Latest / The AI Supercycle / The Memory Constraint: Why HBM Has Become AI's New Oil

Transcript

The Memory Constraint: Why HBM Has Become AI's New Oil · The AI Supercycle

• The Memory Constraint: Why HBM Has Become AI's New Oil Semiconductor Bifurcation, Strategic Partnerships and the Fifth Binding Constraint 2026, Tim Hardwick / QF-MI http://qfmi.substack.com Automatic Shownotes Chapters 0:17 HBM and the AI Supply Chain 5:54 Capital Becomes the Constraint 10:05 The Three HBM Producers 12:52 Demand Is Accelerating 15:57 Market Rally, Then Correction 19:30 Correction or Top? 23:16 The HBM Base Case 26:58 Watching IPOs and Capital Flows Long Summary The AI supercycle is now being defined by physical bottlenecks rather than software releases. This episode focuses on high bandwidth memory, or HBM, and argues that the semiconductor rally is being driven by a genuine supply constraint that has become strategic rather than transactional. NVIDIA’s expanded partnership with SK Hynix is presented as evidence that memory supply is being locked up years ahead of demand. The supply chain is described as concentrated across a small number of firms and geographies. HBM is produced at scale by SK Hynix, Samsung and Micron, with SK Hynix holding the largest share. The memory then depends on TSMC’s advanced packaging capacity in Taiwan, while leading-edge chip production also depends on ASML’s lithography tools in the Netherlands. The episode emphasizes that these linked bottlenecks make the entire AI build-out dependent on a narrow industrial stack. Demand for HBM is said to be rising from three directions at once. NVIDIA’s Blackwell chips use more memory than previous generations, the next Rubin chips are expected to use even more, and hyperscaler custom silicon programs from Google, Amazon and Meta are adding further demand. Inference workloads are also growing faster than training, and the newer reasoning and long-context systems require more memory per query. The market backdrop is presented as a sharp but normal correction within a larger bull trend. Semiconductor ETFs and Korean equities had risen aggressively, then sold off on a mix of earnings, macro data, geopolitics and capital reallocation. The episode argues that the sell-off reflects crowded positioning rather than a change in the underlying AI thesis. The base case is that HBM remains structurally tight through 2027, even as supply slowly improves. Pricing power is expected to remain with suppliers, and strategic supply agreements are likely to become more common. The main risks are a slowdown in hyperscaler capital spending, tighter financial conditions, or geopolitical disruption around Taiwan or Korea. Brief Summary The AI supercycle is being defined by physical bottlenecks rather than software releases, with the episode focusing on high bandwidth memory as a strategic supply constraint. NVIDIA’s expanded partnership with SK Hynix is presented as evidence that memory capacity is being secured years ahead of demand, within a concentrated stack that also depends on TSMC packaging and ASML lithography. HBM demand is said to be rising from NVIDIA’s Blackwell and Rubin chips, as well as custom silicon programs from Google, Amazon and Meta, while inference workloads are also increasing memory intensity. The recent semiconductor and Korean equity pullback is framed as a correction within a broader bull trend, and the base case is that HBM remains tight through 2027 with pricing power staying with suppliers unless capital spending, financial conditions, or geopolitical risks disrupt the market. Transcript [0:10]Hello and welcome to the QFMI Market Intelligence Podcast, where we provide [0:17] HBM and the AI Supply Chain [0:14]independent market intelligence on the AI supercycle. Semiconductors, data centers, AI factories, energy and power, critical materials, and the capital flow shaping the biggest industrial build-out in a generation. Today is Friday the 12th of June, 2026. I'm Tim Hardwick, founder of QFMI, and in today's show, we're going to take a closer look at the semiconductor industry and the binding constraint of high bandwidth memory or HBM. [0:42]As you know, semis have been enjoying a gravity-defying bull market driven by the supply shortage of high bandwidth memory. However, this week we've seen things cool off a little bit across a large part of the sector, and perhaps, as mentioned last week, there has been a little bit too much exuberance in these markets that has warranted a correction. And that's perfectly normal, right? That's healthy, even in the strongest bull market. And the question being bounced around the financial media this week is, of course, is this a correction or is it a market top? Well, before we get into that discussion, there have been two very significant developments in the last few days that reinforce exactly why we are discussing memory today. And I think they tell a much bigger story than the headlines suggest. The first comes from South Korea, where NVIDIA CEO Jensen Huang spent last weekend meeting virtually every major industrial champion in the country. NVIDIA has announced a multi-year technology partnership with SK Hynix, committing to jointly develop the next generation of memory technologies for AI data centers and AI factories. And Jensen Huang made a statement that I think is extraordinary when you really think about it. He said that NVIDIA already purchases billions of dollars of memory from SK Hynix and that those purchases will grow substantially. Let's step back for a minute and think about what that means. [2:09]This isn't a procurement announcement. This is the world's most valuable semiconductor company effectively locking up future memory supply years in advance through a strategic partnership. Memory is no longer simply a component that can be ordered from a catalog. It is becoming a strategic resource with supply increasingly secure through long-term partnerships rather than spot purchases. And that is a fundamental shift in how the AI supply chain operates. But perhaps even more interesting is that NVIDIA's discussions in Korea went well beyond memory. SK Telecom announced plans to build a gigawatt-scale AI cloud using NVIDIA technology, with the first data center coming online in 2027. There were also collaborations with LG, Hyundai, Naver and Doosan spanning robotics, autonomous vehicles, power delivery and cooling. Jensen Huang even referred to Hyundai's planned data centre at Seiman Gum as an AI valley, comparing it directly to Silicon Valley, and I hope I got the pronunciation right of Seiman Gum. Now, I really found that comparison fascinating, because what we are really watching here is AI evolving from a software industry into a national industrial strategy. [3:31]South Korea is not just selling components into the AI supply chain. It is positioning itself as an integrated AI infrastructure economy. That is a very different thing. At almost exactly the same time, another story emerged that deserves just as much attention. Reuters reported that Alphabet had placed an order with Intel Foundry to manufacture more than 3 million Tensor processing units in 2028. And here is the part that really caught my attention. NVIDIA is also reportedly evaluating Intel's foundry capabilities for future processors. Now think about that for a second. For years, virtually all leading-edge AI chips had depended on Taiwan Semiconductor Manufacturing Company, or TSMC, in Taiwan. The entire AI build-out, every graphics processing unit or GPU, every custom accelerator, has flowed through a single manufacturing ecosystem. But the AI boomers stretch that ecosystem to its limits. [4:34]What we are now seeing is hyperscalers actively diversifying fabrication capacity. Not because TSMC is failing, far from it, but because demand has become so extraordinary that relying on a single manufacturing ecosystem creates unacceptable concentration risk. This is exactly how binding constraints evolve. Companies do not diversify suppliers because they have excess capacity. They diversify because capacity has become scarce. [5:04]And this is not just Google. Reuters noted that Intel had also landed Tesla as the first major customer for its next generation 14A manufacturing process for Elon Musk's TerraFab project. Apple is reportedly in discussions as well. The diversification away from TSMC concentration is becoming an industry-wide theme. Both of these stories reinforce one of the central themes of QFMI. AI is increasingly becoming an infrastructure economy constrained by physical assets rather than software innovation alone. What started as a technology cycle is evolving into something closer to a global industrial mobilization program. [5:54] Capital Becomes the Constraint [5:48]And that brings us to today's subject, the first of those physical constraints, high bandwidth memory. But first, I want to flag something else that's been developing rapidly and I think this deserves its own episode. It is the sheer scale of capital formation now building around the AI supercycle. Look at what has happened in just a matter of a few months. Alphabet raised $85 billion in equity. [6:14]OpenAI closed a $122 billion private round at an $852 billion valuation. The largest private financing on record Anthropic completed a $30 billion Series G And is reportedly discussing another round near a $900 billion valuation. [6:34]SpaceX goes public later today at a valuation of roughly $1.7 trillion in what will be the largest IPO in history. Add it up and you're looking at almost $400 billion in fresh equity and debt issued or teed up in a very short window, sitting alongside a hyperscaler capital expenditure wave of roughly $725 billion, for 2026. Now, the surface question is whether equity markets can absorb that kind of supply without affecting valuations But I think the real question is deeper than that. [7:13]Is AI entering the phase where the constraint is no longer compute or power, but capital itself? Money does not magically appear It must come from somewhere, right? Assets are sold to buy new assets Much of what we're seeing right now may not be new capital entering the markets at all. It may be rotation. Institutions selling defensives, selling utilities and staples and healthcare to buy AI infrastructure. Now if that's true, AI valuations rise while the rest of the market quietly derates. I should say though that the bullish case is actually very compelling. This is not speculative capital chasing meme stocks. [7:54]This is infrastructure financing, data centers, power generation, fiber, semiconductor capacity. These are productive assets that generate future cash flows. History suggests that transformational technologies, from railroads to electricity to cloud computing, always requires enormous amounts of upfront financing. And markets generally absorb issuance well when investors believe returns exceed the cost of capital. So the question I keep coming back to is not, can markets absorb 400 billion? It is, can the AI economy generate sufficient returns on the next 400 billion, then the next 700 billion, and eventually the next trillion? That is fundamentally a return on investment question, not a financing question. And if those returns disappoint, the constraint shifts from physical infrastructure to financial discipline. Equity issuance becomes more dilutive, debt becomes more expensive, hurdle rates rise, boards become more cautious, and AI CapEx slows. [8:56]From a QFMI perspective, I think we are watching the emergence of a fifth binding constraint, capital formation and capital absorption capacity. That transition from narrative-driven capital allocation to ROI-driven capital allocation could become one of the defining market themes of the next several years. It is, in my view, one of the most underappreciated risks to AI equity valuations today. We're going to dedicate a full episode to this topic because it deserves the space. For now, keep it in the back of your mind as context for everything else we discuss. And the first place where all of this capital hits a physical wall is memory. Memory is a constraint that the market is now paying attention to. Power is also a constraint the market is paying attention to. Critical materials are another constraint that we need to consider. [9:46]So this episode and the next two are each doing different analytical work on a different time horizon against a different consensus position. So today we're going to start with the constraint that has already moved. The question is not whether it is real, the market has answered that. The question is, what happens from here? [10:05] The Three HBM Producers [10:06]So let's start with who actually makes this stuff. There are three companies in the world that manufacture high bandwidth memory at scale. SK Hynix in South Korea, Samsung also in South Korea, and Micron in the United States. Between them, they make essentially all of it. SK Hynix is the dominant player, it holds around half the global HBR market, and as we covered at the top of the show, that dominance has just been reinforced by the multi-year strategic partnership with NVIDIA. HBM is no longer commodity product. It is a strategic industrial resource, with future supply secured years in advance. Samsung and Micron share most of the rest of the market, with the second and third positions moving around quarter to quarter depending on the qualification cycles. There is no fourth meaningful supplier. The Chinese memory industry is structurally a generation behind on this specific product, locked out by the equipment side, which we'll come back to in a moment. The dependency runs deeper than the market shares suggest. NVIDIA, the company most people associate with AI, sources roughly 90% of its HBM from SK Hynix alone. [11:19]That is one company, in one country, supplying the memory for the chips that are driving most of the AI capex in the world. But making the memory is only the first step. The memory has to be packaged together with the logic die that does the actual compute. And that packaging happens at one place, TSMC in Taiwan, on a process called Chip-on-Wafer-on-Substrate, or KOWOS. KOWOS is the second bottleneck. Without KOWOS capacity, the memory and the logic die cannot be assembled into a usable accelerator. TSMC has been expanding KOWOS aggressively, but the queue remains full well into 2027. And as we discussed earlier in the show, the fact that the hyperscalers are now actively diversifying fabrication away from TSMC tells you just how stretched that ecosystem has become. And then upstream of all that, every leading-edge chip in the stack relies on lithography machines, made by one company, ASML, in the Netherlands. The only company on Earth that builds extreme ultraviolet lithography systems. [12:27]Export controls on those machines are now part of US foreign policy. So we have memory in South Korea, packaging in Taiwan, lithography in the Netherlands, three countries, a handful of companies. The entire AI build-out runs through them. And what the last week has shown us is that the biggest companies in the world [12:52] Demand Is Accelerating [12:48]know it, and they are scrambling to secure their positions. The supply side is just one half of the story. The other half is what is happening to demand. So the demand side has three layers. All of these are accelerating. The first layer is the GPU. NVIDIA's Blackwell generation, now ramping, uses substantially more HBM per accelerator than the previous Hopper generation. The next generation, Rubin, is expected to push that consumption higher again. Each new chip is more memory-hungry than the last one. So even if the number of accelerators shipped were flat, HBM demand would still be rising. The number of accelerators shipped is not flat. The CapEx commitments from the major hyperscalers, Microsoft, Meta, Google, Amazon, are at levels we have never seen before. The four of them collectively are spending in the same range as small national infrastructure programs, and the spending is still going up. [13:49]The second layer is custom silicon, Google's TPU or Tensor Processing Unit, Amazon's Tranium, Meta's MTIA, the Meta Training and Inference Accelerator. Each of these is a chip designed by the hyperscaler for its own internal workloads, manufactured at TSMC, packaged with HBM and competing with NVIDIA for the same memory supply. The custom silicon ramp is genuine. It's not a marginal story at all. And it's adding to the HBM demand curve rather than replacing it. And this layer is expanding faster than most people realize. Google's reported order with Intel Foundry for more than 3 million TPUs in 2028 highlights just how aggressively hyperscalers are scaling their custom silicon programs. They are not just designing alternative chips. They are now diversifying manufacturing capacity to ensure those chips can actually be built. That tells you the demand curve is large enough and sustained enough that a single foundry ecosystem cannot serve it. The third layer is inference. Most of the conversation about AI compute has historically been about training. Training is the big expensive one-off cost of building the model. [15:04]Inference is the ongoing cost of running it. As enterprise AI adoption scales out, inference workloads are growing faster than training workloads. And inference, on the latest generation of models, is heavy on memory. The reasoning models, the agentic systems, the long-context workloads, all of them push HBM consumption per query. So the demand is coming from three directions at once. More memory per chip, more chips, more inference workloads on top of training. All of this is hitting a supply curve that, as we saw a moment ago, is constrained at three different points. The memory itself, the packaging, the lithography. HBM is sold out for the rest of this year. Visibility into 2027 is improving, but it's still tight. That is the supply and demand picture as it stands today. [15:57] Market Rally, Then Correction [15:57]With that framework in place, let's take a look at what the equity market is doing. [16:02]The semiconductor rally has been parabolic. There is no other word for it. SOXX, the main US semiconductor ETF, has rallied over 100% from its late March low to its early June high. SMH, another major semiconductor ETF, is up nearly 80% over the same period. [16:21]South Korea's main stock index, the COSPI, has gone close to vertical, with Samsung and SK Hynix between them now making up over 40% of the entire index. SK Hynix has overtaken Eli Lilly. Samsung has overtaken Walmart. Two Korean companies now sit in the global top 15 by market cap. And then this week, things cooled off sharply. The KOSPI closed down over 8%. Samsung fell more than 10%. SK Hynix dropped nearly 8% also. Now, the financial media tend to point to a single trigger, but this was really a confluence of factors hitting at the same time. We had the Broadcom earnings report resetting expectations We had the non-farm payrolls number coming in strong enough to raise the prospect of a Fed rate hike which repriced the entire rate-sensitive end of the market, We have escalating tensions in the Middle East with tit-for-tat strikes making the ceasefire Well, not much of a ceasefire And then of course we have SpaceX going public later today in what will be the largest IPO in history, When you are asking the market to absorb that much new equity supply in a single event, it's not unreasonable to expect capital reallocation across the broader complex in the days leading up to it. [17:38]But I think the triggers only explain the timing. They do not explain the velocity of the move. To understand that, you have to look at the conditions that were already in place underneath. The cost of financing is higher. Real yields are higher also. That changes the maths on every leveraged position in the market. We have extreme concentration risk, with a handful of names driving the indices. And we have seen an explosion of leveraged ETFs amplifying moves in both directions. Put all of this together and you have a crowded, leveraged, euphoric market that is highly sensitive to any negative catalyst, That combination is characteristic of mid-cycle or late-cycle bull markets, It does not mean the bull market is over It means the market has reached the phase where corrections become sharper and more violent than people expect How far does this one go? 8%? 10%? 15%? Nobody knows But the fact that this is happening is not a surprise and is not a reason to panic. Corrections in this kind of environment are normal and healthy. [18:44]The last point is worth pausing on. We published an In the Spotlight article this week titled The Gravity Trade, SpaceX IPO and Capital Rotation. And the question we asked was whether SpaceX would become a black hole for capital. When a single IPO is large enough to pull institutional flows away from other positions, the effect shows up elsewhere as selling pressure in the days before and after the listing, and that is what we may be seeing across the semiconductor complex this week, Capital flows matter and we'll be watching them very closely over the coming days and into next week, But the capital flows question is secondary to the bigger one, At the top of the show, I ask the question, is this a correction or is it a market top? [19:30] Correction or Top? [19:31]The PRISM framework that we use here at QFMI assigns a high probability, roughly 80%, that this is a mid-cycle correction within a structural bull market rather than a market top. And the reasoning for this is straightforward. A market top typically requires a change in the fundamental story. Either demand is rolling over, or supply is caught up, or the macro backdrop has shifted enough to reprice the entire complex. None of those conditions are present today. [20:02]HBM is sold out for the rest of the year. Hyperscalar CapEx guidance is still rising. NVIDIA just signed a multi-year supply partnership that tells you its own demand visibility is extending, not contracting. What I do see is a trade that got ahead of itself. The semiconductor complex moved so far so fast that a confluence of catalysts, any one of which might have been absorbed in a less extended market, was enough to trigger some pretty rapid profit-taking. And that is the hallmark of a crowded positioning on wind, not a fundamental repricing. The structural thesis remains intact. The short-term risk is elevated because the market front-loaded too much of the move. Those are two very different statements and it's important not to confuse them. The 20% probability I'm holding open for a market top is not trivial and I want to be specific about what would shift it. [20:58]If hyperscalar CAPEX guidance starts being revised downward, if the ROI narrative cracks, or if tighter financial conditions begin to choke the capital formation pipeline, that we discussed earlier, then the correction versus top question gets revisited in a much more serious way and that 80-20 split moves toward 50-50 pretty quickly. But on the evidence available today, the framework reads this as a healthy reset within a structural bull market. What is particularly revealing about this week's price action is not the sell-off itself, but what happened inside it. This is not a uniform liquidation of the semiconductor complex. It is a bifurcated market. ASML, the only extreme ultraviolet lithography supplier on earth, made all-time highs this week, up nearly 10% in a single session. Micron rallied nearly 12%. Meanwhile, NVIDIA, the name most associated with the AI trade, is down over 13% from its highs and barely participating in the rebound. [22:00]Think about what that pattern is telling us. If the market was losing faith in the AI thesis, everything would be selling off. But it isn't. The structural scarcity names are being bought aggressively. What the market appears to be repricing is not the AI supercycle itself, but the most crowded expressions of it. And that is a positioning story, not a fundamental story. It suggests to me the correction is concentrated, where the leverage and the retail euphoria were most heavily stacked. The broader semiconductor ETFs are oscillating rapidly between sell-offs and sharp rebounds on individual headlines. That kind of two-way volatility is exactly what our statistical framework flags as characteristic of a market in mid-cycle correction mode. The trend is still up, but the momentum has tapered off. In fact, it's broken, and the market has not yet found its footing. Whether the rebound we saw yesterday on the Iran news holds or whether it proves to be a dead cat bounce will depend on what comes next. The SpaceX IPO will be a significant test of capital absorption, so we're watching that very closely. So this is the price action picture. Now the question is how we put fundamentals and technicals together in a forward view. [23:16] The HBM Base Case [23:16]So here's our base case, and this is the part of the episode where the framework commits to a reading. So my base case on the HBM constraint over the next 12 to 18 months is that the constraint persists, but gradually eases at the margin. It doesn't break, it doesn't have to resolve, but it shifts from acute shortage to structural tightness. On the supply side, the trajectory is one of significant but insufficient expansion. Samsung is working to qualify HBM4 and close the gap with SK Hynix Micron is adding capacity TSMC is scaling KOWO's packaging, All of that adds supply But the ramp curves on advanced memory and advanced packaging are measured in quarters, not weeks You cannot fast-track the physics of chip stacking and thermal management, right? Supply will grow, but it will grow on its own schedule. [24:08]Demand tells a different story, but equally an important story. Blackwell is ramping. Rubin is coming. Hyperscalar custom silicon programs are expanding aggressively. As the Google Intel Foundry orders have underscored, right? And inference workloads are scaling in ways that most investors are still underestimating. Every new reasoning model, every agentic system, every long-context application pushes HBM consumption per query higher. The demand curve is not slowing, if anything, it's steepening. [24:40]So when I put these two sides together, my read is that supply growth will be meaningful but demand growth is likely to outpace it through 2027, The result is persistent structural tightness rather than acute shortage, Pricing power should remain with suppliers Capacity reservations will continue years in advance, Strategic partnerships like the NVIDIA-SK Hynix agreement announced this week will become increasingly common as the industry's largest buyers seek to guarantee supply in a constrained market. The equity layer reflects much of this already, but the market has not yet fully priced the duration of the constraint. There is a difference between pricing a shortage and pricing a multi-year structural regime of tightness. I think we're still in the early stages of the latter. Now, the principal risk to this view would be a material slowdown in AI infrastructure spending. Driven by deteriorating return on investment or tighter financial conditions, reducing hyperscaler capex and easing demand for HBM more quickly than we'd expect. That brings us back to the capital formation question we discussed earlier in this episode. If the AI economy cannot demonstrate sufficient returns on capital being deployed, the demand curve shifts down and the constraint eases from the demand side rather than the supply side. [26:02]Conversely, geopolitical disruption around Taiwan or Korea could tighten the market even further and turn structural tightness into acute crisis. [26:13]Okay, so the indicators that I'm watching closely are hyperscaler CAPEX guidance, KOWO's expansion timelines, Samsung qualification progress, also micron capacity additions, and inference growth rates. But I'm also watching the macro layer very carefully. Real rates matter enormously for the sector trading at these multiples. The FedWatch tool is currently pricing in a rate hike for December of this year with odds at approximately 68%. That tells you the market expects tighter conditions ahead, and higher real rates raise the cost of capital across every leverish position in the complex. We also have a new Fed chair in Kevin Walsh, and how he navigates this environment is something we need to keep a close eye on as well. [26:58] Watching IPOs and Capital Flows [26:58]And then there is the capital markets calendar. In the near term, all eyes are on SpaceX, and the sheer scale of that listing makes it the single most important test of capital absorption this year. Later in the year, we have Anthropic and OpenAI expected to come to market as well. Beyond the IPOs, we need to watch follow-on offerings and critically, the expiry of lock-up agreements over the coming months. [27:23]When you've had a run like this, a wave of insider selling as lockups expire can create sustained supply pressure, and the market has to absorb on top of everything else. How the market digests all of that will tell us a great deal about the underlying demand for semiconductor equity, and for AI equity more broadly, and critically, evidence of enterprise AI monetization. That is the thread that ties everything together. Those metrics will determine whether the memory constraint begins to ease or remains one of the defining bottlenecks of the AI supercycle. So this is the base case as it reads today. As new evidence arrives, we'll revise that. So that is it for the memory constraint. A Korean company that almost nobody had heard of 18 months ago is now worth a trillion dollars. Its CEO has just signed a multi-year strategic partnership with the most valuable semiconductor company on earth. The supply chain runs through three countries and a handful of firms. Demand is accelerating from three directions at once. And the base case is that this constraint persists as a structural regime of tightness through 2027 and beyond, with the market still in the early stages of pricing its duration. [28:36]Okay, so you can find all of our work at qfmi.substack.com. The Market Pulse and In the Spotlight articles are free and always will be. The weekly outlook, the weekend debrief and the strategic research reports sit behind a paid subscription. Subscribe and you get the full picture. Next week in episode three, we'll cover the latest market developments and we'll take a closer look at the next binding constraint for the AI supercycle, delivered power. Why the grid is the bottleneck, the market is still underpricing and how the live energy crisis is reshaping the maths. Thank you for listening to the QFMI Market Intelligence Podcast. See you next week.