You can collect a check from the AI boom without owning a single tech stock.

That’s the premise behind the most interesting retirement income argument I’ve seen in years. It’s not about picking the next Nvidia. It’s not about day-trading AI hype cycles. It’s about owning the physical infrastructure that AI depends on and collecting the tolls.

AI is an electricity machine. Every query, every training run, every inference call demands power. Not a little power — jaw-dropping, grid-crushing, power-plant-scale power. The International Energy Agency estimates AI data centers consumed 415 TWh in 2024. By 2030, that number hits 945 TWh. That’s more than Canada and Germany combined.

Someone has to build the pipes, wires, and turbines to deliver that energy. The companies that do collect fees. Those fees flow to investors. That’s the AI income thesis in one sentence.

The toll-collector model

The most efficient vehicle for this is the Master Limited Partnership — an MLP.

MLPs are a specific corporate structure designed for capital-intensive industries like pipelines, storage terminals, and energy infrastructure. Congress created the structure in 1987 to encourage private investment in domestic energy. The trade-off is simple: MLPs pay almost no corporate income tax, and in exchange they must distribute at least 90% of their distributable cash flow to unitholders.

That means MLPs are purpose-built income machines.

A pipeline company that earns $1 billion in cash flow distributes roughly $900 million of it to investors. The yield on a typical midstream MLP ranges from 6% to 8% — sometimes higher. Compare that to the S&P 500’s current dividend yield of roughly 1.3%. The income difference is an order of magnitude.

But yield alone isn’t the story. The question is whether the income is sustainable — and that’s where the AI angle changes the math.

Why AI changes the pipeline story

Pipelines have always been steady businesses. Transporting oil, natural gas, and refined products generates reliable fee-based revenue. But the growth outlook was modest. U.S. oil production plateaus. Pipeline networks are mostly built out. The midstream sector was a yield play, not a growth story.

AI flipped that.

AI data centers need uninterrupted power 24 hours a day, 365 days a year. The grid wasn’t built for this. In Northern Virginia — Data Center Alley — Dominion Energy paused new grid connections in parts of Loudoun County in 2023. Dallas-Fort Worth has data center interconnection requests exceeding 40 GW, more than double Oncor’s current peak load. Dublin has a moratorium on new connections. Grid interconnection queues now run five to ten years.

Hyperscalers aren’t waiting. They’re building their own power. And that power needs pipelines.

Behind-the-meter natural gas generation is the fastest path to gigawatt-scale data center power. Natural gas turbines can be permitted and installed in 18 to 24 months — versus half a decade or more for grid interconnection. SpaceX and xAI’s Colossus cluster in Memphis runs 46 natural gas turbines at 1.2 GW total capacity. That cluster came online faster than any comparable data center in history because they didn’t wait for the grid.

Every one of those turbines needs a gas pipeline connection. Every connection means a long-term transport contract. Those contracts are typically 15 to 20 years, with minimum volume commitments and fixed fee escalators. They are, effectively, annuities.

The scale is staggering. One pipeline company I’ve reviewed has signed data center gas contracts totaling over $25 billion in committed fees, across over 200 data center connection requests in 15 states. The weighted average contract life is 18 years. This isn’t speculative capacity — it’s contracted, fee-based revenue that the AI companies have already committed to pay.

The royalty concept

The most useful way to think about MLP income from AI is as a royalty.

A royalty is a payment for the use of an asset. James Patterson collects a royalty every time someone buys one of his books. A mineral rights owner collects a royalty every time an oil company extracts from their land. An MLP unitholder collects a royalty every time AI infrastructure consumes energy that moves through their pipelines.

Every Nvidia H100 chip sold creates demand for more data center capacity. More capacity means more electricity. More electricity means more gas transport. More transport means more pipeline fees. More fees means larger distributions to unitholders.

The chain is indirect but mechanically sound. Nvidia doesn’t pay MLP investors directly. But the physics of AI scaling does.

This is different from owning Nvidia stock. Nvidia’s revenue depends on competitive moats, product cycles, and market share. Pipeline revenue depends on contracted volume and fee rates — factors that don’t care whether Nvidia or AMD wins the chip race. AI infrastructure needs power regardless of which company supplies the compute. That diversification is the appeal of the toll-collector model.

What realistic returns look like

The marketing around this thesis often promises numbers that defy probability. Claims of 2,000% returns on a $64 billion pipeline company require a $1.3 trillion market cap — a threshold only eight companies in U.S. history have ever crossed. That’s not investing. That’s hoping.

Realistic expectations are different and better.

A well-positioned midstream MLP with AI data center exposure can deliver a 7% to 8% distribution yield today. Add 3% to 5% annual distribution growth — management teams at major pipeline operators have explicitly guided for this — and modest unit price appreciation, and the total return lands in the 10% to 13% annual range.

Over 20 years, 10% annualized turns $100,000 into $672,000. At 12%, it’s $964,000. The income alone — reinvesting nothing and spending the yield — generates $7,000 to $8,000 per year on that same $100,000, increasing with distribution growth.

That funds retirement. Not yachts, but a meaningful income stream that grows faster than inflation.

The risk is that distribution growth doesn’t materialize, or worse, distributions get cut. MLPs are not bonds. They cut distributions in downturns. One of the largest midstream MLPs halved its distribution during the 2020 COVID crash. Plains All American cut by 58% in 2020. Enterprise Products Partners maintained its distribution through the worst of it — and is rewarded with a premium valuation today. Distribution history matters.

The K-1 complication

There’s an ugly practical detail that every MLP investor needs to understand: the K-1 tax form.

MLPs are partnerships, not corporations. Instead of a 1099-DIV at tax time, you receive a Schedule K-1. It’s longer, more complex, and arrives later — often in March or April, right at filing deadline. K-1s frequently require amendments. If you use tax preparation software, expect an hour of extra work. If you use a CPA, expect a higher bill.

More significantly, MLP income in an IRA triggers unrelated business taxable income — UBTI. If your UBTI exceeds $1,000 in a year, the IRA itself must file a tax return and pay tax. This defeats the purpose of tax-sheltered retirement accounts. Most advisors recommend keeping MLPs in taxable brokerage accounts and using their tax-advantaged yield in a different way within IRAs.

There are workarounds. C-corp pipeline companies like Kinder Morgan (KMI) and Williams Companies (WMB) offer similar AI infrastructure exposure without the K-1. They trade lower yields — KMI yields roughly 4.5% versus 7%+ for MLP peers — but the tax simplicity is worth the yield gap for many retirees. Closed-end funds like the Alerian MLP ETF (AMLP) handle K-1 logistics at the fund level, issuing a single 1099 to holders. The trade-off is an expense ratio and potential premium/discount to net asset value.

Other ways to play the AI infrastructure income thesis

MLPs are the most direct vehicle, but they’re not the only one.

C-corp pipeline companies — Kinder Morgan, Williams, ONEOK — offer AI data center gas transport exposure without partnership tax complexity. Yields range from 4% to 6%. Lower than MLPs, but no K-1, no UBTI concerns, and they’re eligible for lower long-term capital gains rates.

Infrastructure CEFs — Closed-end funds like Cohen & Steers Infrastructure Fund (UTF) and Reaves Utility Income Fund (UTG) hold diversified portfolios of infrastructure assets including pipelines, utilities, and renewable energy. They trade at discounts to NAV, offer 6% to 8% yields, and distribute as 1099 income. Less direct AI exposure, but diversified across the entire infrastructure landscape.

Royalty trusts — Structures like Black Stone Minerals (BSM) and Permian Basin Royalty Trust (PBT) own mineral rights and collect royalties on oil and gas production. These are more commodity-price-sensitive than pipeline MLPs. When oil drops, royalties drop. But they offer yields in the 8% to 10% range and trade on the same AI-energy demand thesis.

Direct data center REITs — Equinix (EQIX) and Digital Realty (DLR) own the data centers themselves, not the power infrastructure. Different risk profile — they’re exposed to leasing demand and tenant concentration rather than energy markets. Yields are lower, typically 2% to 3%, but growth has been strong as AI drives data center leasing to record levels.

When the thesis breaks

Every investment thesis has failure modes. The AI infrastructure income thesis is no exception.

AI capex slows. The current buildout assumes AI demand continues growing exponentially. If the next generation of models disappoints, or if inference costs collapse faster than demand grows, data center construction could decelerate. Pipeline contracts already signed are secure — 15-to-20-year minimum volume commitments don’t go away — but the growth trajectory would flatten.

Gas-to-power faces regulatory headwinds. California and New York are hostile to new natural gas infrastructure. Even Texas, the most pipeline-friendly state, faces growing opposition to gas-fired generation from environmental groups. If regulatory pressure limits behind-the-meter gas generation, the addressable market shrinks.

Alternatives arrive faster than expected. Grid-scale batteries, small modular reactors, or long-duration storage could reduce the gas generation buildout. A breakthrough in any of these could change the calculus within five to ten years. That’s not a risk to current contracts, but it caps the long-term growth runway.

Commodity price shock cuts volumes. Pipeline fee-based revenue is insulated from commodity prices, but volumes aren’t. A deep recession that reduces industrial gas demand could lower throughput volumes. MLPs with diversified customer bases — serving residential, commercial, industrial, and power generation markets — are more resilient than those concentrated in one segment.

Distribution cut risk. This is the largest risk for income investors. MLP distributions are not guaranteed. The 2020 COVID crash triggered distribution cuts across the sector. One major midstream MLP cut by 50%. Plains All American cut by 58%. Even well-managed MLPs can reduce payouts when cash flow declines. The strongest operators — those with investment-grade credit ratings, low leverage, and high distribution coverage ratios (above 1.5x) — are the most likely to maintain and grow distributions through downturns.

The bottom third

I’m not writing this to pitch a stock. I’m writing it because the concept itself is worth understanding.

AI represents the largest energy infrastructure buildout since the interstate highway system. Every watt of that buildout passes through pipes, wires, and turbines owned by someone. The companies that own those assets collect fees. The structures that distribute those fees to investors — MLPs, C-corps, CEFs, royalty trusts — are some of the highest-yielding income vehicles in public markets.

The thesis doesn’t require believing AI will take over the world. It only requires believing AI will continue consuming electricity. That’s a much lower bar. AI data centers already exist. Their power demand is already contracted. The buildout is already happening. The question is whether the income from that buildout can support a retiree.

At current yields and expected distribution growth, yes — for the right investor who sizes the position, manages the tax complexity, and holds through the cycles. The 2,000% return hype is marketing. But the underlying mechanism — owning infrastructure that collects fees from the largest technology buildout in history — is not hype. It’s structural.

The AI income thesis works because it doesn’t depend on AI being magic. It depends on AI needing power. Power needs pipes. Pipes pay fees. Fees fund retirement.

That’s the whole chain. Everything else is a number.