Nick Frasse, product manager for the VanEck Semiconductor ETF (SMH), told Full Signal host Phil Rosen that demand for AI chips is still outrunning supply, and that the physical time it takes to build fabs and data centers works against a 1999-style overbuild.
In the episode, Frasse also breaks down how the semiconductor supply chain fits together, why memory is now the key bottleneck and where else he sees opportunity in AI infrastructure.
I sat down with Frasse on Full Signal to pressure test the biggest bear arguments against the semiconductor trade.
What SMH is and why it matters
According to Frasse, SMH launched in 2011 and is the largest US-listed semiconductor ETF, with more than $70 billion in assets under management.
He said the fund began as an institutional trading vehicle because its holdings were so large and liquid, but that the AI boom has pulled in retail investors and financial advisors too. Frasse said semiconductors now sit in everything from toothbrushes and cars to the racks inside data centers.
The semiconductor stack is collaborative, not a free-for-all
Frasse argued that investors often misread the industry as one big fight between chip companies. He described distinct layers, starting with fabless designers such as Nvidia, Broadcom and Qualcomm that design chips but outsource manufacturing.
TSMC is a foundry that only manufactures, and ASML, which Frasse said is the only maker of the EUV lithography machines used in leading-edge production, supplies the equipment.
Arm provides intellectual property, and Cadence and Synopsys supply design software.
The result, he said, is that when one layer wins, the others tend to win with it. Nvidia and Broadcom make different kinds of chips, but TSMC manufactures both, and TSMC buys its equipment from ASML.
Frasse said that because there are "only oligopolies at every layer of that stack," they are "all winning together." He contrasted that with the older, product-driven boom and bust cycle.
He expects less cyclicality this time because the spending is coming from very large companies with long-term capital commitments.
He did not rule cycles out entirely and said he expects reversions to the mean.
Why Frasse does not see a 1999-style overbuild
I asked Frasse about the risk of an overbuild like the dark fiber era of the late 1990s. He rejected the comparison.
"There's not a GPU that's not lit up like a Christmas tree," Frasse said.
He added that earlier hardware is still in heavy use, noting that older Nvidia H100 chips still fetch a decent price for their compute.
He also pointed to physical limits. Frasse said it can take roughly three years to stand up a data center and two to three years to bring a new fab online, which he called an "inherent governor" that could keep a bubble from inflating too fast.
His two credible bear cases were over-regulation in the US while China keeps moving, and a scenario where models become efficient enough to need far fewer tokens, which would leave too much capacity built.
He added that he thinks the industry is still early, with agentic AI only beginning to drive token demand.
Memory is the new bottleneck
Frasse said the constraint has moved over time. GPUs were scarce first, partly because of packaging, until TSMC expanded its packaging capacity.
Now, he said, memory is the primary bottleneck, and over the last six to eight months a lot of investor money has shifted toward it. Building high-bandwidth memory requires new fabs for DRAM and then a separate process to stack the dies together.
Where else Frasse is looking
Frasse also discussed SMHX, VanEck's fabless semiconductor ETF, which targets capital-light chip designers. He said he personally leans toward asset-light businesses because of their flexibility.
Beyond chips, he highlighted the data center supply chain, which he described as everything from the power grid to the rack, including nuclear power.
He also named commercial robotics as an area of interest, while saying humanoid robots are likely 10 to 12 years from broad commercial adoption.
On the software sell-off, Frasse said some software companies will be disrupted by AI, but that businesses with proprietary data and entrenched customers, such as Salesforce, look much harder to replace.
Listen to the full episode
Tune in to the full conversation with Nick Frasse on YouTube, Spotify and Apple Podcasts.
For more on the AI buildout, read why the AI buildout won't bring another 1999-like overbuild, and subscribe to the free Opening Bell Daily newsletter for a daily briefing on stocks, macro and Wall Street.
Opening Bell Daily is for informational purposes only and is not investment advice. The views expressed are those of the guest. Frasse works for VanEck, which manages the funds discussed.