NextFin News - The AI boom is starting to reward a different class of company: not just the firms that design accelerators or fabricate leading-edge wafers, but the suppliers that make advanced chip production physically possible. As foundries and memory makers raise spending on fabrication and advanced packaging, vacuum systems and ultra-high-purity industrial gases are moving from background inputs to front-line bottlenecks. That shift is showing up in hard capital commitments, long-dated supply projects and surging equipment orders, suggesting the next wave of AI winners may sit deeper in the factory than many investors first assumed.
The numbers behind that shift are no longer abstract. Taiwan Semiconductor Manufacturing Co. said in its second-quarter 2026 conference call that, given “continued strong structural demand” including the “newly emerging agentic AI market,” it had raised its full-year 2026 capital budget to between $60 billion and $64 billion from the $52 billion to $56 billion range it had discussed earlier in the year. On July 31, Linde said it would invest $1 billion to expand its on-site industrial-gases complex in Phoenix to support the expansion of one of the world’s largest semiconductor manufacturers, while Linde LienHwa, its Taiwan joint venture, plans about $800 million more for air-separation and hydrogen assets tied to the same customer’s manufacturing and advanced-packaging sites. Air Liquide has announced nearly 200 million euros for an SK hynix advanced-packaging project in South Korea and more than $170 million for the memory maker’s first U.S. fab in Indiana. Atlas Copco, meanwhile, said its Vacuum Technique business posted 59% organic order growth in the second quarter, with record semiconductor-equipment orders and 19% organic revenue growth.
The investment map matters because it changes how the AI supply chain should be read. For much of the last two years, the market’s AI narrative revolved around chip designers, foundries, servers and power demand. That narrative remains correct, but incomplete. Before an AI model runs on a data-center cluster, a wafer has to be processed through vacuum-dependent manufacturing steps, memory has to be packaged at scale, and high-purity gases have to be delivered with reliability levels that conventional industrial supply chains do not require. Once those steps become the rate limiter, the companies providing them stop looking like commodity suppliers and start looking like infrastructure owners.
Broad semiconductor demand explains why this shift is occurring now. The Semiconductor Industry Association said on June 5 that global semiconductor sales reached $110.5 billion in April, up 11% from March’s $99.5 billion and 93.9% from a year earlier, and projected annual global chip sales would top $1.5 trillion in 2026. That does not mean every semiconductor-adjacent business becomes an automatic winner. It does mean that when the industry’s revenue base is compounding at that speed, each new fab module, each memory expansion and each advanced-packaging line requires the utilities layer to scale with it. In an earlier cycle, those supporting inputs might have remained a footnote. In the AI cycle, they are becoming part of the main story.
The central question is whether this move reflects a temporary capex surge or a more durable broadening in AI’s beneficiary list. The answer is not binary. The pace of orders and installations will still behave cyclically, because semiconductor capital spending always does. But the direction of travel now appears structural. AI is changing the mix of chip production toward more advanced fabrication, more high-bandwidth-memory packaging, more regional redundancy and more process intensity per unit of output. That does not erase the cycle. It does change who deserves to be counted inside the theme.
The New AI Constraint Is Physical, Not Conceptual
The first mechanism starts with a simple point that is often lost in market storytelling: AI demand has to move through a physical manufacturing chain before it becomes computing capacity. That sounds obvious, but it matters because the bottlenecks in that chain are not evenly distributed. A leading-edge semiconductor fab is not simply a building filled with tools. It is a tightly controlled process environment where contamination, pressure stability, gas purity and uptime are inseparable from yield. Vacuum systems are required in deposition, etch, chamber transfer and abatement applications. Bulk and specialty gases such as nitrogen, oxygen, argon and hydrogen are not optional consumables. They are foundational utilities embedded in the production architecture.
That is why the latest announcements from gas suppliers are more revealing than a normal industrial expansion would be. Linde’s July 31 release did not describe a small incremental contract or a general increase in customer activity. It described a long-term agreement under which the company will invest $1 billion to expand its existing Phoenix gases complex, adding two new SPECTRA air-separation units to complement three already operating at the site. The expanded complex will increase Linde’s supply of ultra-high-purity nitrogen, oxygen and argon for two new semiconductor fabrication facilities. In the same statement, the company said Linde LienHwa plans to invest approximately $800 million to build additional air-separation units and hydrogen production assets in Taiwan for new semiconductor manufacturing and advanced-packaging facilities linked to the same customer.
“Advanced semiconductor manufacturing depends on the reliable supply of gases at exceptional levels of purity,” Armando Botello, President Linde Gases, US, said in the company’s July 31 statement.
That quote is useful because it captures the underlying economics. Ultra-high-purity industrial gases are not just another line item in a fab budget. They sit inside the operating conditions that determine whether expensive tools run as intended. Once a supplier is built into the site with dedicated infrastructure, the relationship looks less like spot industrial demand and more like utility-style embedded capacity. In other words, the AI manufacturing wave is creating a class of supplier whose revenue is tied not merely to semiconductor output, but to the existence and expansion of the fab itself.
Air Liquide’s recent projects show the same mechanism from another angle: advanced packaging and memory. On June 3, the company said it had signed a major long-term contract with SK hynix and would invest nearly 200 million euros to support the memory producer’s large-scale industrial project dedicated to advanced packaging of high-bandwidth memory in Cheongju, South Korea. The company said it would build and operate a nitrogen production unit and supply high-purity gases and high-purity compressed air to the packaging and testing fab, which is scheduled to begin operations in late 2027. On July 1, Air Liquide said it would invest more than $170 million in Indiana to build and operate two new production units for SK hynix’s first U.S. fab, adding that next-generation memory chips such as HBM require ever-increasing quantities of high-purity gases. Those units are scheduled for commissioning at the end of 2028.
The importance of those projects is easy to miss if AI is still viewed mainly as a GPU story. The memory and packaging layer has become strategic because AI accelerators are only as effective as the surrounding memory architecture allows. HBM is no longer a side category; it is a core performance enabler. That means the companies supporting HBM packaging are not participating in a generic memory cycle. They are tied to one of the most AI-specific parts of semiconductor manufacturing. The same logic applies to the vacuum-equipment side. Atlas Copco said Vacuum Technique posted 59% organic order growth in the second quarter and cited record orders for semiconductor equipment, while revenues rose 19% organically. Those figures suggest customers are not merely restocking routine industrial demand. They are ordering to extend semiconductor-processing capacity.
The first-order takeaway is that AI’s supply chain is broadening. The more interesting second-order takeaway is that the broadening happens in places where replacement risk is low and qualification hurdles are high. That changes the quality of the revenue. A supplier that becomes part of a fab’s gas-delivery backbone or process-vacuum environment is exposed to capital deployment decisions that customers cannot execute casually. That gives these businesses a different profile from more transactional industrial names.
Why Vacuum Pumps and Gases Are Different From Commodity Inputs
It would be a mistake to treat these suppliers as if they were merely selling standardized industrial materials into a hot end market. Their strategic position is stronger than that. In semiconductor manufacturing, purity, reliability and process integration are not preferences; they are operating requirements. A molecule is not just a molecule if a contamination event can impair yields on advanced wafers. A pump is not just a pump if a vacuum failure disrupts a critical process step in a high-value tool set. That creates switching costs that are more structural than the market often credits in industrial valuation frameworks.
Consider the geometry of the recent announcements. Linde is not sending more gas trucks to an existing customer and calling it growth. It is building new air-separation infrastructure on-site, next to a semiconductor complex in Phoenix, and pairing that with parallel investment in Taiwan. Air Liquide is not merely broadening its regional sales force. It is building dedicated production assets around SK hynix facilities in South Korea and the United States. Atlas Copco is not benefiting from a generic upturn in factory maintenance. Its Vacuum Technique segment is seeing record semiconductor-equipment orders and absorbing the costs of ramping up capacity to meet them. Each example points to the same pattern: these are inputs whose value rises because they are integrated into expansion projects at the design stage.
That distinction matters for margins and durability. Commodity suppliers win when volumes rise and lose when volumes normalize. Embedded infrastructure suppliers can retain more of their economic role because the customer’s dependence is operational rather than discretionary. The contract may still be cyclical in volume terms, but it is less exposed to easy substitution. That is especially relevant in an industry where downtime is expensive and qualification takes time. When a fab ramps, reliability is worth paying for. When a fab is running at scale, reliability is worth defending.
There is also a policy dimension. The current AI manufacturing push is not concentrated in one geography. Phoenix, Taiwan, South Korea, Indiana, Idaho and Hiroshima all appear in recent project announcements tied to advanced semiconductors, memory and supporting materials. That matters because regional redundancy lifts the utility intensity of the system. It is one thing to expand a single mega-site; it is another to duplicate specialized support infrastructure across multiple jurisdictions. The more the semiconductor supply chain becomes geographically distributed for resilience or political reasons, the more spending flows to the enabling layer that must be rebuilt around each new site.
This is where the comparison with older chip cycles becomes useful. In earlier upswings, demand often ran hard through tool orders and then slowed once customers had enough capacity or inventories corrected. That pattern can still recur. But AI is changing the denominator. The industry is not simply trying to meet higher unit demand for broadly similar products. It is trying to support a more complex manufacturing mix, with tighter process requirements, heavier packaging needs and more geographically dispersed expansion. Even if the next year produces a familiar digestion period, the system that remains in place afterward is likely to be larger, more utility-intensive and more dependent on specialized infrastructure than the pre-AI semiconductor model.
That is the structural core of the story. The infrastructure layer is not becoming glamorous. It is becoming harder to ignore.
This Is Structural in Direction, Cyclical in Pace
The cleanest way to understand the story is to separate time horizons. In the short term, the spending pulse is still cyclical. Semiconductor customers are capable of compressing large amounts of demand into a narrow window, and suppliers will always feel that surge in orders before the eventual slowdown. Atlas Copco’s 59% organic order growth in Vacuum Technique is almost certainly too strong to be read as a sustainable long-run rate. The company also said margins were negatively affected by capacity ramp-up costs and acquisitions, which is exactly the sort of signal that often accompanies rapid expansion: demand is strong, but execution is catching up. For investors, that is a reminder that near-term operating leverage can look strongest just before growth normalizes.
TSMC’s capex path also shows the cyclical element clearly. Management moved from a January range of $52 billion to $56 billion to a second-quarter range of $60 billion to $64 billion. That kind of revision is bullish, but it is also characteristic of a market moving quickly. Large step-ups in capital budgets encourage suppliers to hire, build and invest against a demand pulse that may not stay this intense forever. The same logic applies to HBM-related projects: when a supply chain scrambles to expand packaging and memory support at once, individual order books can overshoot before the ecosystem reaches balance.
That is the cyclical part. The structural part is different and more important. The current investment wave is not centered on one-quarter utilization gains or a short replacement cycle. It is tied to new fabrication plants, new packaging facilities and new regional manufacturing footprints with timelines stretching into late 2027 and 2028. Linde’s Phoenix buildout adds two new air-separation units to support two new fabs. Linde LienHwa is adding gas and hydrogen assets around manufacturing and advanced packaging in Taiwan. Air Liquide’s South Korea project is aligned with a late-2027 startup, and its Indiana project with end-2028 commissioning. These are not tactical responses to a monthly swing in demand. They are physical commitments built around the premise that advanced semiconductor production capacity must keep expanding.
TSMC made the structural argument explicitly. In its second-quarter call, the company said “continued strong structural demand” from customers, including the “newly emerging agentic AI market,” drove the higher capital budget. That language matters because it links the spending not just to a passing recovery in chips, but to a new demand category that customers believe justifies heavier long-term investment. Management also said it was checking AI data-center progress, locations, demand and rack deployments to make sure chips would not be built into inventory. That does not guarantee perfect allocation. It does suggest the current capex wave is being tested against downstream deployment rather than being driven solely by optimism upstream.
“Given the continued strong structural demand from our customers, including the newly emerging agentic AI market, we have decided to raise our full-year 2026 capital budget to be between USD60 billion and USD64 billion,” TSMC said in its second-quarter 2026 conference call transcript.
The right conclusion, then, is not that semiconductor cyclicality is gone. It is that AI is adding a structural floor under a broader set of industrial activities than the market first modeled. Growth rates can normalize, projects can slip and memory cycles can still bite. But the manufacturing system serving AI now appears larger, more capital-intensive and more dependent on specialized support layers than a standard semiconductor rebound would imply. That is why “cyclical in pace, structural in direction” is the right framing. It allows for volatility without losing the larger shift.
The distinction also helps explain why these suppliers may not need speculative multiples to become meaningful winners. If the market starts to recognize that a greater portion of semiconductor capex is being routed through indispensable utilities and process-support infrastructure, earnings quality improves even without explosive top-line optics. That is the sort of reclassification that can matter more over time than a short-lived momentum trade.
The Second-Order Winner Thesis Rests on a Wider Revenue Pool
The market’s first instinct in an AI cycle is to look for direct beneficiaries: the chip designer with pricing power, the foundry with constrained capacity, the networking company selling into data centers. Those names still sit at the center of the narrative. The second-order thesis is narrower and, for that reason, easier to misprice. It says that when AI demand becomes an industrial buildout rather than just a product boom, the share of spending captured by enabling infrastructure can rise faster than investors expect.
The evidence for that thesis is visible in the project mix. Linde’s combined disclosed commitments in Phoenix and Taiwan amount to about $1.8 billion tied to one major semiconductor customer’s manufacturing and advanced-packaging expansion. Air Liquide’s disclosed South Korea and Indiana commitments tied to SK hynix exceed roughly 200 million euros plus more than $170 million, and its July 28 half-year statement also referenced more than $160 million in Arizona, more than $150 million in Idaho and a 200 million euro investment in Hiroshima to support advanced semiconductor activities. Atlas Copco’s order growth, meanwhile, shows that vacuum demand is not isolated to a single customer geography; it is broad enough to register as record semiconductor-equipment orders inside a major industrial group.
That widening revenue pool matters because it reshapes the marginal dollar of AI capex. The direct winners still collect the attention, but a growing slice of the physical expansion budget is being spent before a chip ever reaches a board. Once spending moves into site utilities, materials infrastructure and advanced packaging support, the beneficiary map broadens. In practical terms, the incremental dollar does not all accrue to the visible computing names. More of it leaks into the process environment underneath them.
That does not mean every gas producer or every pump maker becomes an AI stock. The theme remains selective. The companies that stand out are those with clear semiconductor exposure, high purity or high reliability requirements, on-site or near-site infrastructure models, and evidence of being embedded in leading-edge projects rather than selling into generic industrial demand. The recent public disclosures from Linde, Air Liquide and Atlas Copco meet that test. They tie their opportunity to identified semiconductor expansions, advanced-packaging investments and record sector-specific orders. This is not a story of management teams opportunistically invoking AI in broad industrial portfolios. It is a story of the manufacturing stack itself widening.
The consensus baseline is therefore changing in a specific way. The public already knows AI spending is large. The newer information is where that spending is landing. TSMC’s increased capex range of $60 billion to $64 billion, Linde’s $1 billion Phoenix buildout plus about $800 million in Taiwan, and Air Liquide’s semiconductor-linked investments in South Korea, Indiana, Arizona, Idaho and Hiroshima together suggest that the supporting layer is no longer incidental to the AI trade. It is becoming one of the places where that trade is actually being built.
The Counter-Thesis Is Real, and the Falsifier Is Clear
The strongest argument against this whole reading is that it is still just a dressed-up fab cycle. In that view, suppliers of vacuum systems and industrial gases look like new AI winners because customers are in the middle of a capacity rush that will eventually overshoot. Once installation schedules normalize and memory supply catches up, order growth falls back, utilization softens and the market remembers that these are industrial businesses with capital intensity, customer concentration and project-execution risk. History offers plenty of examples to support that caution. Semiconductor equipment booms have repeatedly produced supplier euphoria before digestion set in.
That objection deserves more than a token answer because it attacks the core thesis directly. It is entirely plausible that current demand is front-loaded. It is also plausible that HBM packaging, now a bottleneck, becomes temporarily overbuilt if multiple customers expand too aggressively at once. If that happens, the infrastructure names now being pulled into the AI conversation could lose their premium quickly. The companies themselves also face real execution risks: construction costs, energy inputs, permitting, ramp timing, acquisition integration and the simple danger of building ahead of a slower revenue curve.
The reason the structural thesis still holds, despite those risks, is that the public evidence so far describes multi-year manufacturing commitments rather than vague optimism. The projects already span multiple regions and extend into late 2027 and 2028. The customers involved are not speculative start-ups; they are central semiconductor and memory producers. TSMC is not merely hinting at opportunity; it has already raised its 2026 capital budget to as much as $64 billion. Linde and Air Liquide are not just forecasting demand; they are funding dedicated infrastructure against it. Atlas Copco is not only talking about AI interest; it has already posted record semiconductor-equipment orders in Vacuum Technique.
The falsifier should therefore be concrete. If TSMC were to cut its 2026 capital budget back below $56 billion, the top end of the earlier range it discussed before the latest increase, that would be a meaningful sign that current AI-related manufacturing intensity had been overstated. Likewise, if announced advanced-packaging or gas-supply projects tied to SK hynix, Phoenix expansion or Taiwan packaging capacity were delayed materially beyond the late-2027 to end-2028 timelines now disclosed, the case for vacuum and gas suppliers as durable AI beneficiaries would weaken. The reason to use those thresholds is simple: they are observable and tied directly to the mechanism the article is arguing. If the spending retreats and the projects slip, the supporting layer loses its strongest proof of durability.
Until then, the burden of evidence is moving the other way. The market no longer has to infer a connection between AI and the industrial utility layer of semiconductor manufacturing. The companies are naming the facilities, the capital budgets and the infrastructure being built. That does not eliminate cyclical risk. It does make the second-order beneficiary thesis far more concrete than it looked a year ago.
What Happens Next Depends on the Time Horizon
In the short term, sentiment can still outrun fundamentals. Equipment orders and project announcements tend to arrive before full revenue conversion, and margins can come under pressure as suppliers add capacity, labor and support infrastructure. That is already visible in Atlas Copco’s commentary about ramp-up costs. Over the next few quarters, the likely base case is continued strength in semiconductor-related order books accompanied by periodic concerns about conversion timing, cost discipline and the durability of customer spending. In other words, the near-term market path can stay volatile even if the strategic direction remains intact.
In the medium term, contract quality and entrenchment become the deciding variables. Gas suppliers with dedicated on-site or near-site assets should be better insulated than broad industrial names with only indirect exposure, because their revenue is tied to specific fabs and packaging plants rather than generalized end-market demand. Vacuum suppliers with strong semiconductor positioning should benefit as long as tool installations continue and service revenue follows the installed base. If AI infrastructure spending remains elevated into 2027, that is the horizon in which supporting suppliers can begin to look less like temporary add-ons to the theme and more like durable beneficiaries of semiconductor industrialization.
In the long term, the structural issue is whether AI is building a permanently larger manufacturing system. The evidence so far suggests that it is. More fabs, more packaging lines, more memory intensity and more geographic duplication all imply a broader layer of process support that has to be financed, installed and maintained. That is why the winners are spreading beyond the obvious names. Once an industry shifts from designing better chips to reproducing the physical conditions needed to manufacture them at scale across multiple regions, the infrastructure providers gain economic weight almost by definition.
The scenario framework reflects that split. In the base case, AI demand stays strong enough for leading-edge fabrication and HBM-related packaging expansions to move ahead broadly on schedule, while quarterly growth rates for the enabling suppliers cool from today’s extreme levels. In the upside case, agentic AI, inference deployment and memory-rich architectures accelerate faster than customers expected, forcing another leg of foundry, packaging and utility expansion that further enlarges the addressable market for vacuum and gas suppliers. In the downside case, AI monetization disappoints, customer capital discipline tightens, and the current wave of projects proves more front-loaded than cumulative, sending these companies back toward their traditional cyclical industrial profiles.
What should readers watch? Not slogans about AI demand. Watch the capital budgets, the commissioning timelines, the advanced-packaging buildout and the willingness of major customers to keep funding physical capacity. As of Aug. 12, 2026, those signals still point in one direction: the industrial backbone of AI is expanding alongside the headline names, and the companies that keep fabs under vacuum and full of ultra-pure gases are moving from overlooked suppliers to necessary participants.
The AI trade is no longer only about owning the chip. It is also about owning the infrastructure without which the chip cannot be made. That is not a side bet on the boom. It is the boom becoming industrial.
Explore more exclusive insights at nextfin.ai.
