Consensus is not a mathematical problem. It is a land problem.
The Oldest Ledger in Texas
About four miles northeast of the Somervell County courthouse in Glen Rose, Texas, there is a limestone shelf in the Paluxy River where sauropods walked roughly 113 million years ago. Their feet pressed into wet carbonate mud, the mud lithified, and the record survived a Cretaceous sea, an ice age, and four generations of amateur paleontologists. You can still read the sequence: a heavy circular print, a second one slightly offset, a tail drag, then three-toed theropod tracks crossing the same surface at an angle that suggests, depending on whom you ask, either pursuit or coincidence. In the drought summer of 2022, when the river fell low enough to expose new shelves, more of that record surfaced and the park service asked visitors to stop walking on it.
That shelf is the oldest provenance technology in North America. It is also, and this is not a metaphor I chose lightly, the first disputed ledger in Texas. In the 1930s a set of elongated markings at the Taylor Site were promoted as human footprints walking alongside dinosaurs. They were not. Decades of re-examination attributed them to metatarsal impressions, erosion, and the ordinary human talent for seeing intention in noise. The argument about what those marks mean has outlived most of the people who made it.
Now drive south on Highway 144. Within sight of the Comanche Peak cooling towers, on land that the reporting describes as belonging to Vistra, Amazon is building a data center campus. Two records of the same thing β who was here, in what order, and can you prove it β sitting a few miles apart, separated by 113 million years and about eighteen months of capital allocation.
What Was Actually Reported
The item, as it reached me, was thin. Four facts: Amazon is building a data center campus in Glen Rose, Texas; the site is adjacent to the Comanche Peak nuclear plant; the land is associated with Vistra; the source is a crypto-native outlet aggregating a local business story. No megawatts. No investment figure. No timeline. No power purchase agreement structure. No indication of whether the facility is general-purpose AWS capacity or an AI training and inference cluster. No named reporter, no primary quotation, no link to a company statement.
I want to be precise about how I treat that. I spent four weeks in 2017 reading Parity's multi-signature wallet code line by line, and the lesson I carried out of that room was not about Solidity. It was that the difference between a finding and a rumor is a citation. So I will mark my confidence explicitly throughout this piece, and where I am inferring from engineering convention rather than from disclosed fact, I will say so in the sentence where I do it.
Because the reporting is thin, the temptation is to dismiss it. That would be a mistake. Thin reporting on a physical asset is different from thin reporting on a token. A token announcement with no details is usually nothing. A land acquisition next to a 2.4-gigawatt nuclear plant with no details is a signal that someone with a balance sheet made an irreversible decision. Concrete does not get un-poured.
The Context That Makes This Legible
Comanche Peak is not a small piece of infrastructure. Two pressurized water reactors, Unit 1 and Unit 2, licensed in the late 1980s and early 1990s, roughly 1,200 megawatts each, about 2.4 gigawatts combined, operated today under the Vistra umbrella. It is one of the largest single sources of carbon-free electricity in the state, and it runs at a capacity factor north of ninety percent β meaning it produces at or near its rated output almost continuously, regardless of weather, regardless of whether the wind is blowing in the Panhandle or the sun is out over the Permian.
That number, the capacity factor, is the whole story. Everything else in this article is downstream of it.
Glen Rose itself is a town of fewer than ten thousand people in a county that has historically run on tourism, small manufacturing, and the plant. The Paluxy feeds the Brazos basin. Water rights there are not theoretical; they are adjudicated, metered, and defended. And Texas property tax abatements of the kind that historically made these deals pencil β the Chapter 313 school district program β expired at the end of 2022. The successor mechanism, the Jobs, Energy, Technology and Innovation Act under Chapter 403, is narrower, requires school board approval, and has become a genuine local political event rather than a formality. If this project is going to make its return targets, the tax treatment has to be negotiated in public, in a county where everyone knows the school board president's first name.
Why should a reader of a blockchain publication care about any of this? Because the megawatt is the shared denominator. The same commodity that secures proof-of-work chains, that runs every rollup sequencer, that powers every inference call behind an AI agent that signs a transaction, is the commodity that Amazon is now buying in twenty-year increments. The crypto industry has spent a decade arguing about whether its energy consumption is justified. It has spent almost no time arguing about what happens when the entities doing the consuming are no longer crypto companies.
The four largest cloud providers have all now moved on nuclear-adjacent power. Microsoft signed a twenty-year agreement to take output from the Crane Clean Energy Center at Three Mile Island. Google committed to small modular reactors through Kairos Power. Meta contracted for the output of a large Illinois unit. Amazon itself already bought the Talen campus adjacent to Susquehanna. Glen Rose is not an opening move. It is a second or third move in a game whose rules are being written by interconnection queues.
Which raises the question I actually want to answer here: when the physical layer of the internet becomes scarce, who holds the admin keys?
Reading the Campus Signal
The word campus is doing quiet work. In data center engineering, campus means a parcel with multiple buildings, phased construction, and a site plan that anticipates expansion beyond the first tranche. It is not a single hall. It is not a colocation tenant footprint. It is a developer-scale commitment that usually starts above one hundred megawatts and is designed so that the second hundred megawatts costs less than the first.
Pair that with adjacency to 2.4 gigawatts and you get a defensible order of magnitude: the initial phase is plausibly in the few-hundred-megawatt range, with headroom toward a gigawatt. I want to flag that as inference, not disclosure. The reporting gives me nothing quantitative, and I have watched enough of these projects to know that announced campuses regularly shrink, slip, or get re-scoped when the interconnection study comes back with a number the finance team did not budget for.
The more interesting analytical question is not how big. It is what kind of load.
A general-purpose AWS region is a relatively forgiving electrical customer. Workloads are heterogeneous, utilization is diurnal, and a facility can ride through moderate supply interruption using uninterruptible power and generator backup. An AI training cluster is a different animal. It draws a flat, brutal, gigawatt-scale load for weeks at a time, and interruption is not an inconvenience β it is a loss of the entire run. An inference cluster is the opposite: spiky, latency-bound, wildly variable.
This distinction determines the cooling design, the rack power density, the network fabric, and, most consequentially, whether the site is a natural fit for baseload nuclear or a natural fit for gas peakers and batteries. Adjacency to Comanche Peak tells you the developer wanted baseload. It does not tell you the developer got it.
Here is the part that the industry press consistently underweights. Being next to a nuclear plant is not the same as being powered by one. Electricity follows physics, not press releases. Unless there is a direct physical connection behind the meter, or a contractually specified delivered product, the campus connects to the grid and receives an undifferentiated pool of electrons generated by whatever was dispatched that hour. Nuclear adjacency becomes a branding asset and, at best, a locational price hedge. Whether it becomes a carbon claim depends entirely on the contractual architecture.
The Part Nobody Wants to Model: Load Shape Against Supply Shape
I spent the 2022 bear market in Frankfurt reading zero-knowledge papers, and the thing that kept me sane was that the math did not care about my mood. A proof verifies or it does not. I have come to rely on the same discipline when reading energy claims: what is the shape of the load, what is the shape of the supply, and does the second actually cover the first?
Nuclear is the flattest supply in existence. It produces at rated output with scheduled refueling outages every eighteen to twenty-four months. Training workloads are also unusually flat across a multi-week horizon. That is a genuinely elegant match, and I suspect it is the real reason hyperscalers gravitated to reactors rather than to the more obvious renewables. It is not primarily about sustainability optics. It is about matching a load that cannot be interrupted with a supply that does not want to be interrupted.
But the training run ends. Then what? The industry's own capital plans assume that after the frontier model is trained, the facility transitions to serving inference at volume, and inference is spiky in a way that is genuinely hostile to baseload economics. You cannot ramp a reactor down for a lunchtime dip in API traffic. So the load-serving arrangement has to absorb shape mismatch on both ends: the campus must either overbuild storage and peaking capacity, accept curtailment obligations, or sign a shaped contract that blends firm baseload with flexible supply.
This is where the economics get decided, and where nearly all public commentary stops. The cost per delivered megawatt-hour at a campus like this is not the headline power price. It is the blend of firm baseload, capacity, transmission or wheeling charges, curtailment risk, and the cost of the flexibility product that makes the whole thing tolerable to the grid operator. Every one of those components is negotiable and none of them are disclosed in a four-fact news item.
Based on my experience auditing infrastructure-adjacent contracts, I would expect the structure to look less like a utility bill and more like a bespoke derivative: a long-dated fixed volume with shaped delivery, indexed escalators, curtailment rights purchased at a discount, and an environmental attribute layer stapled on top with its own contractual life. That is a financial instrument wearing a power contract's clothing. And a financial instrument with a twenty-year tenor and a counterparty who can cancel over a water permit.
The FERC Fight That Explains Everything
To understand why a campus would seek nuclear adjacency rather than simply buy clean energy certificates on the open market, look at what happened at Susquehanna.
Amazon bought a data center campus physically adjacent to the plant and struck an arrangement to take power directly from it. The amended interconnection service agreement went to federal regulators in 2024 and was rejected, on the reasoning that the arrangement raised unresolved questions about cost allocation and grid reliability β essentially, whether a large load co-located behind the meter still free-rides on a transmission system paid for by everyone else. A revised approach later moved forward through a regional process with new large-load rules.
The lesson is not about which regulator won. The lesson is that co-location converts a private commercial deal into a public governance question, and public governance questions get resolved slowly, by people who are not parties to the contract.
I want to connect this to something I have argued for years about decentralized autonomous organizations. Code is law does not survive contact with reality, because every smart contract has an upgrade path, and the upgrade path is always controlled by a small set of keys held by a small set of people. The elegance of the on-chain logic is real, and it is also beside the point. The actual distribution of power lives in the multisig, and the multisig is usually a three-of-five among people who have lunch together.
ERCOT is a multisig. It is a better-designed one than most β an energy-only market with a high offer cap, no capacity market, price signals that genuinely clear, and a long institutional prejudice against central planning β but the interconnection queue is the upgrade path, and the queue is single-file. You can have the most market-driven grid in North America and still have a physical chokepoint where a thousand gigawatts of new load requests wait for study. The Texas market is decentralized in its pricing and centralized in its permissions, exactly like the DAOs I have spent years reading, and for exactly the same reason: somebody has to hold the keys that determine who gets to connect.
The Duration Mismatch That Will Reprice This Entire Sector
Here is the analytical point I most want to put in front of you, because I have not seen it made cleanly anywhere else.
A nuclear plant is a forty-to-sixty-year asset. A data center shell and its mechanical plant is a twenty-to-thirty-year asset. A high-end accelerator is a five-to-six-year asset on a depreciation schedule, and arguably a three-to-four-year asset in practice. A trained frontier model's weights are economically relevant for roughly eighteen months. And a power purchase agreement is typically fifteen to twenty-five years.
Stack those durations and you find the actual risk structure of the AI infrastructure boom. You are underwriting a twenty-year contract with an eighteen-month asset sitting inside a six-year asset sitting inside a thirty-year asset sitting inside a sixty-year asset. Every layer has a different half-life. The financing only works if the shortest-lived layer keeps generating enough cash to service the longest-lived layer before it is obsolete.
This is not a new problem in crypto. It is the identical problem of a DeFi protocol that issues a four-year liquidity mining program to bootstrap liquidity for an asset whose actual demand curve nobody has modeled. In 2020 I drafted governance documentation for a lending protocol launch and spent weeks fighting internally about whether emissions incentives should be framed as bootstrapping or as subsidy. The answer, which nobody wanted to hear, was that the framing was irrelevant and the duration was everything. If the incentive ends before the organic demand begins, you have not built a market. You have rented one.
Amazon is renting confidence from Vistra and Vistra is renting confidence from Amazon, and both are renting confidence from a regulator and a county commissioners court. This is why I keep coming back to a line I have used for years: trust is the new token. Not as a slogan about sentiment, but as a precise statement about what is actually being issued. A power purchase agreement is a token. It represents a claim on future physical delivery, it trades, it has a holder base, and its value depends entirely on whether the issuer can perform. The only difference between it and an on-chain instrument is that the settlement layer is a legal system instead of a virtual machine, and the oracle is a meter reader instead of a node.
Liquidity flows where belief resides. And right now, belief is residing in the interconnection queue.
Where Blockchain Actually Fits β and Where It Does Not
The reflexive crypto answer to all of this is that decentralized physical infrastructure networks will route around the hyperscalers. I want to examine that claim with the care it deserves, because I think it is half right in a way that matters more than being entirely wrong.
Where the claim fails is training. Training a frontier model requires tightly coupled accelerators on a low-latency fabric, coordinated for weeks. Bandwidth and synchronization dominate. You cannot shard that across consumer GPUs in fifteen countries without paying a penalty that no amount of token incentive can offset. Any project claiming otherwise is either not training frontier models or not being honest about what it is training.
Where the claim has genuine merit is at the edges: inference for small models, batch workloads, rendering, and above all verification. And verification is where the crypto industry has a real, defensible, technically grounded contribution to make β one that the hyperscalers are adopting without crediting us.
Consider what a serious provenance stack for AI actually requires. First, energy provenance: not annual matching of renewable certificates, but hourly, location-specific, ideally metered attestation of what was actually generated and consumed. The crypto industry built this. Energy attribute tokens, hourly certificate registries, on-chain settlement of meter data β these were crypto-native experiments when they were unfashionable. The hyperscalers now market the same idea as 24/7 carbon-free energy, and the accounting firms have built practices around it.
Second, compute provenance: cryptographic attestation that a specific workload ran on specific hardware in a specific configuration, unmodified. Trusted execution environments have matured considerably. Confidential computing on modern accelerators can produce signed attestations of what code ran and what it produced. That is not a zero-knowledge proof and it does not eliminate trust β it concentrates trust into a silicon vendor's attestation root β but it is a real reduction in the surface area of what you have to believe.
Third, model provenance: hashing weights, recording training data lineage, and being able to demonstrate that the model answering your query is the model you think is answering it. In 2026, working on agents that transact, this is not academic. When an autonomous agent signs a transaction on your behalf, the question of which model produced that decision, under which system prompt, with which tools, is a question about liability. Right now, in almost every deployment I have seen, that question has no verifiable answer.
Here is the honest limit, and I will not soften it. Zero-knowledge machine learning is not going to verify frontier inference at production speed any time soon. The overhead is orders of magnitude, and the research frontier is still at the scale of small convolutional models and toy transformers. Anyone selling you zkML for a 400-billion-parameter model in a live product is selling you a roadmap. I spent months in 2022 reading Aztec's construction because I wanted the mathematical certainty to be real, and the discipline that period gave me was the ability to distinguish between a proof system that works and a proof system that will work.
What this means practically is that the blockchain layer for AI infrastructure is a settlement and coordination layer, not a compute layer. It is where you record that a job was commissioned, that an attestation was produced, that payment was released, that a penalty was levied, that a dispute was resolved. The compute happens on silicon. The trust happens on chain. Conflating the two is how projects raise money and how they fail.
The Bear Market Read: Who Is Actually Bleeding
I have to be useful to readers who are watching their positions, not just their thesis. So let me translate this into survival terms, because in this market survival is the only performance metric that compounds.
The uncomfortable observation from the last eighteen months is that compute-backed and energy-adjacent tokens have traded at persistent premiums to their underlying cash flows. There is a floor under a protocol that owns generation capacity or has contracted for firm power, and there is no floor under a protocol that owns a narrative about owning generation capacity. The difference shows up in a drawdown. Narrative assets lose eighty percent without changing a single line of code, because nothing was ever backing them except the story, and the story was the collateral.
This is the same consolidation logic that MiCA has imposed on European crypto, and I want to name it plainly because I have watched it happen at close range. The Markets in Crypto-Assets regime gave Europe apparent clarity: a licensing path, a reserve framework for stablecoins, a defined perimeter for service providers. What it actually did was impose a fixed compliance cost β legal, audit, custody, reporting, capital β that does not scale down. A team with a genuinely useful product and four engineers cannot absorb it. A team with a mediocre product and a treasury can. Regulation designed to protect users selected for capital, not for quality, and the same selection pressure is now operating in AI infrastructure. You do not need a superior cooling design. You need to be able to sign a twenty-year contract.
In both cases the mechanism is identical and the mechanism is what matters. Scale begets permission, permission begets scale. The queue is the moat.
So when I look at a decentralized compute protocol in this market, I ask three questions in order. What is the per-unit cost of the thing it actually delivers, measured against the cheapest centralized alternative? Who absorbs the loss when the demand forecast is wrong? And what happens to the token if the physical expansion is delayed by two years in an interconnection study?
Most projects have no answer to the third question. That is not a character flaw. It is a structural consequence of building a financial instrument on top of a physical process and never modeling the physical process.
The Contrarian Turn: The Hyperscalers Are Proving Our Thesis
Here is where I part company with most of my peers in this industry.
The prevailing crypto view is that hyperscale AI buildouts are the enemy: centralized, extractive, energy-hungry, a consolidation of control that makes decentralization look naive. I understand the instinct. I felt it acutely during the FTX collapse, when I had to sit with the possibility that my entire worldview was a coping mechanism. But I have come out the other side with a different reading.
The hyperscalers are not disproving the case for trustless systems. They are pricing it.
Look at what they are actually paying for. They are spending billions on physical proximity to baseload power because they cannot trust the grid to deliver. They are signing twenty-year contracts because they cannot trust spot markets. They are pursuing co-location because they cannot trust a utility's interconnection schedule. They are building proprietary silicon because they cannot trust a supplier's roadmap. They are constructing the most expensive trust stack in industrial history, and the invoice is denominated in gigawatts and decades.
Every dollar of that invoice is a measurement of what it costs to rely on trusted third parties. That is the number our industry has been arguing about abstractly for fifteen years. Amazon just made it concrete, and the concrete is in Somervell County.
If we could measure verification cost per joule β the energy and capital required to establish, with evidence, that a computation happened as claimed β we would have the honest comparison the debate has always lacked. The centralized stack pays for trust with land, steel, uranium, and time. The decentralized stack pays for trust with cryptography, redundancy, and coordination overhead. Neither is free. We have simply never bothered to price ours, preferring to assert that it is cheaper because it feels more virtuous.
This is what I mean when I say code has conscience. Not that code is good. That code encodes choices, and the choices have costs, and pretending the costs do not exist is how you end up with a Parity wallet that self-destructs because someone wanted to be a library.
There is a second contrarian point, and it is less flattering to us. The carbon-free energy accounting that hyperscalers now market aggressively β hourly matching, deliverability, locational attribution β was invented in large part by people adjacent to this industry, and we lost control of the narrative by over-financializing it. We built beautiful instruments for proving where a megawatt-hour came from and then wrapped them in yield farms until nobody could tell a renewable energy certificate from a liquidity pool. The hyperscalers took the accounting discipline, stripped the speculation, and are now the credible face of provenance.
That is a strategic failure, not a technical one. And it should sting more than any regulatory setback.
There is a third point, and it is the one I keep circling. ERCOT, with no capacity market and no central planner, is arguably a more decentralized coordination system than most DAOs in existence. Millions of participants, price-mediated, no governance token, no proposal forum, no delegate politics. And yet it has a single-file permission queue at the front door, and a small set of commissioners with effective upgrade rights over the whole system. Decentralization at the pricing layer coexists comfortably with centralization at the access layer. Every DAO I have ever examined has the same shape. The market is open. The door is not.
What I Am Watching Now
I do not believe in forecasts. I believe in tripwires β specific, observable events that force a revision of the model. Here is what I am watching, and what each signal would change.
The first is the interconnection. Until an ERCOT interconnection agreement with a defined capacity and a milestone schedule exists, any timeline in any article about this campus is fiction, including mine. The queue is the gate, and the gate is where projects die quietly. If I see a signed agreement with a specific megawatt figure, I will raise my estimate of the campus scale and shorten my timeline by years. If I see nothing for eighteen months, I will assume re-scoping.
The second is the water. Somervell County sits over the Paluxy and the Brazos, and cooling is not a footnote in central Texas. If the design is closed-loop liquid cooling with dry coolers, the water controversy mostly evaporates and the project proceeds. If it is evaporative, the county will have a fight on its hands, and I have watched enough local land-use disputes to know that the county commissioners court is a more consequential regulator than most federal agencies for a project like this. Watch the TCEQ permits, not the press releases.
The third is the tax abatement. Chapter 403 requires a public process. That hearing is the first place where the real size of the project and the real shape of the local bargain become visible, because a school district does not grant an abatement without knowing the assessed value and the employment commitment. If that hearing is smooth, the project has local consent. If it produces organized opposition, add two years.
The fourth is the power contract. A disclosed long-term agreement with a defined volume and a defined shaped-delivery profile would tell me more about the economics of this campus than any capital expenditure figure. Absent that disclosure, I assume the arrangement is either a short-dated hedge or something structurally creative enough to be difficult to explain in a press release.
The fifth, and the one I care most about, is whether anyone in this industry builds the measurement I described earlier. A credible, auditable, standardized metric for verification cost per unit of delivered computation would do more for the decentralization thesis than a hundred conference panels. It would let us say, with evidence, here is what trust costs, and here is what it costs us.
The Long View
A hundred and thirteen million years ago, something very large walked through shallow water in what is now Somervell County and left an impression in lime mud. Nobody signed it. Nobody minted it. It survived because the physical process that recorded it was more durable than the forces that would have erased it.
That is the only real test of a ledger. Not how clever the consensus mechanism is, but whether the record outlasts the incentive to falsify it. The Paluxy tracks passed that test for a hundred million years and then failed it at the interpretive layer β the marks survived, the meaning was contested, and the argument is still running.
Amazon is pouring concrete next to a nuclear plant because it has concluded that the thing it cannot buy on the open market is a reliable claim on the future. Not compute. Not capital. A reliable claim.
The blockchain industry has spent fifteen years insisting it could manufacture reliable claims more cheaply than any institution. Some of us were right. Some of us were selling the claim instead of the reliability, and the distinction is now being priced in public, in gigawatts, by companies that do not read our forums.
The question I would leave with you is not whether decentralized infrastructure can compete with a two-point-four-gigawatt campus on cost. It probably cannot, not on training, not this cycle. The question is whether we will do the unglamorous work of proving what our alternative actually costs β per joule, per attestation, per verified claim β before the hyperscalers finish teaching the market that provenance is a service you buy rather than a property you verify.
Code has conscience. But conscience is not a business model. It is a standard, and standards have to be measured before anyone believes them.
Liquidity flows where belief resides. The megawatts are already moving. The question is what we are going to build that deserves them.