October 13, 2026 | Podcast

Chips, Memory, and the Limits of the AI Boom

Everyone can see the demand side of the AI boom. The harder question is what’s actually constraining supply. In this episode of The Active Share, Hugo is joined by Siuchoon Koay, a research analyst covering large-cap global semiconductors at William Blair, for a technical walk through the AI hardware stack—from graphics processing units (GPUs) and high-bandwidth memory (HBM) to substrates and rack-level power.

Siuchoon traces how the industry’s bottlenecks have migrated over the past two years, weighs Nvidia’s and Taiwan Semiconductor Manufacturing Company’s (TSMC) strategic choices around capacity and vertical integration, and assesses the risk that Chinese open-source models pose to Western margins. The real question isn’t whether AI demand is real; it’s who ends up paying for it.

Comments are edited excerpts from our podcast, which you can listen to in full below. This episode was recorded in June 2026.

Why can’t chipmakers make enough products to meet demand?

Siuchoon Koay: I think it’s rooted in how sudden the AI boom all happened. It had been building since around 2023, but demand took off in 2025.

However, supply hasn’t kept pace. Chipmakers, especially foundries such as TSMC, have been much slower to add capacity, particularly once inferencing[1] took off. Demand accelerated too fast for supply to catch up.

It’s not only foundries but memory, substrates, and a lot of other components that go into the server rack as well. The issue is gradually being resolved, but not quickly enough for this demand profile.

Have you ever seen an environment where the supply chain is this tight?

Siuchoon: Never. In the past, demand was driven more by consumer adoption cycles with the internet and smartphones. This time, the consumer isn’t the direct driver. It’s AI frontier models and the cloud, and those have moved much faster than consumer adoption ever did.

ChatGPT’s launch in November 2022 seemed to be the catalyst, when AI in its current form arrived. Was that the moment when demand began stressing the supply chain?

Siuchoon: Yes. The switch flipped almost overnight with generative AI, and demand took off, primarily in AI chatbots. Now, we’re seeing coding as the other major use-case, and I think we’ll see more graphics such as AI-generated photos, and eventually video, going forward.

Where’s the demand bottleneck today?

Siuchoon: GPUs kicked off the AI revolution through training. As it extended into the second and third year, inferencing emerged. And as inferencing grows, a lot more chips will be required—not just GPUs, but central processing units (CPUs), language processing units (LPUs), and application-specific integrated circuits (ASICs)—plus everything around them such as memory, substrate, and racks.

Today, the most acute bottleneck is in memory, particularly HBM. Every GPU deployed needs corresponding HBM, and memory supply wasn’t ready to scale the same way.

Memory is also eating up more of the existing dynamic random-access memory (DRAM) capacity, which has sharply driven pricing up—commodity DRAM has doubled, if not tripled, in the first half of 2026—and that’s looped back to push HBM prices even higher. It’s getting resolved, but not fast enough.

Let’s zoom out. How do these bottlenecks get solved? Is it companies buying more equipment to boost capacity, or is it products emerging to work around the shortages?

Siuchoon: In early 2026, there were still some companies who were unconvinced about AI. They were dragging their feet on capacity decisions, wondering if we were late in the cycle and whether adding more would just create a glut down the road.

But demand strength held up over the following months, and foundries and memory vendors became convinced it would last. By the time they realized that, though, it was a bit too late, and we ended up in this very tight situation.

How the industry resolves this bottleneck will likely come down to cost. Everyone’s focused on reducing costs across chips and chip-making equipment, and that has driven a proliferation of form factors beyond the most expensive options.

ASICs are one example: you don’t always need a GPU, and for some workflows, a CPU, which is much cheaper, will do. The same applies to memory. It’s not always HBM; sometimes DRAM or small outline compression attached memory modules (SOCAMMs) bring the cost down for the hardware stack overall.

How the industry resolves this bottleneck will likely come down to cost.

Profitability across most of the supply chain, whether you look at margins or absolute profit dollars, has never been better. Will companies end up having to spend a lot more on capital expenditures (capex)?

Siuchoon: Right now, demand doesn’t appear to be slowing. Looking at token demand specifically, it’s growing roughly 8 to 10 times a year, which is an extraordinarily strong profile. No one can spend 8 to 10 times more capex every year to keep up with that, so innovation must fill the gap at the hardware level, the software level, and in the algorithms to bring costs down.

For instance, every new generation of Nvidia’s lineup of chips and systems is trying to lower the cost of compute and bring token costs down. That’s how you solve the problem: more tokens at a cheaper price, which drives adoption and lets usage keep proliferating.

Should we be worried that the cure for high prices is a wave of capex that overshoots supply, like every past chip cycle, where makers eventually expand too much?

Siuchoon: No, because I think the risk is less on the demand side. It’s clear the AI use-case is real and will stay relevant, so demand for tokens isn’t the issue. What could break is who’s paying for it, the financing behind the chips, and their cost. That’s where the strain may show up in this cycle.

So far, a lot of this has been funded through private financing. IPOs have helped release some of that pressure, and as frontier model companies continue to go public, the market will take on more of the financing burden for AI infrastructure.

But for the long term, we still need answers on two things: what margins will look like for frontier models, and how large the total addressable market is for what AI can disrupt. Those two questions will determine the valuations these companies get, and ultimately whether they can keep funding chip supply.

What does agentic AI mean for chip demand?

Siuchoon: For agentic AI, we expect a mix of GPUs, ASICs, and some LPUs and CPUs; it won’t look like the relatively pure GPU-driven picture we saw with training. That will help make the competitive picture more complex, with CPU vendors pushing into GPUs, IP makers pushing into CPUs, GPU players diversifying into LPUs, and so on.

China’s AI ecosystem seems happy being a fast follower of frontier models such as OpenAI, Anthropic, and Gemini, at much lower cost. Is that a threat in the medium term?

Siuchoon: China has been pushing open-source models, largely driven by the big internet companies alongside DeepSeek. That’s the direction of travel for Chinese AI efforts. The total cost per token is lower, partly because they’re using older chips and have more abundant energy.

Longer term, I think critical applications in the United States and Europe will still rely heavily on Western frontier models. And that’s not likely to change soon: open source still has trust issues and questions about whether it’s good enough for mission-critical tasks.

But as we move toward a point where every household item could carry some intelligence, I’d expect more adoption of open-source models, though not necessarily Chinese ones. That could challenge the existing narrative for frontier models, since open source puts pressure on those margins.

I’d expect more adoption of open-source models. That could challenge the existing narrative for frontier models.

If China succeeds in building a fully self-sufficient chip supply chain, does that risk pulling demand away from the companies investors focus on today?

Siuchoon: If a Chinese chip ecosystem were to develop, emerging markets could see higher adoption of Chinese models. And that would potentially shift demand away from the Western, Japanese, South Korean, and Taiwanese semiconductor names we have typically focused on, and toward the Chinese hardware ecosystem instead.

Given how China has driven down costs and dominated profit pools in other industrial markets, could it end up doing the same with semiconductors?

Siuchoon: If Western frontier models can’t defend their margins against open source, it becomes untenable, because the question of who’s paying for it can’t be answered.

For hardware investors, commoditization is the real fear over the long term. Things do get cheaper to produce at the margin over a decade or two, as we’ve seen in other industries, and that erodes the competitive advantage today’s players hold. It really comes down to whether frontier models can keep holding their user base in the West. If that falters, and open source becomes more prevalent, the competitive balance could shift further in China’s favor.

What is the central risk to the whole chip picture?

Siuchoon: For the AI stack economics to work, frontier models need to hold their margins, ideally improve them, and not see them decline from where they are today. If they can’t defend that standing, even with volume growth, things will get difficult: valuations will likely decline, and every incremental dollar of cash flow will become more constrained.

I think Nvidia is hedging against this by offering its own LLM, an alternative to both the expensive frontier models and the Chinese open-source models, to companies who don’t want to rely on the latter.
Ultimately, margins need to be defended at the frontier level, or investors will start questioning what terminal value actually looks like.

Given rising supply chain costs and the circularity concerns around vendor financing, could companies such as Nvidia go further and invest directly in suppliers?

Siuchoon: Yes, I think so. Supply is very tight today, and by offering financing for suppliers’ capex, Nvidia can help resolve some of these bottlenecks and buy itself a time-to-market edge over competitors.
So, in short, Nvidia’s strategy of extending supplier financing makes sense, and incumbents need to be more aggressive on capex to defend their position.

If Western frontier models can’t defend their margins against open source, it becomes untenable.

If a supplier such as TSMC stays disciplined on capex, does it risk handing market share to rivals who expand instead?

Siuchoon: Investors have already asked whether TSMC missed the boat by not adding capex sooner, and I think many would agree. That’s part of why Intel and Samsung have recently rallied: they’re picking up new customers, improving utilization, and building a virtuous cycle for the No. 2 and No. 3 foundry players. By delaying capacity, the incumbent effectively hands an advantage to its rivals.

Rather than asking when quantum computing arrives, what signposts are you watching that would raise its probability as a real threat?

Siuchoon: We’re still looking for real use-cases regarding quantum computing. Jensen Huang, the CEO of Nvidia, has said before that quantum processing units (QPUs) will exist alongside GPUs and CPUs, since quantum handles different workloads. So, there’s still a clear differentiation in use-cases, and from that angle, the threat to GPUs and CPUs isn’t imminent.

That said, it’s an evolving landscape. Algorithms are improving, and the range of problems quantum computers can solve is broadening. As the scope of quantum computing widens, it could become more of a threat to classical computing. We’re not at that tipping point yet, but it’s something to keep watching for.

Are there certain quantum computing companies that are now on your radar?

Siuchoon: There are already some smaller quantum computing companies that are public today, reporting sales, who their customers are, and what they’re buying these systems for. We’re tracking that narrative to see whether it starts encroaching on classical computing and whether these systems begin doing exactly what classical computers do. So far, it hasn’t reached the point of being a real replacement.

[1] Inferencing is the process where a trained AI model uses its learned knowledge to analyze new, unseen data and make a prediction or generate a response.

Topics: AI China Technology United States

Subscribe Now

Want the latest insights on the economy and other forces shaping the investment landscape?

Subscribe to our Investing Insights newsletter.

Any investment or strategy mentioned herein may not be appropriate for every investor. There can be no assurance that investment objectives will be met. Products and services listed are available only to residents of this jurisdiction and may only be available to certain categories of investors. The information on this website does not constitute an offer for products or services, or a solicitation of an offer to any persons outside of this jurisdiction who are prohibited from receiving such information under applicable laws and regulations. Nothing on this webpage should be construed as advice and is therefore not a recommendation to buy or sell shares.

Please carefully consider the Funds’ investment objectives, risks, charges, and expenses before investing. This and other information is contained in the Funds’ prospectus and summary prospectus, which you may obtain by calling 1-800-742-7272. Read the prospectus and summary prospectus carefully before investing. Investing includes the risk of loss.

The William Blair Funds and William Blair ETFs are distributed by William Blair & Company, L.L.C., member FINRA/SIPC.

The William Blair SICAV is a Luxembourg investment company with variable capital registered with the Commission de Surveillance du Secteur Financier (“CSSF”) which qualifies as an undertaking for collective investment in transferable securities (“UCITS”). The Management Company of the SICAV has appointed William Blair Investment Management, LLC as the investment manager for the fund.

Please carefully consider the investment objectives, risks, charges, and expenses of the William Blair SICAV. This and other important information is contained in the prospectus and Key Investor Information Document (KIID). Read these documents carefully before investing. The information contained on this website is not a substitute for those documents or for professional external advice.

Information and opinions expressed are those of the authors and may not reflect the opinions of other investment teams within William Blair Investment Management, LLC, or affiliates. Factual information has been taken from sources we believe to be reliable, but its accuracy, completeness or interpretation cannot be guaranteed. Information is current as of the date appearing in this material only and subject to change without notice. Statements concerning financial market trends are based on current market conditions, which will fluctuate. This material may include estimates, outlooks, projections, and other forward-looking statements. Due to a variety of factors, actual events may differ significantly from those presented.

Investing involves risks, including the possible loss of principal. Equity securities may decline in value due to both real and perceived general market, economic, and industry conditions. The securities of smaller companies may be more volatile and less liquid than securities of larger companies. Investing in foreign denominated and/or domiciled securities may involve heightened risk due to currency fluctuations, and economic and political risks. These risks may be enhanced in emerging markets and frontier markets. Investing in the bond market is subject to certain risks including market, interest rate, issuer, credit, and inflation risk. High-yield, lower-rated, securities involve greater risk than higher-rated securities. Different investment styles may shift in and out of favor depending on market conditions. Diversification does not ensure against loss.

Past performance is not indicative of future returns. References to specific companies are for illustrative purposes only and should not be construed as investment advice or a recommendation to buy or sell any security.

William Blair Investment Management, LLC is an investment adviser registered with the U.S. Securities and Exchange Commission.

Issued in the United Kingdom by William Blair International, Ltd., authorized and regulated by the Financial Conduct Authority (FCA), and is only directed at and is only made available to persons falling within articles 19, 38, 47, and 49 of the Financial Services and Markets Act of 2000 (Financial Promotion) Order 2005 (all such persons being referred to as "relevant persons").

Issued in the European Economic Area (EEA) by William Blair B.V., authorized and supervised by the Dutch Authority for the Financial Markets (AFM) under license number 14006134 and also supervised by the Dutch Central Bank (DNB), registered at the Dutch Chamber of Commerce under number 82375682 and has its statutory seat in Amsterdam, the Netherlands. This material is only intended for eligible counterparties and professional clients.

Issued in Switzerland by William Blair Investment Services (Zurich) GmbH, Talstrasse 65, 8001 Zurich, Switzerland ("WBIS"). WBIS is engaged in the offering of collective investment schemes and renders further, non-regulated services in the financial sector. WBIS is affiliated with FINOS Finanzomubdsstelle Schweiz, a recognized ombudsman office where clients may initiate mediation proceedings pursuant to articles 74 et seq. of the Swiss Financial Services Act ("FinSA"). The client advisers of WBIS are registered with regservices.ch by BX Swiss AG, a client adviser registration body authorized by the Swiss Financial Market Supervisory Authority ("FINMA"). WBIS is not supervised by FINMA or any other supervisory authority or self-regulatory organization. This material is only intended for institutional and professional clients pursuant to article 4(3) to (5) FinSA.

Issued in Australia by William Blair Investment Management, LLC (“William Blair”), which is exempt from the requirement to hold an Australian financial services license under Australia's Corporations Act 2001 (Cth). William Blair is registered as an investment advisor with the U.S. Securities and Exchange Commission (“SEC”) and regulated by the SEC under the U.S. Investment Advisers Act of 1940, which differs from Australian laws. This material is intended only for wholesale clients.

Issued in Singapore by William Blair International (Singapore) Pte. Ltd. (Registration Number 201943312R), which is regulated by the Monetary Authority of Singapore under a Capital Markets Services License to conduct fund management activities. This material is intended only for institutional investors and may not be distributed to retail investors.

Issued in Canada by William Blair Investment Management, LLC, which relies on the international adviser exemption, pursuant to section 8.26 of National Instrument 31-103 in Canada.