence — royalty economics for IP holders, backlog visibility for equipment, capacity utilization for foundries, and pricing-cycle timing for memory IDMs. Keep this map open. Every company-specific breakdown that follows in this series will refer back to exactly where its subject sits inside the empire.

Coming Next: The Western Design Empire — A Deep Dive into the Top 30 Global Fabless Stocks
Click here.