AI Runs on Megawatts: Why I’m Buying the Infrastructure Behind the Boom

I’ve opened five new positions, and while at first glance this may seem like a diverse list of small-cap stocks, the truth is that all five are connected by the same thread. I’ll summarize it in one sentence before diving into the analysis of each one: Artificial intelligence isn’t built with software; it’s built with energy and computing power, and almost no one is looking at the part of the supply chain where supply is truly scarce. While the market fights over paying forty or fifty times earnings for the visible layers of the AI business, the physical infrastructure needed to make it all work — the megawatts, the land with access to the power grid, the liquid-cooled data centers, the solar cell factories — is in many cases trading at the multiples of a forgotten company. And that’s where I wanted to position myself.

The five holdings are (APLD), (CRWV), (CLSK), (KEEL), and (TE). Three of them — APLD, CRWV, and KEEL — are, to varying degrees, direct bets on AI computing capacity. CLSK adds the layer of mining as an energy-intensive asset that is being repurposed for that same computing. And TE provides the energy side of the equation: domestic manufacturing that powers all that electricity demand. This isn’t a basket of individual bets; it’s a single thesis expressed through five avenues with different risk profiles, so that if the underlying direction is correct, I don’t need all five to succeed for the whole strategy to work.

I’d like to offer a word of caution that applies to everything that follows. These are small or medium-sized companies, most of which are reporting losses, have high operating leverage, and exhibit a level of volatility that isn’t suitable for everyone. My entry price for some of them is already trading below my cost basis, and that doesn’t particularly concern me, because I’m not chasing a price; I’m buying into a thesis with a time horizon that isn’t measured in trading sessions. The target prices I assign are my own, based on execution scenarios I consider likely, not on market consensus, which in several cases lags behind what I believe will happen. Here is a summary table with my target prices, and let’s begin the analysis.

Applied Digital (APLD): The Landlord of the Hyperscalers

Applied Digital is the company that best illustrates this thesis. It got its start in cryptocurrency mining and has completely pivoted toward building high-performance data centers — so-called “AI Factories” — which it leases to hyperscalers under contracts lasting about fifteen years. The business model is, at its core, highly specialized industrial real estate: the company secures land with access to high-capacity electricity, builds a liquid-cooled facility suitable for AI chips, and leases it to a tenant with high creditworthiness who pays rent for a decade and a half.

Key Fundamentals: From Story to Numbers

The key takeaway from the last quarter (Q3 of fiscal year 2026) is that this is no longer just a promise. Revenue grew 139% year-over-year to $126.6 million, with adjusted EBITDA of $44.1 million and adjusted net income of $33.2 million ($0.09 per share), although on a GAAP basis the company continued to report a loss of $100.9 million due to the impact of depreciation and non-recurring items. The first 100-MW building at Polaris Forge 1 is now fully operational and generating revenue; management cites approximately 900 MW (nearly 1 GW) under construction and a goal of exceeding $1 billion in net operating income (NOI) within five years. The balance sheet closed with approximately $2.1 billion in cash versus $2.7 billion in debt.

The Backlog Is the Keystone

The real gem is the contract portfolio. The Ellendale campus (Polaris Forge 1) has 400 MW fully leased to CoreWeave for approximately $11 billion over fifteen years. Added to that is a 200 MW contract with an investment-grade hyperscaler at Polaris Forge 2, and a deal worth about $7.5 billion for 300 MW at Delta Forge 1 with another highly rated hyperscaler, which also signed a repeat contract just one month after the initial agreement. In total, the company has accumulated between $16 billion and $23 billion in contracted rental revenue, with the stated goal that 70% will come from investment-grade counterparties. When a customer returns so quickly for more capacity, it is usually the best possible sign of the asset’s quality.

Valuation, Risks, and Target Price

Revenue estimates for fiscal year 2027 point to the $540 million to $550 million range, which represents nearly a doubling of the current figure as the campuses reach operational maturity. The flip side is well known: increasing leverage to finance construction and a significant concentration of customers; any delay in ramping up operations or a problem with one of those few large tenants would hit the stock. The consensus on Wall Street has been raising targets to the $45–$70 range as contracts were signed (Needham $66, Citizens $60, Compass Point up to $70). My own price target is $65, based on the gradual conversion of that backlog into recurring NOI and a reasonable NOI multiple valuation for assets under 15-year contracts. I entered at $45.60; the stock has corrected somewhat since then and is trading even below my cost, which for me does not change the thesis one iota.

CoreWeave (CRWV): The AI Highway Toll

If Applied Digital is the landlord, CoreWeave is the toll booth. It rents out GPU-accelerated computing power to those building large AI models, and its client list speaks for itself: Microsoft, Meta, OpenAI, Anthropic, Mistral, Cohere, and quantitative firms like Jane Street and Hudson River Trading. It is one of the most in-demand computing infrastructures on the planet right now, and its ties to NVIDIA are so close that it is often the first to receive and validate the latest hardware.

Hypersonic Growth and a Massive Backlog

The figures for the first quarter of 2026 are nothing short of impressive. Revenue of $2.078 billion, up 112% from a year earlier, with adjusted EBITDA of $1.157 billion and an adjusted EBITDA margin of 56%. But the figure that truly anchors the thesis is the contracted backlog of $99.4 billion, nearly 50% higher than at the end of the previous year, with an average duration of close to five years. That backlog includes a commitment of about $21 billion with Meta through 2032, an agreement of about $6 billion with Jane Street, and the relationship with OpenAI. Management claims to have already contracted more than 75% of its 2027 capacity.

The Trust of Smart Capital

There’s one sign that strikes me as particularly revealing: who is putting money in and at what price. NVIDIA invested $2 billion in CoreWeave shares at $87.20, and Jane Street put in $1 billion at $109. When the chipmaker itself and one of the world’s most sophisticated quantitative firms are buying equity at those levels, they’re telling you something about the asset’s value. Furthermore, CoreWeave has managed to pioneer a new class of financing: GPU-backed loans that have achieved investment-grade ratings (A3 from Moody’s on an $8.5 billion facility), which lowers its cost of capital and validates the model in the credit market.

The Risk Has a Name: Leverage

It’s not all rosy. The company is saddled with more than $50 billion in total liabilities and a brutal pace of investment — with $31 to $35 billion in capital expenditures projected for 2026 — which is squeezing margins and forcing it to raise capital on a recurring basis, including bond issuances at 9.75% that aren’t exactly cheap. Second-quarter guidance was weaker than expected and shifts much of the year’s weight to the second half. But, I insist, the backlog is not a promise; it is contracted revenue waiting to be recognized as capacity comes online. The market consensus ranges widely, with Jefferies at $160, Wells Fargo at $155, and BofA at $140. I set my own target at $160, aligned with the bullish end of that range. I entered at $105.15, below the price at which Jane Street entered, which seems like a reasonable starting point to me.

CleanSpark (CLSK): A Bitcoin Mining Company With a Free Option on AI

At its core, CleanSpark is one of the largest Bitcoin miners in the United States, with an operational hashrate of around 50 EH/s. It was the first publicly traded miner to reach that level using entirely self-operated data centers, and it holds a significant amount of BTC on its balance sheet. But the reason it is in this basket and not another is what it is building on top of that business: the company controls approximately 1.8 gigawatts of contracted power (with about 808 MW actually in use), including a recent approval for 300 MW in ERCOT (Texas). That access to large-scale power, at a time of structural shortages, is precisely what the AI industry needs and cannot find.

Bitcoin as a Productive Asset

One aspect the market sometimes overlooks is the size of its cash reserves. CleanSpark holds more than 13,400 Bitcoin on its balance sheet — most of which was mined in-house — and actively manages them by selling a portion of its monthly output and using BTC-backed lines of credit to finance growth without unduly diluting shareholder value. In May, it produced 671 Bitcoin (about 21.66 per day), selling 654 at an average price of around $80,000. What is relevant from an operational quality standpoint is its production cost, around $43,000 per Bitcoin in marginal cost, which gives it a very wide cushion relative to the market price and places it among the most energy-efficient miners (a fleet efficiency of nearly 16 J/TH). In other words, even a severe correction in Bitcoin would leave it operating with a positive margin.

The Accounting Mismatch That Creates the Opportunity

On paper, the results for the second fiscal quarter were a failure: a loss per share of $1.52 compared to $0.49 the previous year, on revenue of $136.4 million. But here’s the nuance the headline doesn’t capture: most of that loss stems from a non-cash accounting write-down of about $224 million on the value of Bitcoin and collateral, not from a deterioration in the operating business. The gross margin remained above 40%, the company retains liquidity in the range of $1.2 billion, and a comfortable current ratio. In other words, the damage is largely accounting-related, not cash-related. It trades at a price-to-sales multiple of around 10 times and, according to some analysts, at a forward P/E ratio that represents a considerable discount compared to its peers; the market is paying for growth, Bitcoin exposure, and the AI opportunity, not for current profits.

The Nearly Free Option for AI

Management is shifting part of its power capacity toward high-performance data centers; it is redeveloping its 250-MW site in Sandersville, Georgia, to support AI and HPC workloads alongside cloud mining; and it is in ongoing discussions with an investment-grade hyperscaler client, with a development pipeline that the company itself estimates at over 5 GW. The acquisition of GRIID also provided it with quality sites and a TVA-powered pipeline in Tennessee, where its ability to rapidly modulate consumption (demand response) makes it a valuable asset for the power grid itself. The way I see it is simple: the market is letting me buy an efficient, well-capitalized Bitcoin miner at a reasonable price and, on top of that, get an option on its conversion to AI infrastructure almost for free.

Risks and Target Price

The risk is twofold and clear: the stock moves in tandem with the price of Bitcoin, which adds a level of volatility I cannot control and could trigger accounting shocks in any given quarter; and the plan to transition to HPC has yet to be proven with signed contracts — as of now, it remains a possibility, not a certainty. It is as much a trading vehicle as it is a conviction investment, and it should be treated as such. Analysts’ estimates range from $16 to $30 depending on the weight they assign to the pivot (Maxim and Macquarie at $22, B. Riley/KBW up to $26, with the consensus average around $20–$23). My own target is $26, based on the 2027 EBITDA valuation and hashrate growth, with AI potential and the strength of its cash position as additional upside factors. I entered at $17.40.

Keel Infrastructure (KEEL): The Renaissance of Bitfarms

Keel is probably the least well-known of the five, and also the most speculative, so it’s worth taking a closer look at it. Keel Infrastructure is the former Bitfarms, the Canadian Bitcoin mining company, which in April of this year completed a profound transformation: it reincorporated in the United States, moved its headquarters to New York, changed its name and ticker symbol, and pivoted from pure mining toward digital and energy infrastructure for HPC and AI in North America. It is, quite literally, a company with nearly two years of restructuring behind it, now debuting a new identity.

An Energy Portfolio With Guaranteed Interconnection

What makes Keel interesting isn’t a generic “pivot to AI” narrative — half the mining sector is already doing that — but rather the specific asset it holds. The company defines itself as a digital and energy infrastructure operator with a 2.2-gigawatt pipeline and, crucially, established power grid interconnections, concentrated in three of North America’s most strategic and supply-constrained markets: Pennsylvania, Washington State, and Quebec. In the AI data center business, the real bottleneck isn’t land or concrete; it’s access to megawatts with approved grid interconnections, and that can take years to secure from scratch. Keel already has it, and that’s where much of its latent value lies.

Sites and the Transition to HPC

These assets have their own names. The Panther Creek site in Pennsylvania has submitted master plans for an HPC data center campus of up to 350 MW, and the company is moving forward with development at Sharon (also in Pennsylvania) and Moses Lake (Washington), with zoning approvals and environmental permits currently pending. Much of its power generation in Pennsylvania comes from plants that burn waste classified as an alternative energy source, which gives it access to waste tax credits and the ability to sell electricity on the PJM wholesale market. To validate the technical feasibility of the conversion, Keel conducted assessments with specialized partners such as World Wide Technology and Appleby Strategy Group across all its North American sites. CEO Ben Gagnon himself has been explicit: everything built by 2025 — the sites, the equipment, the balance sheet — served a single thesis: that the exponential growth of AI demands world-class infrastructure, and that they intend to build it.

A Streamlined Balance Sheet and Geographic Focus

The company has done its homework on the balance sheet and its strategic focus. It has completed its full exit from Latin America with the sale of the Paso Pesado site, bringing forward two or three years’ worth of cash flows to reinvest them in North American HPC/AI infrastructure, where it expects significantly higher returns on invested capital. As of May 8, it had total liquidity of approximately $533 million ($336 million in unrestricted cash and about $197 million in unencumbered Bitcoin), and is executing an orderly wind-down of its Bitcoin position to finance the transition. In addition, in early June, it bolstered its cash position by issuing $400 million in 1.250% convertible notes maturing in 2032 — a very low-cost financing arrangement that gives it the flexibility to undertake campus development without immediate cash flow pressure.

Re-Rating as a Driver, Risk as a Trade-Off

I won’t sugarcoat it: this is the riskiest one in the group. It remains deep in the red (its TTM net income is around –$284 million), relies on securing those initial lease agreements with investment-grade counterparties to validate its business model, and converting a mining site into an AI data center is neither immediate nor cheap. Analyst consensus is widely scattered, with targets ranging from Cantor’s $3 to H.C. Wainwright’s $5.50, including Chardan’s $4.50. My own target, $8, is deliberately above the consensus, and I justify it this way: I believe the market still values Keel for what it was — a Bitcoin mining company — and not for what it is becoming — an energy-rich infrastructure operator with guaranteed interconnection. The day it signs its first major HPC lease agreement, the market will be forced to apply a data center multiple — not a mining multiple — to its 2.2 GW pipeline, and that is where the re-rating lies. If that reclassification does not occur, that is the scenario that would first make me reconsider the thesis. I entered at $5.45, and precisely because of that dichotomy, it’s the one I’m weighting most conservatively within the portfolio.

T1 Energy (TE): The Energy End of the Chain

T1 Energy completes the supply-side side of the thesis. It is the former FREYR Battery, now a domestic manufacturer of solar modules and cells in the United States, headquartered in Austin. While the previous four companies consume electricity voraciously, T1 is on the generation side, manufacturing solar capacity within the United States at a time of strong political support for domestic energy manufacturing. It is the piece that completes the puzzle: if AI is an energy problem, someone has to produce that energy on American soil.

The Operational Turning Point

The first quarter of 2026 marked a genuine turning point. Revenue surged to approximately $177 million from $53 million the previous year; the company reported its first positive net income from continuing operations (about $3.9 million) and a record adjusted EBITDA of $9.1 million, driven by its fully operational G1 plant in Dallas. The stock reacted with a rise of more than 50% following those results, and yet, even with its accumulated appreciation, it continues to trade at compressed sales and EBITDA multiples relative to its future potential.

The G2 Catalyst and the 45X Credit Leverage

The major pending catalyst is the G2 plant in Austin, a 2.1-GW cell factory scheduled to begin production in the fourth quarter of 2026, whose ramp-up would lead management to target adjusted EBITDA of between $375 million and $450 million in 2027. Added to this is a very specific and unique cash lever: the monetization of 45X tax credits, which reward domestic solar production. The company has already sold $160 million in these credits to an investment-grade financial institution at $0.91 per dollar of credit. Furthermore, it has announced the acquisition of KORE Power for approximately $32 million to align with energy storage and the power needs of AI data centers, which reinforces the fit with the thesis.

Risks and Target Price

The risks are those typical of any capital-intensive construction project. 2026 is a “transition year” with profitability still under pressure; financing the approximately $225 million still needed for G2 remains both the catalyst and the central risk; and the company has had to respond to allegations from a short seller (Fuzzy Panda) regarding compliance with Foreign Entities of Concern regulations, against which it claims to be in good standing and to source its polysilicon from Hemlock. Northland initiated coverage with a target price of $16, and BTIG sets it at $8. I set my own target at $16, based on the jump in EBITDA resulting from the launch of G2 plus the leverage of the 45X credits. I entered at $10.80.

The 5 Theses as a Whole

I could have concentrated everything on the highest-quality option — probably CoreWeave or Applied Digital — and that would have been a perfectly defensible decision. I chose to spread my investment across five because each one represents the same idea at a different point on the risk and maturity curve. CRWV and APLD are the most established plays, with a contracted backlog and real revenue visibility. CLSK adds the Bitcoin angle and a virtually free option on AI. KEEL is the early re-rating play, the one with the greatest potential upside and the highest risk. And TE gives me the other end of the spectrum, power generation, which benefits from exactly the same structural demand but through a completely different mechanism and with its own fiscal leverage.

But there is something more important than simply allocating assets by risk profile, and that is the internal correlation logic of the portfolio. These five holdings are not independent of one another in a strictly statistical sense, and it is important to understand this. If demand for AI accelerates, they all benefit. But the transmission mechanisms are so different that one holding does not replace another: CoreWeave rises when hyperscalers sign more computing contracts; Applied Digital rises when that same growth drives demand for physical data center space; CleanSpark rises due to Bitcoin and its reclassification as HPC; Keel rises when the market applies an infrastructure multiple rather than a mining multiple; and T1 Energy rises when energy demand and domestic industrial policy converge at its Austin plant. Five distinct catalysts, five paths to the same conclusion.

It is also worth considering the stage of the cycle at which this position is being opened. The dominant narrative in the market has been focused for months on the software and application layers of AI — the Palantirs, the Salesforces, the model providers — and paying multiples for them that discount decades of growth. Physical infrastructure, on the other hand, continues to trade at a discount relative to what the demand curve implies, which is already locked in the form of a backlog. There is a reason for that: the market tends to underestimate assets that don’t shine in a demo, and a data center in North Dakota or a solar cell factory in Texas don’t shine in a demo. They shine in fifteen-year revenue projections and tax credits. And that is, precisely, the inefficiency I am trying to capture.

There is an additional factor that makes the basket more interesting than it appears at first glance, and that is the implicit hedge that T1 Energy provides against the others. The other four are net energy consumers; their business models depend on access to cheap electricity in increasing quantities, while T1 Energy is on the supply side, generating the power that meets that demand. If energy prices rise structurally, T1 — and its 45X credit leverage — appreciates at the same time that the operating costs of the other four companies increase. It’s not a perfect hedge, but it introduces a real element of diversification into what would otherwise be a purely one-way bet on electricity consumption.

Finally, and this is what matters most to me as an investor, this portfolio has a characteristic I particularly value: the catalysts are binary and imminent. I’m not waiting for a slow five-year margin expansion. I’m waiting for G2 in Austin to start production in the fourth quarter, for Keel to sign its first lease agreements, for CleanSpark to announce its first HPC deal, for Applied Digital to bring its pending 150 MW facilities online, and for CoreWeave to convert its backlog into recognized revenue in the second half of the year. These are events that will or will not play out over the next twelve to eighteen months, and that gives me a clear time window to reassess the thesis. If the milestones are met, the target prices I’ve set are conservative. If they aren’t met, we’ll know soon and can act accordingly. That’s well-managed asymmetry, not only in the potential return but also in the information the market will provide us regarding whether we’re on the right track.