Let's be honest with each other. You saw the AI hype train leave the station in 2023, and you did what any reasonable person would do: you bought Nvidia. Maybe some Microsoft. Perhaps a little Google on the dip. You felt clever. You were "investing in AI."

Congratulations. So did literally everyone else with a brokerage account and a pulse.

The chips trade is obvious. The software trade is crowded. And both are increasingly priced like the market expects these companies to cure cancer, achieve world peace, and still have bandwidth left over to generate your kid's homework. The forward P/E ratios on some of these names would make a 1999 dot-com investor blush, and those guys thought pets.com was going to the moon.

Here's the uncomfortable truth nobody at your Thanksgiving table wants to hear: AI doesn't run on hype. It doesn't run on vibes. It doesn't even run on the latest transformer architecture.

It runs on electricity.

And electricity, unlike venture capital or FOMO, cannot be printed on demand.

Power is scarce. Grid access is slower than a DMV appointment. And the companies that locked up megawatts years ago while everyone else was arguing about prompt engineering may quietly dominate AI economics well into 2026 and beyond.

This isn't anti-NVIDIA. Think of it as going one layer deeper than Nvidia. Because someone has to keep the lights on for those H100s to do their thing.

Here's a fun fact to drop at your next cocktail party: a single AI query can consume up to 1,000 times more electricity than a traditional web search. That's not a typo. One thousand times. Your innocent ChatGPT question about the best taco spots in Austin just consumed more power than your grandma uses browsing Facebook for a week.

Modern AI workloads are power-dense, always-on, and latency-sensitive. They're basically that one roommate who leaves every light on, runs the AC at 62 degrees, and somehow still complains about the electric bill.

But here's where it gets interesting for investors: scaling AI isn't just about throwing more GPUs at the problem. Data centers require grid interconnection approvals (bureaucracy), transmission upgrades (more bureaucracy), and environmental and zoning permits (bureaucracy with a side of NIMBYism). The typical timeline for major power infrastructure? Three to seven years.

Gartner predicts that power shortages will restrict 40% of AI data centers by 2027. The International Energy Agency expects data center electricity consumption to more than double to 945 TWh by 2030, roughly equal to Japan's entire annual electricity consumption.

The hyperscalers know this. Microsoft's CEO, Satya Nadella, told investors that data centers "don't get built overnight" after reporting construction delays. Amazon's Andy Jassy admitted, "Pretty much everyone today has less capacity than they have demand for".

The AI arms race isn't constrained by software talent or algorithmic breakthroughs. It's constrained by physics. And physics, unlike a Series C funding round, doesn't care about your growth projections.

Stop following AI models. Stop following chatbots. Stop following whatever app-layer nonsense just got 10 million downloads and zero revenue.

Follow the watts.

The framework is simple:

That's it. That's the thesis.

The logic works because power infrastructure is painfully slow to build and genuinely scarce. Early movers who secured grid capacity years ago now have moats that money alone can't replicate. You can't just throw $10 billion at a utility and get a gigawatt of power next quarter; it doesn't work like that. Hyperscalers know this, which is why they prefer contracts over speculation. They'd rather lock in 15-year leases with boring infrastructure operators than risk the grid won't be ready when they need it.

Here's the kicker: the market hasn't fully priced this layer yet. While everyone fights over who makes the best chatbot, the companies controlling the power to run those chatbots are quietly signing multi-billion-dollar contracts with decade-plus terms.

This is AI's railroad phase, not its app-store phase. The fortunes aren't going to the people building the apps; they're going to the people laying the track.

Let's compare two investment universes:

Traditional AI trades give you high multiples, narrative-driven price action, and the constant risk that some Stanford PhD just built something better in a garage. The stocks move on vibes, tweets, and demo days.

AI infrastructure trades give you asset-heavy businesses with contracted revenue, long visibility windows (5-15 years), and genuine hard-to-replicate moats. They're less sexy, which is precisely the point.

The unsexy trade is often the right trade. Nobody gets invited on CNBC to discuss grid interconnection timelines and take-or-pay contract structures. But when hyperscaler capex is projected to exceed $600 billion in 2026, with roughly 75% ($450 billion) going directly to AI infrastructure, maybe boring deserves another look.

These stocks look boring right until they aren't.

This is the part you actually came for. Seven names. Seven different ways to play the thesis. Let's dig in.

If there's a poster child for the "Follow the Watts" thesis, it's Applied Digital.

The company just signed a $5 billion lease with a U.S.-based investment-grade hyperscaler at its Polaris Forge 2 campus in North Dakota for 15 years, 200 megawatts of critical IT capacity. Combined with their CoreWeave deal at Polaris Forge 1, total contracted lease revenue now stands at approximately $11 billion.

That's not a typo. $11 billion in contracted revenue for a company that most retail investors still think is "some crypto thing."

With 600 MW of total leased capacity across two campuses, home to two of the world's largest hyperscalers, APLD has positioned itself as one of the fastest-scaling builders of AI infrastructure in the United States. They also secured $5 billion in infrastructure financing from Macquarie Asset Management and claim an "active pipeline" of 4 gigawatts.

The bull case: Long-term, predictable cash flows from creditworthy counterparties. Execution has improved materially. Pipeline suggests continued scaling.

The bear case: Concentration risk with two major customers. Capital-intensive buildouts. Execution still matters.

TeraWulf pulled off something remarkable: they got Google to write them a check.

The company signed two 10-year colocation agreements with AI cloud provider Fluidstack, worth approximately $3.7 billion, with Google backstopping $1.8 billion of the lease obligations. In exchange, Google receives warrants representing approximately 8% of TeraWulf's equity.

Read that again. Google, one of the most sophisticated capital allocators on the planet, took an equity stake in a former Bitcoin miner to secure AI infrastructure capacity.

The deal covers 200+ MW at TeraWulf's Lake Mariner facility in New York, with expected net operating income of $315 million annually. The first phase (40 MW) goes online in the first half of 2026, with full deployment by year-end. And they've already expanded the partnership: a new 168 MW AI compute joint venture with $9.5 billion in contracted revenue over 25 years.

The bull case: Google alignment provides credibility and financial stability. Clean energy narrative. Multiple decades of contracted revenue.

The bear case: Execution on buildout timelines. Competition from more established data center operators. Revenue concentration.

Cipher Mining's transformation from "Bitcoin miner" to "AI infrastructure landlord" happened so fast it gave investors whiplash.

In Q3 2025, the company announced a $5.5 billion, 15-year lease agreement with AWS to deliver 300 MW of capacity in 2026. They also signed a 10-year AI hosting deal with Fluidstack (backed by Google). Combined, that's $8.5 billion in contracted revenue from two of the most creditworthy counterparties on Earth.

AI/HPC now represents 67% of Cipher's operating and contracted gross capacity, compared to 33% for Bitcoin mining. Revenue hit $72 million in Q3 2025, up 65% from Q2. They raised $1.3 billion via a convertible offering, taking cash from $63 million to approximately $1.2 billion in one quarter.

CEO Tyler Page put it bluntly: "In my 25-year professional career, I have never witnessed anything close to what is going on in the market right now".

The bull case: AWS and Google-backed contracts provide revenue visibility. Successful pivot execution. Strong balance sheet after capital raise.

The bear case: Transition execution risk. Still generating cash from Bitcoin mining, which is cyclical. Customer concentration.

Riot is the "wait and see" play in this basket.

The company is evaluating 600 MW of remaining power capacity at its Corsicana Facility for AI/HPC use, in addition to the 400 MW already used for Bitcoin mining. The total approved capacity at the site is 1 gigawatt. Across two Texas facilities, Riot now owns over 1,100 acres and 1.7 GW of power capacity.

They've engaged Altman Solon, a leading data center consultant, for a feasibility assessment and are "expanding and accelerating outreach to potential partners in the AI/HPC sector". In January 2026, they signed their first AI deal: a 10-year, 25-MW data center lease with AMD at their Rockdale facility.

It's a toe in the water. But Riot's asset base, particularly its Texas power position, offers significant optionality if it executes on the pivot.

The bull case: Massive power footprint in Texas (ERCOT market is favorable). AMD deal validates AI pivot strategy. The upside of conversion is substantial.

The bear case: Pivot is early-stage. No major hyperscaler contracts yet. Bitcoin mining is still the dominant revenue source.

Galaxy Digital is the wild card, a crypto-native company that's pivoting hard into AI infrastructure.

The $1.7 billion Helios project in West Texas is the main event. Galaxy secured a $1.4 billion project financing facility for Phase 1, which will deliver capacity to CoreWeave by mid-2026. The company is essentially converting its Bitcoin mining expertise (and power positions) into AI data center infrastructure.

Q3 2025 numbers were impressive: $505 million in net income, $629 million in adjusted EBITDA, $3.2 billion in equity capital. The crypto trading business remains strong, but the data center pivot represents a structural shift in the business model.

Analysts see upside of 60% or more if the Helios project executes as planned. The company is positioning itself as a "neo-cloud" provider, following the CoreWeave playbook.

The bull case: A strong balance sheet to fund the buildout. Helios power position is valuable. Dual revenue streams from crypto and AI infrastructure.

The bear case: Still early in data center transition. Crypto volatility affects perception. Execution risk on a new business line.

Nebius is experiencing the kind of growth that makes venture capitalists cry tears of joy.

Revenue surged 625% year-over-year to $105.1 million in Q3 2025. But the real story is the contracts: a $17.4 billion deal with Microsoft (through 2031) and a $3 billion partnership with Meta (five years). Potential revenue from major contracts could reach $20 billion.

The company is targeting $7-9 billion in annual recurring revenue by 2026 and plans to build out more than 1 gigawatt of power capacity by the end of 2026. Current data center capacity is fully sold out.

With $2 billion in planned capex for 2025 and aggressive expansion in the U.S. and Europe, Nebius is betting big on the AI infrastructure boom being a "once-in-a-generation opportuni”

The bull case: Microsoft and Meta contracts provide extraordinary revenue visibility. Hypergrowth trajectory. Demand exceeds supply.

The bear case: Aggressive expansion requires flawless execution. High capex intensity. Less established than U.S. peers.

CoreWeave is the benchmark for this entire sector.

The company just reported revenue of $1.36 billion in Q3, up 134% year-over-year and $70 million above analyst estimates. The backlog? A mind-bending $55.6 billion. Contracted power capacity: 2.9 gigawatts, enough electricity to power roughly 2.2 million homes, all dedicated to AI.

Stock is up 164% since its March 2025 IPO. Major clients include OpenAI, Meta, and Microsoft. NVIDIA is both a strategic ally and a significant investor. Analysts project $5.1 billion in revenue for 2025 and $12 billion for 2026.

The company has become the Airbnb of AI infrastructure, matching supply and demand in a market where both are growing exponentially.

The bull case: Market leader in purpose-built AI cloud infrastructure. Backlog provides years of visibility. Deep Nvidia partnership.

The bear case: Valuation is stretched. A capital-intensive model requires constant investment. Execution risk on buildout. The stock has already run significantly.

Let's put some numbers on this thesis.

Hyperscaler capex for the "big five" (Microsoft, Amazon, Google, Meta, Oracle) is projected to exceed $600 billion in 2026, a 36% increase over 2025. Approximately 75% of that ($450 billion) is going directly to AI infrastructure.

The AI infrastructure market itself is expected to reach $90-101 billion in 2026, growing to $202-465 billion by the early 2030s at a 14-24% CAGR.

The stocks profiled above have combined contracted revenue in the tens of billions, yet their current market caps, in some cases, represent a fraction of that contracted value. Markets often misprice asset-heavy transitions. They did it with railroads. They did it with telecom. And they may be doing it again with AI-powered infrastructure.

Infrastructure trades lag narratives until they don't.

We're not cult members here. Every thesis has risks, and intellectual honesty requires acknowledging them. Here's what could go wrong:

Execution Risk: Building gigawatt-scale data centers is hard. Delays happen. Cost overruns happen. Some of these companies are essentially construction projects with stock tickers. Not everyone will execute cleanly.

Counterparty Risk: These contracts look bulletproof until a hyperscaler decides they overbuilt. Renegotiations happen. Projects get cancelled. Even investment-grade counterparties can pull back if the AI revenue equation changes.

Regulatory and Grid Delays: Local politics, environmental hurdles, and utility bureaucracy can kill projects. A NIMBY lawsuit in the wrong jurisdiction can delay a campus for years.

Technology Shifts: What if efficiency gains reduce power consumption per compute unit? What if someone invents a chip architecture that's dramatically more efficient? The power thesis weakens if AI becomes less power-hungry.

Market Mispricing Persistence: These stocks could stay cheap longer than your portfolio can stay solvent. Being right eventually isn't the same as being right now.

These aren't reasons to avoid the thesis entirely; they're reasons to size positions appropriately and diversify across multiple names. Being concentrated in one execution story is a recipe for pain.

Let's be clear: this is not an "all-in" trade.

You don't wake up one morning, sell your index funds, and YOLO into a basket of former Bitcoin miners pivoting to AI. That's not investing, that's gambling with extra steps.

A reasonable structure might look like:

Within the "Follow the Watts" allocation, consider diversifying across:

This thesis fits as a picks-and-shovels AI allocation, a complement to, nota replacement for, exposure to the software layer. If you own Nvidia and Microsoft, this is the diversifier that benefits from the same macro trend without the same valuation risk.

Here's your homework: a checklist to monitor weekly:

Contract Announcements: New hyperscaler deals, MW capacity commitments, contract extensions. These are the events that move stocks and validate the thesis.

Grid Interconnection Progress: Approvals, timeline updates, regulatory clearances. The bottleneck that defines who wins.

Capex Guidance Changes: When companies increase capex guidance, it signals confidence in demand. Watch both infrastructure operators and hyperscalers.

Earnings Commentary from MSFT, META, GOOGL: Listen for language about "capacity constraints," "demand exceeding supply," or "power limitations." These comments validate the scarcity thesis.

Construction Progress: Updates on data center buildouts, energization timelines, and phase completions. Execution is everything.

AI isn't limited by intelligence. GPT-5, GPT-6, whatever comes next, the algorithms will keep getting better. Talent will continue to flow into the sector. Venture money will keep pouring in.

But none of that matters without electricity.

AI is limited by electricity, land, and permits the boring, physical, unglamorous inputs that make all the software magic possible. The companies that secured those inputs early, while everyone else was debating model architectures, built moats that can't be replicated on any reasonable timeline.

The "Follow the Watts" thesis reframes AI as infrastructure first, software second. It's a bet that the market is overpricing chatbots and underpricing the power plants that run them.

Early power winners may quietly outperform headline AI stocks over the next cycle. They won't get the magazine covers. They won't trend on Twitter. But they might generate the kind of boring, contracted, predictable returns that actually compound wealth.

In the AI gold rush, the real money may be in owning the power plant, not the chatbot.