Let's cut straight to it. No, Nvidia does not operate its own lithography machines or semiconductor fabrication plants (fabs). The question "Does Nvidia use lithography?" touches on one of the biggest misconceptions in tech. The answer is a fascinating yes and no. Yes, Nvidia's graphics processing units (GPUs) and systems on a chip (SoCs) are absolutely created using advanced lithography processes. But no, Nvidia doesn't do any of the physical manufacturing itself. I've spent years following the semiconductor supply chain, and this distinction between design and fabrication is the single most important thing to understand. Nvidia is a fabless semiconductor company. They are the brilliant architects who design the blueprints for some of the world's most complex chips, like the H100 and the RTX 4090. Then, they hand those blueprints over to specialist manufacturing partners—primarily TSMC (Taiwan Semiconductor Manufacturing Company) and Samsung Foundry—who own the multi-billion-dollar fabs where the actual lithography, etching, and deposition happen.

What You'll Discover

Fabless vs. IDM: The Two Chip-Making Worlds

To get why Nvidia's model works, you need to see the whole playing field. The semiconductor industry is split into two main camps, and the choice between them defines a company's entire strategy, risk profile, and cost structure.

Model What It Means Key Examples Biggest Advantage Biggest Challenge
Fabless (Nvidia's Model) Designs chips only. Outsources all manufacturing to foundries. Nvidia, AMD, Qualcomm, Apple (for chips) Agility & focus. Can chase the best manufacturing tech without capital risk. Supply dependency. Competing for foundry capacity with rivals.
IDM (Integrated Device Manufacturer) Handles both chip design and in-house manufacturing. Intel, Samsung (for memory/logic), Texas Instruments Control. Vertical integration from design to finished wafer. Colossal capital expenditure. Must master both design and process tech.

I remember talking to a veteran engineer who'd worked in both models. He said the mindset is completely different. In a fabless firm like Nvidia, the entire culture is geared towards pushing the limits of chip architecture—more cores, better AI tensor units, smarter memory hierarchies. The fab is a black box they send files to. In an IDM like Intel, there's a constant, intense dialogue between the design teams and the manufacturing process teams. They're solving problems together, which can be powerful, but it also means your design is locked to your own factory's capabilities.

Nvidia's choice to stay fabless wasn't an accident. It was a strategic bet that paid off massively, especially as manufacturing costs skyrocketed.

How a Chip is Actually Made: The Lithography Steps

So, when Nvidia sends its GPU design to TSMC, what happens? The process is a modern marvel, and lithography is just one critical step in a very long chain. Calling chip making just "lithography" is like calling baking a cake just "putting it in the oven." It's crucial, but there's a lot before and after.

The Core Idea: Lithography is the process of projecting the chip's intricate circuit patterns onto a silicon wafer using light. It's like a super-precise, microscopic photographic process. The machine that does this is called a stepper or scanner, and the most advanced ones are made by a company called ASML. A single latest-generation EUV (Extreme Ultraviolet) lithography machine from ASML can cost over $150 million.

Here’s a simplified walkthrough of how an Nvidia design becomes physical reality at a foundry:

1. Design & Tape-Out

Nvidia's engineers finish the digital design. This involves billions of transistors. They simulate, verify, and finally produce a set of files called the "tape-out" or GDSII file. This is the digital blueprint sent to the foundry. I've seen these files—they're incomprehensible to the human eye, just layers of geometric data defining where every wire and transistor goes.

2. Mask Making

The foundry takes the GDSII file and uses it to create a set of physical masks (or photomasks). Each mask is a quartz plate with a patterned chrome layer that corresponds to one layer of the chip (a chip can have 50+ layers). Making these masks is an art form in itself, requiring its own specialized e-beam lithography tools.

3. Wafer Fabrication (Where Lithography Happens)

This is the main event in the fab. A blank silicon wafer goes through a cycle repeated for each layer:

  • Cleaning & Preparation: The wafer is meticulously cleaned.
  • Photoresist Coating: A light-sensitive chemical (photoresist) is spun onto the wafer.
  • Lithography (The Key Step): The mask is placed in the lithography scanner. Deep Ultraviolet (DUV) or Extreme Ultraviolet (EUV) light is shone through the mask, projecting the pattern onto the photoresist-coated wafer, hardening it in the exposed areas.
  • Etching & Deposition: Chemical or plasma processes etch away material not protected by the hardened resist, or deposit new materials (like copper for interconnects).
  • Stripping & Repeating: The remaining photoresist is stripped off, and the cycle starts again for the next layer with a new mask.

4. Testing, Slicing, and Packaging

After all layers are built, the wafer is tested. Functional chips are sliced into individual dies (the little square you think of as a chip). These dies are then packaged—placed onto a substrate, connected with microscopic wires or bumps, and sealed. This creates the finished product you can hold, like an RTX 4090 GPU chip.

Nvidia's Manufacturing Partners: TSMC & Samsung

Nvidia's success is tied directly to its foundry partners. They don't just pick a factory; they form deep, collaborative partnerships.

TSMC (The Primary Partner): For its most advanced chips, like the H100 and Blackwell AI GPUs and the high-end RTX 40-series, Nvidia relies almost exclusively on TSMC. TSMC's process nodes (like 4N, 5nm, 3nm) are industry-leading. The "N" in Nvidia's "4N" process is a custom variant co-developed with TSMC. This isn't a client-vendor relationship; it's a co-engineering effort. I've heard from contacts that Nvidia's design teams work in lockstep with TSMC's process integration teams for over a year before a chip goes into mass production, optimizing for power and performance.

Samsung Foundry (The Secondary Source): For certain product lines and for geographic/pricing diversification, Nvidia also uses Samsung. The GeForce RTX 30-series (like the RTX 3090) was famously manufactured on Samsung's 8nm process. Using a second foundry is a smart hedge against supply chain shocks and gives Nvidia negotiating leverage, but it also introduces complexity—a chip designed for one foundry's process can't just be ported to another without significant redesign work.

Why Doesn't Nvidia Build Its Own Fabs?

This is the billion-dollar question—actually, the hundred-billion-dollar question. With all its cash, why doesn't Nvidia just vertically integrate like Intel? The reasons are compelling and explain the dominance of the fabless model.

The Mind-Boggling Cost: Building a state-of-the-art fab today costs well over $20 billion. That's before you buy the equipment. A full production line with EUV lithography machines is a capital sinkhole. Nvidia would have to divert virtually all its R&D and profit for years into building just one factory, while its core design business would stagnate. It's a different business entirely.

The Pace of Obsolescence: Process technology advances every 2-3 years. By the time you finish building a fab for the 3nm node, the industry is already qualifying 2nm. If you stumble on a process transition (as Intel did), your entire product line suffers. As a fabless company, Nvidia can simply jump to whichever foundry has the best, most mature process for its needs. They let TSMC and Samsung bear the financial and technical risk of advancing lithography.

Focus is a Superpower: Nvidia's core competency is designing accelerated computing architectures for graphics and AI. That's hard enough. By not owning fabs, they can pour all their talent and money into software (CUDA), systems (DGX), and chip design. Trying to also master the physics of plasma etching and EUV light sources would dilute that focus dramatically.

An analogy I often use: Imagine a world-class restaurant. Should it also own the farm, the slaughterhouse, and the dairy to ensure ingredient quality? Some might (the IDM approach). But the most innovative chefs often prefer to source the best ingredients from dedicated, expert suppliers (the fabless model), so they can focus entirely on creating an unforgettable meal.

An Insider's Perspective on the Supply Chain

After years of watching this industry, I see a few things that most generic articles miss.

First, the notion that "design is easy" is completely wrong. The physical design of a multi-billion-transistor chip that will be manufactured by a third party is an exercise in extreme constraint management. Nvidia's engineers must design to the exact electrical and physical rules provided by TSMC's "process design kit" (PDK). They have to anticipate signal noise, power delivery, and heat dissipation across a massive die. A single suboptimal circuit block can ruin yield (the number of good chips per wafer) for the entire design, costing millions. The pressure is immense.

Second, the real bottleneck isn't always the lithography machine. While EUV tools are scarce, the bigger logjams can be in other areas: the availability of advanced packaging (like TSMC's CoWoS technology for H100), specific chemicals, or even the power and water needed to run a fab. Nvidia's supply chain team isn't just managing one factory order; they're navigating a global web of material and capacity constraints.

Finally, the partnership is not always smooth. Foundry capacity is allocated during times of shortage (like the post-pandemic chip crunch). Even a giant like Nvidia has to negotiate, commit to large purchase volumes, and sometimes wait. I've heard whispers of tense meetings where product launch timelines hinge on securing a few more thousand wafer starts per month. This dependency is the inherent trade-off of the fabless model.

Your Burning Questions Answered

If Nvidia doesn't manufacture chips, what do they actually do?

They are the architects and system designers. Their work is entirely in the realm of intellectual property and software. Teams of engineers design the microarchitecture (how the transistors are organized), create the logic circuits using hardware description languages, physically lay out the chip, and develop the massive software stack (drivers, CUDA, libraries) that makes the hardware usable. The final output is a set of digital files, not a physical object.

Will Nvidia ever buy or build its own fabs to control supply?

It's highly unlikely in the foreseeable future. The capital intensity and operational complexity are prohibitive. A more plausible scenario, which some analysts speculate about, is Nvidia taking an equity stake in a foundry like TSMC or making massive, multi-year prepayments to secure dedicated capacity. This gives them influence and security without the burden of directly running a manufacturing business with its own cyclical downturns.

How much does the lithography/foundry process contribute to a GPU's performance?

It's foundational. A better process node (e.g., moving from 8nm to 4nm) allows for more transistors in the same area, higher clock speeds, and significantly better power efficiency. You could have the world's best GPU design, but if it's built on an old, inefficient process, it will be hot, slow, and uncompetitive. Nvidia's design and TSMC's process are co-dependent. The design exploits the process's capabilities, and the process is tuned to meet the design's demands.

What's the difference between ASML, TSMC, and Nvidia's roles?

Think of it as a hierarchy of tools. ASML makes the most advanced printing presses (EUV lithography machines). TSMC buys these presses and many other tools to build the most advanced printing factory (the fab). Nvidia creates the intricate design (the "text" and "images") that gets printed in TSMC's factory using ASML's presses. ASML doesn't make chips. TSMC doesn't design chips for the open market. Nvidia doesn't own the factories. Each is a master of its specific domain.

Is the fabless model risky for Nvidia given geopolitical tensions?

This is the single biggest strategic risk highlighted by investors. Over 90% of the world's most advanced chip manufacturing capacity is in Taiwan (TSMC). Any disruption there would severely impact Nvidia's ability to produce its flagship products. This is why Nvidia (and the industry) is actively diversifying. They are qualifying more chips at Samsung (South Korea) and are likely to use TSMC's new fabs being built in Arizona, USA. However, geographic diversification of advanced fabs will take years and comes at a higher cost.

The bottom line is clear. Nvidia uses lithography in the most critical way possible—by designing chips that push the limits of what lithography can achieve. They just don't operate the machines themselves. This fabless partnership model, separating the brain (design) from the brawn (manufacturing), has been the engine behind the rapid innovation in computing. It allowed a company focused on graphics to pivot and power the AI revolution, all without ever having to worry about the physics of an EUV light source. That, in my view, is their real genius.