What’s Holding AI Back? The Companies Behind the Bottlenecks
What’s Holding AI Back? The Companies Behind the Bottlenecks
YouTube25 min 37 sec
Watch on YouTube
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights
  • TSMC (TSM) is a direct way to access AI chipmaking and advanced-packaging constraints; monitor whether new capacity and customer qualification keep pace with demand.
  • Broadcom (AVGO) offers exposure to custom AI accelerators and networking, with AI semiconductor revenue up 221% year over year to $16.7 billion in the quarter ended August.
  • Lumentum (LITE) has strong optical demand and reported quarterly revenue growth above 100%; treat the proposed Marvell (MRVL) acquisition as pending, with closing expected in the first half of 2028.
  • For diversified exposure, consider SMH (0.35% expense ratio) over the more concentrated bottleneck ETF described, which charges 0.75% and has nearly 70% in its top 10 holdings.
  • Keep memory and infrastructure positions sized for cyclical risk: capacity expansion, customer pricing power, and easing bottlenecks could weaken today’s supply-driven profits.
Detailed Analysis

SK Hynix (000660.KS)

  • SK Hynix is one of the three leading suppliers of high-bandwidth memory (HBM), which is packaged beside AI processors to move data quickly.
  • The transcript says its third-quarter 2026 operating margin was about 80%. Long-term agreements with 14 major customers cover more than half of its expected 2027–28 production, at prices within agreed ranges.
  • The company expects demand to exceed supply through 2030, but is accelerating capacity expansion and expects investment to rise significantly.

Takeaways

  • HBM scarcity and contracted sales support the bullish case, but high margins can attract new capacity and memory booms have historically been cyclical.
  • Watch whether demand, supply growth, and pricing remain favorable; strong current margins alone do not establish how long they can last.

Micron Technology (MU)

  • Micron is a major HBM, DRAM, and NAND supplier. Its first HBM4 samples went to customers in June 2025, with production ramping planned for 2026 alongside customers’ next AI platforms.
  • The transcript attributes Micron’s strong revenue growth to both DRAM and NAND, not HBM alone. It also notes that neither Micron nor SK Hynix reports HBM profit separately.

Takeaways

  • Micron offers exposure to several memory markets that may benefit from AI demand, but its company-wide results are not a clean measure of HBM profitability.
  • Consider memory-cycle risk: capacity expansion and changes in product mix can affect results even if AI demand remains strong.

Samsung Electronics (005930.KS)

  • Samsung is another major HBM supplier, but its memory business sits alongside phones, displays, and contract chipmaking.

Takeaways

  • Samsung provides broader technology exposure, but its results may be less directly tied to HBM scarcity than those of more concentrated memory suppliers.

Kioxia Holdings (285A.T)

  • Kioxia makes NAND flash, which is used for data storage rather than the high-speed, chip-adjacent memory role of HBM.
  • The transcript links NAND to AI because data centers need storage for training data and saved work.

Takeaways

  • Kioxia is an indirect AI infrastructure play through storage demand, not a direct HBM bottleneck exposure.

Coherent (COHR)

  • Coherent makes indium phosphide wafers, laser chips, and optical transceivers used to move data through data centers.
  • It is shifting from 3-inch to 6-inch wafers. The speaker said the larger wafers can yield more than four times as many chips at less than half the cost, but scaling production takes time.

Takeaways

  • Larger wafers could ease laser-supply constraints and improve costs, but successful scaling may also increase supply and reduce scarcity-related pricing power.

Lumentum (LITE)

  • Lumentum makes laser chips and optical transceivers. Its CEO said in October that the company could meet roughly 70% of demand for some products into 2027, while shortages for other products could continue through 2028.
  • The transcript reports that quarterly revenue more than doubled year over year to $1.26 billion, while gross margin rose from 34% to 51%.
  • Marvell agreed to acquire Lumentum, pairing its lasers and modules with Marvell’s silicon. The deal requires shareholder and regulatory approval and is expected to close in the first half of 2028.

Takeaways

  • Current demand and supply constraints are supportive, but the transcript highlights risks from new capacity and large cloud customers’ ability to negotiate prices.
  • The proposed acquisition could combine complementary products, but it is not yet completed.

Marvell Technology (MRVL)

  • Marvell designs chips used in optical modules and custom chips for large cloud companies. It agreed to buy Lumentum, a move intended to combine optical components with Marvell silicon.
  • The transcript notes that cloud companies design some networking components themselves and can split orders among suppliers.

Takeaways

  • Marvell has exposure to data-center networking and custom silicon, but customer bargaining power and in-house design efforts could limit suppliers’ pricing power.
  • The Lumentum deal remains subject to approvals and is expected to close in the first half of 2028.

Arista Networks (ANET)

  • Arista makes data-center switches that connect processors and optical links. The transcript describes networking as essential to keeping expensive processors from sitting idle while they wait for data.

Takeaways

  • Arista is exposed to AI data-center networking demand, but the transcript provides no company-specific outlook or valuation assessment. Treat the broader infrastructure opportunity as distinct from a conclusion about the stock’s price.

Eaton (ETN)

  • Eaton supplies switchgear, power distribution, and backup systems that help deliver electricity from the grid to data-center racks.
  • The transcript emphasizes that a site can face a power bottleneck even when the wider region has sufficient electricity, because utility connections and local equipment may not be ready.

Takeaways

  • Grid connections and power equipment are potential AI infrastructure bottlenecks. The key question is whether demand translates into sustained orders and profits for equipment suppliers.

GE Vernova (GEV)

  • GE Vernova makes gas turbines and grid equipment. The transcript says customers can reserve production slots before placing firm orders.
  • In the third quarter of 2026, reservations fell while customers converted reservations into firm orders; firm backlog rose, and the combined total increased.

Takeaways

  • The transcript presents the backlog conversion as a sign that orders are moving through the pipeline, but it does not establish how much of the company’s results comes from AI data centers.
  • Equipment demand is tied to broader power markets as well as data centers.

Vertiv (VRT)

  • Vertiv sells power equipment and thermal-management systems for high-density racks. The transcript notes that AI racks can generate more heat than air cooling can efficiently handle, increasing the need for liquid-cooling systems.

Takeaways

  • Power delivery and cooling are both necessary parts of data-center deployment. However, Vertiv’s company-wide results include business beyond AI, so headline results may not isolate this opportunity.

Bloom Energy (BE)

  • Bloom Energy makes on-site fuel cells that can generate electricity from natural gas or hydrogen. The transcript says this can help some data-center developers avoid waiting for a grid connection.
  • On-site generation requires a fuel supply and air permits. Natural-gas operation also involves recurring fuel costs and carbon emissions.

Takeaways

  • Bloom may benefit where grid delays make on-site power attractive, but fuel availability, permitting, operating costs, and emissions are material constraints described in the discussion.

ASML (ASML)

  • ASML is described as the only company with extreme ultraviolet (EUV) lithography technology used to print the finest features on leading-edge chips.
  • The transcript says ASML’s production is limited in part by Zeiss’s capacity to make optical components. Zeiss’s latest expansion took about four years from groundbreaking to its first new building.
  • ASML integrates the system with components from suppliers including Trumpf and Zeiss. The transcript also describes lengthy preparation and installation requirements for EUV equipment.

Takeaways

  • ASML’s unique technology and the complexity of expanding its supply chain support its strategic importance, while also illustrating that semiconductor capacity can take years to increase.
  • The transcript does not provide a valuation or stock-price assessment.

TSMC (TSM)

  • TSMC manufactures chips designed by other companies, including leading AI processors. Its advanced packaging combines processors and HBM stacks in a single package.
  • In July, TSMC’s chief executive said packaging capacity was so tight that it was limiting customers’ growth. New packaging capacity requires equipment, precision assembly, and customer qualification, so it can take quarters or years to arrive.

Takeaways

  • TSMC has exposure to both chip manufacturing and a packaging bottleneck. Monitor whether added capacity and customer qualifications keep pace with demand.
  • The transcript distinguishes production capability, capacity, yield, and customer qualification; progress in one does not ensure the others are sufficient.

Broadcom (AVGO)

  • Broadcom co-designs custom AI accelerators for cloud companies and also sells networking chips.
  • The transcript reports AI semiconductor revenue of $16.7 billion in the quarter ended in August, up 221% year over year. It says about 30% of revenue comes from infrastructure software, whose connection to AI is indirect.

Takeaways

  • Broadcom has direct exposure to custom AI chips and networking, but its overall business includes a significant software segment that is not as directly linked to AI infrastructure demand.

CoreWeave (CRWV)

  • CoreWeave rents computing capacity. It buys processors and supporting equipment, leases data centers, and charges customers to use the computing resources.
  • Its economics depend on keeping equipment busy. The transcript says the company finances much of its build-out with debt and relies heavily on its top customer.

Takeaways

  • CoreWeave can benefit when computing capacity is scarce and rental prices are high, but carries the risk that demand or utilization weakens while debt payments remain due.
  • Customer concentration and financing needs are key risks identified in the discussion.

Cerebras Systems (private)

  • Cerebras makes a wafer-scale processor designed for AI training and especially inference. Keeping more work on one piece of silicon may reduce some chip-to-chip traffic, although large deployments still connect multiple systems.
  • The transcript reports third-quarter revenue of about $211 million and $26.1 billion of signed work not yet delivered as of the end of September, with OpenAI among its important customers.
  • The transcript cautions that signed work is not yet revenue or cash; it must be delivered, and the contracts must hold. It also says Cerebras’s operating loss was slightly larger than revenue and that it depends on a small number of large customers.

Takeaways

  • The reported backlog is substantial, but execution, losses, and customer concentration make this a high-risk opportunity. Cerebras is identified here as a private company, not a publicly traded stock.

Oklo (OKLO)

  • Oklo is developing small nuclear plants that could supply data centers directly.
  • The transcript says the Nuclear Regulatory Commission approved one of Oklo’s reports on principal design criteria in May 2026, but that report is not an operating license. In June, NRC staff noted that Oklo had no licensing application before the agency for its Idaho plant, which is proceeding through a separate Department of Energy authorization process.
  • The company has no revenue yet, according to the transcript.

Takeaways

  • On-site nuclear power could offer a way around some grid constraints, but Oklo remains at a pre-revenue stage and still faces authorization and licensing steps.

Sivers Semiconductors (SIVE.ST)

  • Sivers makes indium phosphide laser chips designed for AI optical systems and is developing designs with GlobalFoundries.
  • The transcript warns that its opportunity pipeline represents possible business, not signed orders.

Takeaways

  • Sivers offers exposure to AI optics, but the distinction between a prospective pipeline and contracted sales is important when assessing the opportunity.

Everspin Technologies (MRAM)

  • Everspin sells magnetoresistive random-access memory (MRAM), which retains data without power, for industrial, medical, data-center, and other applications.
  • The transcript reports third-quarter revenue of about $21 million and says a larger AI role would likely come later.

Takeaways

  • Everspin is a smaller, more indirect AI memory opportunity; the discussion does not describe it as a current major AI bottleneck supplier.

NVIDIA (NVDA)

  • NVIDIA’s processors are central to the transcript’s AI data-center example. Its rack-scale systems link processors using high-speed connections and are designed around liquid-cooling manifolds.
  • An NVIDIA engineer is cited describing the response-generation phase of AI as “memory bound”: time can be spent moving model data from memory rather than doing arithmetic.

Takeaways

  • The discussion emphasizes that processor purchases alone do not ensure useful AI capacity; memory, networking, power, cooling, and packaging must also be available.
  • NVIDIA is not among the named holdings in the fund snapshot discussed, although the fund’s prospectus permits GPU makers that meet its criteria.

Advanced Micro Devices (AMD)

  • AMD is mentioned as a GPU designer that was not in the fund’s November holdings snapshot. The transcript says the fund’s prospectus permits GPU makers if they meet its requirements.
  • AMD is included among the holdings of the competing AIS fund discussed below.

Takeaways

  • The transcript provides no company-specific investment case or outlook for AMD; its relevance here is as an example of a GPU exposure that some AI funds include and the featured fund did not hold at that time.

Intel (INTC)

  • Intel is mentioned as an example of the complexity of installing EUV equipment: the transcript says Intel described its first EUV system in Ireland as having 100,000 parts, and that preparing the building took 18 months of design and construction.

Takeaways

  • The example illustrates how long and complex semiconductor-factory deployment can be, but the transcript does not provide a specific investment thesis for Intel.

AI-Bottleneck ETF (referred to as “NAC” in the transcript)

  • The transcript describes an actively managed fund holding 20 companies across memory, optics and networking, chipmaking and compute, power and cooling, and future bottlenecks.
  • Its November 10 snapshot allocated about 29% to memory, 24% to optics and networking, 21% to chipmaking and compute, 19% to power and cooling, and less than 8% to future bottlenecks. Its 10 largest holdings made up almost 70% of the fund.
  • More than a third of the portfolio was in companies headquartered in South Korea, Taiwan, Japan, the Netherlands, and Sweden; close to a quarter was in shares priced in won, yen, or krona. The fund’s exposure therefore includes regional and currency swings.
  • The transcript says the fund excludes GPU designers such as NVIDIA and AMD from the snapshot, as well as cloud companies and utilities. Under normal conditions, at least 80% is allocated to companies meeting its definition of AI-bottleneck companies, with up to 20% for future bottlenecks.
  • The expense ratio is 0.75% per year. The transcript also reports a median bid-ask spread of about 0.45% over the prior 30 days; that is a trading cost estimate, not an annual fee.

Takeaways

  • The fund offers concentrated exposure to the transcript’s bottleneck thesis, but many holdings depend on the same AI data-center customers and could fall together if that spending slows.
  • Active management may help the portfolio respond as bottlenecks shift, but it adds a fee and does not prevent the fund from being wrong about which constraints will persist.

VanEck Semiconductor ETF (SMH)

  • SMH tracks an index of 25 semiconductor companies. The transcript says NVIDIA made up about one-fifth of the fund at the end of October.
  • Its expense ratio is 0.35%.

Takeaways

  • SMH is a lower-fee, semiconductor-focused alternative, but it does not target the broader power, cooling, and data-center infrastructure bottlenecks in the same way as the featured fund.

Defiance AI & Power Infrastructure ETF (AIPO)

  • AIPO follows an index of AI and power infrastructure stocks. The transcript says its largest positions are grid builders and electrical-equipment companies.

Takeaways

  • AIPO offers a more infrastructure-oriented way to access the theme, with emphasis on power and grid equipment rather than the full range of memory, optics, and chipmaking bottlenecks.

VisaShare AI ETF (AIS)

  • AIS is actively managed and, according to the transcript, is also heavily weighted toward memory at the top of the portfolio.
  • It holds 64 companies, including AMD, and charges a 0.75% expense ratio.

Takeaways

  • AIS is broader than the 20-company bottleneck fund discussed, while charging the same stated annual fee. The transcript’s comparison suggests portfolio breadth and holdings—not just the fee—are important differences to examine.

AI Infrastructure Bottlenecks — Investment Theme

  • The transcript identifies four connected areas: memory, connections, power and cooling, and the factories that make and package the components.
  • A constraint may exist inside a working system, at a factory, or during deployment—for example, when equipment is ready but the site is waiting for a power connection.
  • The proposed investment logic is that limited supply may give suppliers bargaining power and improve profits. But each step can fail: supply may catch up, large customers may negotiate prices down, and strong expectations may already be reflected in share prices.

Takeaways

  • Assess each company by asking what the specific constraint is, why it is difficult to expand or replace, who has the bargaining power, and what could ease the shortage.
  • The transcript’s overall tone is cautiously constructive on AI infrastructure demand, but emphasizes that bottlenecks can move or disappear and that a strong industrial story does not automatically make a stock attractive at its current price.

Ask about this postAnswers are grounded in this post's content.
Video Description
Follow me on LinkedIn: https://www.linkedin.com/in/leo-cui-ph-d-cfa-59531439/ I am a venture capital investor at FoundersX Ventures, an AI and deep-tech VC fund based in Menlo Park, Cambridge and NYC. I previously did my Ph.D. at Harvard, postdoc at MIT. 📊 Want the infographics from this video? Join the channel membership for downloadable infographics, ad-free videos, early access, and the ability to propose research topics. Picture a company whose first racks of AI servers have just arrived. The company is made up. The problems aren't. Those racks are worth a small fortune, and right now they cannot do anything useful. The building still needs electricity delivered to the site, a way to carry the heat back out, and thousands of high-speed connections so the racks work together as one machine. And the next shipment depends on something farther upstream — how much high bandwidth memory and advanced packaging the industry can turn out. You can buy the processors. But what else has to come together before they do useful work, and who gets paid when those things are scarce? This is the whole dependency map, ring by ring — then an honest look at NECK, the actively managed fund organized around exactly that question. 🎯 You'll learn: • The three places a bottleneck shows up — inside a running system, at the factory, at deployment — and why fixing one just reveals the next • The chain of reasoning — scarcity → bargaining power → profit → shareholders — and every link that can break, including what NECK's own risk disclosures warn • Capacity vs bandwidth with a pantry door; DRAM, NAND and HBM; why answering a question is memory-bound; and what the KV cache does as conversations get longer • Why HBM needs roughly 3× the wafer area of standard server DRAM — so every HBM wafer is a wafer that didn't become ordinary memory • Stacking yields: twelve steps at 99% each, and why every die is tested before it goes into a stack • SK Hynix's ~80% operating margin, long-term agreements covering more than half of 2027–28 capacity, and why Micron's tripling revenue is not a pure HBM story • Scale-up vs scale-out, 5,000+ copper cables in one rack, why copper gives way to light, and what's inside an optical link • Indium phosphide, the tightest spot right now: Coherent moving from 3-inch to 6-inch wafers, and Lumentum's CEO saying it can't fill ~70% of demand for some products into 2027 • The grid queue: the 2025 FERC rulemaking on large loads, five-year median interconnection timelines, and ~128-week lead times for large transformers • Why nearly every watt becomes heat, and why water carries more than 3,000× as much heat as the same volume of air • EUV from molten tin to mirrors in a vacuum, why ASML's output is limited by ZEISS optics capacity, and the 18 months Intel spent preparing a building for one tool • How a bottleneck eases, with Cerebras and Oklo as examples — including precisely what Oklo's NRC milestone is and isn't • Capability vs capacity vs yield vs qualification at a foundry, CoWoS packaging, and TSMC saying packaging capacity was limiting customers' growth • NECK's actual November 10 holdings: 20 companies, the sleeve weights, ~70% in the top ten, and the currency exposure that comes with it • The 0.75% expense ratio, the spread, and how it compares with SMH, AIPO and AIS ⏱️ CHAPTERS 0:00 Racks That Can't Do Anything Yet 1:27 The Map, and What a Bottleneck Is 3:24 Memory: HBM, and Why It's Hard to Make 7:44 Connections: Copper, Light, and Lasers 11:55 Power, Cooling, and the Grid Queue 15:40 Factories, Packaging, and Routing Around It 21:32 The Fund: Holdings, Costs and Alternatives 🏢 Discussed: SK Hynix, Samsung, Micron, Kioxia, Coherent, Lumentum, Marvell, Arista, Eaton, GE Vernova, Vertiv, Bloom Energy, ASML, ZEISS, TRUMPF, Intel, TSMC, Broadcom, CoreWeave, Cerebras, Oklo, Sivers, Everspin ⚠️ Disclaimer: Paid promotion, commissioned by xETFs. For educational and informational purposes only — not financial advice, and not an offer to sell or a solicitation to buy any security. Companies are named as examples of how the map works, not as recommendations. NECK is actively managed, concentrated and non-diversified, and holds foreign securities, so it carries currency, regional and concentration risk; holdings are as of November 10 and will change. A bottleneck may prove temporary or shift elsewhere, and there is no assurance it produces pricing power or a higher share price. Consider the prospectus before investing, and do your own research. Some infographics are AI-generated and may contain errors; corrections pinned in the comments. 👍 If this was useful, subscribe and turn on the bell — I publish deep dives like this every week. #AI #ETF #NECK #Semiconductors #HBM #Memory #DataCenter #Optics #SKHynix #Micron #ASML #TSMC #Investing #AIInfrastructure
About Leo Cui, Ph.D., CFA
Leo Cui, Ph.D., CFA

Leo Cui, Ph.D., CFA

By @leo-cui-phd-cfa

I’m Leo Cui, General Partner at FoundersX, a Venture Capital firm focused on deep tech, AI, biotech, healthcare, semiconductors, robotics, quantum, fintech and other frontier technologies. Previously did my Ph.D. at Harvard, postdoc at MIT. You can also follow me on LinkedIn: https://www.linkedin.com/in/leo-cui-ph-d-cfa-59531439/ This channel is for founders, technical builders, scientists, engineers, and students who want to understand how startups are built and how investors think. I share practical advice on fundraising, pitch decks, venture capital, customer validation, market sizing, founder storytelling, and turning strong technology into a real business. I also discuss deep-tech trends, startup mistakes, and what makes a company fundable. LinkedIn: https://www.linkedin.com/in/leo-cui-ph-d-cfa-59531439/ X: @Leo_Cui_PhD_CFA For business inquires, please contact me at: partnerships@lyagroup.org .