In the 1990s it was assigned the country code .ai. For a quarter century, no one cared. In July 2018 there were about 50k .ai domains, generating $2.9 million a year, ~4% of the national budget.

Then ChatGPT launched. By 2023 registrations had jumped to 354k and revenue to $32 million, driving 20% of government revenue. In 2025, $85 million, or roughly 47% of the government’s total revenue. The island now funds half its government by selling two letters.

Anguilla has a monopoly on .ai domains, and the monopoly is an example of a bottleneck. We look for these because that’s where the money is (Willie Sutton). And while Anguilla is an example of one, it is not the type of bottleneck we would like to find. If .ai disappeared tomorrow, nothing would change. Datacenters would still get built, chatbots would still chat. What we want to find (and build the framework for) are bottlenecks by necessity. You can replace the domain name, but you can’t replace the transformer, memory, or power.

Hyperscalers have been pouring capital into AI infrastructure since 2022 (see my thesis on the AI supercycle). That money settles in nodes (compute, memory, optics, cooling, power) and creates bottlenecks along the way. This is a walk down one of them.

Supply

The US generated 4,430 TWh of electricity in 2025. A record, up 2.8% from 2024, which was also a record.

Before that, generation was flat from the mid-2000s through the early 2020s, and consumption actually fell about 1% across the 2010s. The assumption was that demand had stopped growing. And for fifteen years it was the correct one.

Demand

Datacenters consumed 176 TWh in 2023 — 4% of US electricity. By 2025, roughly 245 TWh, about 5%.

Lawrence Berkeley National Laboratory projects 325 to 580 TWh by 2028. Between 7% and 12% of everything the country generates. Incrementally, this means: +80 TWh in the low case, +335 TWh in the high case, over three years.

Another estimate (from EPRI, looking to 2030), puts datacenters’ share of electricity consumption at 9% to 17%. Two more years, roughly double the increment at the top end. EPRI’s current estimate is about 60% higher than EPRI’s own estimate two years ago. The forecasts are being revised upward.

What the country can add

The US plans to add 86 GW of generating capacity in 2026. That is a record — about 7% of the entire national fleet in one year, against 53 GW actually delivered in 2025. Apply capacity factors to the 86 GW and you get roughly 163 TWh of new annual generation. Note the mix: solar and storage are 79%, 7% gas, and the remaining is wind. But 163 TWh is planned, not delivered. Last year about 102 TWh was delivered (53 GW at 22% blended capacity factor), that’s a more realistic, long-term figure. Plants also retire, subtracting about 35 TWh annually.

The gap

Demand growing at the observed 2.8% requires +383 TWh over three years. (2.8% is historical, from 2025. As the number of datacenters grows, the overall growth in demand will exceed 2.8%, so this is highly unlikely and unreasonably conservative.)

At record buildout, net new supply is about +384 TWh ((163 - 35) x 3). At realistic buildout, about +201 TWh ((102 - 35) x 3). So at best, the country covers total demand growth from datacenters only and exactly, with nothing spare. Otherwise it falls roughly 182 TWh short.

Now put datacenters inside that:

• Low case: +80 TWh (datacenter increment) → 21% (share of all US demand growth)

• High case: +335 TWh (datacenter increment) → 87% (share of all US demand growth)

In the high case, datacenters take 87% of everything the country adds. Residential, industrial, EV charging, heat pumps, and reshored manufacturing split the remaining 13%, for the entire rest of the economy.

Nobody runs out of electricity. The marginal gigawatt gets slow and expensive, and every other buyer is bidding against someone with functionally unlimited capital.

Gas is the right answer and it is not available

Datacenters draw continuously. Solar does not. Serving 1 GW of continuous load takes about 0.75 GW of nuclear, 1.2 GW of gas, or 2.8 GW of solar plus storage.

Gas is the obvious fit. It is also sold out. GE Vernova reported 116 GW of gas equipment backlog and slot reservations in Q2 2026, guiding to at least 125 GW by year end. Lead times run about three years. Pricing on new orders is 10 to 20 points higher per kilowatt than six months earlier. GE Vernova, Siemens Energy and Mitsubishi make roughly 75% of large-frame turbines between them, and Siemens’ gas book-to-bill is above 2.5 — booking faster than it can ship.

Nuclear has the right shape and the wrong decade. About 10 GW is committed across thirteen hyperscaler deals. First electrons arrive in 2027 with the Crane restart. Most SMRs land after 2030.

Which leaves renewables, mostly solar. Solar’s problem is storage. Charge midday, discharge overnight. The US installed 57.6 GWh of storage in 2025 — a record. Cumulative grid-scale capacity is 137 GWh (so 73% increase YoY), and all of it is already working, balancing the grid we have.

US battery pack manufacturing capacity is 79 GWh. US battery cell manufacturing capacity is 22 GWh. Wood Mackenzie estimates domestic cells met 6% of US demand in 2025. So 94% of the cells would have to come from somewhere else, and that somewhere is overwhelmingly China — about 60% of global battery additions in 2025. FEOC rules are now in effect specifically to restrict that. So gas turbines are a bottleneck, batteries as a storage are a bottleneck, nuclear is a bottleneck. Energy is a bottleneck bonanza.

Where this leaves the chain

The CapEx splits various ways: site, shell, interconnection, generation, fuel transport, transmission, distribution equipment, thermal, compute, foundry, packaging, memory, servers, interconnect.

Most of these industries supply a generic product and absorb the shock. Server assembly is fragmented, capital-light, scalable in months. Site work and plumbing have no constraint worth naming.

Other industries are hard to scale. HBM has three suppliers, all sold out through 2026 — but they are spending over $54 billion on new fabs, and that capacity lands in 2027 and 2028. Foundry and semi packaging were the binding constraint in 2023 and are now well understood and well priced. Gas turbines are sold out through 2030. But GE Vernova says the turbine is often not what stops a project. Turbines, transformers, and switchgear are separate supply chains, and these are complementary when you need to power a datacenter.

Of roughly 16 GW of US datacenter capacity announced for 2026, only about 5 GW was actually under construction. Sightline Climate estimates 30 to 50% of the pipeline gets delayed or cancelled — on power constraints, transformers, and switchgear.

Large power transformers now run about 128 weeks on average, with high-capacity units quoted at three to five years. Before 2020 it was closer to one. Switchgear is effectively sold out through 2028. Interconnection in datacenter growth zones takes 36 to 48 months against a federal target of 8 to 11 months, and PJM projects reaching operation in 2025 had averaged eight years in queue.

The US large power transformer market is about $1.16 billion a year. Hyperscaler capex in 2026 is roughly $650 billion. The industry slowing the buildout is under two-tenths of one percent of the spending it has to serve.

And the constraint sits one layer deeper still. Transformers require grain-oriented electrical steel. Cleveland-Cliffs is the only US producer. New annealing furnace capacity takes three years or more, and the industry is lobbying for Defense Production Act support to finance it. In the meantime the manufacturer ration sheet width to their largest customers — not by price, by allocation.

That is what a bottleneck by necessity looks like. This is where the money is.

What this was

Map the chain. Ask which node cannot scale against the demand arriving at it. That is gate two.

The answer moves every couple of years, GPUs, then packaging: now transformers and turbines. This shortage has an end date: 2027 for transformers and 2030 for turbines. So this is a two-to-four year window (as of 2026). And if transformers are short, some of the announced datacenters don’t get built on time. The shortage is bullish for one side but not for the other. The existence of the bottleneck gets priced in (see efficient market hypothesis), but magnitude and duration are often not.

The bigger point is to build the chain, look for bottlenecks, and update the map as bottlenecks migrate. The map might get stale but the method doesn’t.

APPENDIX

Everything above, with the arithmetic shown. The assumptions here.

A1. Definitions

Power is watts — a rate, what something draws at a moment. Energy is watt-hours — power times time. A 700W GPU running all year uses 700 × 8,760 = 6.1 MWh.

PUE (Power Usage Effectiveness) — total facility power divided by IT equipment power. 1.25 means 25% overhead for cooling and conversion losses.

Capacity factor — actual output divided by nameplate times 8,760 hours. Solar ~25%, wind ~35%, gas combined-cycle ~60%, nuclear ~93%.

TWh — terawatt-hour, one billion kWh.

A2. The demand unit

A GB300 NVL72 rack draws about 135 kW of IT load. At PUE 1.25 that is ~169 kW at the meter. At 70% utilization: 169 × 8,760 × 0.70 = ~1 GWh per rack per year — about 95 US homes, continuously.

Industry-average rack density is 7.6 kW. An AI rack is roughly 18 times that.

Per-GPU draw by generation: A100 400W, H100 700W, GB300 ~1,200–1,400W. Efficiency per unit of compute improved every generation. Absolute power per chip rose every generation. Different things, routinely conflated.

A3. Reconciling the demand forecasts

Goldman models ~95 GW of US datacenter capacity by end-2027 at 70% utilization → 66 GW average draw → 66 × 8,760 = 578 TWh.

LBNL’s independent bottom-up 2028 high case: 580 TWh.

Different institutions, different methods, half a percent apart.

A4. Capacity to energy

2026 planned additions:

• Solar: 43.4GW at 25% → 95TWh/yr

• Wind: 11.8GW at 35% → 36TWh/yr

• Battery: 24.3GW at 0 (stores, does not generate)

• Gas + other: ~6GW at 60% → 32TWh/yr

• Total: 86GW at ~22% blended ~163TWh/yr

Realistic rate of 60 GW/yr: ~114 TWh/yr. Retirements ~8 GW/yr at ~50% CF: −35 TWh/yr.

A5. The three-year accounting, 2025→2028

Demand growth at 2.8% compounded: 4,430 × (1.028³ − 1) = +383 TWh

• Record sustained (163 × 3): 489 (gross), -105 (retired), +384 (net)

• Realistic (102 × 3): 306 (gross), −105 (retired), +201 (net)

Datacenter share of demand growth: 80 ÷ 383 = 21%; 335 ÷ 383 = 87%.

A6. The gas capacity factor concession

The 60% figure is economic dispatch, not a physical limit. Combined-cycle plants can run 85–90%. A plant built specifically for datacenter baseload would deliver roughly 50% more energy per GW than the table assumes.

This does not rescue the supply side — the binding constraint through 2030 is turbine availability, not utilization. But it is the strongest counterargument available and it should be conceded rather than discovered.

A7. Battery arithmetic

If 80% of new supply is intermittent and some fraction needs time-shifting:

• 50% (time-shift assumption), ~367 GWh (high-case storage needed), ~7 (years at 2025 rate)

• 25% (time-shift assumption), ~184 GWh (high-case storage needed), ~3.5 (years at 2025 rate)

• 10% (time-shift assumption), ~73 GWh (high-case storage needed), ~1.5 (years at 2025 rate)

Reasonable people pick different numbers and the deployment answer moves a lot.

The manufacturing answer does not. US cell capacity is 22 GWh against 6% of domestic demand. At every assumption in the table, the overwhelming majority of cells are imported, and the dominant supplier is under FEOC restriction.

A8. Why plants retire, and why they suddenly aren’t

Coal and older gas units retire for four reasons: age (much of the US coal fleet is over 50 years old), dispatch economics, emissions compliance costs, and rising maintenance.

In 2025 operators planned 12.3 GW of thermal retirements and executed 4.6 GW — the least since 2008. Coal specifically: 8.5 GW planned, 2.6 GW retired, least since 2010. 4.8 GW pushed to later years, two plants cancelled retirement outright, another 1.2 GW scheduled for 2027 cancelled. DOE issued emergency orders keeping large coal plants running — 90-day orders, reissuable indefinitely.

Forecasts are opinions. Cancelled retirements are revealed behavior.

A9. Known weaknesses

• The 2025 datacenter figure (~245 TWh) is extrapolated from LBNL’s 2023 baseline at ~18% CAGR, not published.

• The time-shift assumption in A7 is crude. A real answer needs hourly load and generation profiles.

• Capacity factors are national averages. Texas solar and New England solar are not the same.

• Holding non-datacenter demand growth at 2.8% while datacenter share doubles is internally inconsistent, as noted above.

• The $1.16B transformer market figure comes from a single market research firm.

• The 16 GW / 5 GW construction split is reported secondhand.

• Anguilla revenue figures vary by source: US$85.3M (Ministry of Finance via Anguilla Focus), $93M (TechFlow), ~$70M (Sherwood). The Ministry figure is used here.

A10. Sources

EIA (generation, capacity additions, retirements, AEO 2026) · Lawrence Berkeley National Laboratory LBNL-2001637 · EPRI · Goldman Sachs Research · GE Vernova Q1/Q2 2026 filings and calls · Siemens Energy · Wood Mackenzie (transformer lead times, battery cell capacity) · SEIA · American Clean Power Association · Sightline Climate via Bloomberg · Carbon Direct · FERC and ISO queue data · IMF (Anguilla) · Anguilla Ministry of Finance

When COVID hit, the Fed cut rates to zero and flooded the economy with liquidity. Excess capital, like excess alcohol, leads to risky decisions. The money rushed to venture capital, which raised $330 billion in 2021, double the prior year. Locked-down consumers were bored, app engagement went vertical, and analysts extrapolated the curve. Clubhouse hit a $4 billion valuation (what happened to that app?). Hopin reached $7.75 billion and was later sold for parts. Fast raised $100 million and shut down in two years.

Those startups needed a bank. Silicon Valley Bank (SVB) was the default.

SVB then did the boring, prudent thing. It put roughly $91 billion into long-dated government-backed securities — Treasuries and agency mortgage bonds. Zero credit risk. The safest instruments in the world. Weighted-average duration of 6.2 years, with most of the mortgage paper maturing in ten years or more.

Then the Fed started hiking.

Bond prices fall when rates rise. Every macro textbook says so, and it doesn’t matter if you hold to maturity. High rates led to low venture funding in 2022 and deposits walked out the door. SVB had to sell. It liquidated $21 billion at a realized loss of $1.8 billion, announced a capital raise, and depositors requested $42 billion in withdrawals in a single day.

And the rest, as they say is..., the bank died of duration.

What the gate measures

Treasury yield is the return on lending money to the US government. It is the floor under every other return in finance, and every asset gets priced off it.

Stock yield is the inverse of the price-to-earnings multiple. If the S&P trades at 20 times forward earnings, you are buying five cents of earnings for every dollar.

Stocks are riskier than Treasuries. So the market demands a premium for holding them - equity risk premium:

Equity risk premium = S&P earnings yield − 10-year Treasury yield

It answers one question. Are you being paid enough to take the risk? There is no fixed threshold. The threshold moves because the multiple moves. Two moving parts. Track the spread, not the level.

This is gate one because it is regime-level. Treasury yields are gravity. When gravity increases, everything gets heavier.

Where we are right now

The ten-year sits at 4.67 percent, having just touched its highest level since January 2025.

S&P forward P/E is 19.8, having just fallen below 20 for the first time this year. Earnings yield: 5.06 percent.

Equity risk premium: +0.4 percent.

Thin, but positive. You are being paid forty basis points to accept equity volatility.

The multiple compressed this year because earnings grew faster than prices — back-to-back quarters above 20 percent growth. Earnings are doing the work. Not price discipline.

The pressure on the other side comes from three separate places.

1. Middle East conflict pushed Brent above $100, which increases inflation expectations, which lifts yields.

2. Government debt issuance adds supply.

3. AI capex is increasingly debt-financed, adding more.

Markets now price roughly a 35 percent chance of a rate hike next week and near 80 percent by September.

What I am simplifying, and why

1. Earnings grow with inflation. Bond coupons do not. You are comparing a real number to a nominal one. I use the formula anyway, as a regime signal rather than a valuation model. It tells you what the alternative pays.

2. Second problem, and this one is worse. Forward P/E runs on analyst estimates. Analysts are pro-cyclical and they are most wrong at turning points. In a downturn, estimates get cut, the multiple looks higher, the premium looks thinner — precisely when you should be buying. The gate is least reliable exactly when it matters most.

3. Third. The premium sat near zero or below for most of 1997 through 2000, and again through much of 2017 to 2021. Both times equities kept climbing. A closed gate can stay closed for years while the market ignores it.

This is a risk filter. It is not a timing tool.

Gravity is not distributed evenly

Higher discount rates hit distant cash flows harder. That’s why high-multiple names get punished first.

But duration is not a fixed property of a company. It is a function of what management does with the cash.

Google’s advertising business is short duration. Cash today, proven, repeatable. Same for Amazon retail and Microsoft’s software franchise. Yet nobody prices these companies on those businesses anymore. They are priced on an AI payoff that demands enormous capex now for returns that land years out.

Every dollar of AI capex converts a near-term cash flow into a long-dated claim. The hyperscalers are lengthening their own duration in real time, and the headline multiple has not caught up.

Rising rates do not just punish expensive stocks. They punish companies converting near-term cash into long-dated bets, whatever the multiple says.

How this gate weights against the others

The premium is not another vote. It is regime-level, while the remaining gates are company-level. It modifies position sizing. It does not overturn a thesis.

The asymmetry matters. A closed gate can hold back an excellent company. An open gate cannot rescue a bad one.

Premium wide, company gates pass: full position.

Premium thin, company gates pass: initiate small, or stage the entry.

Premium closed, company gates pass: thesis stays live, capital stays home. Keep the research current so you can move when the gate reopens.

Premium wide, company gates fail: nothing. Macro does not save you.

Gates one and six are also coupled. When the premium compresses, the valuation gate tightens on its own, because a thin premium is exactly when long-duration multiples are most exposed.

And when the gate does close, the money goes into bills and short-duration paper. Not long bonds. Don’t be like SVB.

Most retail investors act like morons. They see a company they like (for whatever reason). They buy it. Then they assemble reasons why they were right. The story feels like analysis. But the conclusion was reached before the work began.

This is different. Everything here starts with a macro trend and works its way down to a company. The direction matters: it changes what you own and why, which changes what you do when things go wrong.

I. The Problem With How Most People Invest

The rationalization trap has three symptoms.

1. No macro anchor. If you bought a semiconductor stock because you heard AI was going to be big, you own it for a story, not a thesis. A thesis specifies which macroeconomic and structural conditions must hold for the investment to work. A story doesn’t. When the Fed raises rates and growth multiples compress across the board, the story investor is surprised.

2. No exit criteria. If you didn’t define what would make you wrong before you bought, you will not know when to sell. You’ll hold through deteriorating fundamentals telling yourself the market doesn’t understand the long-term vision. Sometimes you’ll be right. More often you’ll be sitting on a 40% loss waiting for a recovery the underlying business no longer supports.

3. No understanding of the big picture. A company can execute perfectly and still be the wrong investment if the macro turns. Thesis-level thinking forces you to understand what conditions the company needs to thrive — and whether those conditions are still intact.

Rigorous institutional research costs $10,000 to $50,000 a year and is built for professional allocators. Retail research services are affordable but mostly stock-first and narrative-driven, with little systematic process and no macro framework. The serious, framework-based thematic research space designed for sophisticated individual investors is largely empty. That’s where Conviction Stack lives.

II. The Workflow

Macro Trend → Industry → Segment → Company.

Each level removes noise. Each level gets more granular. And each level must earn its place before the next one opens.

It starts with a macro trend — a durable, structural force. Not a cycle. Something demographic, technological, or geopolitical that is already in motion and will remain in motion. AI infrastructure buildout. Defense spending driven by geopolitical fragmentation. The repatriation of supply chains after decades of offshoring.

The question becomes: which industry sits in the path of this force and captures the rent? Trends don’t reward uniformly. Identifying where value concentrates within a trend is as important as identifying the trend.

Then it gets more granular: within that industry, which segment? The AI supercycle is real. But the returns from it have not been distributed evenly. You have to know which node in the value chain captures the margin.

Only then does the company selection begin. The stocks that survive to this level aren’t picked. They’re surfaced. Most investors start at the bottom and work up. The cascade forces the opposite: start at the top, earn each level, and let the framework do the filtering.

III. The Filters

A gate is a filter that a thesis must pass before moving to the next level. Any gate can kill the idea — and killing ideas early is the point.

1. Equity risk premium filter. Is the macro environment compensating equity investors enough to justify the risk over risk-free Treasuries? A dynamic relationship between the S&P earnings yield and the 10-year Treasury rate. When that spread approaches zero, the case for owning volatile equities weakens. “Don’t fight the Fed” as your grandpa used to say.

2. Value chain filter. Where in the chain do the dollars flow, and who actually captures them? This filter identifies where pricing power sits before any company analysis begins.

3. Bottleneck filter. Is there a structural chokepoint in the supply chain? Chokepoints create inelastic demand and pricing power. They are the mechanism by which trends translate into stock returns.

4. Margin extraction filter. Does the bottleneck position actually translate to expanding and sustainable margins? A full analysis of product economics, industry structure, competitive dynamics, and cost structure, all to answer one question: can this company hold the toll booth and keep raising the toll?

5. Management filter. Does management have skin in the game? Do they believe in their own company?

6. Valuation filter. Is the entry point reasonable, or has the market already priced in perfection? Parabolic chart patterns are automatic disqualifiers regardless of thesis quality. No margin of safety means no asymmetric return.

Each filter gets its own post. Positions are re-evaluated continuously. Quarterly earnings, policy changes, competitive developments — each one triggers a gate re-evaluation. The question is always the same: has anything materially changed?

IV. What This Is Not

This is not a stock tips newsletter. What it is: a documented, systematic workflow for filtering noise and stress-testing a thesis before capital is committed. Human judgment handles thesis construction, logical integrity, and the final call. AI handles the rest.

V. Coming Next

The next posts will walk through filters and active theses in detail.

Conviction Stack publishes systematic thematic research for serious investors. The thesis is human. The work is automated.