The Bottleneck Economy

Large-scale AI is best analyzed as an industrial conversion process, not a software product. Inputs of energy, silicon, and capital enter a physical system; the output is computation, sold as intelligence. This framing is the organizing model for the rest of this report.

The Model

Seven layers convert raw energy into economic value. Each layer depends entirely on the layer beneath it — abundant accelerators with no power are stranded capital; abundant power with no advanced packaging is underused generation; abundant capital with no interconnection rights buys nothing but a queue position.

The Governing Principle

Marginal value accrues where supply elasticity is lowest. At any point in time, one layer of the stack expands more slowly than demand for it grows. That layer sets the marginal price for everything built on top of it, and its owners capture the resulting scarcity rent for as long as the constraint binds. Because capacity investment is itself elastic over a multi-year horizon, every bottleneck is eventually relieved; the analytical task is estimating how long relief takes, not whether it arrives.

Three corollaries follow directly from this principle, and each is used repeatedly in the sections that follow:

• Rents are positional, not permanent. A company's advantage is a function of which layer is currently binding, not of any fixed technological superiority. IBM in mainframes, Intel in client CPUs, and Cisco in networking equipment each earned extraordinary margins while their layer was the binding constraint, and each saw those margins compress once the constraint moved.
• Capital intensity is a symptom, not a cause. The AI buildout looks capital-intensive because the current bottleneck (power, packaging) happens to require large, illiquid, long-lived physical assets. If the bottleneck moves to a layer with different capital characteristics — land, water rights, skilled labor — the capital profile of the winning positions moves with it.
• The stack should be read top-to-bottom for who captures value, and bottom-to-top for who enables it. Energy and power infrastructure enable everything above them but, in isolation, capture the least value per unit of demand growth; applications capture the most value per unit of deployed capital but are also the most easily replicated layer in the stack.

Scarcity migrates. As each layer of the stack attracts capacity investment, the binding constraint moves to the next-least-elastic input. The table below ranks the nine layers most relevant to AI infrastructure by the severity and expected duration of their current scarcity — from power, the most severe and longest-lived constraint today, to algorithms, the layer where scarcity has already been competed away.

Reading the Ladder

Three patterns are worth isolating from the table. First, the most durable bottlenecks are the most physical ones. Power and advanced packaging both involve capital equipment or transmission infrastructure with multi-year lead times and limited geographic substitutability; algorithms and, increasingly, accelerator-level compute, are the opposite — knowledge and manufacturing capacity that diffuses relatively quickly once demand justifies the investment.

Second, the ladder is not a strict queue — several rungs bind simultaneously. A hyperscaler in mid-2026 is typically power-constrained in its preferred market, packaging-constrained on its accelerator order book, and only lightly constrained on data-center shell construction or algorithmic capability. Positioning around “the” bottleneck is less useful than positioning around whichever rung is binding for a specific project, in a specific region, at a specific time.

Third, duration is the variable that actually prices the rent. A five-year power constraint and a twelve-month packaging constraint may look similarly severe on a supply-demand chart today, but they justify very different capitalization rates. This is the central reason Section 3 treats the power layer in more depth than any other: it is not just the tightest constraint on the ladder, it is also the longest-lived one by a wide margin.

Electricity has moved from a line item to the pacing variable of the entire AI buildout. Every other bottleneck in this report — chips, packaging, capital — can in principle be relieved by writing a large enough check. Firm power cannot; it is bound by physical transmission infrastructure that is built on a schedule measured in years, not quarters.

The Demand Side

Global data-center electricity consumption is on a path to roughly double this decade, from approximately 485 TWh in 2025 to approximately 945–950 TWh by 2030 under the IEA's base case — a trajectory that has held broadly steady across the agency's last two annual updates despite the intervening surge in AI investment. AI-accelerated servers are the entire story: their electricity consumption is on pace to roughly triple over the same period, growing at roughly 30% annually versus roughly 9% for conventional servers, such that data centers alone are expected to account for over one-fifth of all global electricity demand growth through 2030.

The United States and China absorb most of this growth. US data centers are projected to draw roughly 430 TWh by 2030 — enough, on IEA figures, to exceed the combined consumption of the country's aluminum, steel, cement, and chemicals industries — while China adds a further roughly 280 TWh. Locally, the concentration is starker still: data centers already account for approximately 25% of Virginia's total electricity supply and could reach 46% by 2030 on Dominion Energy's own resource planning, and Ireland's data centers already exceed 20% of national demand.

The Supply Side: Grid Limitations and Interconnection

The binding constraint is not generation capacity in aggregate — it is the ability to deliver firm, high-reliability power to specific sites on a usable timeline. The Lawrence Berkeley National Laboratory's most recent interconnection queue survey found roughly 2,061 gigawatts of generation and storage capacity nationally awaiting grid access, with a median wait now approaching five years and, for some large-load data-center projects, running as long as twelve. Two regional examples illustrate the scale of the mismatch:

• ERCOT (Texas): a large-load interconnection queue of approximately 410 gigawatts as of April 2026 — roughly five times the grid's current installed capacity — of which data centers represent the large majority. CenterPoint Energy alone saw large-load interconnection requests jump from 1 GW to 8 GW in a single year.
• PJM (Mid-Atlantic): the operator missed its own reliability target by 6,625 MW in its December 2025 capacity auction — the first shortfall of its kind in the market's history — with natural gas supplying 43% of the cleared resource mix and a further shortfall of roughly 6.6 GW projected for 2027–28.

Equipment lead times compound the queue problem. Large power transformers, switchgear, and grid-scale batteries all sit behind their own multi-year backlogs; utilities such as Dominion Energy now quote 36-plus months for new substation service and roughly seven years on average for a 100 MW connection in the most constrained corridors.

The Generation Response

Nuclear: from stranded asset to strategic PPA

Every major hyperscaler has now signed at least one nuclear power agreement, and the sector's combined commitments total roughly 13 deals and approximately 10 GW as of mid-2026 — a scale of corporate nuclear procurement with no historical precedent. Microsoft's $16 billion, 20-year power purchase agreement to restart Three Mile Island Unit 1 (now the Crane Clean Energy Center, 835 MW, targeted for 2027) was the opening move; Google followed with a 500 MW small modular reactor fleet agreement with Kairos Power, the first of its kind; Amazon has committed $700 million to X-energy for up to twelve Xe-100 reactors alongside a separate $20 billion-plus conversion of its Susquehanna campus; and Meta has assembled the largest single commitment, up to 6.6 GW across TerraPower, Oklo, Vistra, and Constellation. Nuclear's appeal is structural, not sentimental: unlike solar or wind, it offers the 95%-plus capacity factor that a continuously utilized AI training cluster actually needs.

Natural gas: the bridge fuel, itself supply-constrained

Where nuclear cannot be built quickly enough, developers are turning to gas and running into the same bottleneck dynamic one layer downstream. GE Vernova's gas-turbine order backlog has grown from 83 GW at the end of 2025 to 116 GW by the second quarter of 2026, with management guiding to 125 GW by year-end and now booking delivery slots into 2031; the company's data-center-linked electrification orders alone reached $2.4 billion in a single quarter, more than all of 2025 combined. Only three manufacturers — GE Vernova, Siemens Energy, and Mitsubishi Power — build heavy-duty gas turbines at scale, and their combined manufacturing capacity (roughly 10 GW per year today, targeted to reach 20–30 GW annually by 2028–30) cannot keep pace with the backlog. The result is a wave of behind-the-meter gas generation: roughly 56 GW of announced US data-center capacity, spanning 46 projects, is now designed to run on-site generation rather than wait for grid interconnection, and nearly three-quarters of that capacity is gas-fired.

Renewables and the intermittency mismatch

Solar and wind remain the cheapest incremental generation in most markets, but their 25–35% capacity factors are a poor structural match for a load that needs to run above 90% utilization to earn back its capital cost. Renewables are increasingly paired with storage and used to offset grid draw rather than to directly power training clusters — a supporting role in the AI power mix rather than the primary one.

Geography and Site Selection

Firm power availability, not land cost or fiber access, is now the dominant variable in data-center site selection. Markets with power headroom and fast permitting — parts of Texas outside the Dallas/Houston core, the Ohio Valley, the Gulf states, and sovereign-backed campuses in India and the Middle East — are absorbing disproportionate new capacity, while incumbent hubs such as Northern Virginia are increasingly capacity-constrained despite deep existing infrastructure. This is the geographic corollary of the Bottleneck Ladder: capital follows whichever region can deliver the currently-binding input fastest.

The practical consequence for underwriting is that power availability should be treated as the first constraint checked, not the last. An accelerator order, a packaging allocation, or a capex commitment that is not backed by a confirmed interconnection date or firm power contract is, by the logic of the equation above, an option on future compute rather than compute itself and should be priced with the corresponding discount.

Four companies sit at the choke points of the AI silicon supply chain, and each captures scarcity rent through a different mechanism — ecosystem lock-in, manufacturing concentration, capital-intensive capacity, or outright technological monopoly. Distinguishing the mechanism matters, because each is vulnerable to a different kind of relief.

NVIDIA: Ecosystem Rent, Not Just Silicon Rent

NVIDIA's data-center segment generated $193.7 billion in fiscal 2026 (ended January 2026), up 68% year-over-year, and the growth rate has since accelerated: fiscal Q1 2027 data-center revenue reached $75.2 billion, up 92% year-over-year, at a non-GAAP gross margin of roughly 75%. This is not simply a hardware margin. NVIDIA's moat has three layers that reinforce one another — the CUDA software stack that makes switching costly for developers, the NVLink/Spectrum-X interconnect that makes NVIDIA systems, not just NVIDIA chips, the unit of competition, and a packaging and HBM allocation (NVIDIA holds an estimated 60% of TSMC's CoWoS capacity) that a challenger cannot simply out-order.

The competitive risk is real but should be read as margin compression on a larger base, not share collapse. Estimates of NVIDIA's share of merchant AI-accelerator revenue range from roughly 70% to the mid-80s depending on methodology and time window — down from effective near-monopoly levels two years ago — with AMD holding a mid-single-digit share and custom silicon (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) collectively approaching 15–20% and growing roughly three times faster than merchant GPU shipments. Because the overall data-center accelerator market has itself roughly quadrupled since 2023, to above $200 billion in 2026, NVIDIA's revenue continues to grow even as its share erodes — both things are true at once, and investors conflating the two are likely to mistime the position.

The custom-silicon threat is structural rather than cyclical because it originates with NVIDIA's largest customers. Roughly 40% of NVIDIA's revenue is concentrated in four hyperscalers, each of which is simultaneously NVIDIA's biggest buyer and its most capable potential substitute. AWS now runs more than 60% of its machine-learning instances on its own silicon; Google runs over three-quarters of Gemini on TPUs and is reportedly exploring external TPU sales; and Bloomberg reported in June 2026 that Amazon is in early talks to sell Trainium chips outside AWS for the first time — a historic break from a closed-distribution model that, if it proceeds, would mark custom silicon's transition from an internal cost-reduction tool to a direct commercial competitor. Broadcom, as the leading design partner behind the TPU, MTIA, and Maia programs, disclosed a $73 billion AI backlog and guided to $100 billion of annual AI chip revenue by 2027 — a figure now approaching the scale of NVIDIA's own data-center business.

Margin sustainability rests on NVIDIA continuing to sell systems, not chips — the more of the rack (networking, software, reference architecture) it controls, the more defensible the ecosystem rent becomes even as the underlying silicon share declines.

TSMC: The Manufacturing Toll Both Sides Must Pay

TSMC's position differs from NVIDIA's in an important respect: it profits regardless of which accelerator architecture wins, because NVIDIA, AMD, and virtually every hyperscaler's custom silicon program are fabricated on the same TSMC process nodes. This gives TSMC a more diversified, lower-beta exposure to the AI cycle than any single chip designer, expressed instead through capacity allocation — 2026 CoWoS demand is estimated at roughly 1.0 million wafers, up from roughly 370,000 in 2024, with NVIDIA alone absorbing an estimated 60% and the top three customers together over 85%.

TSMC is scaling CoWoS-class packaging capacity from roughly 35,000 wafers per month in late 2024 to a targeted 115,000–140,000 by the end of 2026 and roughly 170,000 in 2027, with the reported supply-demand gap narrowing from approximately 20% to approximately 10% over the course of 2026. That narrowing gap is the clearest quantitative signal available that the packaging bottleneck — while still binding — has a visible, dated relief path, unlike the power bottleneck discussed in Section 3.

Geopolitical risk is the one variable that does not show up in a capacity chart. Taiwan still produces more than 90% of the world's most advanced logic, and TSMC's most advanced nodes (2nm today, 1.4nm on the roadmap) remain in Taiwan even as the company has now committed a cumulative $265 billion to its Arizona campus, up from an initial $65 billion — a rapid escalation that reflects both AI demand and an explicit hedge against Taiwan Strait risk. The Arizona fabs currently run several nodes behind the Taiwan frontier and, by design, dilute rather than eliminate the geographic concentration that analysts have called the industry's “silicon shield.” A disruption scenario — ranging from expanded military exercises and shipping-insurance cost increases to outright blockade — remains a low-probability, high-impact tail risk that is structurally under-priced in most AI infrastructure models precisely because it has not yet occurred.

High-Bandwidth Memory: The Second Packaging Bottleneck

HBM is functionally a second packaging constraint, since it is the component that CoWoS-class technology exists to attach. SK Hynix holds the leading position, with an estimated 50–55% of the HBM market and roughly half to two-thirds of NVIDIA's HBM4 allocation for the upcoming Rubin platform; Samsung, after lagging through 2024–2025 on qualification issues, is mounting a credible recovery via HBM4 and is targeted to hold roughly a quarter of NVIDIA's allocation; Micron holds the remainder and has stated it will exit consumer memory and storage entirely to concentrate on AI and data-center customers. The global HBM market is projected to grow from approximately $38 billion in 2025 to approximately $58 billion in 2026, and Bernstein estimates SK Hynix's DRAM gross margins could peak near 93% in the fourth quarter of 2026 — a level that signals how acute pricing power in memory has become during this cycle.

ASML: The Irreplaceable Layer

ASML is the cleanest monopoly in the stack: it is the sole global supplier of extreme ultraviolet lithography systems, without which no leading-edge logic or memory chip can be manufactured, and holds roughly 94% of the lithography equipment market overall. There is, at present, no credible commercial alternative — Nikon and Canon exited EUV development years ago, and China's domestic effort (SMEE) is estimated to be more than a decade behind. ASML has raised its 2026 revenue guidance twice this year, from an initial €36–€40 billion to €43–€45 billion, at a gross margin approaching 54–56%, backed by a year-end order book of €38.8 billion; its own long-range targets for 2030 span €44–€60 billion in revenue, with Goldman Sachs modeling a bull case above €70 billion. Because every unit of TSMC, Samsung, Intel, and SK Hynix capacity growth requires ASML tools, the company's order book is a leading indicator for the entire semiconductor capex cycle — and its near-total insulation from any single customer's competitive outcome makes it the purest expression of the scarcity-rent thesis in this report.

A modern AI data center is closer to a power plant with a compute payload than to a conventional real-estate asset. Underwriting it as the latter — on lease comparables and cap rates — misses the variables that actually determine whether the asset earns its cost of capital.

Power Density Has Broken the Old Model

The industry-average data-center rack draws approximately 7.6 kW, comfortably within the 8–25 kW ceiling of air cooling. NVIDIA's GB200 NVL72 rack draws 120–140 kW and requires mandatory direct-to-chip liquid cooling; heat flux at the chip level (500–600 W/cm²) is now comparable to a nuclear reactor fuel rod. The trajectory does not level off: NVIDIA's Rubin platform, expected in 2027, is projected to require 250–900 kW per rack, and vendors including NVIDIA, Vertiv, Schneider Electric, and Eaton are jointly developing 800-volt DC power architectures to support megawatt-scale racks by 2027. A data hall engineered for the previous generation's density is not merely suboptimal for the next — it is often structurally unable to host it, which is why liquid-cooling retrofit, not new construction, is increasingly the binding constraint at the facility layer of the Bottleneck Ladder.

Rack Economics and Utilization

Because both the chips and the power/cooling infrastructure supporting them are expensive, utilization is the single variable that separates an attractive AI data-center investment from an unattractive one. A facility running at 90%+ utilization on a multi-year, take-or-pay contract earns a fundamentally different return than the same facility running at 50% while waiting for demand, or for its own power allocation, to catch up. The asymmetry is severe: a fully energized, fully leased facility captures the scarcity rent described throughout this report; a partially utilized one earns closer to the cost of capital tied up in a depreciating, increasingly obsolete asset.

The Depreciation Question

Hyperscalers depreciate GPU-based infrastructure on schedules of roughly five to six years. Critics — and, increasingly, some sell-side analysts — argue the real economic life of a given GPU generation, given the pace of architectural improvement (roughly a 3–4x jump in rack-level performance density between the Hopper and Blackwell generations alone), is closer to two to three years before a workload's economics favor migrating to newer silicon. On the more conservative estimate, industry-wide AI revenue only first exceeded quarterly depreciation expense in the fourth quarter of 2025 — a break-even that arrived only recently, and before the far larger 2026 capex vintage has begun working through the depreciation schedule. This is not an accounting footnote: it is the assumption on which the entire debate in Section 6 turns.

Stranded Asset Risk

Section 3 established that silicon without power is stranded capital. The reverse also holds: power and shell infrastructure built ahead of a specific accelerator commitment, or built for a rack density that the next chip generation outgrows, is stranded in the other direction. Both failure modes are visible in the current cycle — fully built shells waiting on interconnection studies, and older air-cooled halls unable to host liquid-cooled next-generation racks without a capital-intensive retrofit. The practical underwriting discipline is to treat power contracts, accelerator commitments, and offtake agreements as a matched set; any AI infrastructure asset that has secured only one or two of the three legs carries materially more risk than the headline capex figure suggests.

This framing also clarifies why the same headline capex number can support two opposite conclusions. Bulls emphasize a numerator that is growing (rising AI demand, contracted pricing, expanding utilization as backlogs convert to live workloads). Bears emphasize a denominator that is growing at least as fast (capex per rack rising with density, power costs rising with scarcity, and a depreciation clock that, on the more conservative view, is running faster than the assets are earning it back). Section 6 takes up this tension directly.

Every argument in this report about bottlenecks, rents, and duration is downstream of one larger question: does the economic value AI generates justify the capital being committed to build it? This is the central risk underneath the entire theme, and it is genuinely unresolved.

The Scale of the Bet

Combined capital expenditure guidance from Amazon, Microsoft, Alphabet, and Meta for 2026 has been revised upward three separate times this year — from roughly $690 billion in February, to roughly $725 billion by mid-year, to roughly $760 billion following second-quarter results — an increase of more than 80% over 2025's already-record base of roughly $410 billion. Goldman Sachs' baseline scenario puts cumulative AI-related capital expenditure at approximately $7.6 trillion between 2026 and 2031. Set against this, the plainest expression of the bear case is a single widely cited comparison: OpenAI's estimated annual compute spend of roughly $60 billion against estimated annual revenue of roughly $13 billion — a gap that, if representative of the broader industry's unit economics, implies the financial architecture underneath the buildout is not yet self-sustaining.

The Bull Case

• Productivity revolution. Enterprise cloud growth tied to AI workloads is compounding: Google Cloud revenue grew 63% year-over-year and AWS 28% in the most recent reported quarter, with Google Cloud's contract backlog reaching approximately $460 billion, roughly double the prior year — a leading indicator of contracted, not merely hoped-for, future demand.
• Enterprise adoption is still early. Inference token generation reportedly rose roughly tenfold year-over-year at NVIDIA's largest customers, a volume signal that sits well ahead of revenue recognition and suggests usage is scaling faster than monetization has caught up.
• AI agents and automation represent a second wave of demand — inference-heavy, continuous, and only beginning to be deployed at enterprise scale — layered on top of the training-driven demand that built the current infrastructure base.
• The buildout is structurally hard to reverse quickly. Construction lead times of 18–30 months and accelerator refresh cycles of three to four years mean that even a demand disappointment converts into obsolescence over several years rather than into an immediate, visible glut; and a large share of guided capex is already contracted, not discretionary, at the point it is announced.

The Bear Case

• Unit economics that do not yet close. The OpenAI compute-to-revenue gap cited above is the most quoted data point, but the broader pattern — widely cited 2025–2026 enterprise surveys reporting that a large majority of corporate AI pilots show no measurable ROI — suggests monetization is lagging deployment industry-wide, not just at frontier labs.
• The financing structure is more levered than headline capex suggests. Moody's estimated in early 2026 that hyperscalers carry approximately $662 billion in signed-but-not-yet-commenced data-center lease commitments sitting off balance sheet — a figure larger than these companies' combined on-balance-sheet debt — while Amazon's trailing capital expenditure has begun to exceed its operating cash flow outright, pushing free cash flow negative.
• Falling inference costs compress the very asset being built. Efficiency gains — architectural, algorithmic, and at the model layer — lower the cost of delivering a unit of intelligence over time. That is unambiguously good for AI adoption and unambiguously bad for the pricing power of infrastructure sized for yesterday's cost curve; see Section 9 for this risk treated in isolation.
• Roughly two-thirds of recent quarterly hyperscaler capex is going into short-lived assets — GPUs and CPUs depreciated over five to six years by accounting convention, but which some analysts argue have a real economic life closer to two to three years, as discussed in Section 5. If the shorter estimate is closer to correct, the effective capital intensity of the buildout is meaningfully understated by reported depreciation schedules.

The AI Infrastructure Cycle

Every capacity cycle in industrial history moves through the same three phases — shortage-driven pricing power, capacity expansion, and eventual oversupply — but different layers of a multi-layer stack rarely occupy the same phase at the same time. That is precisely the situation in mid-2026.

Where Are We Today?

The honest answer is phase-dependent, not stack-wide. Power remains firmly early-cycle: interconnection queues are still lengthening, not shortening, and gas-turbine bookings now extend to 2031. Packaging and merchant compute sit in the middle of the cycle: capacity is visibly catching up to demand, and the CoWoS supply-demand gap is on a credible, dated path toward balance. The capital layer is showing the earliest markers associated with a late-cycle transition — not in utilization or physical oversupply, which remain the more commonly cited late-cycle signals and are not yet visible, but in financing-structure strain: compressing free cash flow, rising off-balance-sheet commitments, and a widening gap between disclosed capex and disclosed AI-linked revenue at the most exposed pure-play buyers of compute.

This is a materially different conclusion from either “it's a bubble” or “it's not.” It implies that the physical infrastructure layers of this report (power, packaging, memory) are not yet overbuilt and are unlikely to be for several years, even as the financing layer supporting that buildout shows genuine, measurable strain today. The most probable resolution, consistent with the majority of the institutional commentary reviewed for this report, is a plateau in the rate of capex growth rather than an abrupt reversal — a deceleration that would compress equity multiples on the capital-intensive names well before it shows up as physical stranded capacity.

Translating the Bottleneck Ladder into a thematic map. Each row below is a category, not a security recommendation — durability of the underlying scarcity, not the enthusiasm currently attached to a theme, should govern position sizing and duration.

How to Use This Map

Three portfolio-construction implications follow from the map above. First, duration should set position horizon, not conviction. A one-to-two-year theme (packaging) and a five-to-ten-year theme (power) both score as high-conviction bottlenecks today, but a portfolio that sizes them identically is implicitly betting that the shorter-duration theme's relief — already underway — will not compress its own returns before the position is exited.

Second, the themes with the longest duration (power, utilities/transmission) are also the ones with the most unfamiliar risk profile for investors used to underwriting semiconductor or software cycles — regulatory approval risk, interconnection-queue risk, and construction risk dominate over the technology risk that governs the compute and memory rows.

Third, capital providers sit structurally last in the loss-absorption order across nearly every row above them; the map should be read alongside Section 6's assessment of financing-layer strain before sizing exposure to that theme specifically.

The Bottleneck Ladder is geographically uneven — no single country holds the least-elastic input at every layer simultaneously. Comparing six regions on the same four dimensions makes the resulting division of labor explicit.

Every framework in this report is conditional. The following seven risks are the ones capable of invalidating it outright, rather than merely delaying it — ranked here by the discipline of naming a probability, an impact, and a specific indicator to watch, not by rhetorical severity.

The Two Risks That Matter Most

Ranked by the product of probability and impact, two risks dominate the register above. AI efficiency breakthroughs carry both high likelihood and high impact because they attack the thesis at its foundation: every framework in this report — the Intelligence Factory, the Bottleneck Ladder, the Power Equation — assumes a roughly stable relationship between useful intelligence and physical inputs. A discontinuous improvement in that relationship (a new architecture, a training technique, or a widely adopted efficient-inference method) does not just dent demand at the margin; it resets the denominator of every return equation in this report simultaneously, which is precisely why Section 6 treats it as the single largest risk to the entire capital stack rather than as one line item among several.

Semiconductor disruption, concentrated in Taiwan Strait risk, carries the opposite profile — low near-term probability but the highest impact of any risk on the list, because more than 90% of the world's most advanced logic manufacturing has no near-term substitute. This is the risk most likely to be under-priced in conventional portfolio construction precisely because its low annual probability makes it easy to discount, even though its impact, conditional on occurring, would exceed every other risk in this section combined.

The remaining five risks are best read as rate-of-relief modifiers rather than thesis-breakers in isolation: open-source disruption and model commoditization compress value at the applications layer without necessarily reducing physical infrastructure demand; hyperscaler overinvestment and regulatory constraints affect the pace and distribution of the buildout without changing its ultimate direction; and power demand disappointment, while framed as a bear risk throughout this report, is assessed here as the least probable of the seven — every demand signal reviewed in Section 3 points toward under- rather than over-estimation of near-term power needs.

The Bottleneck Test

Every position in this report — and every position a reader might consider outside it — can be scored against the same seven criteria. This is the proprietary framework underlying every claim made in Sections 1 through 9, made explicit.

The Seven Criteria

• Supply elasticity — how quickly new supply of the input can be brought online. Score 10 for inputs that take years and specialized capability to expand; score 1 for inputs any competent team can add in months.
• Replication difficulty — how hard it is for a well-capitalized competitor to reproduce the position from scratch. Score 10 where decades of accumulated know-how or IP make replication effectively impossible; score 1 where capital alone is sufficient.
• Geographic concentration — how few hands, sites, or jurisdictions control the input today. Score 10 for extreme concentration (a fortification of pricing power for the incumbent, though also a risk flagged separately in Section 9); score 1 for an input available broadly and redundantly worldwide.
• Capital intensity — how large a check a challenger must write, and how long before it pays back, to contest the position. Score 10 for the most capital-intensive barriers to entry; score 1 where minimal capital is required.
• Pricing power — realized gross margin and the demonstrated ability to raise price without losing volume. Score 10 for sustained mid-70s%-plus gross margins with full demand absorption; score 1 for commodity, cost-plus economics.
• Duration of scarcity rent — the credible time horizon, in years, before the specific constraint is substantially relieved. Score 10 for constraints likely to persist beyond five years; score 1 for constraints already visibly closing.
• Risk of displacement — the probability that a substitute technology, architecture, or competitor bypasses the position entirely rather than merely competing within it. Score 10 for the lowest displacement risk; score 1 for imminent, structural displacement.

A composite score is the simple average of the seven criteria, each scored 1–10. Scores above 7.5 indicate a durable bottleneck position likely to sustain scarcity rent for several years; scores of 4–7.5 indicate a real but time-limited or partial bottleneck; scores below 4 indicate a position that, whatever its current narrative, is not economically a bottleneck at all.

Worked Example

Applying the framework to six illustrative categories referenced throughout this report — not individual securities — produces a ranking consistent with, and a useful cross-check on, the qualitative conclusions of Sections 3 through 5:

Three results are worth isolating. EUV lithography and firm power score highest, confirming the qualitative conclusion reached independently in Sections 3 and 4: the two most durable bottlenecks in the current cycle are a true technological monopoly and a physically-bound infrastructure constraint, not a demand-driven or ecosystem-driven position. The merchant GPU ecosystem scores well but not at the top of the range — exceptional pricing power is offset by real, structural displacement risk from the hyperscalers that are simultaneously its largest customers, exactly the tension described in Section 4. A generic, power-unsecured data-center shell scores well below every semiconductor or power position — the clearest numerical illustration in this report of Section 5's central claim that data centers are not, by themselves, a bottleneck; the power and offtake contracts attached to them are.

Conclusion: Owning the Bottlenecks

The expansion of AI infrastructure is, at its core, an industrial contest over inputs that cannot be manufactured, permitted, or shipped fast enough to meet demand — specialized silicon, firm electricity, advanced cooling, and large-scale capital. It is not, in the end, a contest about which model is smartest. It is a contest over the rate at which the physical and financial system can convert energy and capital into useful computation, and over who is positioned to capture the rent while that conversion stays constrained.

Three conclusions carry across every section of this report. First, constraints migrate: the bottleneck that defines the winners of 2026 will not be the bottleneck that defines the winners of 2029, and a portfolio anchored to today's snapshot of the Bottleneck Ladder rather than to its trajectory will systematically mistime the cycle. Second, physical scarcity dominates financial engineering: the two most durable positions identified by the Bottleneck Test — lithography and firm power — are durable because they are bound by physics and permitting, not by any balance-sheet advantage a well-capitalized competitor could simply replicate. Third, the open question is not whether AI adoption continues; it plainly is. It is whether the economic value being created will cover the capital already committed to building it — and Section 6's answer, that the physical layers of the stack remain under-supplied even as the financing layer already shows genuine strain, is the single most useful takeaway in this report for sizing a position today.

The tools built here — the Intelligence Factory, the Bottleneck Ladder, the AI Power Equation, the Data Center Return Equation, and the Bottleneck Test — are designed to be applied forward, not merely to explain the present. A position that would score well on the Bottleneck Test in 2029 is, by definition, not the same position that scores best on it today. Identifying that gap early is the actual opportunity this report is describing, not a restatement of which companies are large now.

The companies and countries that understand where the physical bottlenecks sit — and that control them for as long as the constraint lasts — will capture a disproportionate share of the economic value AI creates over the next decade. Everyone else will be renting capacity from them.

Sources & Notes

This report synthesizes public company disclosures (SEC filings, earnings releases, and investor presentations from NVIDIA, TSMC, ASML, GE Vernova, SK Hynix, Samsung, Micron, Broadcom, and the four largest US hyperscalers), International Energy Agency publications (Electricity 2026 and the Energy and AI series), Lawrence Berkeley National Laboratory interconnection-queue research, regional grid-operator disclosures (ERCOT, PJM, Dominion Energy), and industry, trade, and financial press reporting current through early August 2026. Figures presented as estimates, projections, or ranges reflect genuine variance across public sources and should be read as directional rather than precise.

• International Energy Agency — Energy and AI / Electricity 2026

  Global data-centre electricity demand and AI-related power growth.

  iea.org/reports/energy-and-ai/energy-demand-from-ai
• Lawrence Berkeley National Laboratory — Queued Up: 2026 Edition

  U.S. generation and storage interconnection queues and project timelines.

  emp.lbl.gov/queues
• ERCOT — Large Load Update, April 2026

  Large-load interconnection requests in Texas.

  ercot.com/files/docs/2026/04/09/ERCOTLargeLoadUpdate…
• PJM — Powering Reliability Through Market Design, 2026

  2027/28 capacity-market reliability shortfall and grid pressures.

  pjm.com/…/20260506-powering-reliability-through-market-design.pdf
• Goldman Sachs — Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out

  AI infrastructure capex framework and $7.6 trillion 2026–2031 baseline.

  goldmansachs.com/insights/articles/tracking-trillions…
• NVIDIA — Q4 FY2026 and Q1 FY2027 Results

  FY2026 data-centre revenue and Q1 FY2027 data-centre revenue / margins.

  investor.nvidia.com/news/press-release-details/2026/…
• ASML — Q2 2026 Results

  2026 revenue guidance and lithography capacity outlook.

  asml.com/en/news/press-releases/2026/q2-2026-financial-results
• GE Vernova — Q2 2026 Results

  Gas-turbine backlog and electrification demand.

  gevernova.com/news/press-releases/ge-vernova-reports-second-quarter-2026…
• Moody's Ratings — Hyperscaler AI Investments Drive Unprecedented Data-Center Lease Commitments

  Undiscounted lease commitments and leases not yet commenced.

  dkf1ato8y5dsg.cloudfront.net/uploads/52/504/…