The New Proof of Work: How Chevron and Williams Are Betting That AI Never Compresses Its Energy Problem

Here is the raw fact that every AI narrative analyst keeps tripping over: PJM's capacity auction — the mechanism that prices the right to keep the lights on across 13 mid-Atlantic and Midwest states — jumped roughly 900 percent from one delivery year to the next. Not a stock. Not a token. A physical market that was priced at $28.92 per megawatt-day and now commands $269.92. GE Vernova, meanwhile, just logged its strongest gas turbine order book in fifteen years, and the only reasonable reading is that somebody — a consortium of somebody — has decided that the scarcest resource in the digital age is not the GPU, the chip, or the model. It is the electron. It always was. Now Chevron and Williams, two of the most conservative names in American hydrocarbons, are reportedly betting billions on gas-fired power plants designed explicitly to feed AI data centers. The first reaction of most crypto natives will be: so what, that is an energy story, not a Web3 story. That first reaction is exactly the kind of dismissal that causes people to miss the cycle. Because what Chevron and Williams are doing is not an energy story at all. It is a settlement layer story. And the industry that has the deepest experience with settlement layers, with concentration, with the collapse of abstraction into physical constraints, is ours. Scarcity is a narrative we agreed to believe. But open the hood of that narrative and underneath it sits a second, less flexible truth: every block, every rollup, every inference, eventually invoices the same thermodynamic debt collector. Let's establish what actually happened. Chevron, a crude-centric exploration and production giant, and Williams, a natural gas pipeline and midstream behemoth, are both moving downstream into power generation — specifically, combined-cycle gas plants sized to run behind the meter, adjacent to, or in dedicated service for, hyperscale AI facilities. The "billions" figure is not an accident; it signals a multi-year capital commitment that, for Williams, represents a meaningful percentage of its market capitalization. The strategic translation is blunt: these companies were in the business of selling molecules, and they now want to sell electrons. Better still, they want to sell firm, dispatchable, grid-quality electrons at a time when hyperscalers are desperate for exactly that. I have spent my career watching abstraction layers eat the world and then hit a physical wall. In 2017, when I audited Raiden Network and state channels as part of a six-week deep dive into early Ethereum Layer 2, I wrote a fifteen-page thesis arguing that off-chain payment channels lacked economic security guarantees. The L2 pitch back then was pure efficiency — move computation off the mainnet, move collateral off the mainnet, everything becomes frictionless, throughput becomes infinite. It was a beautiful story, and it missed the same thing the AI-energy story is missing today: efficiency does not remove constraints, it only relocates them. State channels relocated the constraint to liquidity providers and watchtowers. AI relocates the constraint to the grid. And the grid is where reality bats. Here is the technical canvas. A hyperscale AI cluster is not a website. It is a machine that needs 500 megawatts to a gigawatt, with a capacity factor requirement above 90 percent, because modern GPU utilization (MFU) is brutally sensitive to voltage sag, thermal events, and frequency deviation. Renewable generation delivers capacity factors of 30 to 40 percent without storage. Natural gas combined-cycle (NGCC) generation delivers over 90 percent, starts from cold in roughly thirty minutes, and can modulate output faster than any baseload alternative. Its levelized cost of electricity (LCOE) sits around $40 to $60 per megawatt-hour under the current Henry Hub range of $2 to $4 per MMBtu — compared to $100 to $180 for newly built nuclear. It deploys in two to three years, not eight to fifteen. In the "time, cost, scalability" triangle, gas is the only resource that wins on all three vertices today. That is the entire thesis. It is also, I suspect, a trap — but I will get to that. The first analytical frame I want to offer is the one nobody in the AI world will say out loud: the interconnection queue is the new block space. The U.S. grid is a shared, permissioned settlement layer, and the right to plug into it is increasingly allocated by queue, not by market. In PJM, in ERCOT, in CAISO, interconnection requests stack up in queues that stretch three to five years. The physical block space of civilization is full at peak hours, and the "gas limit" is measured in gigawatts, not in millions of units. This is the same dynamics I observed in 2020 during DeFi Summer, when I spent three months modeling the Compound-Aave-UNI flywheel of collateralized debt positions. Back then, the consensus was that liquidity was infinite and synthetic assets could be redeemed forever. I published a warning that leveraged yield farming would draw down 40 percent when the cascade hit. The mechanism I was modeling was a leveraged loop built on a shared, finite base of collateral. Every participant believed the base was deep enough, right up until the moment it wasn't. AI data centers are running that exact play on the physical layer. Every hyperscaler is modeling the shared grid as an elastic resource. Every energy company, in response, is monetizing the true scarcity. Chevron and Williams are not betting on AI models. They are betting that the base of this particular collateral pool — firm, dispatchable power — is the thing that cannot be synthetically expanded in the next five years. That is a rational trade in the short term. But the word I keep coming back to is "composition." The financial abstraction of DeFi found its breakdown when all participants tried to exit the loop at once. The physical abstraction of "AI power" will find its breakdown when all participants try to enter the same capacity auction at once. That transition is already visible in the PJM numbers. The +900 percent capacity price is just the liquidation event of the old equilibrium, happening in slow motion. Yields are merely attention taxes in disguise. I mean that literally, not poetically. A 10 to 15 percent internal rate of return on a gas-fired plant built to serve AI capacity is not compensation for innovation; it is compensation for having noticed that future AI growth far outpaces the grid's expansion rate. The attention tax is paid by every hyperscaler, every GPU allocator, every token project that assumed compute would simply remain cheap. The gas plant is the validator of this cycle. It charges the fee, produces the block, and confiscates the priority fee — the mev, if you will — of every faction competing for the same precious slot. The data center that signs a 15-year power purchase agreement (PPA) with a Chevron- or Williams-backed merchant plant is paying upfront for inclusion in the physical mempool of the 2030 grid. And that is exactly why I see a fractal repetition of the Bitcoin mining story. Tracing the fractal logic beneath the chaos: years ago, small miners with a few thousand ASICs could compete with anyone. Then the margin compressed, the capital requirements swelled, and hash power concentrated into three or four giant pools, leaving the pretense of decentralized consensus hollow. The same arc is now visible in AI compute. First, anyone could rent GPUs. Then the hyperscalers bought the GPUs. Then Microsoft bought a nuclear reactor restart at Three Mile Island, Amazon poured billions into X-energy's small modular reactors, Google signed geothermal deals, and Meta bought a gigawatt of solar-plus-storage. Now Chevron and Williams are buying gas turbines. The market is simply deciding that energy providers are the next validators — and, like every validator set, it is consolidating fast. The "three pools" thesis for Bitcoin is becoming the "three power corridors" thesis for AI: ERCOT, PJM, and the Southeast. If you want to know where the value flows next, follow the electrician, not the prompt engineer. My second core insight is about the carbon ledger, because this is where blockchain actually enters the building. A 1-gigawatt gas plant running at 85 percent capacity factor emits somewhere between three and four million tons of CO2 per year. Under the Inflation Reduction Act, if that plant bolts on carbon capture and sequestration (CCS), its operators can claim Section 45Q tax credits of up to $85 per ton of CO2 — potentially transforming a politically unpalatable asset into a policy-subsidized cash machine. But here is the part I find genuinely interesting as a Web3 researcher: the buyers of power from these plants are hyperscalers who have publicly committed to 100 percent clean energy. Microsoft, Google, Amazon, Meta — all of them have pledged their datacenters to carbon neutrality. They cannot buy gas-fired electrons and maintain those promises without a massive volume of environmental attribute credits. The demand for tokenized RECs, digital carbon credits, and provable clean-power attestations — the on-chain infrastructure that has been dismissed as a niche for years — is about to become the mandatory compliance theater of the AI economy. This is the deep irony I want to sit in for a moment. Web3 spent its entire adolescence being mocked for not having "real utility." DeFi was called a casino, NFT trading was called wash trading by numbers — and in 2021 I spent eight weeks examining on-chain behavior of early crypto art collectors and found that a stunning share of high-value PFP sales were wash trades designed to inflate social signals. I called that essay "The Illusion of Ownership." Yet here we are in 2025, and the largest, most creditworthy institutions in the world — hyperscalers — are about to become the biggest buyers of tokenized environmental assets on the planet, not out of ideological fealty, but because they have painted themselves into a carbon corner and need provable digital scarcity to escape. The bug is the feature they didn't see coming: the immutable ledger was never the product; it was the eschatology, and the end of days for renewable-energy accounting is now. The narrative is shifting in exactly the direction I outlined in my agent-sovereignty research. In 2024, after the Bitcoin ETF approvals, I spent three months analyzing the tokenomics of decentralized compute networks like Akash and Render, and argued that the next dominant story would not be currency or art but agent sovereignty — the capacity of autonomous agents to hold wallets, execute trade-offs, and secure their own resources. But an AI agent does not need a wallet nearly as much as it needs a power contract. The most important existential dependency of the machine economy is not code, it is thermodynamics. An agent that cannot prove access to firm power is an agent without uptime. An agent without uptime is a god without believers. So when Chevron and Williams commit billions to gas generation, they are, whether they know it or not, building the first meaningful sovereignty infrastructure for agents — not the crypto-native, wallet-in-the-cloud version, but the physical grid-scale version. The question is whether that infrastructure will remain centralized, fossil-fired, and bank-owned, or whether it becomes distributed, auctioned, and tokenizable. I place low odds on the latter, but that is exactly why the former is a good investment and a bad civilization. Which brings me to the contrarian position I feel compelled to defend, because every bull narrative has a spectral twin. The consensus read of the Chevron-Williams move is straightforward: AI energy demand is real, gas is the only bridge, and these titans are inserting themselves into the tollbooth. My read is less charitable. I believe this massive capex is not evidence of AI's robustness; it is evidence of AI's admission of frailty. For years, the AI industry has run on the same faith that powered algorithmic stablecoins: the belief that an engineered abstraction can outrun its physical collateral. Terra's UST promised to manufacture stability from a mint-and-burn reflex. It worked in theory, it worked on the chart, and then the anchor detached. The parallel does not need to be perfect to be instructive. AI models promise to manufacture intelligence from data and flops. They are now discovering that the substrate of that intelligence — electricity — has an anchor price determined by generators, not by models. By pouring tens of billions into gas plants, the energy industry is effectively shorting the AI abstraction: they are betting in steel and concrete that AI's spectacular growth will cheerfully pay fossil rents for a decade or more, and that no fusion, no advanced nuclear, no hundred-billion-dollar storage breakthrough will arrive in time to undercut them. That is a rational trade. It is also the single most powerful counter-signal the climate world has seen in a decade, because it announces that the transition is being underwritten by transition-fuel pragmatists, not transition idealists. Still, the risk of stranded assets is enormous. If SMRs achieve commercial scale by 2032, if battery costs fall another 60 percent, this new gas fleet looks like the LNG carriers of 2020: too expensive to retire, too dirty to justify, too embedded to escape. And the burden of that misreading will not fall on Chevron's treasury. It will fall on the ratepayers, the data center host communities, and the grid users who never signed a PPA, the way the burden of Terra's failure fell on the holders who arrived last. The lens I keep returning to is one I first polished in 2022, during the LUNA collapse forensics, when I helped reverse-engineer the UST de-pegging mechanism with three other researchers and we built an open-source simulation of the death spiral. That period taught me something that has only sharpened since: the more spectacularly an ecosystem abstracts its own constraints, the more violent the moment of reification becomes. The crypto market learned that the peg is the product. The AI market is about to learn that the grid is the peg. The same applied to the regulatory dynamic. Hong Kong, where I now work, issued virtual asset licenses not because it loves innovation but because it wants Singapore's position as Asia's financial hub — pragmatism wearing a flag. The Chevron-Williams gas bet is the same shape of pragmatism: it is not a bet that gas is the future. It is a bet that the future is too slow to arrive on schedule, and that being the last credible bridge builder is the most profitable position in the queue. Let me articulate the taxonomy I am working with, because I think it helps the Web3 audience internalize this. The AI power market is currently separating into three distinct layers. The first is the physical layer: generation assets, turbines, fuel supply, and interconnection rights. The second is the financial layer: PPAs, capacity auctions, hedging instruments, and carbon credits. The third is the speculative layer: equities, project debt, tokenized energy assets, and every derivative flavor in between. Historically, the crypto industry only ever competed in the third layer — we tokenized this, we securitized that, we built lending protocols for things that did not yet exist. And we repeatedly got burned because the first layer, the physical layer, was somebody else's control surface. This time, the playbook is different. The biggest opportunities won't come from another exchange listing or another idle-sidechain narrative. They will come from building the accounting and settlement plumbing that connects layer three to layer one — registry systems that verify a tokenized REC actually retired a real credit; smart contracts that settle PPA payments against measured output; provenance graphs that defeat double-counted capacity claims. I am also watching for the securitization moment. The next casualty cycle, in my view, begins when carbon-heavy gas assets are packaged into green bonds using PPA-backed revenue streams, and the collateral is a 2040 power contract whose counterparty credit depends on a hardware company, a software company, and a grid that hasn't been built yet. As someone who has spent 29 years in this industry, I can tell you that the most dangerous sentence in finance is not "this is risk-free"; it is "the physical load will for sure be there."" A power contract is a collateralized position, and the collateral is a fuel source that nobody successfully shorted because the market to short it doesn't exist in a liquid form. That asymmetry will produce a sequence of surprises, and each surprise will be called an "unexpected commodity price spike" rather than what it actually is: the discovery that energy markets are deeply inefficient, and that inefficiency is a toll booth for those who can wait. Q: So what do we actually take away from this, beyond the chart maps and LCOE tables? The takeaway is that the attention of the digital economy is shifting from the block to the busbar. The block was how we ordered value claims; the busbar is how we enforce them. The next several years of AI value creation will be determined less by model architecture and more by, literally, whether you can hold a load. The contrarian trade to the Chevron-Williams consensus is not to short natural gas; it is to notice who holds the scarce, weird, under-appreciated asset: the right to plug in. Entities sitting on old industrial sites with 500 megawatts of interconnection capacity are holding a call option on AI itself. This is exactly the dynamic that made Bitcoin miners into high-performance-computing landlords earlier this year, when they discovered their power contracts were worth more than their chips. The signal through the noise floor is unambiguous — the future does not belong to the most capable model, nor to the most efficient miner, nor to the fastest rollup. It belongs to whoever can demonstrably present firm power at the right location, and then extract the narrative arbitrage between the way the grid is valued and the way the world is run. History rhymes and code compiles, but thermodynamics in one direction. I write often about narrative arbitrage — the distance between what markets believe and what physics requires. Right now, that distance is enormous. Chevron and Williams have closed part of it with concrete and turbines, and their reward is a discounted claim on AI's future tax base. The rest of us — the token protagonists, the agent builders, the infrastructure apostates — should be asking ourselves a different question entirely. If we are really building a machine economy, where is the machine's meter? Who reads it? And most importantly — do you hold a position in the settlement layer of the physical world, or are you still paying the attention tax of pretending it does not matter? The next narrative will be "proof of power” — not proof of work, not proof of stake, but the demonstrated capacity to secure, transport, and settle electrons in a network that values exactly two things above all else: who pays for the load, and who can switch off the light. The consensus of the disconnected is that AI will solve energy. The collision of opposites is already showing us otherwise. Follow the signal through the noise floor, and the signal is a PJM auction number that went vertical while the rest of the world kept discussing whether a GPU has a bottleneck. The gas turbine is the new ASIC. Chevron and Williams just became the miners of the AI era. And we all know, from painful experience, what eventually happens to mining margins when every player in the market rushes to build the same rig at the same time. Chasing the horizon of the next paradigm, one question still burns: who will be left holding the stranded thermal asset when the horizon finally arrives? That is not rhetorical. It is the question every cycle asks, and this time, the answer may well be inscribed in steel, buried in a PPA, and buried again in a token. Start your search there.