AI Bottlenecks and Why They Matter

Key Takeaways:

• Supply-side bottlenecks provide a holistic view of how quickly the AI industry can grow, and where the risks are today
• AI runs on two physical ingredients: advanced computer chips and electricity. How fast it grows depends on how much the world can produce.
• Chips are scarce at nearly every step, from the most advanced manufacturing to memory, machines, and raw materials.
• Specialty materials are just as constrained, and many come from companies better known for ordinary products like MSG, PVC pipes, and luxury toilets.
• Power is equally tight. Data centers wait years for a grid connection, and generators and electrical equipment are in short supply.
• Shortages are normal in an investment boom. High profits attract capital, industries overbuild, and today’s scarcity becomes tomorrow’s surplus.
• Across our portfolios, we own the companies benefiting from these shortages, but we underwrite each one knowing today’s biggest winners often become tomorrow’s losers.

Why bottlenecks matter

Demand for artificial intelligence is growing at a breathtaking pace, and new capabilities and models routinely capture the headlines. Beneath the excitement sits a much simpler question: can the physical infrastructure supporting AI keep up?

The newest models and the scramble for hardware get plenty of attention, while the longer-term structural risks surrounding physical machinery, materials, and power are often overlooked. We think these supply-side inputs offer the clearest and most measurable view of the landscape. They act as a critical barometer, showing not just how quickly the industry can scale today, but how much systemic risk is building within the AI value chain.

Stripping the technology down to its most basic ingredients

The cost of an economic unit of AI (a token) really comes down to the price of the advanced chips that do the calculating, the electricity required to run them, and the margin each provider across the technology stack extracts along the way.

In other words, you turn silicon into intelligence by running electricity through it. If you want to understand where AI can and cannot go, follow the chips and the power. That is also where many of the most measurable opportunities and risks are likely to emerge for investors.

Most of today’s bottlenecks do not stem from delivering AI to customers. They stem from building capacity ahead of demand, and that pre-emptive scaling introduces significant forecasting risk, which creates a giant game of telephone.

Model developers and enterprises forecast end-user demand. Hyperscalers translate those forecasts into massive compute requirements, and suppliers must ultimately deliver specific volumes of chips, equipment, and power.

Because upstream players sit several layers removed from the end-market, they feel the sharpest impact of any forecasting error in either direction. Watching these bottlenecks tells us how constrained supply chains really are, and, more importantly, where the industry is most vulnerable to overbuilding once those constraints finally ease.

The chip bottleneck

Advanced chips are the basic building blocks of AI. No AI model gets trained, and no AI service reaches a customer, without a chip switching on inside a data center somewhere.

Investors often focus on the companies designing AI models, but the more immediate constraint is whether the semiconductor ecosystem can supply enough computing power to meet demand.

Bottlenecks exist at nearly every stage of chip production, from manufacturing and advanced packaging to memory, equipment, and specialized materials. Understanding where these constraints sit offers a useful lens on both the opportunities and the risks building across the AI value chain.

1. Making the chips

Only one company can manufacture the most advanced AI chips, Taiwan's TSMC, and it is deliberately holding back. Building and operating a leading-edge semiconductor fab rank among the most complex engineering challenges in the world. It demands nanometer-scale precision, around the clock operations, inside cleanrooms that are 100 times cleaner than hospital operating rooms.

These facilities require up to $20 billion in upfront investment and take years to build, which makes overbuilding painful to unwind. TSMC learned this the hard way after the pandemic, when the entire industry rushed to expand just before demand collapsed and its new factories sat half-empty.

Today, it builds below customer orders, producing to what it believes is sustainable demand rather than producing to overly optimistic forecasts, even as AI buyers plead for more.

2. Two further chokepoints sit within the manufacturing process

Advanced packaging (CoWoS), is the first choke point. This process allows manufacturers to fit the main processor and its memory onto a shared sliver of silicon, so the two can pass information quickly. Before AI arrived, almost no capacity existed for this work, so the industry built it from scratch. TSMC has since expanded that capacity by roughly tenfold, yet it remains fully booked into 2027.

The second choke point is the high-speed memory these chips depend on, made by stacking ordinary memory chips into tall blocks. Only three companies in the world produce it; Samsung, SK Hynix, and Micron, and all three have sold out their entire production through the end of 2026.

Converting factories to produce this memory also leaves less capacity for standard memory, tightening supply for non-AI markets such as PCs, mobile phones, and industrial and medical equipment. Like TSMC, the memory players have stayed cautious since the last boom and bust, favoring profit over volume.

3. The machines that build them

Even where fabs want to expand, equipment can be a bottleneck of its own. Stacking those memory chips requires drilling thousands of microscopic holes straight down through the layers so the chips can talk to each other, and the one tool that does this comes almost entirely from a single supplier, Lam Research. The next generation of memory needs roughly 6 times as many holes, and a single flaw ruins the entire product. Supply for these tools is expected to remain tight until 2028.

The advanced lithography machines shown above, which print chip designs onto silicon, are an even more extreme case. Built by a single Dutch company, ASML, they rank among the largest and most expensive precision instruments ever made, weighing 200 tons and costing up to $400 million each.

ASML builds only 50 to 60 of the most advanced machines a year, and each takes two to three years to assemble, dismantle, ship, rebuild, and calibrate before it can run properly inside a manufacturing line. Supply simply cannot scale up quickly.

4. Piping, MSG, balloons, and toilets

Further upstream sits an even more peculiar mix of constraints, and they have emerged in the most unexpected places.

What do PVC piping, MSG flavoring, balloons, and toilets have in common? Historically, not much.

Today, companies in otherwise ordinary businesses find themselves at the center of the semiconductor supply chain, because they produce highly specialized materials the industry cannot function without.

• Silicon wafers: These originate mostly from two disciplined Japanese suppliers, one of which is the world’s largest provider of PVC piping for houses. Shin-Etsu applied the chemical engineering knowledge he built in PVC to purify raw materials and develop silicon wafers, the foundational canvas for printing semiconductor circuits.
• Insulating film: The small circuit board that sits directly under each chip relies on a special insulating film that comes almost entirely from Ajinomoto, the Japanese company best known for inventing MSG seasoning. It entered semiconductors by applying amino acid science, normally used in food products, to create durable films and resins for other end markets.
• Ceramic plates: TOTO is globally famous for its ultra-luxurious bathroom ceramics, namely its automated intelligent bidets and toilets. Their competence in ceramics technology has allowed them to expand into ceramic plates, used to hold silicon wafers in place during the chip-making process.
• Helium: Helium, yes, the stuff in balloons, works as both a cooling agent and a detector, protecting semiconductor wafers from extreme heat and contamination. Supply should stay short for years, after conflict in the Middle East knocked out roughly 40% of global production at the world's largest hub in Qatar. Industry participants have responded quickly, learning how to recycle helium at reuse rates of 80-90%.

The energy bottleneck

AI chips are extraordinarily powerful computing engines packed into a tiny form factor, but they cannot produce a single answer without a continuous flow of electricity. Connecting that power has become just as hard as securing the chips themselves, and the difficulty shows up in three distinct places.

1. Connecting to the grid

The first hurdle is simply plugging into the electrical grid. The United States electric grid has been notoriously underinvested in, and prior to the start of the AI boom, electricity load growth was flat for a decade.

Before a new data center can connect, the local grid operator must study whether the surrounding infrastructure can handle the enormous new load.

The backlog is staggering. More than 2 terawatts (TW) of projects now sit in these queues, which is enough to power one billion homes, and the typical wait stretches up to five years. Not all of it is real. Many developers file speculative or duplicate requests simply to hold a place in line, which clogs the queue for projects behind them. In the largest U.S. grid region, data centre projects approved in 2025 are not expected to be operational for another seven to eight years.

Even once approved, construction hits a wall. The large power transformers needed to connect now take roughly two to five years to arrive, compared with well under a year before the pandemic. Step-up transformers and heavy cabling are in severe shortage too, and heavy customization compounds the problem. Skilled workers are in short supply too, with the Bureau of Labor Statistics projecting a shortage of at least 349,000 workers in various trades across the electrical, HVAC, and construction trades.

2. Generating power on site

To get around these grid delays, some operators build their own power on site, but that path comes with its own constraints. The market for heavy-duty gas turbines, the standard solution for large-scale on-site power generation, is essentially spoken for, with the leading manufacturer carrying a backlog of well over 100 gigawatts (GW) stretching toward the end of the decade.

Data centers with gas turbines also prefer to sit next to existing natural gas resources and infrastructure, which creates another crowding effect in key regions like Virginia, Texas, Georgia, and Ohio. New nuclear projects, meanwhile, face heavy regulations, fuel shortages, and even longer build times. A typical two-year data center build-out now takes an additional one to three years when a new generation is part of the plan.

3. The equipment inside the building

Finally, even ordinary electrical gear in and around data centers is in short supply.

The switching equipment, circuit breakers, and backup power systems that move electricity from the property line down to each individual chip now carry wait times of roughly one to three years, several times longer than the few months they took before the boom.

These unglamorous components bridge the gap between the grid and the server. Without them, the power cannot reach the chips, no matter how much of it is available.

Bottlenecks are a feature, not a bug

Bottlenecks are a normal and even healthy feature of every investment boom, and they tend to be self-regulating. The pattern is remarkably consistent. Years of under-investment create shortages. Shortages push prices sharply higher, higher prices attract a flood of new investment, and eventually that delayed supply arrives at once and turns scarcity into surplus. The pattern is even more pronounced during a technology boom, for a few reasons worth keeping in mind:

• No history to lean on: Brand new technology has no track record of demand, so forecasts are hard to get right and tend to stretch endlessly into the future, regardless of whether that future arrives.
• Suppliers assume the good times last: Companies enjoying unusually high profits often believe those profits are permanent, and that belief shapes how aggressively they expand.
• Winner-take-all (or most) markets: First-mover advantage, network effects, and economies of scale are decisive in these markets, so players commit significant resources, for fear that not doing so is an existential risk. When everyone believes they must win at all costs, management teams and investors alike abandon their discipline around prudent capital allocation.
• Wall Street cheers it on: Banks and advisors earn handsome fees by promoting an optimistic story and helping companies raise capital, with little stake in whether the build-out makes sense over the long run.

What history teaches about bottlenecks

Every technology boom has run into its own bottlenecks: steel rails during the railroad era, power equipment as the electrical grid grew for lighting, refining capacity in oil, fiber during the dotcom boom, to the chips behind the personal computer. What looked like an impossible obstacle at the time was eventually solved through a familiar mix of human ingenuity and the simple power of profits and economics, and each solution became table stakes as whole industries adapted.

Not every bottleneck in AI will be solved quickly or easily, and some may take years. The industry has produced its own predictions for when the market tilts from undersupplied –to oversupplied, but we avoid developing our own timeframes because we view the AI demand curve as impossible to forecast. We can be reasonably confident, though, that as these industries scale up and money flows in, bottlenecks will clear. When they do, the world often finds itself with far more capacity than it ever thought needed.

For investors, the challenge is not simply identifying where shortages exist today but understanding how those shortages evolve over time. Companies earning extraordinary profits from current constraints can be attractive investments, but history suggests today’s bottleneck beneficiaries do not always remain tomorrow’s winners. That distinction shapes how we evaluate opportunities throughout the AI value chain. The takeaway is to treat these constraints as the real signals to watch, to enjoy the scarcity phase without mistaking it for permanence, and to stay alert for the moment the shortage quietly turns into a surplus.

Our holdings in perspective

As it relates to bottlenecks, here is a quick rundown of how Nicola Wealth is positioned across our public and private funds:

• Semiconductors: These are the obvious picks-and-shovels winners in AI today.  Our exposure runs across the semiconductor supply chain, including Nvidia (GPUs), Broadcom (custom chips/ASICs), Synopsis (chip design software), TSMC (outsourced manufacturing), ASML (lithography equipment), Samsung and SK Hynix (memory). These are all the necessary tools needed to build the underlying infrastructure that powers AI in the data centers.  We also have exposure to specialized semiconductor design companies including Tenstorrent, Groq, and SambaNova.
• Power and energy infrastructure: Here we own Schneider, Siemens, Hammond Power, and Itron. Each of these component-manufacturers benefit from the rising requirements needed to deliver and manage power within utilities and data centers. Within our infrastructure LP, we own a portfolio of federally regulated electric transmission utilities that benefit from data center demand as those projects connect to the grid.
• Distribution and delivery: We invest across the AI delivery ecosystem, pairing equity growth potential with steady credit income. We capture Big Tech's massive distribution footprint through core public holdings in the primary cloud hyperscalers (Microsoft, Google, and Amazon). Our private funds are co-invested in Nscale's advanced compute cloud services and back SpaceX's frontier play to build orbital data centers and Terafab chip manufacturing. We are also a senior lender to a leading data center service provider positioned to benefit from growing datacenter demand.

We own companies positioned to benefit while these bottlenecks persist, as well as companies we believe are likely to benefit if these constraints ease over time. This flexible approach allows us to actively manage our exposure as these scenarios unfold, while maintaining a balanced footprint across the AI value chain.

To learn more about how Nicola Wealth evaluates opportunities across the AI value chain, speak with your Advisory Team.