There is a phrase circulating with increasing frequency in conversations about AI infrastructure:

The bottleneck.

Compute—specifically graphics processing units (GPUs)—was the dominant variable in AI progress during the early years of AI, but it has become abundant enough that new constraints have moved to the foreground. The limitation today is less often whether a model can run and more often whether:

• Data can move fast enough to feed it
• Memory can store and retrieve at the speed the model demands
• The physical layer connecting servers across a data center can handle the bandwidth that frontier AI workloads require

These are engineering problems masquerading as investment themes, and identifying which strategies have oriented themselves toward the genuine constraints, as opposed to the general excitement around AI, is now one of the more meaningful distinctions in thematic investing.

Two bottlenecks dominate the current conversation:

2. Memory bandwidth and capacity

High-bandwidth memory, or HBM, has become the critical substrate for GPU performance. When Nvidia ships an H100, a Blackwell or a Rubin system, the memory that determines how fast it can operate typically comes from SK Hynix, Micron or Samsung. Storage more broadly, like the NAND flash from the companies that became SanDisk and Kioxia, feeds the data pipelines that AI training and inference consume continuously.

2. Optical networking

This represents the fiber, transceivers, and switching equipment that move data between GPUs at the speeds AI workloads require. Companies like Lumentum, Coherent, Corning, and Credo Technology are building the connective tissue of the AI data center, and the demand signal coming from hyperscalers has been among the strongest in the hardware ecosystem over the past two years.

Software?

Software, by contrast, is not a bottleneck in this sense. It is an opportunity, a competitive arena, and it could be a platform for monetization, but it is not the scarce resource. The observation is not a slight against software companies; it is simply a structural point about where AI's scaling constraints actually live in 2026.

Against this backdrop, three AI-oriented ETF strategies offer instructive contrasts.

• The Global X Artificial Intelligence & Technology ETF (AIQ) and the iShares A.I. Innovation and Tech Active ETF (BAI) are among the largest by assets under management, as of May 29, 2026. Both were above $10 billion, and BAI was actually above $16 billion.¹
• The WisdomTree Artificial Intelligence and Innovation Fund (WTAI) is a strategy where we are seeking to directly think about the bottlenecks when we go through our quarterly rebalances of the WisdomTree Artificial Intelligence and Innovation Index. The Index Committee also evaluates four focal categories of exposure, namely Semiconductors, Other Hardware, Software and Innovation.

WTAI has positioned itself most explicitly around infrastructure and bottleneck exposure. As of May 29, 2026, major exposures included Micron (4.70%), Samsung Electronics (4.55%), Nvidia (4.23%), Broadcom (3.36%), SanDisk (3.19%), and Kioxia (3.18%). Together, the memory complex alone represents a substantial plurality of assets. The optical networking thesis also appears with names like Lumentum (2.02%), Coherent (1.69%), Corning (1.23%), Credo Technology (1.18%), and Ciena (0.63%), each of which derives meaningful revenue from AI data center buildout. The pattern across the fund is consistent, placing exposure in companies that supply the physical and semiconductor infrastructure that AI requires to actually function, rather than companies that build applications on top of it. Software and consumer-facing platforms are notably slimmed down in exposure to only the highest conviction names.

BAI, an actively managed strategy, takes an approach that recognizes the bottleneck thesis but holds it alongside a broader AI growth bet. SK Hynix led the fund at 7.4%, followed by Broadcom (5.48%), Micron (5.40%) and AMD (5.03%). The optical networking layer appeared through Lumentum (2.69%), Fabrinet (2.05%), and Credo (1.83%). BAI also carried allocations to Snowflake (1.64%) and Palantir (1.10%). Being actively managed, BAI appears designed to capture both the infrastructure constraint narrative and the software monetization narrative simultaneously, accepting the tradeoff that might come with that breadth.

AIQ, as of this moment in time, had taken the widest view. First, it is tracking the total return performance, before fees and expenses, of the Indxx Artificial Intelligence & Big Data Index, which selects companies from developed markets, plus China-domiciled ADRs/GDRs,² that demonstrate direct, meaningful exposure to AI. Candidates are sorted into two buckets:

• AI developers and application companies (Category 1, up to 60 names)
• AI hardware and quantum computing companies (Category 2, up to 25 names)

Each company receives a proprietary "Exposure Score" based on how central AI is to its business, and only those clearing a positive score threshold are eligible. Weighting is modified market-cap with individual security caps of 3% (or 1% for lower-scoring names). The index is rules-based with no discretionary committee cited in the methodology.

SK Hynix (7.12%) and Micron (5.76%) were the largest holdings by weight, and Samsung Electronics (4.79%) was also significant. Alphabet (2.72%), Meta Platforms (2.30%), Microsoft (2.35%), Netflix (2.43%), Amazon (2.73%), Apple (2.97%), Salesforce (1.31%), Tesla (2.29%) and Palantir (2.23%) also constitute a substantial portion of the portfolio, and they represent something different from what we might think of as ‘bottlenecks’. Fujikura (0.39%), the Japanese optical fiber manufacturer, and Marvell Technology (1.27%) did appear as less substantial positions. Of the three strategies in this piece, AIQ is effectively a broad technology strategy with an AI narrative, which may serve investors seeking wide exposure to how AI reshapes the economy, but it is a different bet than the infrastructure thesis.

The distinction matters because the bottleneck thesis and the AI adoption thesis can diverge sharply in performance over specific periods. When the market focuses on AI capex cycles and infrastructure spending, memory and optical names tend to lead. When attention shifts to monetization, software multiples could expand. Knowing which thesis a given fund is actually expressing is the starting point for any honest evaluation.

Conclusion: Who is Leading in Terms of Performance?

To be fair, we have to note that performance in this space can be volatile. We are looking at a snapshot as of May 29, 2026. The future is unknown. What’s clear, particularly in the year-to-date 2026 period, is that WTAI and BAI are dramatically ahead of AIQ. We noted above their focus on what we term ‘the bottlenecks’. AIQ doesn’t ignore the bottlenecks by any means; it simply distributes its weight across a broader set of companies.

It is a short-term period in Figure 1a due to the short, live history of BAI, which only began October 21, 2024.

Figure 1a: The Performance Picture So Far

Figure 1b: Standardized Performance

Our bottom line is this—if investors are looking for AI strategies that are ‘minding the bottlenecks,’ we believe the evidence is clear, as of this point in time, regarding where those exposures may be found. If people are focusing on other areas, that is also clear.

Figure 2: Additional Information

Sources: WisdomTree, iShares and Global X, with data from each respective fund page as of May 29, 2026. Subject to change.

¹ Sources: iShares and Global X, specific BAI and AIQ fund pages. Assets under management as of May 29, 2026. Subject to change.

² ADR stands for American Depository Receipt. GDR stands for Global Depository Receipt.