GM. This is Milk Road AI, the only newsletter that knows the difference between investing in AI and investing in what AI actually runs on.
Here’s what we’ve got for you today:
- ✍️ The power company running the AI economy.
- 🎙️ The Milk Road AI Show: AI Is Where the Internet Was in the Early 2000s w/ Puru Saxena.
- 🍪 Travis Kalanick’s robotics comeback.
Shareland enables users to trade real estate like stocks with direct access to US housing markets. Trade real estate just like stocks with Shareland.
Prices as of 10:00 a.m. ET.
THE COMPANY POWERING THE AI REVOLUTION
In the late 1870s, John D. Rockefeller didn’t become the richest man in America by building the best car.
He became the richest man because he owned the oil.
While everyone else argued about which automobile company would win, Rockefeller cornered the one thing every car on Earth needed to move.
And by 1880, Standard Oil controlled roughly 90% of U.S. refining capacity.
He didn’t bet on which company would win the race, he bet on the fuel, and it’s one of the smartest investment decisions in American history, built on a simple idea:
When you don’t know who wins the war, own the ammunition.
Fast forward 140 years.
Today, we’re not arguing about cars but about Claude vs ChatGPT vs Gemini, which GPU wins, which hyperscaler spends the most, and which startup gets acquired for billions before it ever turns a profit.
While the internet debates the tools, a modern-day Rockefeller might be sitting in Irving, Texas.
Its name is Vistra Corp (VST), and it powers the entire AI race.
Every AI model you've ever used requires a staggering amount of electricity, and not just any electricity but constant, reliable, always-on electricity.
A traditional data center uses 30-50 MW of power, while an AI hyperscale facility requires 200+ MW, approaching a full gigawatt.
That's enough to power 750,000 American homes from a single campus, and we're building hundreds of them.
And the scale of the power demand is hard to overstate:
- Goldman Sachs: Data center power demand up 175% by 2030 vs 2023.
- IEA: Data centers reach 945 TWh by 2030, up from 415 TWh today.
- EPRI: U.S. data centers use 9–17% of electricity by 2030, up from 4.5% today.
Those numbers have only gotten more alarming.
BloombergNEF now projects global data center demand hits 1,596 TWh by 2035, a 255% jump from today, with the U.S. alone surging 144% to 430 TWh.
Just the data centers coming online between now and 2030 will need 600+ TWh to run enough to power roughly 60M homes.
EPRI recently revised its U.S. data center power forecast 60% higher than it projected just 18 months ago, driven by record development and an IRA tax credit rollback that weakened wind and solar economics and pushed the build mix toward natural gas.
Their new numbers show U.S. data centers rising from 4–5% of American electricity today to 9–17% by 2030, with nominal IT capacity (power used by servers and computing equipment) reaching 56–132 GW, up from 35–44 GW today.
And the grid isn’t ready, not even close.
Peak spare capacity has been declining for years, and grid planners across the country are already warning about tightening conditions.
The threshold for what utilities consider a critically tight grid is roughly a 15% reserve margin, meaning the system only has about 15% more power capacity than expected peak demand, and some projections suggest the U.S. could reach that level by 2030.
The Department of Energy estimates the U.S. will need roughly 209 GW of new power generation by 2030, according to a July 2025 report.
Yet only about 22 GW of firm, dispatchable capacity is currently expected to come online, even though the country needs more than 100 GW of reliable power within the next five years.
At the same time, building new power infrastructure moves slowly.
A new nuclear plant can take 10 to 15 years to complete, while major transmission lines often require 6 to 10 years just to navigate permitting and approvals.
Meanwhile, the demand pipeline continues to explode.
ERCOT (Electric Reliability Council of Texas) alone has roughly 233 GW of projects waiting in the interconnection queue, up nearly 300% year over year, with more than 70% tied directly to new data center development.
The implication is clear: electricity demand is rising rapidly while new supply is coming online far more slowly.
And as the AI race intensifies, the outcome will not be determined solely by which company builds the most advanced model, but by which companies can secure enough reliable electricity to keep their infrastructure running.
That is exactly where Vistra Corp VST enters the story.
Meet the company powering AI
Vistra is the largest competitive power generator in the United States, with 41 GW of capacity, enough to power roughly 22M homes.
They serve about 5M retail customers across 20 states, with the retail segment generating a record $1.6B in adjusted EBITDA in FY2025 and total company revenue of $17.7B.
Their fleet is roughly 60% natural gas and 15% nuclear, with coal being phased out, while solar and battery storage make up the rest.
They operate across every major U.S. power market: ERCOT in Texas, PJM in the Mid-Atlantic, ISO-NE in New England, NYISO in New York, and CAISO in California.
They are everywhere, and right now, that geographic reach matters enormously.
But what truly turns this from a utility story into an AI infrastructure thesis is nuclear power.
Vistra operates the second-largest competitive nuclear fleet in the country: 6,400 MW across four plants:
- Comanche Peak in Glen Rose, Texas: 2,400 MW on ERCOT.
- Perry in Perry, Ohio: 1,260 MW on PJM.
- Davis-Besse in Oak Harbor, Ohio: 916 MW on PJM.
- Beaver Valley in Shippingport, Pennsylvania: 1,872 MW on PJM.
And why does nuclear matter for AI? Because it comes down to one number: 92.5%.
That is the average capacity factor for a nuclear plant, meaning it generates electricity roughly 92.5% of the hours in a year.
Solar? 25% capacity factor, and it doesn't work when it's dark.
Wind? 35% capacity factor, and it’s great when it blows, less great when it doesn't.
Meanwhile, nuclear runs 24 hours a day, seven days a week, 52 weeks a year, rain or shine, even at 3 a.m. on a Tuesday when nobody is thinking about electricity.
AI data centers cannot pause for weather, the servers run continuously, and the workloads never stop.
Goldman Sachs estimates 85-90 GW of new nuclear is needed by 2030 to satisfy projected data center demand.
The amount actually expected to be available globally? Less than 10% of that.
Existing nuclear plants have become some of the most valuable pieces of infrastructure in the modern energy system.
Which makes what happened next extremely significant.
Last year, Amazon Web Services signed a 20-year power purchase agreement with Vistra for up to 1,200 MW from Comanche Peak in Texas.
The customer was initially undisclosed (a very polite way to build suspense for a nuclear plant) and confirmed as AWS in Vistra's Q4 2025 earnings report.
Delivery begins Q4 2027, ramping to full capacity by 2032.
Then Meta showed up in January 2026, signing 20-year agreements for 2,609 MW across three PJM plants:
2,176 MW of operating generation from Perry and Davis-Besse, plus 433 MW of nuclear uprates across all three plants, including Beaver Valley.
And in total, Vistra has roughly 3,800 MW of nuclear capacity under 20-year contracts with two of the largest AI spenders on the planet.
The entire nuclear output of some countries is smaller than what Vistra just contracted to Amazon and Meta.
The entire nuclear output of some countries is smaller than what Vistra just contracted.
While other companies are still pitching hyperscalers, Vistra already signed the leases, and the biggest revenues from these deals haven't started flowing yet.
AI-POWERED REAL ESTATE AGENT
Wouldn’t it be cool if you could buy a slice of a New York home in the morning and then sell it in the evening?
That’s exactly what Shareland is building.
Tycoon is Shareland’s AI-powered real estate agent that trades real estate just like stocks.
Here are their key USPs:
- Trade real estate without owning the property
- Use their AI-powered agent to inform your investment strategy
- Tracks housing markets at both the city and neighbourhood level
And you don’t need large capital for this, you can start with as little as $1.
Trade real estate just like stocks with Shareland.
THE COMPANY POWERING THE AI REVOLUTION (P2)
Vistra trades at a forward EV/EBITDA of roughly 9.8–10x (a valuation metric comparing the company’s total value to its expected operating earnings).
Now let's talk about why the market hasn't fully figured this out, because here is where it gets genuinely, head-scratchingly absurd.
For context: Constellation Energy, the pure-play nuclear darling of financial Twitter, trades at approximately 15–19x trailing EV/EBITDA and roughly 13x forward.
NRG Energy trades at roughly 11–15x trailing and around 9x forward.
Vistra is priced cheaper than Constellation on either metric, and roughly in line with NRG despite carrying what many analysts consider a stronger AI power thesis than either, given its scale, geographic reach, and the AWS and Meta contracts already signed.
So why is the stock cheap? Two words: accounting noise.
Most power companies hedge part of their electricity output, but Vistra hedges almost all of it.
The company typically locks in 96–98% of the electricity it expects to generate years in advance, agreeing today on the price it will sell power for in the future.
That strategy stabilizes cash flow and protects the business from swings in energy markets.
The complication comes from accounting.
Under GAAP rules, those hedge contracts have to be revalued every quarter based on current market prices, even though the electricity will not actually be sold until later.
If power prices move, the accounting value of those hedges moves too, and those changes show up as gains or losses on the income statement.
That is what happened in FY2025.
Falling forward power prices created about $808 million of non-cash hedge losses, which pushed reported earnings down and made the stock appear expensive at roughly 74x trailing earnings.
But those losses did not reflect the actual health of the business.
Vistra still generated $5.91B in adjusted EBITDA, and the company expects $6.8–$7.6B in 2026, implying strong growth.
In other words, the gap between reported earnings and the company’s real cash generation largely comes from how the hedges are treated in the accounting.
And the company has been aggressively expanding its power portfolio.
In under two years, Vistra spent $9.3B to add 12,100 MW of new capacity:
- Energy Harbor (2024): ~$3.4B for ~4,000 MW of nuclear in PJM, including Perry, Davis-Besse, and Beaver Valley.
- Lotus Infrastructure (2025): $1.9B for ~2,557 MW of natural gas across PJM, NYISO, ISO-NE, and CAISO.
- Cogentrix (pending): ~5,500 MW natural gas portfolio across PJM, ISO-NE, and ERCOT.
Beyond acquisitions, Vistra is building 2,000+ MW of new natural gas capacity in the ERCOT Permian Basin and pursuing 433 MW of nuclear uprates tied to the Meta power agreements.
They are assembling the most strategically positioned power portfolio in America, in real time, during the most important infrastructure buildout of our generation.
The Rockefeller framework
The AI value chain has three layers, and the market has priced each one very differently.
Layer one is the model companies: OpenAI, Anthropic, Google DeepMind, and Meta AI.
Their valuations are astronomical. If you want proof, just look at the valuation of OpenAI.
Layer two is the chip companies: Nvidia, TSMC, and Broadcom.
NVIDIA alone is up roughly 10x over the last two years. Much of the obvious money has already been made.
Layer three is the power companies, the ones that make layers one and two physically possible to operate.
This is the layer the market still seems to be ignoring.
Every gigawatt of AI compute requires a power plant behind it. Electricity cannot be virtualized.
You cannot prompt engineer your way around thermodynamics.
Servers need power, cooling systems need power, and all of it has to run twenty-four hours a day with no interruptions.
Vistra sits at the intersection of two of the tightest constraints in AI infrastructure:
Existing nuclear baseload capacity and existing grid interconnection rights.
The United States could face a power capacity shortfall of roughly 175 GW by 2033.
The plants Vistra already owns cannot be replicated quickly within any investment timeframe that matters.
That is not just a competitive advantage, it is a control of a bottleneck that the AI economy cannot function without.
The risks, because we are not here to cheerlead.
Net debt sits at $18.5B, and interest coverage is roughly 2.1x, which is adequate but not exactly comforting.
The Cogentrix acquisition adds another $1.5B of assumed debt.
Management expects leverage to fall below 3x EBITDA by 2027.
That is achievable, but achievable is not the same thing as guaranteed.
GAAP earnings will continue generating confusing headlines because hedge contracts must be marked to market every quarter.
Power prices move, hedge values swing, and those swings run through the income statement even when the underlying cash generation has not changed.
Regulatory risk is also real. PJM capacity prices surged to $329 per MW-day for the 2026–2027 delivery year, roughly ten times higher than previous levels.
Data centers accounted for about 63% of that increase.
Residential electricity bills in PJM states are already rising by roughly $16–18 per month.
PJM responded by extending a price cap through 2030, limiting upside on non-contracted capacity.
The AWS and Meta agreements are insulated from this.
The rest of the fleet is not.
One more thing: the Iran wildcard.
Upgrade to PRO to see whether this Iran wildcard strengthens the Vistra thesis or breaks it:
- The scenario that could send AI power stocks much higher.
- Why hyperscalers may start locking up power plants, not just GPUs.
- The macro risks that could derail the thesis.
- What the Iran situation actually means for electricity markets.
- My full verdict on whether Vistra is a buy right now.
- How I’m positioning in the AI power trade.
Hackers are using prompt injections to break into crypto users accounts. Don’t be a victim! Use private, safe and encrypted chat Okara. Use the code MILKROAD to get a 20% discount.
Summ (formerly Crypto Tax Calculator) is a tax software built specifically for crypto. Get started for free with Summ.
Nexo is back in the U.S. - and new clients get 30 days of Wealth Club Premier perks! Higher yields, lower borrowing rates, and crypto cashback - start here.
BITE-SIZED COOKIES FOR THE ROAD 🍪
Uber founder Travis Kalanick is back with a new robotics startup called Atoms. He’s also planning to acquire autonomous trucking startup Pronto.
Apple’s new MacBook Neo is the most repairable MacBook in about 14 years, according to iFixit. The redesign includes a screw-mounted battery that’s far easier to replace.
Meta is reportedly considering layoffs that could impact 20% or more of its workforce. The cuts could help offset heavy spending on AI infrastructure.
Tax season is just around the corner. If you’re not sure how to go about it, SUMM is a tax software built specifically for crypto.
MILKY MEMES 🤣
ROADIE REVIEW OF THE DAY 🥛
