AI Has a Pipeline Problem: Part 2
Capital, jobs, energy stacks, and the local constraints that decide the winners
By Bryan J. Kaus
“The first principle is that you must not fool yourself, and you are the easiest person to fool.” — Richard P. Feynman
Part 1 made the case that AI demand is showing up in the physical infrastructure layer: utility capital plans, midstream backlogs, financing packages, construction labor, and the storage and pipeline footprint behind it. The demand is real. The buildout is real. The system is being asked to deliver more than it has in decades.
This is the part where discipline matters.
The Problem
The market tends to compress complexity into a label.
AI infrastructure.
Power demand.
Data centers.
Natural gas.
Grid modernization.
Transmission.
Nuclear.
Cooling.
Transformers.
Each label contains real opportunities. Each also contains ways to lose money.
That is where discipline matters.
The market can be directionally right about a theme and still overpay for the wrong assets inside it. Railroads. Telecom fiber. Shale. Renewables. The first wave of LNG. E-commerce logistics. Cloud. Every major buildout attracts capital. Every major buildout creates winners. Every major buildout also creates overcapacity, failed projects, and narrative-driven investments that do not earn their cost of capital.
The operators know this.
When asked whether Kinder Morgan would consider another pipeline expansion into the Northeast given how tight gas supply into New England has become, CEO Kim Dang did not soften the answer. The need was clearly there, she said. But the company would have to have certainty on state permits and the commercial support to underwrite the project. “We’ve gone down that road once. We wrote off a fair amount of capital and I think that’s not something that we are interested in doing again.”
That is the operator’s lesson made explicit.
A project can have real demand sitting in front of it and still destroy capital if the permitting environment, the contract structure, or the commercial backing is not there. Discipline is not abstract. It is what an operator who has actually written off capital sounds like when asked to do it again. And this happens all the time.
AI infrastructure will not be exempt from this pattern.
The buildout is real.
But the buildout will not be evenly distributed.
Some projects will be backed by investment-grade customers, long-term contracts, strong interconnection positions, credible power supply, reasonable leverage, and assets with value beyond one customer or one use case. Others will depend on speculative load, weak tenant credit, unrealistic timelines, high leverage, uncertain interconnection, rising construction costs, and a perfect capital market.
Those are very different investments.
This is why the Richard Kinder frame is so useful.
He is not telling the market to buy the slogan.
He is pointing to the constraint.
And in infrastructure, constraint is often where value is created.
Constraint as Opportunity
A constraint is not merely a problem.
A constraint can be an investable bottleneck.
If power is scarce, the assets that can deliver power become more valuable.
If gas deliverability is scarce, pipelines with the right location and capacity become more valuable.
If transmission is scarce, utilities and developers with viable projects, regulatory support, and cost recovery mechanisms become more important.
If transformer supply is scarce, procurement discipline and equipment access matter.
If interconnection is scarce, projects already positioned in the queue have an advantage.
If credible load is scarce, counterparties and contracts become decisive.
If capital is scarce, balance sheets matter.
This is where investors need to separate exposure from advantage.
Exposure is easy.
Advantage is harder.
A company can mention data centers on an earnings call and have no durable economic edge. Another company may never use flashy AI language but own the corridor, pipe, substation, generation site, water rights, or contracted position that actually matters.
That is the difference between thematic investing and underwriting.
Smart capital does not chase “AI infrastructure” in the abstract. It asks who the customer is, whether the load is contracted, whether the counterparty is creditworthy, who pays if the project is delayed, who absorbs cost overruns, whether generation and transmission and gas deliverability are actually available, whether the interconnection is real, whether the permitting path is credible, whether the equipment is available, whether the return is regulated or contracted or merchant or speculative, and whether the asset still has value if the original demand case is reduced.
That last question may be the most important one.
The best infrastructure investments usually have multiple ways to win. They serve durable demand. They sit in constrained corridors. They have credible customers. They benefit from real physical scarcity. They generate cash flow even if the most aggressive version of the theme does not materialize.
The weaker investments depend on the perfect version of the story. Perfect load growth. Perfect financing. Perfect timing. Perfect technology adoption. Perfect permitting. Perfect utilization. Perfect customer economics.
Perfect stories are expensive.
And fragile.
The Financing Layer
The capital requirement is also where this cycle connects to the credit markets directly.
If the world needs trillions of dollars of AI-related data center and hardware investment, and hyperscalers fund only part of that from internal cash flow, the rest has to come from somewhere. Credit markets, asset-backed finance, partnerships, project finance, leases, private capital, and structured vehicles all become part of the AI capital stack.
That changes the risk profile.
A model may be software. The financing is not.
Debt has a maturity.
Projects have budgets.
Equipment has lead times.
Power has availability constraints.
Counterparties have credit quality.
Capital has a cost.
There is also a structural timing mismatch that is forcing the system to adapt. Rabobank framed it clearly this week: grid interconnection timelines of 36 to 84 months are structurally incompatible with data center build cycles of 12 to 24 months. With grid connections taking up to seven years, hyperscalers are increasingly moving behind the meter, with more than 130 GW of capacity now proposed across the United States. Gas accounts for more than 80% of that announced behind-the-meter total.
That is the tangible version of “AI has a pipeline problem.”
The model may scale quickly.
The grid does not.
When the AI story is told only through the lens of productivity, software adoption, and model capability, it can feel abstract. When it is told through the lens of power, pipelines, utilities, and project finance, it becomes more concrete.
That is healthier.
The more capital-intensive the cycle becomes, the more important it is to ask whether the cash flows justify the investment.
There is also a tenant-credit issue. Not all AI demand is created equal. A lease or project backed by a major hyperscaler with a fortress balance sheet is different from one backed by a newer AI cloud provider, a speculative tenant, or a customer whose economics depend on future capital markets remaining wide open.
The market will increasingly need to distinguish between real contracted demand and aspirational demand, between investment-grade counterparties and weaker tenants, between regulated cost recovery and merchant exposure, between essential infrastructure and theme-adjacent assets, between durable bottlenecks and temporary shortages.
That is where winners and losers will separate.
The Jobs Story
There is another dimension that deserves attention.
This buildout is not just financial. It is industrial.
If even a meaningful portion of the planned AI and power infrastructure buildout occurs, it will create work across the physical economy: pipeline construction, electrical work, civil engineering, data center construction, gas generation, transmission and distribution, substations, control systems, cooling, maintenance, field services, project management, welding, heavy equipment, environmental review, land work, and operations.
Meta’s El Paso project is one example. Meta has said the facility is expected to support over 4,000 construction workers at peak and more than 300 permanent jobs once completed.
That does not make every project good.
It does not eliminate environmental concerns, community concerns, affordability concerns, water concerns, reliability concerns, or capital discipline concerns.
But it does mean the AI story is not confined to Silicon Valley, Seattle, Austin, or Wall Street.
The digital economy is creating demand for people who build things.
That matters.
One of the recurring mistakes in modern markets is treating “technology” as if it floats above the real economy. It does not. Technology eventually lands somewhere. It lands in land use, energy use, water use, labor markets, municipal planning, utility bills, capital budgets, and local politics.
AI is no different.
The model may be virtual.
The buildout is not.
The Energy Stack
This is also an energy-stack story.
The market likes to talk about energy transition as if sources simply substitute for one another. In practice, systems stack. Demand grows. Reliability matters. Intermittency matters. Dispatchability matters. Location matters. Existing infrastructure matters. Time matters.
Data centers need reliable power.
That does not mean only natural gas will win. It does not mean renewables do not matter. It does not mean nuclear will not be part of the mix. It does not mean storage, demand response, efficiency, or grid optimization are irrelevant.
It means the system has to work.
Hyperscalers may want clean power. Utilities may need dispatchable generation. Grid operators may need reliability. Customers may demand affordability. Regulators may demand prudence. Communities may demand water and land-use protections. Investors may demand returns.
All of those pressures will meet in the same place.
The physical system.
Allianz described AI as likely imposing the largest sustained demand shock on U.S. electricity infrastructure in decades, with data-center power consumption expected to nearly double by 2030, lifting the sector’s share of total U.S. electricity demand from roughly 5% to around 9%. Although planned generation additions look sufficient on paper, data centers may absorb nearly half of projected new capacity, leaving thin margins if electric-vehicle adoption or industrial electrification accelerate faster than expected.
That is the energy-stack problem.
AI is not the only demand source.
Manufacturing, electrification, cooling load, transportation, LNG, industrial growth, and normal economic activity are all competing for the same system.
Natural gas is positioned to matter because it can provide dispatchable power at scale, and because the pipeline and generation ecosystems already exist in large parts of the country. But that does not make every gas project attractive. The winners will likely be the projects tied to credible load, strategic locations, and disciplined contracts.
The same applies to transmission. Transmission is essential, but transmission is difficult. It faces permitting, siting, cost allocation, landowner concerns, and regulatory complexity. Everyone wants more deliverable power. Fewer people want the line near them or the cost on their bill.
That is the real constraint.
Not desire.
Execution.
The Local Constraint
There is also a local politics and community acceptance layer.
The proposed Stratos data center in Utah is a useful illustration. The project, developed by Kevin O’Leary’s (aka “Mr. Wonderful”) O’Leary Digital, would span 40,000 acres of private land plus 1,200 acres of military and state-owned property in Box Elder County. At full buildout, the campus would reach 9 gigawatts, all produced on-site through a connection to the Ruby Pipeline, more than double Utah’s current statewide electricity use of roughly 4 gigawatts. The site lies on the northern shore of the Great Salt Lake, and protests have raised concerns over Utah’s water shortages and the struggling Great Salt Lake, which supports millions of migrating birds and is drying into a serious dust problem impacting public health.
Whether that specific project ultimately proves viable is not the point.
The point is that physical infrastructure has a constituency.
It has neighbors.
It has environmental constraints.
It has water constraints.
It has ratepayer implications.
It has local political risk.
That is why the phrase “AI infrastructure” is too clean. The reality is messier. Every serious project has to move through a world of people, permits, power, water, land, cost recovery, and execution.
The market can model the load.
Operators still have to build the system.
Winners and Losers
There will be winners and losers.
Some utilities will grow rate base intelligently and protect affordability. Others may face pushback if customers believe they are subsidizing speculative load growth.
Some midstream companies will convert tight capacity and customer demand into attractive contracted projects. Others will chase marginal expansions that depend on optimistic forecasts.
Some data center developers will secure power early, build in the right locations, and match capacity to real demand. Others will discover that land without power is not a project.
Some hyperscalers will earn attractive returns because AI becomes embedded in high-value workflows, enterprise software, advertising, cloud services, cybersecurity, industrial systems, and consumer products. Others may spend heavily to maintain competitive position without earning proportional incremental returns.
Some investors will own the bottlenecks.
Others will own the narrative.
That is the difference.
The winners will be the ones closest to indispensable infrastructure, credible demand, and disciplined execution. The losers will be the ones that confuse thematic exposure with economic moat.
The Point Taken
Richard Kinder’s comments are useful because they bring the AI conversation back to earth. He is a sage voice - and I was stunned when he appeared on Bloomberg TV and the hosts said they had to look him up - the man is a legend.
The question is not whether AI is real.
It is.
The question is whether the physical system can scale fast enough, cheaply enough, and reliably enough to support the demand being assumed, and which projects, companies, and operators will capture durable value in the process.
That is where the constraint becomes both a risk and an opportunity.
The risk is that markets overbuild the wrong assets, finance weak projects, assume every announced data center becomes real load, and price infrastructure as if execution were automatic.
The opportunity is that credible demand is colliding with constrained systems: gas pipelines, power generation, transmission, substations, equipment, cooling, storage, construction labor, and capital.
Some constraints will become investment opportunities.
Some constraints will become growth bottlenecks.
Some projects will create durable value.
Some will become stranded monuments to a cycle that moved faster on paper than in the field.
That is why this is not a call to chase everything attached to AI.
It is a call to underwrite the physical chain with discipline.
Contracted demand matters.
Counterparty quality matters.
Power availability matters.
Gas deliverability matters.
Permitting matters.
Capital structure matters.
Execution matters.
Time matters.
The AI economy may be built in code, but it will be constrained by concrete, copper, steel, gas, water, wires, workers, and time.
That constraint is not just the risk.
Properly underwritten, it may also be the opportunity.