Phil Rosen Sep 17, 2026 33m 21m saved
With Nick, product manager at VanEck and part of the team behind SMH, the largest US-listed semiconductor ETF
The bottleneck in the AI build-out moved. Graphics chips were the constraint two years ago; after TSMC doubled its packaging capacity, the money went to memory, and a new memory fab takes years rather than quarters.
That timing problem is also the reason the VanEck product manager who helps run SMH does not expect a fiber-optic-style overbuild. A company cannot pour concrete faster than concrete sets, and the physical limit acts as a brake on how far ahead of demand anyone can get.
"There's not a GPU that's not lit up like a Christmas tree."
Nick works on VanEck's semiconductor funds, including SMH, which launched in 2011 and now holds more than $70 billion. He spoke to Phil Rosen about what he is watching at each layer of the supply chain and what he would buy outside it.
The full interview is covered here so you can skip it. 33 minutes of audio, 12 minutes of reading.
Here are the 14 insights that matter.
Key Takeaways
Memory replaced graphics chips as the primary constraint in the last six to eight months
The dark-fiber analogy fails because old H100s still fetch a price, three generations on
Each layer of the chip stack is an oligopoly, so the layers win together rather than compete
Overregulation in the US is the bear case he names first, because China will not slow down
SMH went from an institutional trading vehicle to majority retail
Humanoid robots are 10 or 12 years from broad commercial adoption; commercial robots already work
He expects chip-financing debt to be sliced into products, ETFs included
Proprietary data is what survives vibe-coding, which is why he still rates Salesforce
1. The Biggest Semi ETF
SMH has been running since 2011 and is the largest US-listed semiconductor ETF. Nick described a market where the ordinary rule of a cyclical industry has stopped applying, because demand keeps arriving faster than fabs can be built.
"It's been around since 2011. It's the largest US listed semiconductor ETF, right, over 70 billion dollars in AUM today." — Nick
"So it's a very interesting time where demand for semiconductors has exceedingly outpaced supply for them." — Nick
2. The Bear Case
Asked to make the strongest case against his own asset class, he named two things. Regulation is the first, and he framed it as a competitive rather than a moral question: restricting US developers does not restrict Chinese ones. The second is that model efficiency improves faster than the infrastructure being built to serve it, which he considers unlikely at this stage because agentic systems consume far more tokens than chat did.
"Specifically in the US, I think if we overregulated that could be very problematic from a US perspective, right?" — Nick
"Because China is not going to slow down." — Nick
"And these data centers are essentially token factories." — Nick
"And we are burning more tokens than we even understand." — Nick
3. Institutional to Retail
The fund's holder base inverted. For most of its life SMH was a way for institutions to get liquid exposure to a handful of very large names; the ChatGPT moment in 2022 and the six months after it changed who was buying. Hundreds of millions of dollars still move through it daily, but a growing share of the assets now sit still.
"It's been a whirlwind." — Nick
"No, I would say it's primarily retail I would guess now." — Nick
"We've grown our retail client base for SMH tremendously over the last few years." — Nick
4. Layers of the Chip Stack
Rosen put the common assumption to him, that the semiconductor names inside SMH are competing with one another. Nick broke the industry into five layers instead: fabless designers such as Nvidia, Broadcom and Qualcomm; the foundry, TSMC, which manufactures and designs nothing of its own; equipment, where ASML is the only maker of EUV lithography machines; intellectual property, where ARM licenses architectures; and software, where Cadence and Synopsys sit.
"So you have fabless companies. These are like the Nvidias of the world. They design the chips." — Nick
"So there's only one ASML in the entire world that makes these EUV lithography machines that everybody in the Western world uses, well, I guess in Taiwan, to manufacture these chips, right?" — Nick
He said that is what stands between China and parity, since the machines available there work at larger nanometer sizes. The commercial consequence of the structure is that a win anywhere flows through everywhere: Nvidia and Broadcom build different products, TSMC makes both, and TSMC buys only from ASML.
"And now, because the demand is so far outpacing the supply, and there's really only oligopolies at every layer of that stack, they're all winning together, right?" — Nick
5. Cyclicality, Revisited
Nick did not claim the cycle is dead. His argument is that the old cycle was consumer-product-driven, with chipmakers racing at one new device and crashing prices when demand fell away, and that the buyer today is a handful of very large companies with multi-year capital commitments. He also allowed that investors are already rotating away from the most expensive name in the group.
"They will continue to be cyclical." — Nick
"I think people are going to not want to pay, they don't want to pay up for Nvidia right now, they would rather pay for, you know, a Micron or, SK Hynix." — Nick
"So it's a very different market dynamic than it used to be." — Nick
6. Not the Dark Fiber Story
The comparison investors reach for is the telecom build-out of the late 1990s, when fiber was laid and left unused. Nick's rebuttal is utilization: chips three generations old still command a price for their compute, because smaller models run on them.
"You look back at the 2000s, they called it dark fiber, because it wasn't being used." — Nick
"There's not a GPU that's not lit up like a Christmas tree." — Nick
"They can't build them fast enough because of the use." — Nick
He also said the demand side is not comparable. The internet had to be populated before anyone could use it; the users of AI tools already exist and are already online.
"We went from not having the internet to having the internet, and nothing was online, right, so you had to bring everything online." — Nick
7. Fabless and Capex Light
VanEck runs a second fund, SMHX, holding only the fabless layer plus the IP and software companies. Nick said the reason to separate it is that the early performance of SMH came disproportionately from Nvidia, and Nvidia's pivot from gaming graphics was possible because it owned no factories. Newer designers such as Groq and Cerebras fit the same shape.
"So fabless means that you do not manufacture your own chips, simply, you're designing them." — Nick
"And it was a lot of the attribution of performance for SMH." — Nick
8. Design Over Moore's Law
Asked whether he would rather own asset-light or asset-heavy businesses over the next three to five years, he chose asset-light, and gave a technical reason rather than a financial one. Transistor shrinkage has run into physics, so the next gains have to come from how a chip is laid out. Apple designing its own silicon is the example he gave of where that leads.
"I personally tend to lean towards the asset light." — Nick
"How do you get smaller than an atom, right?" — Nick
"In my mind, it's about the design, right?" — Nick
9. Where the Bottleneck Is
Every physical step in the chain takes years. A new TSMC fab in Arizona has taken two to three years; a data center takes years. The constraint moves as each one is relieved, and Nick traced where it sits now: graphics chips were short because of packaging, TSMC doubled its packaging capacity, and memory became the choke point.
"Standing up a data center takes years, right?" — Nick
"So now that they can push those out, now you're seeing memory as the primary bottleneck, and that's where, you know, a lot of the money has shifted over the last, say, six or eight months, because you have to stand up new foundries, or new fabs rather, to make that memory, right, and then it's a whole other process to take all these DRAM dies and put them together to make high bandwidth memory, and that's a whole other process." — Nick
He treats that slowness as protective rather than frustrating, because it stops capital being deployed faster than it can be absorbed.
"So it's almost like this inherent governor that we've got on it, that hopefully will help a giant bubble from being created, because we're so constrained on compute that it will help us inadvertently slow down and not get ahead of ourselves." — Nick
10. Data Centers, Robots, Space
Asked what he would put capital behind, Nick answered as a product designer looking for exposure that does not duplicate what VanEck already sells. Three areas came up: the data center supply chain, robotics, and space. On robotics he separated the near-term case from the one that gets the attention.
"So data center supply chain is one that we're very focused on." — Nick
"I think we're probably, humanoids are probably 10 or 12 years away realistically from being commercially adopted further for retail I would imagine, but commercial robots, right, have been used forever now, apply AI to commercial robots, that's a very different thing, so that's a very interesting space." — Nick
11. Grid to Rack
Pressed to say what a data center supply chain fund actually holds, he described the span from generation to the server rack: grid infrastructure, transformer makers, and nuclear as the leading answer for power sited next to the building. Chips are in it too, but capped, so a client who owns SMH as well is not doubled up on the same names.
"Yeah, it's grid to rack, right?" — Nick
"Small modular reactors, stand them up behind the meter, they can provide power to the data center and hopefully back to the communities that they're in." — Nick
12. Financing the Chips
The idea he volunteered as genuinely new is the debt side. Nvidia brokering a financing deal points at a market for chip-backed paper, and VanEck already runs collateralized loan obligation and private-lending products. He was careful that nothing exists yet.
"I mean, it's in its infancy, right?" — Nick
"But if it's not an ETF built around it, you can bet people are going to package those up, and they're going to slice them up" — Nick
13. Software's Real Moats
On the enterprise software selloff, Nick said some of it is deserved and most of the large names are not at risk. The dividing line he drew is proprietary data and entrenchment: a customer relationship system a company has built its operations around is not replaced because code became cheap to write.
"But are there wide-moat, you know, companies within software that have intangible assets that can't be replicated or recreated because you can vibe-code some software?" — Nick
"Like, I think Salesforce is a great example of that." — Nick
"But I think there's probably a lot of software companies out there that might have been slight grifters on what they're offering, or different things, right, that, you know, they may not make it through, because you could vibe-code their app, or they just don't have any proprietary data or anything proprietary underneath." — Nick
Rosen offered FactSet as the same argument in a single stock, a pick from his Best Ideas Club that has risen by double digits since publication, and raised the claim that Claude Code could reproduce a Bloomberg Terminal.
"No, impossible. Like, you cannot do that." — Nick, on rebuilding the Bloomberg Terminal, adding that its messaging network is part of the product
14. Asset Light Goes Heavy
Rosen's last line of questioning was about companies that were capital-light and are now building physical infrastructure. Nick is relaxed about the hyperscalers and less so about smaller operators changing business model, naming Bitcoin miners converting to AI compute as the interesting middle case.
"I just, like, these are very large companies that know what they're doing, like even if they don't have the expertise they have the ability to bring in expertise, right?" — Nick
"Like a lot of these Bitcoin miners have kind of added that to their business model." — Nick
Robotics is the apparent contradiction, since building machines is capital-heavy, and his answer was that the software layer is what changes the economics.
"I mean, I think in that case, like, AI makes robotics better, right?" — Nick
Bonus Insights
SK Hynix was excluded from SMH because the index requires US listings. Nick said the company has since issued an American depositary receipt and should qualify at the September review if its size holds
Rosen described a tool called Instinct AI that runs only inside iMessage, with no website, as an early example of a personal-assistant agent, and said he expects the category to become ordinary
Nick declined to confirm on air whether Marvell designs its own chips: "Sorry, don't quote me on that." Rosen said he likes the stock
Permitting is part of why data centers take so long, and Nick said the politics are local — "depending on where you live, they either hate or love data centers" — while noting that Elon Musk has been unusually fast by retrofitting existing buildings
On a company changing its identity to an AI business, both men reached for the same example and neither could place it. It was Allbirds, not TOMS
VanEck runs somewhere between 70 and 80 ETFs, with a heavy launch schedule over the past 12 to 18 months: "I think we're over 70, we're probably closer to 80, I would say, but we've really launched a bunch in the last year, so I've kind of lost track a little bit."
Asked how it feels to sit on the firm's largest fund, he said "Sometimes it's better to be lucky than smart."
Nick's bottom line is that the physical constraints of the chip industry are doing the work investors expect a bubble to do: every layer is sold out, the choke point keeps moving, and the time it takes to build anything is what stops capital running too far ahead of demand.
Products, Companies & Tools Mentioned
VanEck Semiconductor ETF (SMH) (Launched 2011, market-cap weighted, over $70 billion; the firm's largest fund and now majority retail)
VanEck Fabless Semiconductor ETF (SMHX) (The capex-light slice: designers, IP and software, separated because Nvidia's pivot showed what that model can do)
Nvidia, Broadcom and Qualcomm (The fabless designers; Broadcom builds application-specific chips, Nvidia builds general-purpose graphics processors)
TSMC (The foundry that manufactures for both and designs nothing of its own; doubled packaging capacity, which moved the bottleneck to memory)
ASML (Sole maker of EUV lithography machines, which Nick called the reason China has not caught up)
ARM, Cadence and Synopsys (The IP and design-software layers of the stack)
Micron and SK Hynix (Memory; where investors are rotating, and where the current constraint sits)
Groq and Cerebras (Newer designers Nick pointed to as evidence the fabless model keeps producing entrants)
Salesforce (His example of a software moat built on proprietary data and switching costs rather than code)
FactSet (Rosen's Best Ideas Club pick, up double digits since publication on the same data-moat logic)
Apple (Designs its own chips in house, which Nick used as the case for design as the source of future gains)
If this was worth your time, send it to someone who has to have a view on this.
Get the latest market chatter as it happens: