Artificial intelligence is usually discussed in terms of models and compute. But the harder problem is physical. Before a model can be trained or served, someone has to build the facility that runs it, secure the power to feed it, and manufacture the memory, optics, and silicon inside it. Each of those can carry a lead time, sometimes measured in years.
The scale is easiest to see in electricity. On its second quarter 2026 earnings call, Eaton put the U.S. data center industry backlog at 307 gigawatts, roughly 15 years of construction at 2025 build rates.1 In April 2026, GE Vernova told investors it expects its power equipment and services backlog to reach $200 billion in 2027, a year earlier than it had previously indicated.2
Figures like those describe an industrial program rather than a product cycle. Capacity arrives when the slowest input catches up.
In any large construction program, output is governed by the scarcest component rather than the most abundant one. AI infrastructure works the same way, and that is where the investment question narrows. Capital spending is broad, but the leverage concentrates in specific places.
A bottleneck is a current demand and supply constraint that limits the development, deployment, or scaling of AI infrastructure; it reflects today’s shortage.
Memory. Training and serving large models requires moving enormous volumes of data between memory and processors. When memory cannot keep up, the processor waits. That has made high bandwidth memory one of the scarcest inputs in the buildout. In July 2026, SK Hynix CEO Kwak Noh-Jung said, “We still forecast that customer demand will remain higher than our supply capacity even beyond 2030.”3
Optics and networking. Data that leaves one chip has to reach the next. As AI scales, that traffic increasingly moves optically instead of across wires. Sales of optical transceivers and active optical cables were estimated at approximately $10 billion in the first quarter of 2026, more than 90% higher than a year earlier.4
Compute. Orders for AI silicon have been running ahead of what suppliers can deliver. On its second quarter fiscal 2026 earnings call, Broadcom reported more than $30 billion of AI semiconductor bookings against $10.8 billion shipped in the quarter.5
When a company holds a constraint, it shows up in the numbers: backlogs grow, lead times stretch, and prices rise. That makes future revenue easier to count on.
But that position isn't permanent. Constraints shift as new capacity comes online, substitutes emerge, and spending priorities change. A company that's supply-constrained today may not be tomorrow, so a portfolio built around these companies needs to be able to shift with them.
The xETFs AI Bottlenecks ETF (NECK) is an actively managed ETF that seeks to invest in companies we believe are positioned at critical points of constraint across the AI stack, all through a single ticker. Rather than focusing only on the biggest AI companies, NECK looks across the infrastructure required to build and scale AI for areas where demand may be growing faster than supply or where companies hold critical strategic positions.
The portfolio is expected to generally include between 15 to 25 companies globally.
As bottlenecks ease, migrate, or emerge across the AI stack, the Fund expects to periodically rebalance its investments to reflect changing market conditions.
Broad AI exposure tends to concentrate in the same handful of names. NECK is built around a different question: what companies supply the critical inputs that AI cannot scale without?
Learn more and view the prospectus at www.xETFs.com/NECK
Sources
1 Eaton Corporation plc management, Q2 2026 earnings call, July 31, 2026.
2 GE Vernova Inc., Q1 2026 earnings call, April 22, 2026.
3 SK hynix CEO Kwak Noh-Jung, quoted by Reuters, July 10, 2026. Company forecast; not a guarantee of future results.
4 LightCounting Research, Quarterly Market Update, June 2026. 5 Broadcom Inc., Q2 fiscal 2026 earnings call, June 3, 2026.