Capital does not allocate evenly across an industry undergoing a structural break. It funnels. The companies that own the narrowest part of the funnel are the ones that capture the economic rent of the rotation, regardless of which user-facing applications eventually win.
This is the most important fact in equity investing during a paradigm shift, and it is consistently misunderstood by capital that allocates at the surface layer rather than the substrate. In the railroad mobilization of the 1860s, the dollars that flowed through the system enriched not the railroads themselves — most of which went bankrupt within a generation — but the firms that controlled the steel, the locomotive forging, the track-laying patents. In the dot-com buildout, the lasting wealth was created not by the websites but by the firms that sold them servers, routing infrastructure, and database licenses. In the smartphone rotation, Apple captured roughly forty percent of the total profit pool of the global handset industry through the device, but the hidden compounders — TSMC, ASML, ARM Holdings, Qualcomm — captured most of the remainder by owning the substrate.
The artificial intelligence rotation has the same shape. Of the $660-to-$690 billion of 2026 hyperscaler capital expenditure, most of it does not stay in the United States, does not stay in software, and does not stay in the application layer. It funnels through five specific physical chokepoints — pieces of equipment, patented processes, or geographic positions that cannot be replicated within the time horizon of this fund. The companies that own those chokepoints sit at the narrowest point of the funnel.
This Movement walks through the five. For each, it states the chokepoint, the company or oligopoly that owns it, the financial signature of that ownership in current filings, and the specific differentiated position it generates in the fund.
Every leading-edge artificial intelligence accelerator manufactured in 2026 — Nvidia's Blackwell and Rubin, Google's TPU v6, Apple's M-series, AMD's Instinct, Meta's MTIA, Amazon's Trainium — is etched at the atomic scale by a single piece of equipment manufactured by a single company in a single country. The machine is the extreme ultraviolet (EUV) lithography scanner. The company is ASML Holding ASML, headquartered in Veldhoven, the Netherlands. There is no second supplier. There is no domestic alternative. There is no Chinese alternative. There is, in the entire history of industrial civilization, no other firm that has solved the engineering problem of focusing 13.5-nanometer light through a multi-element mirror system precisely enough to print transistors smaller than a virus.
The total worldwide installed base of EUV scanners through 2027 will be approximately 66 machines. Each carries a price exceeding $200 million. Each requires roughly two years from order to delivery. ASML's management has guided that company production capacity will roughly double by 2027, but the ceiling on global frontier-node semiconductor capacity in any given year is set by the production schedule of one factory in Veldhoven, full stop.
The financial signature of this position is unambiguous. ASML's 2025 revenue was €32.7 billion. The company is guiding 2026 revenue of €36 to €40 billion, with gross margins of 52 to 53 percent — figures that historically belong to software companies, not heavy industrial equipment manufacturers. Management has publicly sized its addressable opportunity over the next several years at $2.5 trillion.
The next-generation High-NA EUV system (the TWINSCAN EXE:5200B), required for sub-2-nanometer production, is shipping in tiny quantities at a unit price approaching $400 million. Intel and TSMC are competing for the first units. ASML is the gatekeeper.
One layer beneath, the broader semiconductor wafer fabrication equipment market is itself an entrenched five-firm oligopoly. Applied Materials AMAT, Lam Research LRCX, Tokyo Electron 8035.T, and KLA KLAC — together with ASML — control approximately 70 percent of the global market. None can be displaced by capital. None can be replicated in the time horizon that matters. This is the foundation of every dollar of differentiated AI compute, and it is owned by five firms across three friendly jurisdictions.
The bill of materials of an Nvidia B200 Blackwell graphics processor — the workhorse of the 2026 frontier training run — is dominated not by the silicon logic die at its center, but by the high-bandwidth memory (HBM) stacks soldered around it. The logic die accounts for less than 15 percent of the unit's manufacturing cost, roughly $850. The HBM3e stacks account for approximately 45 percent — between $2,800 and $3,100 per accelerator. The economic center of gravity of the AI hardware stack has migrated, quietly, from logic to memory.
The reason is physical. Frontier large language models are not bottlenecked on theoretical compute. They are bottlenecked on the rate at which weights and activations can be moved between memory and the processor. The "memory wall" — a phrase semiconductor engineers have used for two decades to describe the diverging speeds of logic and memory — has become the operational ceiling on every AI training run on the planet. The faster the memory, the faster the model trains. The wider the bandwidth, the larger the context the model can hold.
The supply of high-bandwidth memory at the scale required by frontier AI is controlled by exactly three firms. SK Hynix 000660.KS, the dominant supplier, holds 62 to 69 percent of the high-end HBM market and supplies an estimated 90 percent of Nvidia's HBM consumption. Micron Technology MU holds 21 to 25 percent. Samsung Electronics 005930.KS, despite its enormous DRAM scale, has lost share to 13-to-17 percent due to repeated yield problems on the cutting-edge HBM nodes.
This three-firm structure is producing a financial signature that is, in the historical context of the memory industry, anomalous. Memory has historically been the most cyclical and lowest-margin business in semiconductors. In 2026, it is not.
The transition to HBM4, slated for production in 2026, is making the chokepoint tighter, not looser. The industry's required leap to a 16-layer stacked architecture inside a 775-micron envelope is forcing a transition from traditional thermocompression bonding (using micro-bumps) to direct copper-to-copper hybrid bonding. Early industry yield estimates on hybrid-bonded HBM4 are reportedly around 10 percent. Commercial viability requires roughly 60 percent. Samsung has already pushed its HBM4 mass-production timeline into deep 2026 because of these yield struggles.
The economic consequence is that HBM4 contract prices for 2026 orders are running 40 to 50 percent above HBM3e — a dynamic that hyperscalers, having secured GPU allocation but not memory allocation, are willing to pay rather than miss the training cycle. The chokepoint compounds.
Producing a leading-edge logic die and a stack of HBM is necessary but not sufficient. They must be physically integrated into a single accelerator package — connected by a silicon interposer routing terabytes per second of data between them. This step is called advanced packaging, and the dominant technology is Taiwan Semiconductor Manufacturing Company's Chip-on-Wafer-on-Substrate (CoWoS) process.
For roughly two years, the binding constraint on the global supply of frontier AI accelerators was not the printing of the silicon. It was TSMC's CoWoS capacity, which had to scale from approximately 35,000 wafers per month in late 2024 to a projected 130,000 by the end of 2026 to keep pace with hyperscaler orders. Even at that throughput, Nvidia is expected to consume 60-to-63 percent of all global CoWoS capacity in 2026 — nearly 595,000 wafers — primarily for its Rubin generation. Broadcom and AMD account for most of the remainder. Second-tier AI startups and ASIC designers are functionally locked out of the leading-edge packaging market.
TSMC's broader position is more concentrated than the packaging-specific story implies. The company holds greater than 90 percent of all global sub-7-nanometer fabrication capacity. Its 2-nanometer process — required for the next generation of frontier accelerators — is reportedly fully booked for the entirety of 2026. Wafer pricing for 3nm and 2nm is 25 to 50 percent higher than the 5nm generation, and TSMC has not faltered in passing the increase through to its customers, who have no alternative.
One layer beneath TSMC's CoWoS sits an even narrower bottleneck: Ajinomoto Build-up Film (ABF) substrates, a complex multi-layer resin film required to mount the silicon interposer onto the server's printed circuit board. The global ABF substrate market is held by five firms — Unimicron, Ibiden, AT&S, Nan Ya PCB, and Shinko Electric — that together command 74 percent of the market. As frontier AI packages grow physically larger to accommodate more compute and more HBM, ABF substrate yields decrease and global supply chronically lags demand. The chokepoint behind the chokepoint.
A new AI data center campus must source compute, memory, networking, and electrical infrastructure simultaneously. The compute supply chain has scaled aggressively. The electrical supply chain, governed by heavy industrial manufacturing and regulated grid interconnection, has not. The disparity between the two is the largest governing constraint on the rate of AI buildout from 2026 through 2030.
The single most significant structural shift in the AI infrastructure thesis between 2024 and 2026 is the migration of the binding constraint. For roughly two years, the limiting factor on AI deployment was silicon. From 2026 forward, it is electricity.
U.S. data center electricity demand is projected to rise from 176 terawatt-hours in 2023 to between 325 and 580 terawatt-hours by 2028 — figures that exceed the total annual electricity consumption of most developed nations. Morgan Stanley models a 49-gigawatt shortfall in available U.S. data center power against existing grid capability by 2028. The Goldman Sachs forecast for global data center electricity by 2030 has been revised upward by 160-to-165 percent in the past year.
The bottleneck is not electricity generation in the abstract sense. Generation can be built. The bottleneck is the heavy electrical equipment required to deliver that generation to the rack: large power transformers, generator step-up units, medium-voltage switchgear. These items are not mass-produced. Each is a specialized industrial unit requiring electrical-grade steel, copper wire windings, and intensive skilled labor. The United States imports 80 percent of its power transformer supply. Producer prices for transformers have risen 86 percent since 2019. Lead times have stretched from a historical baseline of 12 weeks to a current standard of 128 weeks for large units, with generator step-up units running 144 weeks.
The administrative bottleneck is worse than the manufacturing bottleneck. The median time required to study, approve, and complete the upstream transmission upgrades for a new large-scale grid interconnection in the United States is five to six years, with the queue-to-commercial-operation timeline now averaging 2,100 days. As of late 2024 there were approximately 10,300 projects waiting in U.S. interconnection queues, representing 1,400 gigawatts of capacity. Historically, only 13 percent of capacity that enters these queues actually reaches commercial operation.
The economic manifestation of this constraint is the most lucrative structural arbitrage of the past decade: the pivot of the publicly traded Bitcoin mining sector to AI hyperscaler hosting. Bitcoin miners possess one asset that hyperscalers cannot procure on any timeline that matters — energized, interconnected gigawatt-scale grid rights, secured during the cheap-power era of 2018 to 2022. By converting interruptible mining capacity into N+1 redundant high-density compute hosting, these firms bypass the five-to-ten-year interconnection queue entirely. The arbitrage is so compelling that hyperscalers are paying a 33× capital expenditure premium per megawatt to retrofit existing mining facilities for AI hosting, accepting the cost simply to acquire the time machine of an existing utility tie-in.
As of early 2026, more than $70 billion of cumulative AI and HPC contracts had been signed across the public Bitcoin mining sector. CoreWeave's $9 billion all-stock acquisition of Core Scientific CORZ; the $12.8 billion contracted HPC pipeline at TeraWulf WULF; the Google-anchored $7 billion 15-year lease at Hut 8 HUT; the rebrand of MARA Holdings MARA from a Bitcoin pure-play into an AI infrastructure platform — these are not isolated transactions. They are the public equity market repricing of secured grid interconnection as the most valuable real asset in digital infrastructure.
The infrastructure suppliers themselves carry the same financial signature. Vertiv Holdings VRT generated $10.2 billion in 2025 revenue, up 27.7 percent year-over-year, and entered 2026 with a $15 billion order backlog. The company holds the number-one global position in both data center thermal management and large three-phase uninterruptible power supply systems — both of which are required prerequisites for any hyperscaler GPU cluster. Bloom Energy BE has emerged as the dominant supplier of natural-gas-powered solid oxide fuel cells deployable in 90 days, against utility tie-in timelines exceeding five years. Constellation Energy CEG, Talen Energy TLN, and Vistra VST have signed more than 10 gigawatts of nuclear power purchase agreements with hyperscalers in the past eighteen months, repricing dispatchable baseload nuclear from a stranded asset to the most-coveted long-duration energy contract in the United States.
A frontier large language model is not trained on a single GPU. It is trained on tens of thousands of GPUs operating in continuous, deterministic synchronization across a single physical facility. If a single packet drops, if latency exceeds a threshold measured in microseconds, if any link in the network suffers jitter — the entire training run incurs cost without producing capability. The network, in a frontier AI training cluster, becomes the computer.
The result is an explosion in spending on a layer of the technology stack that was previously low-margin commodity hardware. Networking and optical interconnect revenue is currently growing faster than the underlying AI semiconductor market itself. The dominant economic positions cluster around four highly specialized choke points within this layer:
Switch silicon. Broadcom AVGO holds an estimated 80+ percent merchant market share in data center switching silicon. Its Tomahawk and Jericho product families form the literal backbone of the 800-gigabit and emerging 1.6-terabit Ethernet revolution that hyperscalers are deploying to escape Nvidia's proprietary InfiniBand network. Broadcom's AI-specific revenue ran $12.2 billion in 2024, up 220 percent year-over-year.
Custom silicon design. When a hyperscaler decides to build its own AI accelerator — Google's TPU, Meta's MTIA, Amazon's Trainium, Microsoft's Maia — they cannot do it alone. They partner with Broadcom or Marvell Technology MRVL. Together, these two firms control more than 80 percent of the custom AI ASIC design services market. The hyperscalers' attempt to bypass Nvidia simply migrates the chokepoint, rather than eliminating it.
Lasers. Lumentum Holdings LITE commands an estimated 50-to-60 percent market share in high-end electro-absorption modulated lasers (EMLs) — the physical components that generate the precise light pulses required for 800G and 1.6T optical transceivers. Lumentum is, by industry consensus, the only firm currently shipping 200-gigabit-per-lane EMLs in commercial volume. Coherent COHR serves the secondary position. The two firms together rate-limit the entire global optical-network buildout.
Active electrical cables. For shorter intra-rack connections, optical transceivers are prohibitively expensive and consume too much power. Active electrical cables — copper interconnects with embedded signal-conditioning silicon — fill this gap. Credo Technology CRDO holds an estimated 73 percent market share in this niche. A near-monopoly in a category most equity investors do not yet know exists.
Switch deployment. Arista Networks ANET sits downstream of Broadcom's silicon, building the integrated network switches that enterprises and hyperscalers actually deploy. Arista holds 19-to-21 percent of the overall data center Ethernet switch market and is growing faster than the broader segment, particularly in AI-cluster deployments where its Spine-Leaf architecture has become a de facto reference design.
The five chokepoints described above are not theoretical. Each has been operating, with the financial signature described, for at least four consecutive quarters as of the date of this document. ASML's gross margin is 52.8 percent. SK Hynix's operating margin is 72 percent. TSMC's CoWoS capacity is fully sold through 2026. Vertiv's order backlog is $15 billion. Lumentum's EML capacity is rate-limited and contractually sold forward. PJM capacity prices have hit the federal cap.
None of these facts is contested. None is a forecast. Each appears, as a hard line item, in the most recent ten audited financial filings of the firm in question or in the publicly published auction results of the regulator in question. The thesis of this Movement is not that these companies will become profitable. The thesis is that they already are, that their profitability reflects ownership of structural positions that cannot be replicated within the time horizon of this fund, and that public equity prices have not yet fully integrated the durability of those positions into the market capitalization of the firms that hold them.
The differentiated picks of this fund — at the substrate layer — emerge directly from this analysis. ASML for the lithography monopoly. SK Hynix and Micron for the memory oligopoly. TSMC for the foundry-and-packaging citadel. Vertiv for the thermal-and-power chokepoint. Constellation, Talen, and Vistra for secured baseload nuclear. The Bitcoin miner cohort for secured grid interconnection arbitrage. Lumentum and Coherent for the optical-laser duopoly. Credo for the AEC monopoly. Broadcom and Marvell for switch and ASIC duopolies. Arista for the deployment layer.
These names are the conservative core of the book — the positions where the rotation thesis pays off even if no specific application-layer winner emerges, because the dollars must flow through these firms en route to whatever does win. The aggressive expressions of the same thesis — the embodiment-layer companies in robotics, the launch and constellation owners in space, the long-call options structures that compound in the leveraged cases — are presented in Chapter V.
Chapter III is the proof that the thesis described here has been traded, with capital at risk, for forty months prior to this document's date. It is the personal-account record of Dean Gallagher, presented in full, with reconciliation to brokerage statements available under separate cover. It is what tells you whether the manager of this fund identified the rotation in time to position for it, or is articulating it now in retrospect.
The answer, to be plain about it, is the former.
The standard objection to a thesis as widely discussed as artificial-intelligence infrastructure is that, by 2026, it must already be priced into the equities of the firms that benefit. The chokepoint multiples below address that objection directly. They are not, in aggregate, extreme. In several cases — given the underlying growth rates and operating margins — they are surprisingly compressed.
| Chokepoint | Tk | Mkt Cap | Op Mgn | Fwd P/E | The reading |
|---|---|---|---|---|---|
| ASML — Lithography | ASML | $385.2B | 35.3% | 38.2× | Fair. Sole EUV supplier, 66 total scanner deliveries through 2027. The multiple already credits durability without yet crediting High-NA upside. |
| TSMC — Foundry & CoWoS | TSM | $1.80T | 53.3% | 21.0× | Compressed. 53% operating margin and the only foundry capable of leading-edge manufacturing trades at a market multiple. The geopolitical discount is the spread to capture. |
| Micron — HBM oligopoly | MU | $145.2B | 10.5% | 8.5× | Sharply compressed. Single-digit forward earnings while HBM4 contract prices are running 40-50% above HBM3e. The cycle skepticism in the multiple has not absorbed the structural memory shift. |
| Broadcom — Switch silicon & ASIC | AVGO | $1.92T | 45.2% | 36.9× | Fair. Custom AI ASIC dominance (Google TPU) plus 80%+ switch silicon share. The multiple credits the position; the durability of custom silicon is the spread. |
| Vertiv — Thermal & power | VRT | $38.5B | 16.2% | 38.5× | Premium. The only chokepoint name where the multiple already prices a meaningful share of the thesis. $15B backlog is real; the multiple credits it. Position sizing is more cautious here. |
| Nvidia — Frontier compute | NVDA | $5.06T | 61.1% | 24.9× | Compressed for the position. 61% operating margin and 90%+ training-workload share at a market-multiple forward P/E. The multiple expresses skepticism about durability; the chokepoint position is the spread. |
| Lumentum — EML photonics | LITE | $4.5B | N/A | 45.2× | Fair. The 200G/lane EML supplier; 50-60% market share. Currently unprofitable on a GAAP basis as the company invests through the optical-interconnect transition. |
| Applied Materials — WFE | AMAT | $175.4B | 29.5% | 22.4× | Compressed. Industry's broadest WFE portfolio with high exposure to packaging at a market-multiple forward P/E. The cyclical fade in the equipment cycle is over-discounted. |
The reading: ASML at 38× forward earnings, against a 35% operating margin and an effective monopoly on the most expensive piece of capital equipment in the manufactured world, is approximately fair. TSMC at 21× forward earnings, against a 53% operating margin and the only foundry capable of leading-edge manufacturing, is materially compressed. Micron at 8.5× forward earnings — at a moment when HBM4 contract prices are running 40-to-50% above the prior generation — is not a thesis that has been priced in. Vertiv at 39× forward earnings on a $15 billion backlog is the only chokepoint name where the current multiple already reflects much of the thesis, and the position sizing in Tier I reflects that.
The interpretation is not that every chokepoint is mispriced. It is that the spread between durable substrate-layer compounding and the multiple at which it currently trades has not yet closed. Capital allocated against the thesis at present multiples captures that spread. Capital allocated after the spread closes captures the consensus, which by then has been compressed back into the index. The thesis-vintage window described in Chapter IV is also visible here — in the multiples themselves.