Roundhill Neocloud ETF ($NCLD): the first concentrated US bet on GPU-as-a-Service

Roundhill filed a registration statement on May 22, 2026 for a new thematic ETF - the Roundhill Neocloud ETF, ticker $NCLD. Concentrated holdings, single sleeve, built around one idea: the companies that buy AI accelerators by the tens of thousands and rent the compute back out. It's the first US-listed fund purpose-built around the neocloud layer of AI infrastructure, and the newest addition to Roundhill's thematic shelf next to $DRAM (memory) and $LYTE (photonics).

The filing isn't effective yet - no expense ratio, no NAV, no holdings disclosed - but the prospectus framing ("Neocloud Companies": GPU-as-a-Service platforms, high-density AI data centers, supporting power infrastructure, high-speed AI networking) tells you most of what the basket will look like. This piece walks through what $NCLD is, what a "neocloud" actually is, which listed names are likely in scope, the private capacity leaders you can't buy, and how the fund compares to the existing AI-infrastructure ETF universe.

What is the Roundhill Neocloud ETF (NCLD)?

$NCLD is a concentrated thematic fund. Roundhill's filing defines "Neocloud Companies" as businesses focused on GPU-as-a-Service platforms, high-density AI data centers, the power infrastructure that feeds them, and the high-speed networking that ties tens of thousands of GPUs together. That is a deliberately broad net - it spans the pure GPU-cloud operators, the datacenter developers hosting them, and the power/networking picks-and-shovels around them.

Roundhill's pattern is familiar: file a concentrated, single-theme sleeve early in a narrative, then let the basket track the obvious pure-plays. They did it with $CHAT (generative AI), $DRAM (memory), and $LYTE (photonics). $NCLD is the neocloud entry in the same "foundational AI layers" collection - and, as usual, Roundhill filed first.

Until the listing is effective there is no confirmed holdings list, no expense ratio, and no AUM. Everything below is scope inference from the filing language plus the investable universe, not a published portfolio.

What "neoclouds" actually are

A neocloud is an AI-native cloud: a company that buys NVIDIA (and increasingly AMD) accelerators at scale, racks them in high-density datacenters, and rents GPU-hours to AI labs and enterprises. The difference from a hyperscaler is the stack depth. AWS, Azure and Google Cloud run a full enterprise-software cloud - hundreds of managed services, global regions, decades of lock-in. A neocloud runs a thin, compute-first stack optimized for one thing: training and inference density at the lowest cost per GPU-hour.

That narrowness is the whole thesis. Neoclouds spin up GPU capacity faster and cheaper than a hyperscaler can, because they aren't carrying the overhead of a general-purpose cloud. When demand for accelerators outruns hyperscaler supply - which it has, repeatedly, through the 2024-2026 buildout - the neoclouds absorb the overflow. Microsoft itself has struck roughly $60B of capacity commitments across neoclouds (CoreWeave, Nebius, Nscale) rather than build it all in-house. When a hyperscaler rents from a neocloud, the structural picture is clear.

The bear framing writes itself too, and it matters: neoclouds are capital-intensive, debt-financed, single-customer-concentrated, and levered to the price of the exact chips they resell. More on that below.

Likely $NCLD holdings

The listed pure-plays split into two tiers.

Tier 1 - the clean GPU-cloud pure-plays:

• $CRWV (CoreWeave) - the flagship listed neocloud, the largest independent NVIDIA capacity outside the hyperscalers, and an anchor of the Microsoft commitments. Track it live at /stocks/crwv.
• $NBIS (Nebius) - European-rooted AI-native cloud (a Yandex spinout), full-stack GPU cloud plus its own datacenters, also inside the Microsoft deals.

Tier 2 - the miner-to-neocloud converts:

• $APLD (Applied Digital) - HPC/AI datacenter developer-operator leasing high-density capacity to neocloud and hyperscaler tenants.
• $IREN - a bitcoin miner converting owned power and sites into GPU cloud and AI hosting.
• $CORZ (Core Scientific) - HPC hosting anchored by a multi-year CoreWeave contract; the landlord to a listed neocloud.
• $WULF (TeraWulf) and $CIFR (Cipher Mining) - earlier-stage converts, more power-and-land optionality than full GPU-cloud stacks.

The full listed cluster - ranks, rationale, and how it moves together - lives on the Neoclouds theme.

The catch: the capacity leaders are private. Crusoe (~$10B, reportedly in talks near ~$30B), Nscale ($14.6B, March 2026), Together AI ($8.3B, July 2026), and Lambda (~$5.9B, IPO-track) are driving much of the actual buildout, and none of them are buyable in the public market yet. An ETF can only hold what's listed - so $NCLD, like every neocloud vehicle, is structurally over-weighted toward CoreWeave and Nebius and the miner converts, and under-exposed to the private leaders. That gap is the thing to watch: the first big neocloud IPO reshapes the basket.

How $NCLD compares to existing AI-infrastructure ETFs

There is no other pure-play neocloud ETF - that's the point. The nearest proxies are all compromises:

• WGMI (CoinShares Bitcoin Miners) holds the miner-converts (IREN, CORZ, WULF, APLD, CIFR) but drags in full bitcoin-price beta - you're buying crypto risk on top of neocloud risk.
• DTCR / SRVR (datacenter and digital-infra ETFs) hold the physical real estate and REITs, not the GPU-cloud operators.
• Broad AI ETFs (AIQ, IGPT, BOTZ) hold CRWV and NBIS at small weights buried in 40-80 lines.

$NCLD's pitch is that it strips all of that away: no bitcoin beta, no REIT dilution, no mega-cap ballast - just the neocloud cluster in one ticker. Whether concentration is a feature or a risk depends on your read of the theme.

What to track until $NCLD launches

• The effective date + holdings disclosure. The filing has no NAV, expense ratio, or portfolio yet. The day the listing prints, the basket weights tell you how much is CoreWeave/Nebius vs the miner-converts.
• The first neocloud IPO. Lambda is closest; Crusoe and Nscale are large enough to reshape the index. A listing pulls a private leader into the investable set and rebalances the whole theme.
• Hyperscaler capacity commitments. More Microsoft/Oracle/Amazon deals with neoclouds confirm the overflow thesis; a pullback would signal the hyperscalers are catching up on their own capacity.
• GPU pricing + utilization. Neocloud margins live and die on GPU-hour pricing and how full the fleet runs. Falling rental rates compress the whole cluster at once.

Risks and the bear case

The neocloud model is genuinely fragile in ways the memory or photonics themes are not.

Customer concentration. Several listed neoclouds lean on one or two anchor tenants. CoreWeave's Microsoft exposure is the canonical example - a single renegotiation moves the stock.

Capital intensity + leverage. Buying GPUs by the tens of thousands is debt-financed against the collateral value of depreciating chips. If accelerator prices fall or a faster generation obsoletes the installed base, the asset side of the balance sheet reprices fast.

The hyperscaler squeeze. The overflow thesis works because hyperscalers are supply-constrained. The day they aren't, the neocloud's structural reason to exist weakens.

Cyclicality dressed as secular. Much of the demand is training capex, which is lumpy. A pause in frontier-model training - or a shift toward cheaper inference - hits GPU-rental demand directly.

This is why $NCLD is a concentrated thematic bet, not a diversified holding. The theme is real; the individual names carry real balance-sheet and customer risk.

Where this fits in the QA taxonomy

Neoclouds sit at the seam of three of QuantAbundancia's bubbles: AI Compute Accelerators (the chips they resell), Datacenter Power (the electricity that gates their growth), and Hyperscalers / Cloud (both their competitors and their customers). Because the listed basket beta-drifts across all three, we track it as the Neoclouds theme rather than a 13th empirical bubble - the same call we made on robot foundation models.

For the broader map, the 12 AI bubbles ranked by empirical realness places compute and power inside the supercycle taxonomy, and the Datacenter Power bubble piece covers the electricity leg that ultimately caps how fast any neocloud can grow. To trade the listed names from a US-retail account, see /stack/ibkr. Bubble shifts and rule-based alerts on the cluster are part of /pro.

$NCLD is, in short, the cleanest single-vehicle expression of "I think GPU-as-a-Service is a durable layer, not a bubble" that the US market has offered. Whether that thesis is right is a separate question; that it's now ETF-wrappable is the structural fact, and Roundhill, as usual, filed first.

What to watch

• $NCLD effective date and first holdings disclosure (weights = CoreWeave/Nebius vs miner-converts)
• The first large neocloud IPO (Lambda / Crusoe / Nscale) - reshapes the basket
• Hyperscaler capacity deals with neoclouds vs in-house buildout
• GPU-hour pricing and fleet utilization - the margin signal for the whole cluster
• A shift from training capex to inference - would re-rate demand assumptions

Live data on this ETF: /etfs/ncld - issuer, expense ratio, holdings, and bubble mapping, updated the day the listing prints.

Theme context: /themes/neoclouds - the listed pure-plays and the private capacity leaders driving the buildout.

QuantAbundancia is educational research. Nothing here is investment advice. See /disclosures.