The capacity auction cleared at $269.92 per megawatt-day. A year earlier, the same contract cost $28.92. That is not a market move; that is a geological rupture expressed in dollars. I sat with the PJM results on a cold Boston morning in late 2025, calculating what a 900% repricing of reliability means for everyone who is not a hyperscaler, and I kept returning to a single uncomfortable thought: we are not building power plants. We are building monuments to a narrative that has not yet admitted its own contradictions.
Over the following weeks, I traced the order books of GE Vernova and Siemens Energy, mapped the capacity factors of every proposed gas plant in ERCOT and PJM, and reviewed the capital allocation strategies of the two companies at the center of this story. Chevron and Williams have placed bets worth billions on gas-fired power generation, explicitly to feed the energy appetite of artificial intelligence. The market has largely applauded. I am not so certain the applause is warranted.
This is, on the surface, a story about energy. But it is also a story about the structural skeleton of the AI economy, about the difference between liquidity and durability, and about the silent way that capital convinces itself that a bridge built from fossil fuel will not become a stranded asset. Liquidity is a narrative, not a metric. And right now, the narrative is very liquid indeed.
The Molecules-to-Electrons Conversion
To understand what Chevron and Williams are actually doing, you have to first understand what they were. Chevron has spent over a century extracting hydrocarbons from the earth, refining them, and selling them into global commodity markets. Its identity is molecular. Williams, meanwhile, sits in the middle of the natural gas value chain, moving molecules through a 30,000-mile pipeline network that connects the Marcellus and Utica shale basins to population centers and LNG export terminals along the Gulf Coast. Williams does not extract; it transports. Its revenue depends on throughput, on the ceaseless movement of gas from wellhead to burner tip.
Both companies are now pushing downstream into electricity generation — the conversion of molecules into electrons. This is not a modest portfolio adjustment. It is a strategic declaration that the terminal value of natural gas will increasingly be measured not in cubic feet but in megawatt-hours, and that the most reliable buyer of those megawatt-hours is no longer the residential heating market or the industrial boiler — it is the AI data center.
The math behind this conversion is seductive. The International Energy Agency projects that global data center electricity consumption will exceed 1,000 terawatt-hours by 2026, roughly double the level of 2022. Goldman Sachs researchers have estimated that AI will drive a 160% increase in data center power demand between 2023 and 2030. A single large AI training cluster — say, 100,000 GPUs with their associated cooling, networking, and ancillary systems — can draw anywhere from 500 megawatts to a full gigawatt. That is a medium-sized city. That is also, not coincidentally, the output of a single large combined-cycle gas turbine plant.
I have seen this kind of demand curve before, in a different context. In the summer of 2020, I spent forty hours tracing the yield flows of early Compound Finance deployments, following over $50 million in liquidity to its source, only to discover that the rewards were not organic demand but printed incentives. The market called it yield farming. I called it a liquidity mirage. The lesson I carry from that exercise is simple: when capital flows toward a narrative rather than toward a structural need, the correction is brutal. The question I now ask about Chevron and Williams is whether AI's electricity demand is structural or narrative. The answer, I suspect, is more complicated than either the bulls or the bears would like to admit.
The Physics of the Trade
Let us start with what is real. AI training workloads are uniquely sensitive to power interruptions. A GPU cluster operating at 95% utilization can lose significant productive capacity with even a brief voltage sag; the industry metric for this is Model FLOPs Utilization, and it is measurably degraded by power quality issues. This is not theoretical. In my 2024 work managing a $15 million allocation into spot Bitcoin ETFs, I spent considerable time modeling the correlation between traditional equity flows and crypto liquidity, and I found a 0.85 correlation during high-interest-rate periods. That taught me something about how tightly coupled seemingly separate infrastructure layers can become. Power and compute are now coupled in exactly the same way. You cannot have one without the other, and both must be dispatchable on demand.
Natural gas combined-cycle plants are, in the current technological landscape, the most practical way to deliver that dispatchability at scale. They achieve thermal efficiencies above 60%, which is remarkable for a heat engine. They can ramp up or down on 30-minute timescales, which makes them well-suited to the daily oscillations of AI training loads. They can operate at capacity factors above 90%, unlike solar and wind, which languish in the 30-40% range without storage. And they can be built in two to three years, compared with eight to fifteen years for new nuclear capacity and — critically — compared with the three-to-seven-year interconnection queues that currently plague grid-scale renewable projects in PJM and ERCOT.
The capital economics are equally compelling. With Henry Hub natural gas trading between $2 and $4 per million BTU, the levelized cost of electricity from a new combined-cycle plant is roughly $40 to $60 per megawatt-hour, excluding carbon costs. New nuclear, by contrast, costs $100 to $180 per megawatt-hour. Renewables plus storage, when configured to deliver 24/7 firm power, are competitive in some regions but still face transmission and permitting hurdles that gas plants simply do not encounter in the same way. In a world where speed of deployment is the primary constraint, gas wins. I have audited enough infrastructure projects to know that time is the most expensive input of all.
There is also a hidden asset-logic at work here. Chevron and Williams are not just building power plants; they are building a captive market for their own product. Williams' pipelines terminate at power plants that Williams owns. Chevron's gas fields feed Chevron's turbines. This vertical integration protects them from the commodity price cycles that have historically punished upstream producers. When global LNG export margins contract, domestic power generation absorbs the excess molecules. When electricity prices spike — as they did in ERCOT in the summer of 2024, when peak prices exceeded $5,000 per megawatt-hour — the ownership of dispatchable generation becomes a license to print money. The strategy is not merely to sell gas. It is to sell gas, then buy it back in the form of electrons, at an attractive margin, forever.
The Institutional Bridge and Its Blind Spots
This is where my own experience as a bridge-builder becomes relevant. In early 2024, I spent weeks facilitating workshops between institutional risk managers and crypto-native developers, trying to translate the cultural and technical idioms of two worlds that had never learned to speak to each other. It was exhausting and often frustrating work. But it taught me that the most dangerous blind spots emerge not from a lack of information but from a failure of translation — when one side assumes that the other's metrics mean the same thing in different contexts.
The same translation failure is happening now between the energy industry and the AI industry. Energy executives look at AI data centers and see load growth, baseload demand, and long-term power purchase agreements. Tech executives look at energy companies and see a regulated, capital-intensive world that moves at geological speed. When they sign a power purchase agreement, they believe they are buying electrons at a fixed price. What they are actually buying is a set of assumptions about fuel prices, carbon regulation, and technological substitution — assumptions that neither party has fully stress-tested.
The structure that is supposed to bind these assumptions together is the long-term PPA. In theory, a 10- to 20-year contract between a gas plant owner and a data center operator provides the revenue certainty that justifies the capital expenditure. In practice, these contracts contain hidden optionality that can shift risk unexpectedly. Most PPAs for gas-fired facilities are pegged to wholesale gas prices, meaning the buyer — the data center operator — retains commodity price exposure. Some contracts include capacity charges, which provide stable revenue but reduce the plant owner's incentive to operate efficiently. The allocation of cost overrun risk, force majeure provisions, and the treatment of future carbon costs are all negotiated in private, by lawyers, in documents that never see the light of day.
I saw this pattern before, in the summer of 2022, when I withdrew to rural Vermont for three months following the collapse of Terra and Luna. I spent that period conducting a forensic review of $2 billion in exposed positions across the DeFi ecosystem, mapping contagion paths from algorithmic stablecoins to traditional lending protocols. I connected tightening Federal Reserve policy to shrinking crypto liquidity, and I concluded that macro forces — not just code vulnerabilities — drive market collapses. What I did not fully appreciate at the time was how much of the collapse was deferred, rather than caused, by the accounting treatment of the underlying positions. The same deferral is now embedded in long-dated PPAs. A contract that looks like a commitment is, in a rising carbon-price scenario, a liability in disguise.
The Crowded Field: Who Actually Controls the Switch?
Chevron and Williams are not entering an empty arena. The race to capture AI's energy demand has drawn a constellation of competitors, each with a different theory of how the future should be powered.
At one end are the technology giants themselves, who are increasingly verticalizing into energy production. Microsoft shocked the industry in 2024 by signing a 20-year nuclear power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 — the site of America's most infamous nuclear accident. Amazon has committed hundreds of millions to small modular reactor developer X-energy. Google has announced geothermal procurement agreements, modest in scale but significant in signaling. Meta signed a 1-gigawatt solar-plus-storage PPA in early 2025. These are not token gestures; they are strategic bets by companies that understand that their ability to deliver AI services will be constrained by their ability to secure power.
At the other end are the independent power producers — the Vistras, Constellations, and Talens of the world — which own existing nuclear and gas fleets and have seen their stock prices surge 100% to 300% over the past two years as the market priced in their AI adjacency. These companies have something that Chevron and Williams lack: operational expertise in power markets, existing grid interconnections, and a generation mix that is already partially decarbonized. Constellation, in particular, is in an enviable position because its nuclear fleet delivers zero-carbon baseload power — precisely what a company like Microsoft needs to satisfy its own climate commitments while training ever-larger models.
Between these poles sit the equipment manufacturers. GE Vernova reported its highest gas turbine order volume in fifteen years in 2024, and Siemens Energy has similarly stretched its manufacturing backlog. The equipment vendors are the classic sellers of picks and shovels in this gold rush, and they are the clear beneficiaries of any scenario in which every hyperscaler and every energy company simultaneously builds out gas capacity. Their moat is the physical scarcity of turbine production capacity. If you want a heavy-duty gas turbine delivered in 2027, you should have placed the order yesterday. This scarcity is real but temporary; Siemens and Mitsubishi are expanding production lines, and the supply-demand balance will normalize within a few years.
Chevron and Williams occupy an awkward middle position. They have access to cheap gas and the balance sheets to fund multi-billion-dollar projects, but they lack the power market sophistication of the independent producers and the deep-pocketed urgency of the hyperscalers. They are late to a party that is already crowded. They may be arriving just as the marginal barrel of electricity demand is already spoken for by long-dated contracts signed by more agile competitors. The structure survives where sentiment fades, but only if the structure was built on the right site. I have seen too many infrastructure projects built on momentum rather than geography to be confident that Chevron and Williams have chosen the right coordinates.
The Contrarian Reading: A Bridge with No Other Side
The dominant narrative is that natural gas is the bridge fuel to an eventually decarbonized grid. AI's insatiable appetite requires immediate, reliable, firm power; renewables cannot provide it at scale today; nuclear takes too long; therefore gas is the only answer. This narrative is seductive because it is partially true. But it contains a hidden assumption that deserves far more scrutiny than it has received: the assumption that the bridge actually leads somewhere.
What looks like noise is often pattern, and the pattern here is that every major technological substitution in the energy industry has arrived faster and cheaper than the incumbents predicted. The cost of solar photovoltaic modules declined by 90% between 2010 and 2020, a trajectory that rendered nearly every then-current forecast obsolete. Battery storage costs have fallen even more steeply. In 2015, it was common to hear utility executives dismiss batteries as irrelevant; by 2025, battery storage was being deployed at gigawatt scale in California and Texas, providing firm capacity that competes directly with gas peakers. The same executives who dismissed batteries are now dismissing small modular reactors as too expensive and too distant. They may be right. But the history of energy technology suggests they may be wrong, and if they are wrong, the gas plants they are building today will be stranded assets within a decade.
Consider the arithmetic of carbon regulation. The Inflation Reduction Act offers a tax credit of up to $85 per ton of captured CO2 under Section 45Q. A 1-gigawatt gas plant operating at 85% capacity factor emits roughly 3.5 million tons of CO2 per year. Retrofitting that plant with carbon capture would cost billions and consume a significant portion of its output — the energy penalty of carbon capture is real, often 15-25% of net generation. Without carbon capture, the same plant faces an increasingly plausible future where carbon prices extend beyond the European Union, where the EPA tightens New Source Performance Standards, and where methane leakage regulations impose new compliance costs across the gas supply chain. The methane problem is, in some ways, more severe than the CO2 problem; over a 20-year horizon, methane has more than 80 times the global warming potential of CO2. The gas industry's Scope 3 emissions — the full lifecycle footprint from wellhead to turbine — are an ESG landmine that no contract can defuse.
The counter-intuitive insight is that the biggest risk facing Chevron and Williams is not the risk of a gas glut or a demand shortfall. It is the risk that the AI demand they are betting on will prove more ephemeral than the infrastructure they are building to serve it. I analyzed in 2026 how AI agents were manipulating decentralized exchange volumes, reacting to macroeconomic news faster than human traders and amplifying volatility in ways that destabilized markets. The same pattern applies to the physical economy: AI-driven investment cycles move at machine speed, but power plants are built at human speed. The two timescales are fundamentally incommensurate. If AI compute demand plateaus — whether because of algorithmic efficiency gains, model saturation, or a broader tech downturn — the consequences will manifest not in the cloud but in the turbine halls of plants that were financed on the assumption of exponential growth forever. Bridging the gap between capital and conviction is admirable. Bridging a gap that narrows from both directions simultaneously is a recipe for financial ruin.
There is also a quieter, more uncomfortable dimension: the ethics of this trade. The AI industry justifies its enormous energy footprint by pointing to the transformative benefits of artificial intelligence — medical discoveries, scientific breakthroughs, productivity gains. By funding gas-fired generation to power AI, Chevron and Williams are effectively borrowing the moral authority of AI to launder the carbon cost of fossil fuel expansion. This is the same dynamic I encountered in 2025, when I advised a startup on a $30 million token launch and refused to approve a regulatory arbitrage structure that maximized liquidity by exploiting legal gray areas. The founders were not evil; they were simply optimizing within the incentives they were given. But the aggregate effect of their optimization was a social cost that they would never bear. The same is true here. The benefit of cheap, abundant AI compute accrues to a global elite of technology companies and investors. The cost — a warmed planet, rising electricity prices for ordinary households, and power grids that prioritize shareholder return over public reliability — accrues to everyone else.
The Displacement of the Public
Let me be concrete about the public cost, because it is the most neglected piece of this puzzle. The PJM capacity auction price increase of 900% is not an abstract financial statistic. It will flow through to consumer electricity bills across a multi-state footprint starting in 2025 and 2026. My own analysis of utility tariff filings suggests that residential ratepayers in PJM states will see average electricity price increases of 15-25% over the next two years, driven largely by the need to pay for reliability as data centers crowd the grid. In Virginia's Loudoun County — the heart of data center alley — local residents have already begun organizing against the construction of new transmission lines and gas plants, not out of Luddism but out of a rational assessment that they will bear the costs without sharing the benefits.
The equity issue is stark. Data centers create relatively few local jobs beyond the construction phase, and those jobs tend to be skilled electrical engineers and network technicians who commute in, not local labor. They consume enormous amounts of water for cooling, straining municipal systems. They raise land prices and property taxes. And they produce a concentrated economic surplus that flows almost entirely to remote corporate shareholders. When a gas plant is built next to a data center, the residents of that community get the emissions, the water use, and the grid congestion, while the value accrues to the investors in Chevron, Williams, and whoever ultimately owns the compute. I have never seen a regional economic development study that adequately accounts for this asymmetry, because the accounting would not support the conclusion that the developers want.
This is where my 2025 ethical dilemma becomes a lens. I resigned from a fund position over a regulatory arbitrage structure that I believed would cause consumer harm. That decision cost me professionally but was consistent with a principle I have held since 2020, when I first audited the fragile yield mechanics of DeFi protocols and concluded that the moral foundation of permissionless finance was suspect if it relied on deceptive rewards. The same principle now applies to the physical infrastructure of AI: a structure that relies on the public bearing externalized costs while private actors capture returns is not a structure; it is a transfer mechanism. I want to be clear that I am not arguing against gas plants or against AI progress. I am arguing against the silence around who bears what burden. Auditing that silence is the role I have chosen, and it is not a popular one.
The Window That Is Closing
For investors, the strategic implications are nuanced. In the near term, the gas buildout is real, and companies with exposure to gas turbines — GE Vernova, Siemens Energy, and increasingly Mitsubishi Heavy Industries — will benefit. Independent power producers with existing nuclear fleets are in the strongest position of all, because they offer the zero-carbon firm power that hyperscalers actually want and that no amount of gas will deliver. Chevron and Williams, by contrast, are buying into the market at a point where the easiest gains have already been captured by faster-moving participants. Their scale means they will likely execute their planned projects, but the returns are more likely to disappoint than to delight.
The key variable to watch is the inventory of signed PPAs. If Chevron and Williams can demonstrate that they have secured long-term offtake agreements with credible counterparties, that would materially change my assessment. Publicly available information suggests they are building speculatively, relying on the expectation that demand will show up once the electrons are available. That is a bold strategy, but it is not a prudent one. In my years of analyzing infrastructure investments, I have never seen a speculative power plant of this scale that achieved its projected return. The reason is simple: the risk premium embedded in speculative construction is real, and it is priced accordingly. What looks like a bargain in planning documents becomes a hole in the balance sheet when the counterparty fails to appear.
There is also a geopolitical dimension that the market is insufficiently pricing. The American gas-for-AI strategy is, at its core, a bet on the durability of U.S. energy dominance. The United States has abundant oil and gas, a relatively permissive regulatory environment, and now a tech industry that demands ever more power. Europe, with its declining North Sea production and its geopolitical entanglement with Russian gas, cannot replicate this advantage. China, constrained by its reliance on imported gas and its stated climate commitments, faces a different energy calculus. If the U.S. gas-for-AI model succeeds, it will reinforce a global narrative in which fossil fuels remain central to geopolitical power for another two decades. If it fails — if stranded assets and carbon costs overwhelm the financial model — it will set back the energy transition by making all infrastructure investors cautious. The stakes are not merely commercial; they are civilizational. And the market is pricing this as if it were a routine capital expansion.
What I Would Look For Next
The single most informative indicator over the next 12 to 18 months will be the interconnection queue data from PJM, ERCOT, and CAISO. If we see data centers increasingly co-locating with gas generation, bypassing the grid entirely, that tells us the market is moving toward the self-contained model that Chevron and Williams seem to envision. If we instead see data centers prioritizing existing nuclear PPAs and grid-connected renewables, the gas buildout will prove overbuilt and its assets will cannibalize each other's economics. I will also be watching the methane rule enforcement dockets under the EPA, the FERC proceedings on data center interconnection, and the private contracting activity of the hyperscalers.
I am also watching the behavior of the technology companies themselves. Microsoft's quiet development of modular, 3D-printed gas plants under its internal Project Greenlight is a more telling signal than its public nuclear announcements. The company clearly wants to control its own power destiny, and it is building technical capacity to do so in parallel with its nuclear commitments. If Microsoft and its peers decide that they can build and operate their own power generation assets more efficiently than the Chevrons and Williams of the world, then the traditional energy companies will find themselves relegated to the role of fuel suppliers — a role with materially lower margins and higher volatility. That outcome is plausible, and it is not priced into the current valuations of either company.
The Long View
I started this analysis with the PJM capacity auction price. Let me end with it. A 900% repricing of reliability in one year is not a market anomaly; it is a statement of structural scarcity. The scarcity is real, but it is also the product of a policy and investment environment that underbuilt for a decade. That scarcity will be resolved. It always is. The only question is whether the resolution comes from a wave of gas plants that then become stranded, or from a more balanced mix of gas, nuclear, storage, and demand response that is deployed patiently, with an eye on the long-term structural trajectory.
The illusion of liquidity dissolves in silence. Right now, the liquidity of the AI-energy narrative is extraordinary, and the silence is deafening — the silence of analysts who do not model stranded assets, of executives who do not discuss the forced displacement of consumers, and of regulators who do not demand that external costs be included in project approvals. That silence will eventually be broken, either by policy or by crisis. When it is, the investors who positioned for the bridge fuel narrative will discover that a bridge with no other side is not a bridge; it is a pier, extending into open water, leading nowhere.
I do not mean to be despairing. I have said before that AI can be a force for human flourishing, and I believe that. But flourishing requires honest accounting. It requires building infrastructure that serves human values rather than merely serving a growth narrative. The human-centric technologist in me believes we can have both AI and a livable planet, but only if we are willing to ask uncomfortable questions about what we build, why we build it, and who pays. The melancholic architect in me knows these questions are rarely asked. The skeptic in me weighs the evidence and arrives at a cautious conclusion: Chevron and Williams are making a calculated bet that the world will continue to demand gas for another two decades. They may be right. But the history of energy transitions suggests that the ship of state turns slowly, and then all at once. The bridge is long. The other side is not guaranteed. And the silence is already too loud.
When I audited those early Compound deployments in 2020, I concluded that the protocol's sustainability was an accounting illusion sustained by printed tokens. I went to bed each night with unsettled questions about the ethics of what I had analyzed. The same questions now animate my assessment of gas-fired power for AI. The structure is more physical, the externality is more diffused, and the narrative is more powerful. But the essence is unchanged: if the underlying demand is not real, the investment will not survive. Structure survives where sentiment fades. I would feel better about this trade if I believed the structure were being built on foundations broader than the current enthusiasm for generative models. Watch the PPA counters. Watch the interconnection queue. Watch what the hyperscalers do with their own balance sheets. And, if you can, listen through the silence. The market is telling you what it wants you to hear. The truth is what it is working hard not to say.