Safeguard
Supply Chain Security•

How Are Computer Chips Made? From Sand to Software

How are computer chips made? Follow a chip from sand to fab, package and firmware, see who makes each part, what it costs and where software risk enters.

Safeguard Security Team
70 min read

A computer chip starts as quartz rock. It is refined into ultra-pure silicon, grown into a single crystal, sliced into wafers, patterned layer by layer in a fab over roughly three months, cut into dies, packaged, tested, mounted on a board and finally brought to life by firmware and drivers. The companies whose names are on the best-known chips, NVIDIA and AMD among them, do none of the physical manufacturing themselves. They design the chip and write the software, and a chain of specialist suppliers does the rest.

That last point is the reason this guide exists. A fabless chip company ships two things: a layout file that someone else turns into silicon, and a large body of software that makes the silicon useful. The second one is where most of the company's own supply chain risk lives.

This guide is written for two readers. If you simply want to know how chips are made, who makes them and what it costs, the first two thirds walk through every stage from sand to server rack in plain language, with diagrams. If you lead security, product security or engineering at a semiconductor company, a design services firm or anyone shipping drivers, firmware or SDKs, the final third covers where software supply chain risk enters that flow and a practical program for securing it.

Contents

How are computer chips made? The short answer

Computer chips are made in ten broad stages. Quartz is refined into 99.9999999 percent pure silicon, grown into a crystal and sliced into wafers. A fab then builds transistors and wiring on each wafer through hundreds of steps. The wafer is tested, cut into dies, packaged, tested again and mounted on a board, where firmware and drivers make it work.

Two industries meet in that sentence. One is materials and manufacturing: mines, chemical plants, crystal growers, the fabs themselves, and the lithography machines that print circuit patterns. The other is design and software: the engineers who decide what the chip does, the electronic design automation (EDA) tools they use, and the firmware, drivers and toolkits that ship with the finished part.

The diagram below lays out the whole journey. Every stage is explained in detail later in this guide, with the inputs it needs, where it usually happens and what it costs.

How computer chips are made: ten stages from quartz sand and polysilicon to wafer fabrication, packaging, test and firmware, with typical suppliers

From sand to software: the ten stages of making a chip and who typically does each one.

A few numbers help set the scale. A front-end fab runs a 300 mm wafer through 300 or more processing steps, which take 11 to 13 weeks on average at advanced nodes. The lithography machine that prints the finest layers costs about 170 million euros for a standard EUV system and about 350 million euros for the newest High-NA version. A leading-edge fab costs tens of billions of dollars. And after all of that, the chip is still useless until software tells it what to do.

What is a fabless company?

A fabless company designs and sells chips but owns no wafer factory, called a fab. It writes the hardware design and the software, then pays a foundry to manufacture the wafers and outside firms to package and test them. NVIDIA, AMD, Qualcomm, Broadcom and MediaTek all work this way, which lets them spend on design instead of factories.

The word describes a business model, not a size. Some of the largest semiconductor companies in the world are fabless; Wikipedia's ranking of fabless companies by revenue puts NVIDIA, Qualcomm, Broadcom, AMD and MediaTek at the top in 2024. The model became practical once dedicated foundries existed to manufacture other companies' designs, and the foundry that made it mainstream was TSMC, founded in Taiwan in 1987.

Fabless vs foundry vs IDM business models compared across chip design, wafer fabrication, packaging and test, and firmware and software

Who does what: fabless designers, foundries, integrated device manufacturers and OSATs, by stage of work.

Does NVIDIA make its own chips?

No. NVIDIA designs its chips and writes the software for them, but other companies manufacture them. Its fiscal 2026 annual report says it uses "a fabless and contracting manufacturing strategy" and names TSMC and Samsung as the foundries that produce its wafers, with memory from SK Hynix, Micron and Samsung.

The same section of the NVIDIA 2026 annual report says the company uses CoWoS technology for packaging and engages contract manufacturers such as Hon Hai Precision Industry (Foxconn), Wistron and Fabrinet for assembly, testing and packaging of final products. The report also notes that suppliers are "responsible for procurement of most raw materials," while NVIDIA may directly buy certain items such as memory and substrates. In short, NVIDIA's own factories produce designs and software, not silicon.

Does AMD make its own chips?

AMD used to, and now does not. It announced in October 2008 that it would spin off its manufacturing arm, which launched as GlobalFoundries in March 2009, and AMD sold its last stake in 2012. Today AMD relies on TSMC for all of its 7 nm and smaller CPU and GPU wafers and mainly on GlobalFoundries for larger nodes.

Those details come from two places. Wikipedia's GlobalFoundries article records the 2008 announcement, the March 2009 launch with investor ATIC holding the majority stake, and AMD's exit in March 2012. AMD's own fiscal 2024 Form 10-K says it uses "third-party wafer foundries to fabricate the silicon wafers for all of our products," relying on TSMC for 7 nm and smaller and primarily on GlobalFoundries above 7 nm, and that it also uses TSMC, UMC and Samsung for its programmable logic chips.

What is a foundry, and what is an IDM?

A foundry manufactures wafers from other companies' finished designs, using its own process technology and factories. An integrated device manufacturer, or IDM, designs chips and also builds them in its own fabs. TSMC is the archetypal foundry; Intel, Samsung, Texas Instruments, Infineon and STMicroelectronics are well-known IDMs.

TSMC became "the world's first dedicated semiconductor foundry" when Morris Chang founded it in Hsinchu, Taiwan, in 1987. The term "pure-play foundry" describes companies that only manufacture for others, such as TSMC, GlobalFoundries and UMC. Samsung sits in both camps: it designs and builds its own memory and processors, and it also runs a foundry business for outside customers.

The IDM definition on Wikipedia lists Intel, Samsung, Texas Instruments, Infineon and STMicroelectronics as examples, and notes that the label has blurred as the industry changed. Intel is the clearest case of that blurring. It has historically designed and built its processors in-house, and it now also sells manufacturing to outside customers through Intel Foundry, which offers wafer manufacturing, assembly and test, and 2D, 2.5D and 3D packaging.

The IDM model keeps more of the physical supply chain under one roof. A fabless company trades that control for focus. It spends its capital and its best engineers on design and software, and buys manufacturing as a service.

What is an OSAT?

An OSAT, short for outsourced semiconductor assembly and test, is a company that cuts finished wafers into individual dies, packages them and tests them for chip designers. ASE, based in Kaohsiung, Taiwan, is the largest. Amkor, headquartered in Tempe, Arizona, runs factories across Asia and in Portugal. Foundries increasingly offer advanced packaging too.

Wikipedia's ASE article cites Gartner in naming ASE the largest OSAT. Amkor's article lists its plants in China, Japan, Korea, Malaysia, the Philippines, Portugal, Taiwan and Vietnam, and a planned advanced packaging site in Arizona for chips made by TSMC. OSATs matter to this story for a reason that is easy to miss: the test programs and the key provisioning software that run in their facilities are part of the chip's chain of trust.

What does TSMC do, and what does ASML make?

TSMC manufactures chips designed by other companies. It takes a customer's finished layout and turns it into patterned silicon wafers using its own process technology, and it offers advanced packaging. ASML builds the lithography machines that print circuit patterns onto wafers inside fabs. ASML makes no chips and no wafers; TSMC designs no customer chips.

These two companies are the most frequently confused in conversations about chips, so it is worth stating their roles slowly.

What does TSMC do?

TSMC is a contract chip manufacturer. A customer such as NVIDIA or AMD sends TSMC a finished design, and TSMC fabricates it on one of its process nodes, then often packages it as well. TSMC supplies the process design kit customers design against, but it does not decide what the chip does.

The handoff runs in both directions. Before design starts, TSMC gives the customer a process design kit (PDK) describing what its process can build. After design ends, the customer gives TSMC a layout file. In between, the customer's engineers do all of the architecture, logic design and verification. TSMC's work is everything physical: photomasks, wafer fabrication, often wafer test, and increasingly advanced packaging such as CoWoS, the 2.5D technology that places processor dies and HBM memory stacks side by side on a silicon interposer.

What does ASML make?

ASML makes lithography systems: the machines that project a circuit pattern from a photomask onto a light-sensitive coating on a silicon wafer. It is the only company that makes EUV (extreme ultraviolet) lithography machines, and it also builds DUV (deep ultraviolet) machines, a market it shares with Nikon and Canon. It is headquartered in Veldhoven, the Netherlands.

Wikipedia's ASML article describes it as "the world's only manufacturer of extreme ultraviolet (EUV) lithography machines." On the DUV side, Nikon released its NSR-S636E ArF immersion scanner in December 2023, and Canon's lithography lineup centers on KrF scanners and i-line steppers, plus nanoimprint systems. ASML's customers are the fabs: TSMC, Samsung, Intel and others buy its machines and install them in their factories.

Why is ASML important?

ASML matters because lithography sets how small a chip's features can be, and only ASML builds EUV machines, which print the finest layers of every leading-edge chip. Any company making the most advanced logic or memory chips needs ASML's systems. That makes one Dutch company a single point of dependence for the entire leading edge.

That position took decades to build. ASML's own EUV product page says that during early EUV development only ASML, with its partners and suppliers, kept working toward a viable system, and EUV reached high-volume manufacturing in 2019. The machines are covered in detail in the EUV section below.

Who hands what to whom

Put the companies together and a chip's life looks like a relay race. The designer runs most legs, but hands the baton to EDA vendors, IP vendors, the foundry, equipment makers, an OSAT and finally an OEM or ODM that builds the system. The swimlane below shows who does what, phase by phase.

Chip supply chain swimlane diagram showing the chip designer, EDA vendors, IP vendors, foundry, tool vendors, OSAT and OEM across design, tape-out, wafer fabrication, test, packaging and software

The chip flow as a swimlane: the designer's own work (shaded) starts the chain and ends it, with firmware and drivers shipped for years.

Notice where the designer's lane ends. It does not stop at tape-out. It picks up again at test and provisioning and then runs for years after the chip ships, in the form of firmware, drivers and SDK releases. Hold onto that picture; it is the heart of the security discussion later.

Step one: how silicon wafers are made from sand

Silicon wafers are made by reducing quartz with carbon in an arc furnace to make 98 to 99 percent pure silicon, converting it to trichlorosilane and distilling it, depositing ultra-pure polysilicon in a Siemens reactor, melting that and pulling a single crystal ingot by the Czochralski method, then slicing the ingot into wafers and polishing them flat.

This part of the chain gets little attention, yet every chip depends on it. It is also almost entirely a chemistry and materials business, run by a small number of specialist companies.

From quartz to metallurgical silicon

The raw material is silicon dioxide, mined as quartz or quartzite. In an electric arc furnace at about 2,000 °C it is reduced with carbon, which strips away the oxygen and leaves "raw" or metallurgical-grade silicon at roughly 98 to 99 percent purity, according to MicroChemicals' explanation of the process.

That sounds pure. It is nowhere near pure enough. A transistor's behavior is set by adding tiny, carefully controlled amounts of dopant atoms, so stray impurities at the percent level would swamp the design entirely.

Quartz shows up again later in the chain, as the crucible that holds molten silicon during crystal growth. High-purity quartz is a specialty product in its own right. When Hurricane Helene hit western North Carolina in late September 2024, the Associated Press reported that Sibelco and The Quartz Corp had halted their Spruce Pine operations, which produce high-purity quartz used for making semiconductors, solar panels and fiber-optic cables. It was a reminder that even the first step of the chain has concentrated suppliers.

Polysilicon and the Siemens process

To purify it, metallurgical silicon is reacted with hydrogen chloride at about 300 °C to form trichlorosilane, a liquid at room temperature. Many impurities, iron among them, do not form volatile chlorides and stay behind. Repeated distillation then brings the trichlorosilane to purities of up to 99.9999999 percent, the level the industry calls "nine nines" or 9N.

The purified trichlorosilane is then turned back into solid silicon in the Siemens process. Hydrogen and trichlorosilane flow into a bell-shaped reactor containing thin silicon rods heated electrically to roughly 1,100 to 1,150 °C. Silicon deposits on the rods, which grow to about 120 mm in diameter over a cycle of about 46 hours. The result is polysilicon, made of many small crystals, with impurity levels below one part per billion for electronic grade.

Polysilicon producers named in that article include Wacker Chemie in Germany, Hemlock Semiconductor in the United States and Tokuyama in Japan.

Growing a single crystal: the Czochralski method

Polysilicon cannot be used directly, because the boundaries between its tiny crystals disrupt the flow of electrons. Chips need a single, near-perfect crystal, and most of it is grown by the Czochralski method.

Polysilicon chunks are melted at about 1,425 °C in a crucible, usually made of high-purity quartz, under an inert gas such as argon. A precisely oriented seed crystal is dipped into the melt and slowly pulled upward while rotating. By controlling the temperature, pull rate and rotation, the grower draws out a cylindrical single-crystal ingot. Precise amounts of boron or phosphorus can be added to set the silicon's electrical type. According to Wikipedia's Czochralski article, ingots can reach about two meters long and several hundred kilograms, and roughly 90 percent of modern semiconductor devices use material made this way.

Slicing and polishing the wafer

The ingot is ground to an exact diameter and cut into thin discs with a diamond wire saw. Each disc is machined flat, chemically etched to remove the damage left by sawing, and then polished to a mirror finish. A small notch on the edge marks the crystal orientation. The modern standard is the 300 mm wafer, introduced around 1999, which is 775 micrometers thick.

Who makes wafers, and what it costs

Silicon wafers are a concentrated, largely Japanese business. Shin-Etsu Chemical holds the largest global share of semiconductor silicon, and SUMCO, also based in Tokyo, was the second largest wafer producer with about 30 percent of the market as of 2013, per Wikipedia's SUMCO article and Shin-Etsu's.

On cost, the honest answer is that a bare polished wafer is cheap compared with what happens next. The value gets added in the fab, where a finished leading-edge wafer is reported to sell for $30,000 or more at the 2 nm node. Reliable public prices for raw 300 mm prime wafers are scarce, so we do not quote one here. The cost picture for the materials stage is driven by energy (the arc furnace and the Siemens process are both power-hungry) and by purity, which is expensive to reach and to prove.

How is a chip designed? From specification to tape-out

A chip is designed by writing its logic as code in a hardware description language, verifying that code with simulation, formal proofs and hardware emulators, converting it into billions of placed and wired transistors with EDA software, and checking the layout against the foundry's rules. The finished layout file is then sent to the foundry, a step called tape-out.

This is the part of the chain a fabless company owns completely. It is also, by headcount, mostly a software engineering activity.

Specification and architecture

A chip starts as a product requirement: performance targets, power budget, interfaces, cost and the process node it will be built on. Architects model candidate designs in software, often in C++ or Python performance models, to decide how many cores the chip needs, how much cache, which memory interface and which accelerators. This is also when the team decides which blocks to build and which to license from intellectual property (IP) vendors, such as processor cores from Arm, which licenses its designs as IP and does not manufacture chips, or interface controllers for PCIe, DDR memory and USB.

RTL design

Engineers then describe the chip's logic in a hardware description language, usually SystemVerilog or VHDL, at what is called the register transfer level (RTL). RTL describes how data moves between storage elements on each clock tick. It is code in every practical sense. It lives in version control, it depends on licensed IP blocks, and it is built and tested in continuous integration pipelines, just like application software.

Verification, the biggest job in chip design

A software bug can be patched next week. A logic bug in silicon can mean a new set of masks and months of delay, so verification takes a large share of the engineering effort.

Teams use several techniques at once. Simulation runs the RTL against testbenches, commonly written with the Universal Verification Methodology (UVM), a standardized SystemVerilog class library that Accellera approved in 2011 and that is now IEEE 1800.2. Formal verification uses mathematical techniques to prove that certain properties always hold. Hardware emulation compiles the design onto large purpose-built systems that run it thousands of times faster than simulation, such as Synopsys ZeBu, Cadence Palladium Z3 (announced in April 2024) and Siemens Veloce. FPGA prototypes let firmware and driver teams start writing and testing software long before real silicon exists.

That last point has a consequence for security. The firmware and drivers that will ship with the chip are being written and built during the design phase, often by the same organization, on the same infrastructure, months before tape-out.

Synthesis, place and route, and signoff

Logic synthesis converts RTL into a gate-level netlist built from the foundry's library of standard cells, the pre-designed building blocks such as logic gates and flip-flops. Physical design then plans the layout of the die, places millions or billions of cells, builds the clock distribution network and routes metal wires between everything. Static timing analysis checks that every signal arrives on time across the full range of process, voltage and temperature conditions the chip will see.

Physical verification comes last. Design rule checks (DRC) confirm that the layout obeys the foundry's manufacturing rules, and layout versus schematic (LVS) checks confirm that the layout matches the intended circuit. Tools such as Siemens Calibre are used by foundries, IDMs and fabless companies for this work.

The best-known EDA vendors are Synopsys and Cadence, both US companies, and Siemens EDA. Siemens EDA is the former Mentor Graphics, based in Wilsonville, Oregon, which Siemens acquired in 2017 and renamed in January 2021. Wikipedia's EDA overview lists Cadence and Synopsys as the two largest pure-play EDA companies by market value. Open source tools play a growing role too: Verilator is an open source simulator, and the OpenROAD project has released an automated RTL-to-GDS flow that has taped out real designs.

Around all of these tools sits a large body of in-house code. Tcl, Python and Perl scripts glue the tools together, transform design data and launch thousands of jobs on compute farms. Those scripts touch the company's most valuable asset, the design database, and they usually run with broad access to shared file systems and license servers.

The process design kit (PDK)

The PDK is the main technical handoff between designer and foundry. A process design kit is a bundle the foundry supplies for each process node. It typically contains transistor and device models for circuit simulation, technology files for the layout tools, DRC and LVS rule decks, parasitic extraction rules, and standard cell and I/O libraries. Much of it is text, configuration and script.

Designers install the PDK into their flows, wrap it in their own scripts and update it when the foundry releases a new version. Most commercial PDKs are confidential, but open examples exist. The SkyWater SKY130 PDK, released in 2020 as a collaboration between Google and SkyWater Technology Foundry under the Apache 2.0 license, lets anyone see what a PDK contains: design rule documentation, EDA tool support files, primitive cell libraries and models, and several standard digital cell libraries.

Tape-out

When the design passes signoff, the team "tapes out," handing the final layout to the foundry. The term dates from the era when paper and later magnetic tape reels carried the files to the factory that made the photomasks. Today the file is usually in GDSII, long the de facto industry standard for layout data, or its successor OASIS, and it travels over an encrypted transfer. It is the single most valuable artifact a fabless company produces, and the most damaging one to lose.

How much does it cost to design a chip?

Designing a leading-edge chip costs from tens of millions to several hundred million dollars, depending on the node and how much IP is reused. IBS has estimated about $51 million at 28 nm and $650 million at 3 nm. SemiAnalysis argues those figures run high and says startups have built 7 nm chips for $50 to $75 million.

These are estimates, and the spread between them is the most useful thing to understand. Semiconductor Engineering reported in April 2020 that IBS put the design cost of a 3 nm chip at $650 million, against $436.3 million at 5 nm and $222.3 million at 7 nm. A 2023 Semiconductor Engineering analysis shows IBS's 2018 series at about $51 million for 28 nm and about $106 million for 16/14 nm, and argues that newer estimates should be discounted, extrapolating roughly $160 million at 7 nm and $280 million at 5 nm. SemiAnalysis wrote in July 2022 that chip design is expensive but far cheaper than the largest published figures suggest, citing startups that built leading-edge 7 nm chips for $50 million to $75 million including software, design and tape-out.

Chip design cost by process node chart: IBS estimates from about 51 million dollars at 28 nm to 650 million dollars at 3 nm, with a SemiAnalysis range of 50 to 75 million dollars for 7 nm startups

How much it costs to design a chip, by node. All figures are published estimates, and analysts disagree widely.

Why the gap? The high estimates count everything a large company spends on a new flagship: architecture, IP licensing, verification, physical design, software, masks and validation, often for a brand-new design. A startup reusing licensed IP for a focused accelerator spends far less. Notice that software appears in both definitions. Even the cheapest credible estimate counts it as part of the cost of the chip.

How are photomasks made?

A photomask, also called a reticle, is a plate that carries the pattern for one layer of a chip. Masks are written directly from the layout data with electron-beam or laser writers, after software reshapes the patterns so they print correctly. Each patterned layer needs at least one mask, so a modern chip needs a full set of them.

The foundry or a merchant mask shop converts the taped-out layout into masks. In modern scanners, a reticle holds one copy of the circuit layer, which is projected and shrunk four or five times onto the wafer. Before writing, optical proximity correction (OPC) software modifies the pattern so that, after the distortions of printing at tiny dimensions, the shape on the wafer matches the designer's intent. OPC is why masks show zig-zag edges and tiny "assist features" that never print themselves.

Masks for EUV lithography are different again. Because almost every material absorbs EUV light, EUV masks work by reflecting light off a multilayer mirror, rather than by blocking light with a chromium pattern on glass.

Merchant mask makers listed by Wikipedia include Dai Nippon Printing, Toppan Photomasks (now Tekscend), Photronics, Hoya, Taiwan Mask Corporation and Compugraphics, and large chipmakers such as TSMC, Samsung and Intel run their own mask shops.

On cost, SemiAnalysis estimated in 2022 that mask sets cost hundreds of thousands of dollars at 90 to 45 nm, move beyond $1 million at 28 nm and beyond $10 million at 7 nm, and begin to push into the $40 million range at 3 nm. Those are estimates. They also explain why a bug that forces a "respin," a new tape-out with new masks, is so painful.

What happens inside a chip fab? Every process step explained

Inside a fab, each wafer goes through a repeating loop: clean it, grow or deposit a thin film, coat it with photoresist, expose a pattern with a lithography scanner, develop the resist, etch away unprotected material, strip the resist, implant dopants, polish the surface flat and inspect it. The loop repeats for every layer, hundreds of steps in all.

Inside a chip fab: the wafer fabrication process loop of clean, oxidize or deposit, coat resist, expose, develop, etch, strip, implant and anneal, CMP polish, and inspect, repeated for every layer

The fab process loop for one layer. Transistors are built first (front end of line), then the copper wiring stack (back end of line).

Engineers split the flow into two halves. The front end of line (FEOL) forms the transistors directly in the silicon; the back end of line (BEOL) builds the metal wiring that connects them, with up to eleven or more metal layers on a modern chip. The steps below are not a fixed sequence. Each fab uses its own recipe, and the same step appears many times.

The cleanroom: air, water and power

A fab is built around a cleanroom, because a single particle can kill a transistor. Under the ISO 14644-1 standard, an ISO 5 room allows at most 3,520 particles of 0.5 micrometers or larger per cubic meter of air, while ordinary city air holds roughly 35 million. Semiconductor facilities often operate at ISO 7 or ISO 5. Tools and wafer carriers use sealed mini-environments: wafers travel between tools in sealed pods called FOUPs.

Water is the second requirement. Intel says a fab uses nearly 4 million gallons of water a day, with municipal water purified to "ultrapure" levels on site and fed directly to the tools. Semiconductor Engineering reported in August 2025 that advanced-node cleaning and polishing rely on ultrapure water with resistivity above 18.2 megohm centimeters and organic contamination below one part per billion, and that recycling is now a priority: UMC, for example, reports recycling 84.3 percent of its process water.

Power is the third. TechInsights estimates that a single EUV lithography tool can use up to 10.2 gigawatt hours of electricity a year. Even so, EUV tools account for only about 11 percent of a fab's total electricity, which says a lot about how much power the rest of the fab, from air handling to chemical processing, consumes.

Oxidation

One of the oldest steps in chip making grows a thin layer of silicon dioxide on the wafer by exposing it to oxygen or water vapor at high temperature. Thermal oxidation typically runs at 800 to 1,200 °C in furnaces that hold many wafers in quartz boats, and dry oxygen gives a denser, higher quality film than steam. Oxide layers insulate, protect and act as masks for later steps.

Photoresist coating

Next the wafer is coated with photoresist, a light-sensitive polymer that will record the pattern. It is spun onto the wafer as a thin, uniform film in a coater and developer system, often called a track; Tokyo Electron is a major maker of these. The resist chemistry is one of the most concentrated parts of the supply chain. Asia Times reported in 2021 that JSR, Tokyo Ohka Kogyo (TOK), Fujifilm, Shin-Etsu Chemical and Sumitomo Chemical, all Japanese, together hold roughly 90 percent of worldwide semiconductor photoresist sales.

Exposure

The coated wafer moves into a lithography scanner, which shines light through (or, for EUV, off) the mask and projects a shrunken image of the pattern onto the resist. The scanner steps across the wafer, exposing one field at a time, with overlay accuracy measured in nanometers so that each layer lines up with the ones below it. This step decides how small the features can be, which is why it gets its own section below.

Develop

After exposure, a developer solution washes away either the exposed or the unexposed resist, depending on the resist type. What remains is a stencil of resist on the wafer surface, matching the mask pattern.

Etch

Etching removes material wherever the resist stencil leaves it exposed. Modern fabs mostly use plasma etching, where reactive ions bombard the surface. Lam Research, for example, makes reactive ion etch, atomic layer etch and deep reactive ion etch systems. After etching, the leftover resist is stripped and the wafer is cleaned again, ready for the next layer.

Deposition: CVD, PVD and ALD

Deposition adds thin films of insulators and metals. Chemical vapor deposition (CVD) grows a film from reacting gases. Physical vapor deposition (PVD), often sputtering, knocks atoms off a solid target onto the wafer. Atomic layer deposition (ALD) is the most precise: it exposes the wafer to precursor gases one at a time in sequential, self-limiting reactions that add one atomic layer per cycle. ALD is widely used for ultra-thin "high-k" gate insulators such as hafnium oxide. Silicon dioxide leaks too much current once it is thinned to about a nanometer, and a high-k material gives the same effect with a thicker, less leaky layer. Intel reported using ALD for its high-k gate dielectric at 45 nm.

Applied Materials, Lam Research and Tokyo Electron all make deposition systems.

Ion implantation and annealing

To set the electrical properties of the silicon, ion implanters fire dopant atoms such as boron or phosphorus into precisely defined regions. Each ion knocks atoms out of place, so the wafer is then heated in an annealing step, which, as Wikipedia's ion implantation article notes, restores the crystal structure and is necessary to activate the dopants. Implantation and annealing build the source, drain and well regions of every transistor.

Chemical mechanical planarization (CMP)

After many deposition and etch steps, the wafer surface becomes bumpy. CMP polishes it flat using a combination of chemical and mechanical forces: the wafer is pressed against a rotating pad flooded with an abrasive, chemically active slurry. A flat surface matters because the next lithography step can only focus sharply on a flat plane.

Metallization and interconnect

Once the transistors exist, the BEOL builds the wiring that connects them. IBM introduced copper interconnects in 1997, and the industry has since moved from aluminum to copper wires made by the damascene process. Trenches are etched into an insulating layer, lined with a barrier that stops copper diffusing into the silicon, seeded with a thin metal layer and filled with copper by electroplating. CMP then removes the excess copper, leaving metal only in the trenches. Wikipedia's article on copper interconnects describes the dual-damascene variant, which fills a trench and the vertical via below it together. Layer after layer, this builds a three-dimensional city of wires above the transistors.

Inspection and metrology

Between steps, inspection tools hunt for defects and metrology tools measure film thickness, feature width and layer alignment. KLA, based in Milpitas, California, specializes in this process control equipment. The data feeds statistical process control systems that decide whether each wafer continues, gets reworked or is scrapped.

Every one of these machines is a computer as much as a tool. Scanners, etchers and deposition chambers run embedded control software, receive recipes from the fab's manufacturing execution system and stream sensor data into analytics. A modern fab runs on software as much as on machinery.

How long does it take to make a chip?

Fabricating a wafer at an advanced node takes about three months. Wikipedia's fabrication article cites 11 to 13 weeks as the industry average for 14, 10 and 7 nm processes, and up to 15 weeks, across 300 or more processing steps. Packaging and final test add more time, and design before that takes years.

What is EUV lithography?

EUV lithography prints the smallest features on modern chips using extreme ultraviolet light with a wavelength of 13.5 nanometers, more than 14 times shorter than the deep ultraviolet light used before it. Because almost everything absorbs EUV light, the machine uses mirrors instead of lenses and runs in a vacuum. ASML is the only company that builds EUV machines.

What is EUV lithography: simplified diagram of an ASML EUV scanner with tin droplets, CO2 laser, collector and illuminator mirrors, reflective mask, projection mirrors and wafer, compared with DUV and High-NA EUV

How an EUV scanner prints a pattern, simplified, with DUV and High-NA EUV compared.

How an EUV scanner works

The hard part of EUV is making the light. Inside the source, molten tin droplets about 25 microns across are fired from a generator at 70 meters per second. A first laser pulse flattens each droplet, and a second, more powerful pulse vaporizes it into a plasma that emits EUV light. The cycle repeats 50,000 times every second. Wikipedia's EUV article identifies the laser as a CO2 laser.

From there, the light is collected and steered by a series of mirrors. ASML's EUV page explains that EUV systems "use several multilayer mirrors instead of lenses" and that, since EUV light "is absorbed by everything, even air," the light path must be kept in high vacuum. The mirrors use molybdenum and silicon multilayers that reflect EUV through interference between layers. The light reflects off the mask, then through projection mirrors that shrink the pattern onto the resist-coated wafer.

Standard EUV systems have a numerical aperture (NA, a measure of how much light the optics gather, which governs resolution) of 0.33. EUV entered high-volume manufacturing in 2019.

DUV, immersion and multiple patterning

Not every layer needs EUV. Layers with larger features are printed with deep ultraviolet (DUV) light. DUV scanners use argon fluoride (ArF) lasers at 193 nm or krypton fluoride (KrF) lasers at 248 nm.

Immersion lithography squeezes more resolution out of 193 nm light by filling the gap between the final lens and the wafer with ultrapure water. A 1.35 NA immersion tool at 193 nm reaches a resolution limit of about 36 nm, and going smaller requires multiple patterning: exposing and etching the same layer more than once with offset patterns. Multiple patterning works, but every extra exposure adds cost, time and chances for misalignment. That trade-off is what made EUV worth its price.

DUV is where ASML has competitors. Nikon sells ArF immersion scanners such as the NSR-S636E, and Canon sells KrF scanners and i-line steppers used for mature nodes and packaging.

High-NA EUV

The next generation raises the numerical aperture to 0.55. ASML's EXE platform, which it calls High-NA, can print with a resolution of just 8 nm. Intel took delivery of the first unit at its Oregon site in late 2023, and in June 2024 Reuters described it as a 350 million euro tool. Standard EUV machines cost about 170 million euros each, according to Bloomberg reporting in February 2024. ASML does not publish list prices, so treat both as estimates.

What does 3nm mean? Planar, FinFET and gate-all-around transistors

A name like "3nm" is a marketing label for a generation of manufacturing technology, not a measurement of any single feature on the chip. Each new node packs transistors more densely and improves speed or power, but the number no longer matches gate length or any other dimension. Real measurements, such as gate pitch, are much larger.

Wikipedia's 3 nm process article puts it directly: the term "has no direct relation to any actual physical feature" of the transistors. For comparison, TSMC disclosed a 45 nm contacted gate pitch for its N3 process and a 23 nm minimum metal pitch for N3E. The names still matter commercially, because each one marks a step in density, performance and power.

What has changed physically, generation after generation, is the shape of the transistor itself.

Planar vs FinFET vs gate-all-around transistor cross sections, showing how the gate wraps further around the channel at each generation

Three generations of transistor: the gate wraps further around the channel each time, which gives it better control.

A transistor is a switch. The gate turns current on or off in the channel beneath it. For decades, transistors were planar: a flat channel in the silicon surface with the gate sitting on top. As transistors shrank, a gate on one side could no longer fully shut off the channel, and current leaked even when the switch was off.

The FinFET fixed that by raising the channel into a thin vertical fin and wrapping the gate around three sides. Intel announced its tri-gate FinFET at 22 nm in 2011, with chips shipping from 2012, and TSMC began 16 nm FinFET production in 2013. FinFETs carried the industry through roughly a decade of nodes, and TSMC's N3 still uses them.

Gate-all-around (GAA) transistors go one step further. The channel becomes a stack of thin horizontal sheets, and the gate surrounds each one completely. Samsung began initial production of its 3 nm gate-all-around process in June 2022. TSMC's N2 node, its first GAA process, entered volume production in the fourth quarter of 2025, and Intel's 18A process pairs its RibbonFET GAA transistors with PowerVia, which delivers power from the back of the wafer.

How much does a chip factory cost?

A leading-edge chip factory costs more than $20 billion to build and equip, and estimates for the newest nodes run higher. A fab for mature nodes costs around 10 billion euros or more. The biggest single items are the lithography machines: about 170 million euros for a standard EUV scanner and about 350 million euros for High-NA EUV.

Every figure here is an estimate drawn from company statements or major press reporting, because chipmakers and toolmakers rarely publish line-item costs.

How much does a chip factory cost: estimates for leading-edge and mature fabs, EUV and High-NA EUV scanners, 2 nm wafers, mask sets, chip design, and EUV power use

The big numbers in chip manufacturing. Each card names its source; all are estimates, not list prices.

Fabs. In 2017, TSMC founder Morris Chang told Bloomberg that a 3 nm fab would cost $20 billion or more. TrendForce, citing Taiwan's Commercial Times in October 2024, put 3 nm fab construction at no less than $15 billion to $20 billion. At the other end of the spectrum, the fab TSMC is building in Dresden with Bosch, Infineon and NXP, for 28/22 nm and 16/12 nm processes at 40,000 wafers a month, is expected to need total investment of more than 10 billion euros.

Tools. Lithography dominates the equipment bill at the leading edge. Bloomberg reported in February 2024 that a standard EUV system costs about 170 million euros and a High-NA system about 350 million euros, and Reuters used the same 350 million euro figure in June 2024. By revenue, the five largest equipment suppliers in 2024 were ASML, Applied Materials, Lam Research, Tokyo Electron and KLA, according to CINNO Research.

Wafers. TSMC does not publish prices, but TrendForce relayed Commercial Times reporting that 2 nm wafers may exceed $30,000 each, about double the price of 4/5 nm wafers.

Masks and design. As covered above, SemiAnalysis estimates mask sets from about $1 million at 28 nm to about $40 million at 3 nm, and chip design estimates range from $50 million to $75 million for a focused 7 nm startup chip up to IBS's $650 million for a 3 nm flagship.

Running costs. Power and water are large ongoing expenses. A single EUV tool can use up to 10.2 GWh of electricity a year, according to TechInsights, and Intel puts a fab's water use at nearly 4 million gallons a day.

For a fabless company, almost none of this appears on its own balance sheet. It pays through the wafer price. That is the economic logic of the fabless model, and it is also why the designer's own capital goes into engineers, EDA licenses, IP, emulators and software.

How are chips tested? Wafer probe, yield and binning

Chips are tested twice. First, while still on the wafer, automated test equipment touches each die through a probe card and runs electrical tests, and failing dies are marked. After packaging, the chips are tested again, often under heat and stress, and sorted into speed and power grades, a process called binning.

Wafer probe

At wafer probe, also called wafer sort, a prober positions each die under a probe card, a fixture with many fine needles, and a tester runs electrical patterns through it. Test programs are software, written by the designer or the test house, and the test patterns come from design-for-test structures the designer built into the chip, such as scan chains and built-in self-test logic.

Two companies dominate the testers. Counterpoint Research reported in February 2025 that Advantest and Teradyne together accounted for more than 90 percent of the test equipment market in 2024. Advantest is based in Tokyo and Teradyne in North Reading, Massachusetts.

Known good die

Wafer test matters even more for advanced packages. When several dies and memory stacks are bonded into one package, one bad die can ruin an expensive assembly, so each die is tested thoroughly first. The industry calls these "known good dies." Even that is not a guarantee: engineers interviewed by Semiconductor Engineering in June 2025 pointed out that a die that passed on its own can behave differently once heated and stressed inside an assembled package, and that the sheer number of tiny connections makes post-assembly testing hard.

Yield and binning

Yield is the share of dies that work. It drives cost directly: a wafer costs the same whether 90 percent or 30 percent of its dies are good. Wikipedia's binning article notes that semiconductor manufacturing is inherently variable, with yields sometimes as low as 30 percent.

Binning recovers value from that variation. Not every defect makes a chip useless. A part that cannot run at full speed can be sold as a slower model, and a part with one defective block can have that block disabled and be sold as a lower-tier product. Wikipedia's example is Apple's A12X and A12Z processors, which share the same die, with one GPU block disabled on the A12X. Yield and binning data are commercially sensitive, and they flow between foundry, OSAT and designer through software systems that deserve the same protection as the design itself.

How are chips packaged? Wire bond, flip chip, 2.5D, 3D, chiplets and HBM

A chip is packaged by cutting the wafer into dies, attaching each die to a substrate and connecting it with fine wires (wire bonding) or solder bumps (flip chip). Advanced packages place several dies side by side on a silicon interposer (2.5D) or stack them vertically (3D), which is how AI accelerators pair compute dies with HBM memory.

From wafer to die to package to board, with a 2.5D AI accelerator package cross section showing a compute die and HBM stacks on a silicon interposer and ABF substrate

From wafer to board. Below: a simplified 2.5D package with stacked HBM memory beside the compute die.

Dicing

The tested wafer is cut into individual dies. Dies that failed probe are set aside and never packaged.

Wire bonding

The traditional way to connect a die to its package is with fine wires welded between pads on the die and leads on the package. Wire bonding is "the most cost-effective and flexible interconnect technology" and is still used to assemble the vast majority of semiconductor packages. Gold wire is giving way to copper, mainly on cost.

Flip chip

For chips that need more connections and better electrical performance, the die is flipped. Solder bumps are placed on the die's pads while it is still on the wafer; the die is then turned face down onto the substrate and the solder is melted to form every connection at once. The method, also called controlled collapse chip connection (C4), was introduced commercially by IBM in the 1960s. An underfill adhesive strengthens the joints.

Substrates and ABF

Underneath, a package substrate fans the die's dense connections out to the much coarser pitch of a circuit board. High-end substrates are built up layer by layer with an insulating film, and one company supplies most of it. TrendForce reported in April 2025 that Ajinomoto, better known for food seasonings, holds more than 95 percent of the global market for its Ajinomoto Build-up Film (ABF) in GPU and CPU substrates, and plans to raise capacity by 50 percent by 2030.

2.5D packaging and CoWoS

AI accelerators push packaging much further. In a 2.5D package, a large compute die and several memory stacks sit side by side on a silicon interposer, a thin slab of silicon carrying thousands of fine wires between them. TSMC's version is CoWoS, short for chip on wafer on substrate, which has traditionally used a high-density silicon interposer for multi-die integration, typically with HBM memory stacks. NVIDIA's annual report confirms it uses CoWoS for its packaging.

HBM and 3D stacking

High bandwidth memory (HBM) is itself a 3D package. HBM stacks DRAM dies vertically and links them with through-silicon vias (TSVs) and microbumps, and the stack usually connects to the processor through a silicon interposer. SK hynix produced the first HBM chip in 2013, and AMD's Fiji GPU was the first device to ship with it, in 2015. Stacks are commonly 8 or 12 dies high; SK hynix began mass production of 12-layer HBM3E in September 2024. The largest HBM makers are SK hynix, Samsung and Micron, which is also the supplier list in NVIDIA's annual report.

Chiplets

A chiplet is a small die that contains a well-defined subset of a processor's functions. Several chiplets are packaged together to act as one large chip. Wikipedia's chiplet article notes that chiplets can be made with different processes and nodes, each suited to its function, and can be tested before assembly, which improves the yield of the final device. One chiplet design can also be reused across several products. AMD's Ryzen processors based on Zen 2 and later architectures are a common example, and the UCIe standard is one of several efforts to let chiplets from different companies talk to each other.

From a security point of view, chiplets and HBM multiply the number of components inside one package, each with its own firmware, test history and supplier. The bill of materials for a single AI accelerator now spans several companies' silicon.

Final test, provisioning and board assembly

After packaging, each chip is tested again, often including burn-in, a stress test at elevated temperature to weed out early failures. Chips are sorted by speed and power, and identity data, fuse settings and sometimes cryptographic keys are written into them. The tested parts then go to board makers, who mount them in servers, phones and car modules.

Burn-in is frequently run at elevated temperature and perhaps elevated voltage to catch components that would fail in the early, high-failure-rate part of their life, at the cost of some yield.

Provisioning is the point where the physical and software supply chains meet most directly. If a device identity key or a fused security configuration is written at final test, then the provisioning software and the systems that hold key material at the test site become part of the product's root of trust. A designer needs to know who runs that software, where it runs and how its builds are controlled.

Board and system integration is usually done by original design manufacturers (ODMs), contract manufacturers and OEMs. NVIDIA's annual report names Hon Hai Precision Industry, Wistron and Fabrinet among the contract manufacturers for assembly, testing and packaging of its final products. These partners combine the chip with power delivery, memory, networking and cooling, then load firmware and ship systems.

And then the chip's longest phase begins. Boot firmware, management controller firmware, GPU firmware, operating system drivers, user-space libraries and SDKs are installed and updated over the product's life. A chip that ships once may receive software updates for many years.

Who supplies each stage of the chip supply chain?

Each stage of the chip supply chain is dominated by a handful of specialists: polysilicon and wafers from Germany, the US and Japan; EDA software from the US; lithography from the Netherlands and Japan; process tools from the US and Japan; manufacturing in Taiwan, Korea and the US; packaging and test in Taiwan and across Asia.

Who supplies each stage of the chip supply chain: table of stages, key inputs, example suppliers and home regions, from silicon wafers and EDA to foundries, HBM, OSATs and test equipment

Stage, inputs and example suppliers by region. Examples, not a complete list.

StageKey inputsExample suppliersBased in
Silicon and wafersPolysilicon, 300 mm wafersWacker, Hemlock, Tokuyama; Shin-Etsu, SUMCOGermany, US, Japan
EDA softwareDesign, verification, signoff toolsSynopsys, Cadence, Siemens EDAUS
Processor and interface IPCPU cores, PCIe and DDR controllersArm; Synopsys, CadenceUK, US
PhotoresistsLight-sensitive chemicalsJSR, TOK, Shin-Etsu, Fujifilm, SumitomoJapan
LithographyEUV and DUV scannersASML (EUV and DUV); Nikon, Canon (DUV)Netherlands, Japan
Process toolsDeposition, etch, CMP, metrologyApplied Materials, Lam Research, KLA; Tokyo ElectronUS, Japan
FoundryWafer manufacturingTSMC, Samsung Foundry, Intel Foundry, GlobalFoundries, UMCTaiwan, Korea, US
Memory for AI packagesHBM stacksSK hynix, Samsung, MicronKorea, US
Substrate filmBuild-up filmAjinomoto (ABF)Japan
Assembly and testPackaging, final testASE, Amkor; foundry advanced packagingTaiwan, US, plants across Asia
Test equipmentTesters, probe cardsAdvantest, TeradyneJapan, US
Systems and softwareBoards, firmware, drivers, SDKsODMs, OEMs and the chip designerGlobal

A few notes on the table. Applied Materials is headquartered in Santa Clara, Lam Research in Fremont and KLA in Milpitas, all in California, while Tokyo Electron is based in Tokyo. GlobalFoundries moved its headquarters to Malta, New York, in 2021. The foundry row mixes pure-play foundries with Samsung and Intel, which are IDMs that also manufacture for outside customers.

Look at the last row again. Everything above it is a physical input that the foundry, the equipment makers and the materials suppliers manage. The last row is the one the chip designer builds itself.

Every stage runs on software

Count the software in the flow. EDA tools and the scripts that drive them. The PDK, largely a bundle of rule decks, models, libraries and scripts. Optical proximity correction and mask data preparation. The fab's manufacturing execution systems and the embedded software inside every scanner, etcher and deposition tool. Test programs on the testers. Provisioning software at final test. Firmware and drivers at the end.

Wafers, gases and chemicals matter enormously, but they are the foundry's and the equipment makers' operational problem. For the designer, the software in this flow is the part of the supply chain it controls directly, and the part it can secure.

What does a fabless chip company actually ship?

A fabless chip company ships a layout file to its foundry and a large software stack to its customers: boot ROM code fixed in the silicon, firmware for on-chip processors, kernel and user-mode drivers, compilers, libraries and SDKs, framework containers and reference models. Much of it runs with high privileges and depends on open source.

What a fabless chip company ships: the software stack from boot ROM and device firmware through kernel drivers, SDK compilers and libraries, to frameworks, containers and reference models

The software stack above the silicon. Each layer has its own dependencies, build pipeline and signing process.

Take the silicon away, since someone else makes it, and look at what leaves a fabless company's own systems.

Design data. RTL, netlists, the GDSII or OASIS tape-out file and test programs.

Boot ROM. The first code a chip runs is frozen into the silicon at tape-out. A flaw in it can be mitigated but never patched, which is why ROM code deserves the strictest review of anything a designer writes.

Device firmware. Modern chips contain several processors besides their main cores. AMD's Platform Security Processor is an ARM Cortex-A5 core with TrustZone that has been built into AMD processors since about 2013, handling secure boot, early hardware setup and other security functions. NVIDIA GPUs from the Turing generation onward run firmware on a GPU System Processor (GSP), and NVIDIA's open kernel modules "must be used with GSP firmware". Firmware like this is commonly written in C, packaged into signed images and often includes third-party libraries.

Kernel drivers. Drivers run inside the operating system kernel. NVIDIA released open source GPU kernel modules in May 2022 under a dual GPL and MIT license, and AMD's amdgpu driver is an in-tree module in the mainline Linux kernel. On Windows, starting with Windows 10 version 1607, Windows will not load new kernel-mode drivers that are not signed through Microsoft's Dev Portal, which requires an EV code signing certificate to register. On Linux, out-of-tree modules are often rebuilt on the customer's machine by DKMS, which automatically recompiles them when a new kernel is installed. That means the build toolchain on the customer's machine is part of the trust chain.

SDKs and toolchains. NVIDIA's CUDA stack was first made public in February 2007 and includes compilers (nvcc is LLVM-based), libraries and developer tools. AMD's ROCm is an "open-source software stack for AMD GPU computing". Intel ships oneAPI. Each of these is a large software distribution with compilers, math and communication libraries, profilers, debuggers and Python bindings, installed through apt, yum, pip, conda or container pulls.

Reference software. Board support packages, reference designs, sample code and integration guides that OEMs copy straight into their own products.

Containers and models. Prebuilt containers with frameworks such as PyTorch tuned for the vendor's hardware, plus reference models. Containers carry an entire operating system's worth of packages. Model files can carry code too: Python's own documentation warns that it is possible to construct malicious pickle data which will execute arbitrary code during unpickling.

Each of these is a software artifact with its own dependency tree, build pipeline and signing process. Many run with the highest privileges a machine has. That is why a fabless designer's supply chain risk is, in large part, a software supply chain risk.

Where does software supply chain risk enter the chip supply chain?

Software supply chain risk enters a chip company at eight main points: EDA flow scripts, third-party IP with its firmware, open source libraries inside drivers and SDKs, build systems, signing keys, test and provisioning software, distribution channels for SDKs and containers, and the long tail of field updates. Each one has a documented real-world failure pattern.

Where software supply chain risk enters the chip flow: eight entry points from EDA flows and third-party IP to build systems, signing keys, distribution and field updates, each with a risk and a control

Eight entry points for software supply chain risk in a chip company, with the incident pattern and the control for each.

EDA scripts and flows

EDA tools are commercial products, but the flows around them are mostly in-house code: Tcl, Python and Perl that glue tools together and launch jobs on compute farms with broad access to the design database. A tampered flow script could alter design data or quietly copy it out. These repositories deserve code review, CI checks and secrets scanning like any production code, because credentials for license servers and storage tend to end up inside them.

Third-party IP and its software

Licensed IP arrives as RTL, encrypted RTL, hard macros or netlists, often with firmware, reference drivers and verification IP attached. An interface controller might ship with a driver and a firmware blob that go straight into the product. When that happens, the designer inherits the IP vendor's secure development practices, good or bad, along with the block. A firmware blob from an IP vendor is your firmware now.

Open source inside drivers, firmware and SDKs

Toolchains and drivers depend on open source: LLVM, compression libraries, cryptography libraries, protobuf, JSON parsers, Python packages and many more. These projects are maintained by communities on their own schedules. The March 2024 backdoor in xz utils, tracked as CVE-2024-3094, showed how a widely used compression library could be compromised through its release process: malicious code was found in the upstream release tarballs starting with version 5.6.0, hidden so that the liblzma build extracted a prebuilt object from a disguised test file. Any vendor whose stack linked compression libraries had to answer one question fast: do we ship it, and where?

Build systems and CI

Every artifact a chip company ships is produced by a build system: Make, CMake, Bazel, Yocto, internal tooling and the CI runners that execute them. Compromising a build system is worth more to an attacker than compromising a single repository, because the output is trusted and signed by default.

Signing keys

Firmware, drivers and update packages are trusted because they are signed. The signing keys are the crown jewels. If a key leaks, an attacker can produce artifacts that devices and operating systems accept as genuine. Keys can also be mismanaged without any theft at all, which is exactly what the PKfail research described below found.

Test and provisioning software

As covered above, identity keys and fuse settings are often written at final test, in facilities run by OSATs or test houses. The provisioning software and the systems holding key material there sit inside the product's root of trust.

Distribution: SDKs, containers and models

Chip designers increasingly ship complete AI software environments: containers with frameworks, inference servers and reference models, published to public package indexes and container registries. Every package manager and registry in that path is a place where a lookalike or malicious package can slip in, and every model file is a possible carrier for code.

Updates in the field

A web company that ships a vulnerable library can redeploy in an hour. A chip designer's software ships into devices it does not control, through OEMs, cloud providers and system integrators, and it stays there for years. The blast radius is wide and the patch cycle is slow, so knowing exactly which release contains which component matters long after launch.

What incidents show the risk is real?

Several public incidents show the pattern: attackers compromised the SolarWinds Orion build so signed updates carried a backdoor, pushed signed malicious ASUS Live Update installers, planted a backdoor in xz utils release tarballs, published a malicious PyTorch dependency, and researchers found untrusted test keys shipped in Secure Boot firmware across hundreds of PC models.

Timeline of software supply chain incidents: ShadowHammer 2019, Ripple20 2020, SolarWinds 2020, torchtriton 2022, xz utils 2024 and PKfail 2024

Six documented incidents. In every case the target was the software around the hardware, not the silicon.

None of these was an attack on a chip designer's silicon. They are relevant because each one hit a kind of software that chip designers ship every day: update tools, embedded network stacks, signed builds, Python dependencies, compression libraries and boot firmware keys.

ASUS ShadowHammer (disclosed 2019). In January 2019, Kaspersky researchers discovered that the ASUS Live Update utility had been used to push a backdoor. Kaspersky's report notes the trojanized software "was signed with legitimate certificates" belonging to ASUSTeK, that the operation took place in 2018, and that the attackers wanted to surgically target users identified by their network adapters' MAC addresses, more than 600 of which were found in the samples. The injected code was detected on over 57,000 Kaspersky users' machines.

Ripple20 (June 2020). JSOF disclosed 19 vulnerabilities in a TCP/IP library developed by Treck and embedded in devices from over 70 hardware vendors, according to Tenable's analysis. Tenable noted that tracking and identifying all affected vendors and devices was difficult, because the library had spread through the supply chain. Firmware SBOMs exist largely to answer that "do we ship it?" question in minutes rather than months.

SolarWinds Orion (December 2020). Mandiant's analysis found that trojanized Orion updates were digitally signed between March and May 2020 and posted to SolarWinds' official update site, with the backdoored component signed with the company's own certificate. Customers who installed legitimate, signed updates received the SUNBURST backdoor. It remains the reference case for a compromised build.

torchtriton (December 2022). PyTorch disclosed that between December 25 and 30, 2022, Linux pip installs of PyTorch nightly builds pulled a malicious package named torchtriton from the Python Package Index, because PyPI took precedence over PyTorch's own index for a package with the same name. The malicious binary collected system information, SSH keys and other files and sent them out through DNS queries. Users of the stable packages were not affected.

xz utils (March 2024). The backdoor in xz utils 5.6.0 and 5.6.1 was published as CVE-2024-3094 on March 29, 2024. It showed that an upstream release tarball, and not only the reviewed source code, is part of every downstream product's supply chain.

PKfail (July 2024). Binarly found that many devices shipped with a test Secure Boot Platform Key generated by American Megatrends International (AMI) and tagged "DO NOT TRUST." BleepingComputer reported that the issue affected 813 products from vendors including Acer, Dell, Fujitsu, Gigabyte, HP, Intel, Lenovo and Supermicro. Nobody stole these keys. A test key simply reached production firmware.

The common thread is plain. In every case, the attacker or the mistake went through the software that ships with the hardware, because that software is trusted, signed and widely distributed.

How do chip companies secure their software supply chain?

Chip companies secure their software supply chain by generating an SBOM for every firmware image, driver, SDK and container, ranking vulnerabilities by reachability, building on hardened CI with signed provenance, keeping signing keys in HSMs, scanning final binaries and images, enforcing license policy, and holding IP vendors and partners to the same standards.

None of it requires new standards. It applies proven software supply chain practice to the artifacts a fabless company actually ships.

Produce SBOMs for firmware, drivers and SDKs

A software bill of materials (SBOM) is a machine-readable list of every component in a piece of software, with versions and licenses. Generate one for every shipped artifact: each firmware image, each driver package, each SDK release and each container. Use CycloneDX, an OWASP project published by Ecma International as ECMA-424, or SPDX, a Linux Foundation project that was published as ISO/IEC 5962:2021. Many teams produce both, because OEM and cloud customers ask for one or the other.

Three details make firmware SBOMs useful rather than decorative:

  • Generate them from the build, not from a spreadsheet. The build knows exactly which versions it linked. A hand-maintained list drifts within a single release.
  • Include vendored and statically linked code. C firmware rarely uses a package manager. Copies of libraries sitting in a third_party/ folder are exactly what gets missed.
  • Version the SBOM with the artifact. Store it next to the signed image and publish it with the release, so a customer can match one to the other years later.

A minimal CycloneDX component entry for a vendored library in a firmware image looks like this:

{
  "type": "library",
  "name": "zlib",
  "version": "1.3.1",
  "purl": "pkg:generic/zlib@1.3.1",
  "licenses": [{ "license": { "id": "Zlib" } }],
  "properties": [
    { "name": "build:linkage", "value": "static" },
    { "name": "build:source-path", "value": "third_party/zlib" }
  ]
}

Rank vulnerabilities by reachability

A large SDK can match hundreds of known vulnerabilities by version number alone. Most sit in code paths the product never calls. A parser vulnerability in a library your profiler links but never feeds untrusted input is not the same as one in a library your driver uses to parse data from user space, even if both carry the same severity score.

Reachability analysis checks whether the vulnerable function is actually called from your code, and ranks the reachable ones first. Combine that with exploitation signals such as EPSS, which estimates the probability that a published CVE will be exploited in the next 30 days, and with lists of vulnerabilities already known to be exploited. The goal is a short list engineers trust, not a long list they learn to ignore.

Build provenance and signing

Treat the build as a security boundary.

  • Build release artifacts on hardened, ephemeral CI runners, never on developer workstations. Pin toolchain versions, including compilers and SDK build images, and record their digests.
  • Produce provenance attestations that follow SLSA and the in-toto attestation format, recording which source commit, which builder and which dependencies produced each artifact.
  • Keep signing keys in hardware security modules (HSMs), with signing done by an isolated service that verifies provenance before it signs anything.
  • Separate test keys from production keys structurally, so a test key cannot end up in a production image. PKfail is the cautionary tale.
  • Use a transparency log such as Sigstore's Rekor, which provides an immutable, tamper-resistant ledger of metadata, for artifacts where public verification helps your customers.

A minimal provenance check before signing, using the slsa-verifier tool, might look like this:

# Verify the artifact was built by the release pipeline from a tagged commit
slsa-verifier verify-artifact gpu-fw-4.2.0.bin \
  --provenance-path gpu-fw-4.2.0.intoto.jsonl \
  --source-uri github.com/example-org/gpu-firmware \
  --source-tag v4.2.0

On success the tool prints "PASSED: Verified SLSA provenance." The tool's documentation notes that only GitHub URIs are supported with --source-uri.

Scan binaries and images, not only source

Source scanning misses whatever was linked in from a prebuilt blob, an IP vendor's firmware or a base container image. Scan the final artifacts as well: extracted firmware filesystems, driver packages, SDK installers, VM and disk images, and every container image you publish. For packed firmware, extract the filesystem first, from your build's staging directory or with a tool such as binwalk, and scan the result.

Control licenses in SDKs and drivers

SDKs and drivers mix proprietary code with open source under many licenses. Kernel modules raise specific questions about GPL-only kernel symbols. Redistributed libraries carry attribution and notice requirements, and some copyleft licenses carry source-offer obligations. Keep a written license policy, check every release against it and generate the attribution file from the SBOM rather than by hand.

Manage partner and IP vendor risk

Ask IP vendors, EDA flow suppliers, OSATs and test houses for SBOMs of any software or firmware they deliver, plus a description of their secure development practices. Track their components the same way you track your own, monitor them continuously, and write SBOM delivery and vulnerability notification timelines into contracts.

Scan model files and containers before you publish them

Treat reference models like code. Scan model files in formats such as pickle, safetensors, GGUF and ONNX before they go into a published container, prefer formats that cannot execute code on load, and pin the sources your containers pull from so a lookalike package cannot be substituted.

One platform or a stack of point tools?

Most semiconductor security teams running this program assemble it from several products: a developer SCA tool, a binary and license analysis tool, an artifact scanner, a secrets scanner, signing tooling, a GRC tool for audit evidence and a spreadsheet for IP vendors. Each is good at its job. The cost is in the seams between them.

One platform vs a stack of point tools: fourteen capabilities a chip software security program needs, included in Safeguard, compared with the tools typically combined in an assembled stack such as Snyk, Black Duck, JFrog Xray and Vanta

Where each capability usually lives. The named tools are strong products; the difference is how many inventories you have to reconcile.

A typical assembled stack might use Snyk Open Source, which provides software composition analysis for dependencies in source repositories; Black Duck, which can identify open source supply chain risk even without access to the code, for binaries and license compliance; JFrog Xray, an enterprise SCA tool tied to Artifactory, for packages and containers; separate secrets scanning; Sigstore tooling for signing; and a GRC platform such as Vanta, which pulls evidence and maps it to controls, for audits.

A team that already runs Artifactory everywhere, or has a mature Black Duck license program, has real reasons to keep it. The trouble comes when the questions cross tools. Each product has its own inventory, its own severity scale and its own idea of what a component is. Tracing a vulnerable library in a container to the driver release that shipped it, to the OEM that received it, to the control evidence an auditor wants, becomes a manual job.

The alternative is one platform that holds the inventory, scans source, binaries and images, ranks by reachability, manages vendors, produces compliance evidence and opens fix pull requests, all against the same component graph. That is how Safeguard is built.

How Safeguard covers the chip software supply chain end to end

Safeguard covers the whole chip software supply chain in one platform: SBOMs, source, binary, VM image and container scanning, reachability-ranked vulnerabilities, secrets, licenses, signing and provenance, IP vendor risk, AI model scanning, runtime guardrails, compliance evidence and automated fixes, all on one inventory. Everything described here is generally available today.

Start with the scope, because that is the point. A chip company's software estate runs from boot firmware to reference containers, and Safeguard follows it the whole way.

Inventory and SBOMs. Safeguard generates and analyzes SBOMs in SPDX and CycloneDX and exports them in many formats, so every firmware, driver, SDK and container release can carry an SBOM matched to its exact version.

Binary, VM image and container scanning. In the console, teams upload binaries and images under Integrations > Upload App Binary, including QCOW2, VMDK, VHD and OVA virtual machine and disk images and container tarballs, alongside direct container image scanning. That covers SDK installers, driver packages, extracted firmware filesystems, reference VM images and the AI containers you publish.

Application security testing in one place. Software composition analysis covers the full transitive dependency tree, with reachability analysis that ranks the vulnerabilities your code actually calls and typically cuts noise by roughly 60 to 80 percent. First-party SAST covers your own code. DAST tests web surfaces such as management consoles, licensing portals and developer portals. Secrets scanning covers more than 200 patterns with live verification, which is how a leaked signing credential or cloud key gets caught in a repository before it ships. Vulnerabilities are enriched with EPSS and known-exploited data.

License compliance. SPDX license identification, policy enforcement and attribution generation run against the same inventory, so the notice file in your SDK comes from the SBOM, not from memory.

Attestation and signing. Safeguard supports SLSA, Sigstore and in-toto, so provenance and signatures sit next to the SBOMs they describe.

Partner and IP vendor risk. The Third-Party Risk Manager onboards vendors, requests SBOMs from them, scores and monitors their risk, and applies the same policies and gates to their components as to yours. The IP vendor spreadsheet can retire.

Policies and gates. Policies, gates and guardrails apply at build, CI and release, so a release with a reachable critical vulnerability, a forbidden license or an unsigned artifact can be stopped before it ships.

AI model scanning. Model files in pickle, safetensors, GGUF and ONNX formats can be scanned before they go into reference containers.

Runtime guardrails. Guard adds runtime protection on Linux, inspecting syscalls, library calls and network egress, and blocking secret and PII exfiltration on the egress path. It is useful on build farms and inference servers alike.

Compliance evidence. The compliance suite maps controls and collects evidence for frameworks such as SOC 2 and ISO 27001, so the security work you do on firmware and SDKs also produces the evidence your customers' security reviews ask for. Safeguard itself is SOC 2 Type I and Type II audited.

Deployment and AI choices. Safeguard runs as SaaS, in your own VPC or fully on premises, which matters to companies that keep design data close. Its AI features can use Safeguard's own security-tuned models or your own model, or run in Zero AI mode with no LLM in the loop at all. For teams that will not grant repository access, the local runner CLI scans on your own machine and sends only encrypted results.

Fix it without waiting for upstream

Spotting a reachable vulnerability in an SDK dependency is half the job. Safeguard's Griffin models open fix pull requests that upgrade or patch the vulnerable component. Each pull request runs through your own CI, so your tests, your hardware-in-the-loop checks and your reviewers decide whether it merges. You choose which remediation categories run automatically and which stay manual.

When upstream has not shipped a fix, or has fixed only the newest major version while you still support an older SDK branch, Safeguard Gold closes the gap. Gold rebuilds patched versions of open source packages and images from source, signs each build, attaches an SBOM, SLSA provenance and VEX, and serves them from registry.safeguard.sh to your organization. When upstream will not move at all, Safeguard can fork and patch the package so you have a fixed build to depend on. The free Gold Open Source directory at gold.safeguard.sh lets anyone look up packages, vulnerabilities and safe versions without logging in.

For a chip designer, that means a vulnerable compression or parsing library inside a driver or SDK does not have to hold up a release while you wait on a maintainer's schedule.

A fair counter-argument

A reasonable objection: chip companies already run serious security programs, with dedicated product security teams, hardware security architects and long-standing secure boot designs. That is true, and much of the industry's hardware root-of-trust work is excellent. OpenTitan, the open source silicon root of trust stewarded by the nonprofit lowRISC, is shipping in Chromebooks. Caliptra, which began at the Open Compute Project and now lives in the CHIPS Alliance, provides open IP and firmware for a root of trust block inside data center CPUs, GPUs and other SoCs.

The gap is usually not in the silicon security architecture. It is in the sprawl of software around it: dozens of SDK components, thousands of transitive dependencies, containers rebuilt weekly, firmware blobs from IP vendors and CI systems that grew organically. A secure boot chain verifies that an image is signed. It cannot tell you whether the signed image contains a vulnerable library, or was built from a tampered dependency. That is the software supply chain's job.

For partners

Design services firms, EDA resellers and integrators, OSATs, test houses and other companies in the chip ecosystem already sit inside their customers' build and release flows. Many of them are asked for SBOMs, firmware security reviews and audit evidence they have no easy way to produce. Safeguard works with partners who want to deliver software supply chain security to semiconductor and electronics customers. If that describes your firm, see safeguard.sh/partners.

Try it yourself

Click through these interactive Safeguard demos to see the steps described in this guide.

Test yourself

What makes a company fabless, and does NVIDIA make its own chips?

A fabless company designs and sells chips but owns no wafer factory, so it pays a foundry to make the wafers. NVIDIA is fabless: its annual report names TSMC and Samsung as the foundries that produce its wafers.

What is the difference between a foundry, an IDM and an OSAT?

A foundry manufactures wafers from other companies' designs, while an IDM designs chips and also builds them in its own fabs. An OSAT cuts finished wafers into dies, packages them and tests them for chip designers.

Why does the post say a chip is still useless after the fab is done?

After wafer fabrication the chip still has to be tested, packaged and mounted on a board, and firmware and drivers are what tell it what to do. That is why the software around the silicon is part of the supply chain.

Name some of the entry points where software supply chain risk reaches a chip company.

The post lists eight: EDA flow scripts, third-party IP with its firmware, open source inside drivers and SDKs, build systems, signing keys, test and provisioning software, distribution of SDKs and containers, and field updates. Each one has a documented real-world failure pattern.

What went wrong in the PKfail case?

A test Secure Boot Platform Key made by AMI and tagged "DO NOT TRUST" shipped in production firmware. Binarly found it affected 813 products from vendors including Acer, Dell, Lenovo and HP, and nobody had to steal the key for the problem to exist.

How did the torchtriton incident happen, and who was affected?

Between December 25 and 30, 2022, Linux pip installs of PyTorch nightly builds pulled a malicious package named torchtriton from PyPI, because PyPI took precedence over PyTorch's own index. Users of the stable packages were not affected.

Why can't a secure boot chain alone protect a chip company's software supply chain?

Secure boot verifies that an image is signed, but it cannot tell you whether the signed image contains a vulnerable library or was built from a tampered dependency. That is the software supply chain's job.

Frequently asked questions

How are computer chips made?

Quartz is refined into ultra-pure polysilicon, grown into a single-crystal ingot and sliced into polished wafers. A fab builds transistors and copper wiring on each wafer through 300 or more steps over about three months. The wafer is then tested, diced, packaged, tested again and mounted on a board, where firmware and drivers make it work.

Does NVIDIA make its own chips?

No. NVIDIA is fabless. Its fiscal 2026 annual report describes "a fabless and contracting manufacturing strategy," naming TSMC and Samsung as the foundries that produce its wafers, SK Hynix, Micron and Samsung as memory suppliers, and contract manufacturers such as Hon Hai, Wistron and Fabrinet for assembly, testing and packaging. NVIDIA designs the chips and writes the software.

What does TSMC do?

TSMC is a contract chipmaker and was the world's first dedicated foundry, founded in Hsinchu, Taiwan, in 1987. It manufactures wafers from its customers' finished layouts using its own process technology and supplies the process design kits customers design against. It also offers advanced packaging such as CoWoS. TSMC does not design its customers' chips.

What does ASML make, and why is it important?

ASML, based in Veldhoven in the Netherlands, makes lithography machines that print circuit patterns onto wafers. It is the only maker of EUV lithography systems, which print the finest layers of every leading-edge chip, so every advanced fab depends on it. ASML does not make chips or wafers. Nikon and Canon compete with it in DUV lithography.

What is a fabless company?

A fabless company designs and sells chips but owns no wafer fab. It writes the hardware design and the software, then pays a foundry such as TSMC to manufacture wafers and OSATs such as ASE or Amkor to package and test them. NVIDIA, AMD, Qualcomm, Broadcom and MediaTek are fabless.

How much does a chip factory cost?

Estimates put a leading-edge fab at more than $20 billion, the figure TSMC founder Morris Chang gave for a 3 nm fab in 2017. A mature-node fab can cost more than 10 billion euros. A standard EUV scanner costs about 170 million euros and a High-NA EUV scanner about 350 million euros.

How much does it cost to design a chip?

Estimates vary widely. IBS has put a 3 nm chip design at about $650 million and a 28 nm design at about $51 million, while SemiAnalysis says startups have built 7 nm chips for $50 million to $75 million including software. Mask sets alone run from about $1 million at 28 nm to around $40 million at 3 nm.

How do chip companies secure their software supply chain?

They generate an SBOM for every firmware image, driver, SDK and container, rank vulnerabilities by reachability, build on hardened CI with SLSA provenance, keep signing keys in HSMs apart from test keys, scan final binaries, VM images and containers, enforce license policy and require SBOMs from IP vendors. One platform can keep all of this on a single inventory.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

We email you a link to confirm, and only then add you to the list. Unsubscribe in one click. Privacy

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.