SK Hynix Orders $8 Billion in ASML EUV Machines — Largest Single Order in ASML History

SK Hynix just placed the largest single order in the history of ASML — $7.9 billion worth of extreme ultraviolet lithography machines, filed as a regulatory disclosure in Seoul on March 24, 2026.

This is not a vague capital commitment or a letter of intent. It's a binding procurement agreement: 30 EUV scanners, delivered through 2027, deployed across two specific factories. The scale is unprecedented. For context, ASML's total annual revenue is roughly $28 billion. SK Hynix just pre-committed to nearly a third of that in a single order.

What SK Hynix Actually Ordered

The order covers top-end EUV scanners — ASML's most advanced lithography tools, the ones that make sub-5nm chip production possible. At roughly $250 million per High-NA EUV unit, 30 machines represents a complete generational upgrade to SK Hynix's manufacturing capacity.

The machines are going to two facilities:

Yongin: The new megafab currently under construction in Gyeonggi Province, South Korea. SK Hynix has committed 120 trillion won ($88 billion) to build the Yongin Semiconductor Cluster over the next decade. This EUV order is part of the initial tooling phase.

M15X, Cheongju: The existing factory already optimized for High Bandwidth Memory production. This is where HBM3E — the memory stack currently inside every H100 and H200 server — gets made. The new EUV machines here accelerate the transition to HBM4.

Market reaction was immediate: ASML shares rose 4%, SK Hynix closed up 5.7%.

Why This Is About HBM4 and Nvidia Vera Rubin

HBM (High Bandwidth Memory) is not standard RAM. It's a vertical stack of DRAM dies connected by thousands of through-silicon vias, mounted directly alongside a GPU die on the same package. Each Nvidia H100 has 80GB of HBM3. Each Nvidia B200 has 192GB of HBM3E. The Vera Rubin platform — Nvidia's next-generation AI accelerator expected in late 2026 — will require HBM4.

HBM4 uses a 1c DRAM process node that requires EUV at multiple layers. Without EUV machines, you cannot manufacture HBM4 at volume. SK Hynix is buying 30 EUV machines specifically to ensure it has the tooling to supply Nvidia at scale.

TrendForce reporting from March 9 confirmed Samsung and SK Hynix are both tapped as Vera Rubin HBM4 suppliers, with initial shipments targeting Q3 2026. SK Hynix is expected to supply more than 50% of Nvidia's total HBM volume in 2026 across both HBM3E (current) and HBM4 (Vera Rubin).

This $8 billion order is the manufacturing infrastructure that makes that supply commitment physically possible.

The Race Against Samsung

SK Hynix has dominated the HBM market since 2023. It was the first to mass-produce HBM3E and the first to ship it to Nvidia at volume. Samsung has been trying to close the gap. Reports in early 2026 indicated Samsung's HBM3E yield rates were still below Nvidia's qualification threshold.

Samsung is also set to supply HBM4 for Vera Rubin, but SK Hynix's capacity advantage is widening. Doubling the EUV fleet by 2027, as the Yongin investment implies, means SK Hynix is building a lead that Samsung will need 2-3 years to replicate even if it starts ordering EUV machines today.

ASML has an 18-24 month delivery queue for its latest EUV tools. Samsung has not announced a comparable order. That queue position matters enormously — it determines who can manufacture next-generation chips in 2027 and 2028, not just 2026.

What EUV Actually Does (And Why It's a Bottleneck)

For developers used to thinking about chips in terms of nanometers and transistor counts, here's the physical reality: below 7nm, you cannot pattern chip features using conventional deep ultraviolet light. The wavelength is too long. EUV light at 13.5nm wavelength is generated by firing a high-power laser at a droplet of molten tin 50,000 times per second. The resulting plasma emits EUV radiation, which is then focused by mirrors onto a silicon wafer.

ASML is the only company in the world that can build these machines. The High-NA EUV scanner weighs 150 tonnes, requires 457,329 individual parts, and takes a year to assemble and ship. There is no alternative supplier. No country has domestically replicated this technology, including the US, Japan, or China.

This is why ASML's order book functions as a proxy for the global semiconductor roadmap. When SK Hynix books 30 machines, that tells you exactly what the AI memory landscape will look like in 2028.

The Geopolitical Dimension

ASML is a Dutch company, subject to Dutch and EU export controls. Its High-NA EUV machines — and even older DUV models — are export-controlled to China. Chinese chip manufacturers including SMIC cannot receive these tools without special license, which is not being granted.

SK Hynix's $8 billion order lands in this context: it deepens the manufacturing gap between Korea's chip industry and China's. Chinese fabs are currently limited to 7nm processes using older tooling. SK Hynix is buying the machines to manufacture 1c DRAM at under 2nm equivalent density. That gap compounds every year ASML cannot ship to China.

The Supermicro co-founder's arrest on March 19 — where Nvidia B200 and H200 servers were being routed to China through Southeast Asian front companies — illustrates the demand China has for these chips. The hardware embargo is holding. The finished-server smuggling channel is being shut down by federal prosecution.

Developer and Infrastructure Impact

HBM4 matters to infrastructure engineers because it directly determines GPU memory bandwidth — the primary bottleneck in large language model inference at scale.

HBM3E delivers roughly 1.2 TB/s bandwidth per GPU. HBM4 is projected at 1.8-2.0 TB/s — a 50-65% improvement. For a 70B parameter model running on 8 GPUs, that difference can halve token latency at peak throughput.

When SK Hynix books 30 EUV machines to produce HBM4, it's not just a financial story. It's a 2027 infrastructure story: the AI accelerators available in two years will be substantially faster than what's available today, and this order is a significant reason why.

Key Takeaways

• $7.9 billion — SK Hynix's ASML order, the largest single order in ASML's history, filed March 24 2026
• 30 EUV machines deployed at Yongin and M15X Cheongju factories through 2027
• HBM4 production for Nvidia Vera Rubin is the purpose — SK Hynix supplies 50%+ of Nvidia's HBM volume
• ASML +4%, SK Hynix +5.7% on announcement — market confirmed the strategic significance
• Samsung has no comparable announced order — SK Hynix is widening its manufacturing lead
• China cannot receive EUV machines — this order widens the US/ally vs China chip capability gap further
• HBM4 delivers 50-65% more memory bandwidth than HBM3E — AI inference throughput improves materially