Chip Fab Siting Now Depends on Helium and Nanometers

Taiwan's semiconductor sector entered 2026 sourcing 88% of its helium from a single Gulf state, Qatar, up from 46% four years earlier. Helium is not optional in a fab. It cools the superconducting magnets in EUV lithography tools and purges chambers between process steps. When the Strait of Hormuz became a live chokepoint in early 2026, that concentration figure stopped being a procurement footnote and became the clearest signal yet that the geography of advanced chipmaking is being rewritten in real time. Helium is inert, but its supply chain is anything but.

The helium exposure is one data point in a larger recalibration. Over the past eighteen months, the world's three leading-edge foundries, TSMC, Samsung, and Intel, have collectively committed more than $400 billion to build fabrication plants outside their historical home bases. The decisions about where to pour concrete are no longer driven primarily by labour costs, tax rates, or even proximity to customers. They are driven by water rights, grid reliability, rare-gas supply contracts, and an explicit premium placed on political geography. Jingyue Hsiao reported for Digitimes that Taiwan's helium dependency nearly doubled in four years, a concentration that leaves the island's fabs "vulnerable and the outlook uncertain."

$265 billion. That is the total TSMC has now committed to its Arizona campus, after a $100 billion expansion announced on June 16 that added four more planned fabs to the three already under construction, the Milwaukee Journal Sentinel reported. The first Arizona fab is running 4nm wafers for Apple and AMD. The second, structurally complete as of April 2026, will add 3nm. The next six, if all proceed, would make the Phoenix site the largest concentration of advanced logic capacity outside Taiwan. The scale is hard to overstate. TSMC's entire global output in 2025 was roughly 16 million 12-inch equivalent wafers. The Arizona campus at full buildout could add perhaps 1.5 million of its own.

But a fab is not a data centre. It runs 24 hours a day on shift workers who understand vacuum systems, chemical mechanical planarisation, and the difference between a particle count of zero and a particle count of one inside a Class 1 cleanroom. Arizona does not have enough of those people. Amy Cutler of AZFamily reported on July 17 that a new workforce analysis shows Arizona could fall thousands of skilled technicians short each year, even as community colleges and Arizona State University accelerate training programmes. TSMC has already flown in several hundred engineers from Taiwan to keep the first fab on schedule. That is a temporary fix that does not scale to seven fabs. Workforce, not capital, is becoming the binding constraint on American chip sovereignty.

That calculation, what does this site actually have, beyond the subsidy cheque, explains why TSMC chose Kumamoto, Japan, for its first 3nm fab outside Taiwan. The company cited power and water security as primary factors, according to a June 26 report. A 3nm fab with 50,000 wafer-starts-per-month capacity consumes roughly 10 to 12 million gallons of ultrapure water per day and draws 250 to 350 megawatts of electricity. Kumamoto sits on an abundant groundwater basin and is connected to a grid with ample reserve capacity. The site selection was, in essence, a hydrological and electrical engineering decision dressed as industrial policy.

The public money is enormous, and it is now table stakes. TSMC has received or been promised subsidies from the United States ($6.6 billion under the CHIPS Act), Japan (roughly $3.5 billion for the Kumamoto fabs), and Germany (€5 billion for a Dresden specialty-node fab). Samsung secured $6.4 billion from the US CHIPS Act for its Taylor, Texas campus and a record $250 million grant from the state of Texas. Intel has $8.9 billion in committed CHIPS Act funding. Japan's government and 32 private backers poured $1.7 billion into Rapidus, the state-backed 2nm foundry startup. The European Chips Act unlocked sovereign subsidies for Infineon's €5 billion Dresden power-chip fab, Bloomberg reported. None of these projects would exist at their current scale without public money. The question is what that money actually secured.

What the subsidies secured, and what they didn't

Take the European Union. The bloc's 2022 Chips Act set a target of doubling Europe's global semiconductor production share to 20% by 2030. In June 2026, a Reuters Breakingviews analysis concluded that the programme "could do with a reset." The Infineon fab in Dresden makes power semiconductors on mature nodes, essential for electrification, but not the leading-edge logic the Act implicitly targeted. Intel's Magdeburg megafab, once the centrepiece of European chip sovereignty, was postponed in 2024 in close coordination with the German state of Saxony-Anhalt, Tom's Hardware reported. Instead, Intel committed €5 billion to expand its Leixlip campus in Ireland, Tech Times noted, which remains Europe's only site running EUV lithography at volume. The Leixlip expansion is the real story: it is cheaper and faster to expand a working EUV fab than to build a new one from scratch, no matter how large the subsidy.

Intel's broader turnaround, under CEO Lip-Bu Tan, is itself a geopolitical instrument. The company reported Data Center and AI revenue up 22% year-over-year in Q1 2026 and Foundry up 16%. Hyperscalers are quietly funnelling purchase commitments to Intel Foundry as a hedge against TSMC concentration risk, 24/7 Wall St reported, citing $8.9 billion in CHIPS Act backing plus investment from Nvidia and SoftBank. Whether Intel's 18A and 14A process nodes can match TSMC's N2 and A16 in power, performance, and area remains an open question. The geopolitical premium, however, is real and it is denominated in committed wafer volumes, not rhetoric.

That geopolitical premium is also reshaping the memory landscape. Korea Joongang Daily reported on July 12 that SK hynix is in talks to acquire Intel's Ohio One campus in New Albany, a site Intel originally planned as a $20 billion leading-edge logic fab. The deal would let SK hynix meet Washington's demand for US-based HBM production while giving Intel cash for its foundry buildout. It would also mark the first time a Korean memory maker has established advanced packaging and HBM production on US soil. The Ohio site was always a geopolitical asset in search of the right tenant; the SK hynix talks confirm that memory is now subject to the same siting logic as logic.

Samsung's Taylor, Texas fab, meanwhile, has moved from a construction site to a production facility with a name-brand customer. In July 2026, Samsung Foundry completed tape-out of Tesla's AI5 chip for autonomous driving on its SF2P+ 2nm process, Electrek reported, with production starting at the Taylor campus. The Tesla win is significant not just for the volume, AI5 will eventually ship in millions of vehicles, but for what it signals about the node. Samsung's 2nm is competitive with TSMC's N2 in density and superior in some power characteristics. The Taylor fab, backed by $6.4 billion in CHIPS Act money and a $250 million Texas state grant, is now Samsung's most geopolitically important asset outside Korea.

The wildcard in the 2nm race is Rapidus, the Japanese state-backed foundry that began pilot production of 2nm chips in Hokkaido in mid-2026. The Economist called the effort "a model for the sort of new-age industrial policy favoured by Japan's government." Rapidus has disclosed plans to price its wafers below TSMC's, Tom's Hardware reported, a strategy that only works if the yields are competitive and the customer pipeline is real. The company has signed a memorandum of understanding with the UK Semiconductor Centre and raised $1.7 billion from 32 private-sector backers alongside government funding. Whether Rapidus can transition from pilot line to volume production by its 2027 target is one of the highest-stakes questions in the industry. If it succeeds, Japan regains a seat at the leading-edge table for the first time since the 1990s. If it fails, it becomes a very expensive policy lesson.

Every fab siting decision is also a decision about what not to build. TSMC's Arizona fabs will not run the company's most advanced node at any given time; that will remain in Taiwan. Samsung's Taylor fab chose not to pursue the foundry's most aggressive gate-all-around variant, preferring to prove out the 2nm process on a lower-risk schedule. Intel's Leixlip expansion chose not to target sub-2nm nodes, sticking with Intel 4 and Intel 3 for the European customer base that needs those nodes most. The pattern is consistent: offshore fabs are being built for resilience and customer diversification, not for process leadership. The leading edge stays home.

The chokepoints that subsidies cannot fix

The Digitimes helium analysis exposes a structural vulnerability that subsidy cheques do not address. Helium is a byproduct of natural gas extraction, and the global supply is concentrated in a handful of countries, Qatar, the United States, Russia, and Algeria. Taiwan's shift from diversified sourcing to 88% Qatari dependency over four years was driven by price and logistics, not strategy. With the Strait of Hormuz disrupted in early 2026, the cost of that concentration became immediate. ASML's EUV tools, each costing upwards of $200 million, require liquid helium for their superconducting magnet systems. A sustained helium shortage would idle EUV capacity regardless of how many fabs have been built. The fab siting conversation tends to focus on concrete and cleanrooms. The real fragility is in the periodic table.

There is a related chokepoint that siting cannot solve: every leading-edge fab on the planet, regardless of which country pours the concrete, depends on ASML's EUV lithography tools. ASML is headquartered in Veldhoven, Netherlands, and its most advanced High-NA EUV systems contain components from Carl Zeiss in Germany, Trumpf in Germany, and a network of roughly 5,000 suppliers. A single High-NA scanner weighs 150 tonnes and takes 40 shipping containers, three cargo planes, and 20 trucks to deliver. The lead time from order to installation runs 18 to 24 months. No amount of subsidy money can accelerate ASML's production cadence. Fab siting may determine where the tools sit, but it does not determine how many tools exist. That is a Dutch monopoly, and it is the ultimate constraint on semiconductor sovereignty.

Beneath all of this is the original geopolitical driver of fab diversification: the Taiwan Strait. TSMC's advanced packaging capacity for AI accelerators, CoWoS, the chip-on-wafer-on-substrate technology that stacks logic and HBM, remains almost entirely in Taiwan. The Arizona fabs can produce leading-edge logic dies, but those dies still travel to Taiwan for packaging. TSMC is building advanced packaging lines in Arizona and Japan, but they will not reach volume until 2027 at the earliest. The strait premium, in other words, has been partially hedged on the front end of the process and barely touched on the back end. That asymmetry is the next chapter of this story.

The milestone to watch is not the next subsidy announcement. It is Q3 2027, when TSMC's second Arizona fab is scheduled to begin volume 3nm production, Samsung's Taylor fab is expected to ramp Tesla AI5 wafers, and Rapidus is targeting its first commercial 2nm shipments. By that quarter, the industry will have a real answer to the question that the helium dependency figure tacitly asks: whether $400 billion in geographically distributed concrete actually produces a resilient supply chain, or merely a more expensive version of the same one.