Roundhill files for a CPU ETF as its memory fund sits at $27.1 billion

Memory is the franchise: $27.1 billion in the Roundhill Memory ETF CDRAM as of Wednesday, against $6.2 million in a capacitor fund listed September 9. On Wednesday Roundhill asked the SEC for another slice of the same trade, companies whose business is central processing units.

The CPU paper is a post-effective amendment to Roundhill ETF Trust, not a listing, and it arrives two days after the same trust put an AI power ETF in registration as APOW. Intel has returned 245% this year through Thursday's close, AMD 194%. The prospectus is labeled subject to completion, the fee table is blank, and the fund has not commenced operations. The facing sheet checks proposed effectiveness 75 days after filing under Rule 485(a)(2), the path that can make a new series effective unless the SEC comments or the issuer files again.

What the CPU prospectus proposes

The fund would be actively managed, would seek capital appreciation, and would not track an index. Under normal circumstances it would invest at least 80% of net assets, plus investment borrowings, in equities, depositary receipts, or instruments, including swaps and forwards, that provide exposure to CPU companies. The adviser would define those companies by a revenue test: at least 50% of revenue from designing, developing, manufacturing, selling, or licensing CPUs and integrated systems. Standalone microprocessor companies and system-on-chip companies both count. It would concentrate, meaning it would invest more than 25% of total assets, in an information-technology industry, and it would be classified as non-diversified.

The memory fund uses that 80/50 template. So does the Roundhill Neocloud ETF —NCLD.

In the June quarter, Intel Products, the client-computing and data-center segments, reported $15.1 billion of $16.1 billion in net revenue. Intel says that business consists substantially of CPUs and related semiconductor products. AMD does not publish a CPU share. Client revenue, mostly Ryzen processors, was $3.1 billion of $11.5 billion in the same quarter. Data Center, $6.7 billion, mixes EPYC CPUs with Instinct GPUs.

The memory fund that made the slice worth repeating

The memory fund listed on April 2. It is up 119% through Thursday's close, and Roundhill as a firm had $40.17 billion in assets as of Wednesday. The expense ratio is 0.65%.

That product is the argument Roundhill has been making in public. When the firm listed the neocloud fund and the Roundhill Photonics & Optics ETF —LYTE on August 6, chief executive Dave Mazza said, "CDRAM proved that investors want precise exposure to the AI infrastructure trade, not another broad tech fund." He described those two funds as extending the approach to "two critical layers of the AI buildout: the neoclouds providing compute and the photonics companies moving data at the speed of light." On September 9 it listed the Roundhill MLCC & Electronic Components ETF —CCML, aimed at multilayer ceramic capacitor makers.

The four listed funds, assets as of Wednesday:

Where a CPU fund would sit

Holders who want semiconductors already have large, cheap books that mix CPUs with GPUs, memory, and equipment. The VanEck Semiconductor ETF ASMH had $74.2 billion in assets as of Thursday and charges 0.35%. The iShares Semiconductor ETF ASOXX had $49.3 billion as of Wednesday and charges 0.33%.

Those weights are the point of a dedicated CPU product.

Defiance listed the Defiance Inference AI Chip ETF —AINF on August 17, tracking an index of companies tied to AI inference chips, including GPUs, ASICs, FPGAs, and AI-optimized CPUs. That is an inference-chip mandate, not a CPU mandate, and it had $505,000 in assets as of Wednesday.

The industry case for isolating CPUs has been coming from the companies, not from ETF issuers. In May, AMD raised its forecast for the server-CPU addressable market to growth of more than 35% a year through 2030, from 18%, and said a shift toward inference, where trained models are put to work, was opening demand for server CPUs.

Roundhill has shown, with the memory fund, that a narrow AI-infrastructure sleeve can attract real money. This document is the request to try that with CPUs. It is not the try.