Impossible to replace

Published: 27 Jul 2026

We have now named the companies and mapped the money. Which of these companies is genuinely irreplaceable? Not "hard to replace." Not "expensive to replace." Which ones, if they disappeared tomorrow, would shut down the global production of frontier AI?

Let's start with an honest disclaimer about the phrase "impossible to replace."

Given enough money, enough engineers and enough time, almost anything can be rebuilt. The question that actually matters is narrower and more urgent:

Which companies cannot be replaced quickly enough to keep today's AI supply chain running?

That reframe changes everything. A company does not need to be a permanent technological monopoly to be strategically irreplaceable. It just needs to be irreplaceable on the timescale that matters — the years, not the decades. Under that test, the list is short.

The top three #

ASML — Netherlands #

ASML makes the lithography machines used to print circuit patterns onto silicon wafers. More specifically, it is the only company in the world currently manufacturing EUV (extreme ultraviolet) lithography systems for high-volume production.

EUV lithography is required to manufacture the most advanced chips. That means the NVIDIA GPUs running AI training, the processors inside Apple devices, the custom accelerators Google and Amazon build, and the advanced memory SK hynix produces. Without EUV lithography, the leading-edge chips that power modern AI cannot be manufactured at scale.

ASML reported €32.7 billion in revenue in 2025 at a 52.8% gross margin.^[1] For a company making manufacturing machines, those are extraordinary economics. They reflect one thing: if you supply something indispensable that nobody else can produce, the market does not argue about price.

There is no current alternative supplier of EUV machines. If ASML disappeared today:

production from installed machines could continue for some time
maintenance, spare parts and upgrades would become increasingly difficult
capacity expansion at the leading edge would effectively stop

There is a newer generation called High-NA EUV — even more precise, needed for the next wave of chip scaling. ASML is the only supplier of that too.

The dependency: Netherlands.

ZEISS Semiconductor Manufacturing Technology — Germany #

Here is where it gets deeper. ASML's EUV machine depends on components that only one other company can make.

Inside an EUV lithography system, the light that writes circuit patterns onto silicon cannot pass through lenses. EUV light — at 13.5 nanometres — is absorbed by virtually everything, including glass. So instead of lenses, the system uses mirrors. Ultra-precision reflective mirrors, polished to atomic-scale smoothness, coated with specialised multilayer films.

ZEISS SMT in Oberkochen, Germany makes those mirrors.

Replacing ZEISS would require reproducing:

mirror surfaces accurate to fractions of a nanometre
multilayer reflective coatings with precisely controlled thickness
decades of joint engineering with ASML
qualification of every element against actual EUV performance

You cannot commission that component from another precision optics company. The technology and the integration with ASML's system are deeply specific. ZEISS and ASML have operated as an intertwined technical partnership for decades. The relationship is so close that ASML holds a minority ownership stake in ZEISS SMT.

The dependency: Germany.

TRUMPF — Germany #

An EUV machine needs light. But generating 13.5-nanometre extreme ultraviolet light is not straightforward. It requires a very specific mechanism: you fire an extremely powerful infrared laser at a tiny droplet of molten tin, at very high repetition rates. The laser vaporises the tin and creates a plasma. The plasma emits EUV light.

TRUMPF, headquartered in Ditzingen, Germany, makes that laser system.

The CO₂ laser amplifier inside an ASML EUV machine is a TRUMPF product. It must deliver very high average power, with precise timing, into a vacuum chamber, reliably enough to run a chip factory 24 hours a day. No other company currently manufactures a laser that performs this function at commercial scale.

The dependency: Germany.

The hidden chain #

Here is the thing that makes this trio so important. They are not independent chokepoints. They are sequential dependencies:

TRUMPF laser
    ↓ generates EUV light
ZEISS optics
    ↓ shape and project the light
ASML machine
    ↓ exposes the wafer
TSMC / Samsung / Intel fab
    ↓ manufactures the chip
NVIDIA / AMD / Google AI processor
    ↓ powers AI workloads

Remove any one of the first three links and the entire chain above it stops producing. Completely.

This is an unusual kind of fragility. Most supply chains have redundancy somewhere. If one steel supplier shuts down, you find another. If one chemicals company has a problem, you qualify a different source. But for ZEISS mirrors, TRUMPF lasers and the ASML machines they are built into, there is no alternative waiting in the wings.

TSMC — Taiwan #

TSMC is a different kind of case. Samsung Foundry can fabricate leading-edge chips. Intel Foundry is rebuilding that capability. So TSMC is not a strict technical monopoly the way ASML is. But it is effectively irreplaceable in the short and medium term, for a reason that has nothing to do with whether Samsung's process nodes work. The reason is integration depth.

TSMC reported a 59.9% gross margin for full-year 2025^[2] — exceptional for a contract manufacturer — because it combines:

leading-edge logic manufacturing
extremely high manufacturing yields
customer-specific process libraries built up over years
trusted IP handling across hundreds of chip designs
CoWoS advanced packaging, which connects logic dies to HBM stacks
the integration capacity to run all of this simultaneously for dozens of customers

Now consider what it would actually take to move a chip design from TSMC to Samsung:

Redesign the physical layout for Samsung's process rules
Replace TSMC-specific IP blocks with Samsung-qualified equivalents
Recreate the packaging integration
Run silicon validation on the new process
Debug manufacturing problems
Qualify the finished product
Secure production allocation

That process takes years. For a company mid-production of a critical product, it is not a real option on any near-term timescale.

The dependency: Taiwan.

An island of 23 million people that manufactures most of the world's leading-edge AI silicon, sits 180 kilometres from mainland China, and operates in one of the world's most complex geopolitical environments.

Hard to replace — but not impossible #

Below the genuinely irreplaceable tier sits a group where alternatives technically exist, but switching is costly, slow and disruptive enough that these companies are effectively locked in for most practical purposes.

NVIDIA — United States #

AMD makes competitive GPUs. Google, Amazon and Microsoft build their own accelerators. Technically, NVIDIA can be replaced.

Practically, replacing NVIDIA means replacing:

CUDA — a programming environment with a decade of developer inertia
AI libraries and compilers tuned for NVIDIA hardware
InfiniBand networking for tightly coupled AI clusters
rack-scale architecture and thermal management
the engineers who know how to operate it all

NVIDIA reported a 71.1% gross margin in fiscal 2026.^[3] That margin exists because the platform is embedded deeply enough that switching is genuinely expensive.

SK hynix — South Korea #

SK hynix is the leading supplier of high-bandwidth memory for NVIDIA's AI chips. Samsung and Micron also make HBM. So in principle, hynix has competitors.

In practice, individual accelerator designs qualify specific memory products through extensive engineering and testing. Switching HBM suppliers for a deployed product means going back through that qualification process — months of work, at minimum. And with HBM at capacity globally, the question of alternatives is somewhat academic anyway.

SK hynix's 49% operating margin in 2025 is the clearest signal that HBM scarcity created genuine pricing power in what had previously been a low-margin commodity memory business.

Synopsys and Cadence — United States #

You cannot design a modern AI chip by hand. You need electronic design automation software to place billions of transistors, route connections, analyse timing and power, simulate behaviour, prepare manufacturing files. Synopsys and Cadence together dominate this market. Siemens EDA is the third significant player.

Neither company alone is irreplaceable — they compete directly, and switching EDA tools is painful but done. The US-dominated EDA ecosystem collectively, however, is effectively irreplaceable. A country that loses access to Synopsys and Cadence simultaneously cannot design leading-edge chips with any near-term alternative.

KLA — United States #

KLA makes inspection and metrology equipment: the systems that detect defects on wafers and measure whether the manufacturing process is under control. A fabrication plant can physically process wafers without KLA's tools. It cannot do so economically — defect detection directly determines what percentage of expensive dies come out working.

Process control is what separates a commercially viable fab from an expensive research project.

Japanese materials companies — Japan #

Shin-Etsu, SUMCO, JSR, Tokyo Ohka Kogyo — these supply the silicon wafers, photoresists and specialty chemicals that foundries cannot operate without.

None of them is a monopoly. The difficulty is qualification. A semiconductor process depends on materials behaving consistently at atomic scales. Replacing any significant material supplier requires months or years of testing to establish that the new material produces equivalent results in the actual manufacturing process. That stickiness is what makes Japanese materials suppliers harder to replace than their commodity-adjacent pricing would suggest.

Country-level irreplaceability #

Putting it together by country:

Tier	Country	Critical capability
1 — Single-source or near single-source	Netherlands	EUV lithography machines (ASML)
1 — Single-source or near single-source	Germany	EUV optics (ZEISS SMT) and EUV lasers (TRUMPF)
1 — Near single-source at scale	Taiwan	Leading-edge fabrication and advanced packaging (TSMC)
2 — Indispensable concentrations	United States	AI chip design, EDA software, cloud platforms, semiconductor equipment
2 — Indispensable concentrations	South Korea	HBM, DRAM, NAND, Samsung foundry
2 — Indispensable concentrations	Japan	Wafers, photoresists, specialty chemicals, equipment, test
3 — Narrow specialist chokepoints	Switzerland	Semiconductor vacuum systems (VAT Group)
3 — Narrow specialist chokepoints	United Kingdom	Processor IP (Arm)

Here is what this map actually shows.

The United States dominates AI applications, chip design, cloud infrastructure and much of the semiconductor equipment industry. It is the single largest national contributor to the global AI stack.

And yet. The US cannot build frontier AI chips without:

ASML machines from the Netherlands
ZEISS optics and TRUMPF lasers from Germany
TSMC fabrication from Taiwan
SK hynix HBM from South Korea
Shin-Etsu wafers from Japan

The same incompleteness applies in reverse. Taiwan cannot fab chips without ASML equipment. South Korea cannot produce HBM without equipment from ASML, Applied Materials, Lam and KLA. Germany's ZEISS and TRUMPF are essential but cannot run the AI stack alone.

No country controls the complete chain. Not the US. Not China. Not Taiwan. Not any combination of two countries. The frontier AI supply chain is a genuinely multinational system, and every node in the Tier 1 cluster is necessary for the whole thing to work.

That interdependence is sometimes described as a risk. It is also a constraint on any one country's ability to dominate the technology.