When ASML shipped its first EUV lithography system in 2010, few outside the semiconductor industry understood what had just happened. The company had spent three decades and billions in R&D solving a physics problem nobody else could crack: generating extreme ultraviolet light reliably enough to print transistors at single-digit nanometre scales. Today, every advanced chip in the world - every GPU training a large language model, every inference accelerator running one - requires ASML EUV machines to manufacture. There is no alternative supplier. There is no substitute technology at these process nodes. It is, structurally, a monopoly.
Last week, Cerebras Systems listed on Nasdaq under the ticker $CBRS, raising $5.5 billion at $185 per share and closing its first day up 68% at a market cap of roughly $68 billion. The company makes the Wafer Scale Engine (WSE-3): a single chip the size of a dinner plate, manufactured by TSMC on their 5nm process node, containing 4 trillion transistors and 900,000 AI-optimised cores. It is a genuine engineering achievement. It is also priced at 133 times its 2025 revenue of $510 million.
The Cerebras IPO is being covered as a story about AI hardware competition - Cerebras versus NVIDIA, inference versus training, wafer-scale versus GPU clusters. That framing is not wrong. But it misses a cleaner question: in a war between chip architectures, who has the guaranteed contract?
The Supply Chain Nobody Is Talking About #
Every chip in this competitive landscape runs through the same three-step supply chain: ASML builds the lithography machines; TSMC (and to a lesser extent Samsung and Intel Foundry) uses those machines to manufacture wafers; chip designers - NVIDIA, AMD, Cerebras, Google, Amazon - buy manufacturing capacity from those foundries. ASML sits at the top of that stack and sells exclusively to the foundries, not to the chip designers.
This is not a trivial observation. It means that every AI chip company's growth is contingent on TSMC building more capacity, which is contingent on ASML shipping more EUV machines. ASML is the physical rate-limiter of the entire AI hardware buildout. When TSMC expands its advanced node capacity to serve NVIDIA's Blackwell GPU demand, ASML sells more machines. When AMD ramps its MI400 Helios series, ASML sells more machines. When Cerebras needs more wafer starts, ASML sells more machines. The combatants pay the arms dealer regardless of who wins.
What the Market Is Paying #
The chart below compares price-to-revenue ratios across the four main players in this stack, using market caps as of Cerebras's IPO week and the most recent full fiscal year revenues, sourced from each company's earnings releases and Cerebras's S-1/A filing with the SEC.
Cerebras at 133x is being priced as though the company will, over the next decade, take meaningful share from NVIDIA's $5.5 trillion business while maintaining or expanding its margins. That is possible. It requires flawless execution, customer diversification well beyond its current base, and a software ecosystem that can rival CUDA's decade-long head start. The market is pricing zero discount for any of those uncertainties.
ASML at 16x revenue is being priced as a mature industrial company with steady-state growth. The reality is somewhat different: ASML raised its 2026 full-year revenue guidance to €36-40 billion in its Q1 2026 earnings release, up from the prior €34-39 billion range, citing AI-driven demand acceleration. The company is ramping EUV production by 36% this year - from 48 units in 2025 to a target of 60+ in 2026. For a company with a monopoly on the tool that makes all advanced chips possible, 16x revenue with a two-year order book is a different risk profile than 133x revenue with one dominant product and a TSMC single-source dependency.
Revenue Trajectories #
The chart below shows why NVIDIA dominates the valuation conversation: it has genuinely been a different kind of revenue machine. From $27 billion in fiscal 2023 to $216 billion in fiscal 2026 is a 3-year growth trajectory that is historically without precedent for a company of its size. ASML and AMD look flat by comparison, though ASML's €32.7 billion for 2025 represents a 55% increase over 2022 - not nothing for a capital equipment company. Cerebras is growing from a very small base, which makes its percentage growth rate impressive and its absolute revenue small.
The Physical Bottleneck #
ASML's EUV shipment history is one of the most unusual charts in industrial history. The company targets shipping 60+ EUV systems in 2026, up from 48 in 2025 and 44 in 2024. Each system costs roughly €200 million for a standard Low-NA EUV unit, with the new High-NA EUV (the EXE:5200, required for sub-2nm process nodes) at approximately €350 million per unit - a price that TSMC has so far declined to pay at scale, citing cost.
The 2024 dip in units (44 vs 55 in 2022) is attributable to the semiconductor down-cycle, not structural demand destruction. The recovery to 48 in 2025 and the acceleration to 60+ planned for 2026 tracks exactly what the AI infrastructure buildout requires: more leading-edge fab capacity. Each EUV system ASML ships represents a physical constraint on the rate at which the entire chip industry can advance. No company - not NVIDIA, not TSMC, certainly not Cerebras - can route around this bottleneck.
The TSMC hesitation on High-NA EUV is worth flagging separately. TSMC's public position is that the €350 million price point is too high to adopt at scale given current wafer economics. For Cerebras specifically, this creates a timeline risk: the WSE architecture's next generation likely requires leading-edge nodes beyond 5nm, which means High-NA EUV, which means the manufacturing cost and availability of those machines directly constrains Cerebras's product roadmap. This single-foundry dependency - TSMC is the only company capable of manufacturing wafer-scale chips at the required yield - is arguably the most underappreciated risk in the Cerebras investment case.
Gross Margin: What Each Business Is Actually Worth #
The gross margin comparison below is not primarily a quality ranking - hardware companies structurally carry lower margins than software. But within the hardware category, NVIDIA's 71% gross margin on $216 billion in revenue reflects something genuinely unusual: a company that has, through CUDA's software ecosystem and a decade of developer lock-in, turned what should be commoditising hardware into pricing power that looks more like enterprise software. ASML's 52.8% reflects a monopoly supplier of complex precision equipment with long replacement cycles and no competition. AMD's 50% is a normal competitive semiconductor company. Cerebras at 39% reflects a company in early-stage volume ramp, selling into a market it is still developing.
The Groq Acquisition Changes the Competitive Map #
In December 2025, NVIDIA acquired Groq for $20 billion. Groq was, architecturally, Cerebras's closest peer: a company building Language Processing Units (LPUs) optimised specifically for fast inference workloads. Cerebras's WSE-3 and Groq's LPU were the two most credible architectural alternatives to GPU-based inference at scale.
NVIDIA's decision to pay $20 billion for Groq - a company with minimal revenue at the time - was effectively a signal that inference-optimised silicon is a real threat to Blackwell's economics in the data centre. It was also a pre-emptive move to own the competitive response. At GTC in March 2026, NVIDIA announced the Blackwell + Groq LPX combined architecture, merging Groq's inference throughput advantages with NVIDIA's software ecosystem and manufacturing scale. Cerebras is now competing against a NVIDIA that owns both the GPU training stack and a purpose-built inference stack. That is a materially harder competitive environment than existed twelve months ago.
The Investment Angle #
None of this is to say Cerebras is a bad business. $510 million in revenue growing at 76% year-on-year from a company that was doing $25 million three years ago, with a swing to net income in 2025, is a real achievement. The OpenAI partnership cited in the amended S-1 filing suggests genuine commercial traction beyond G42. The wafer-scale architecture produces measurable performance advantages on specific inference workloads. These are real.
The question is whether those advantages are reflected appropriately in a $68 billion market cap. At 133x revenue, the market is effectively assuming that Cerebras will, over the next several years, grow into a business that can justify that multiple - which implies either revenues in the multi-billion range at current multiples, or a compression to more normal hardware multiples at significantly higher revenues. Either path requires NVIDIA not to render the wafer-scale architectural advantage irrelevant, TSMC to continue manufacturing at the required yield and process node, and customer concentration to resolve from the current level without disruption.
ASML's path is simpler to analyse: more AI chips means more TSMC capacity expansion means more ASML EUV machines. The company raised full-year 2026 guidance to €36-40 billion with 51-53% gross margins and confirmed Q1 2026 order intake is strong. The 2026 EUV ramp to 60+ units is already in progress. The High-NA EUV product cycle, whenever TSMC and others adopt it at scale, represents the next significant revenue step-up - each unit at €350 million versus €200 million for the current generation. The €510 billion market cap at 16x revenue is pricing in steady growth from a company with a structural monopoly on the most critical enabling technology in the AI hardware stack.
Investors who want exposure to the AI hardware buildout without betting on which chip architecture wins the next five years have a European option. It is Dutch, it is priced at 16x revenue, and every combatant in the AI chip war is its customer.