In the spring of 2026, three announcements arrived within weeks of each other, and the financial press treated them as three stories.
The first: stablecoins, and USDC in particular, were emerging as the settlement instrument for autonomous software agents. The tokens with “AI” in their names were not the tokens the agents used. Where agents settled, they used the dollar.
The second: Stripe, at its 2026 Sessions conference, laid out a near-complete financial substrate for the agent economy—wallets, card issuance for agents, sub-second settlement, treasury accounts, fraud control—and partnered with or absorbed much of the crypto-native infrastructure that had spent five years building toward the same end.
The third: Goldman Sachs spoke in two voices at once. Its research skeptic, Jim Covello, said in mid-2026 that AI economics looked more questionable than two years earlier, the returns still unproven. Its institute, meanwhile, published a baseline model putting annual AI capital expenditure at $765 billion in 2026 and cumulative spending near $7.6 trillion through 2031—conceding the scale and durability of the buildout even as the bank’s own skeptic doubted its payoff. A running argument with SemiAnalysis continued over which layer of the stack would capture the returns.
Three winners. Three victory laps. Read together they are not three stories. They are one structure seen at three depths.
The structure is old. It has a name. It is the chartered company.
Four layers, one question
Every question about computation resolves into four layers. What physically exists: the infrastructure layer. What rules govern it: the protocol layer. Who controls it: the power layer. What behavior emerges that no one designed: the emergent layer.
The AI-economy commentary of 2026 operates almost entirely at the first layer and occasionally the second. It asks whether the capex is too large, whether the tokens are correctly priced, whether the returns will arrive. These are real questions. They are not the question this essay is about.
The question this essay is about sits at the power layer, and it is the oldest question in the political economy of infrastructure: who holds the charter, and over what territory.
The charter
Begin with the money.
An autonomous agent that buys compute, calls an API, and settles a fee at sub-cent granularity cannot economically use the card networks as its native rail—roughly 2.9 percent plus a fixed per-transaction charge makes sub-cent settlement impossible—and cannot rely on ordinary bank-transfer infrastructure, built for account-to-account settlement rather than open-web, machine-to-machine metering. It needs an instrument that is programmable, near-instant, globally addressable, and cheap at very small denominations. In the open-web implementations now being built, that instrument is overwhelmingly a dollar stablecoin, most often USDC.
Note what this is. The agent economy did not invent a new money. It selected the oldest one. Where autonomous agents are already settling through stablecoin rails, the reference instrument is overwhelmingly the programmable dollar, usually USDC—a token whose value is a dollar held in custody by a regulated American issuer, redeemable through the American banking system, freezable by the issuer under American sanctions and compliance authority.
The currency layer of the machine economy is therefore not outside sovereignty. It is the most deeply embedded part of it. Every settlement is a programmable dollar. Every programmable dollar is a claim the issuer can refuse, freeze, or redeem only under the rules of the dollar system—and the issuer is bound to those rules by American regulation.
Now the protocol.
The standard that lets an agent pay for a resource over the open web is x402, a revival of the long-dormant HTTP 402 “Payment Required” status code. Coinbase originated it in 2025. Cloudflare became a co-founder of the foundation that governs it. Circle, the issuer of USDC, is a member. Google joined the foundation and has supported interoperability between x402 and its own agent-payments work. Stripe joined as a founding member. In April 2026 the Linux Foundation took over its governance.
The protocol’s design has one property that matters more than all the others. It decouples payment authorization from entitlement authorization. It can verify that a buyer paid the quoted price. It cannot verify that the seller had the right to sell the resource. HTTP 402 asks a single question—what is the price—and never asks a second one: does the seller hold the title. The standard transmits money. It does not transmit entitlement.
The consequence is already visible. Third parties wrap other companies’ paid APIs—data services, search results, travel inventory—into x402 endpoints and resell access to agents, often without the originator’s permission. The protocol has no mechanism to detect this, because non-detection is not a flaw in the design. It is the design. A payment rail that also adjudicated property rights would be a court, and the builders did not build a court.
So the protocol layer is a vacuum. Authorization is nobody’s job. It has been externalized to “developer responsibility” and, beyond that, to the ordinary judicial system—which moves at the speed of litigation while the protocol moves at the speed of settlement. A firm can sue an unauthorized reseller. By the time the suit is heard, the resold data has already flowed to a million agents at sub-second latency, generating revenue in a token that clears before the complaint is filed. The gap between judicial reaction time and protocol operating time is not a detail. It is the territory.
Now the pipe.
Stripe’s 2026 Sessions was not a product launch. It was the laying of a financial substrate. Agent wallets. Card issuance for agents. A machine payments protocol co-authored with the Tempo chain. Treasury accounts for agents. Fraud control tuned for non-human actors. Tax and metering. Stripe did not originate the crypto-native pieces. It is becoming the enterprise pipe through which they become legible—the layer through which most of this infrastructure will reach the firms that actually deploy it.
This is not Stripe defeating the crypto industry. The stablecoin issuers, the wallet firms, the settlement chains that appeared on Stripe’s stage are crypto-native companies. Stripe did not beat them. It became the layer they distribute through. The five years of stablecoin development happened elsewhere; Stripe opened the pipe and the flow chose the pipe.
Three layers, then. The currency layer, fully embedded in American sovereignty. The protocol layer, a structural vacuum. The pipe layer, privately re-centralized in a small number of firms. The asymmetry across these layers is the finding. Where the formal state reaches, it reaches completely. Where it does not reach, private consortia have moved into the vacuum and begun performing the functions a sovereign performs.
Consider what those functions are. A frozen USDC address is enforcement. A facilitator verifying a payment is adjudication. The cells of a machine-payments protocol are contract law. A service directory that certifies “authorized first-party” sellers is a registry of legitimate title. None of these is described as governance. All of them are governance.
This is the chartered company.
In 1670, Charles II granted the Hudson’s Bay Company a charter conferring monopoly trade over the entire Hudson Bay watershed, together with the authority to make law, hold courts, and maintain armed force in that territory. The territory—called Rupert’s Land, roughly a third of modern Canada—was one the English state did not and could not yet govern. The Company was not an extension of the Crown’s administration. It was a private entity granted the Crown’s functions in a place the Crown’s administration had not reached. It operated as a quasi-sovereign for two hundred years, until the new Dominion of Canada purchased the territory back in 1870 and folded it into formal state structure.
The consortium assembled around the agent economy—Circle at the currency layer, Coinbase and Cloudflare and the Linux Foundation at the protocol layer, Stripe at the pipe layer—occupies the same structural position. It exercises sovereign functions in a territory the formal state has not reached: the space of machine-to-machine commerce, where transactions complete faster than any court can convene.
The point is not that these firms have received a literal royal charter. None has. The point is that they occupy the chartered-company position: private institutions performing public functions in a territory the formal sovereign has not yet operationally reached. The Crown’s charter to the Hudson’s Bay Company made explicit what the dollar system, the x402 standard, and the Stripe rails make implicit—but the function is the same, and in the modern case it arrived without anyone granting it.
One difference dominates all the others. The Hudson’s Bay Company’s two centuries are compressed, here, into two to five years. The judicial reaction window is not merely slow relative to the protocol. It has been blown through entirely. The chartered authority establishes facts on the ground before the sovereign learns the territory exists.
The rails
Leave the payment stack and descend to the physical layer. Here the commentary is loudest, and here it asks the wrong question most insistently.
The Big Five hyperscalers’ 2026 capital-expenditure guidance runs between roughly $600 billion and $700 billion depending on the count, with about three-quarters—on the order of $450 billion—committed directly to AI infrastructure: chips, servers, networking, purpose-built data centers, and the power to run them. The number is up by more than a third over 2025. Capital expenditure has risen to between 45 and 57 percent of revenue, against 10 to 15 percent at the start of the decade. The spending now exceeds internal cash flow; the gap is bridged with debt. Issuance in 2025 ran past $100 billion, and sell-side estimates put new technology-sector debt issuance near $1.5 trillion over the following three years.
The argument between Goldman Sachs and SemiAnalysis is about whether this is justified—whether the average enterprise will see a return, whether the marginal user already does, which layer of the stack captures the margin. It is a sophisticated argument. It is conducted entirely at the infrastructure and pricing layers, and it accepts a premise it never examines: that the outcome is a market outcome, to be settled by whether the returns arrive.
That is interesting. It is not the real question.
The fiscal question—is this a bubble, will the ROI come—belongs to someone else. The infrastructure question is different and prior: regardless of whether the financing proves sound, what physically gets built, where does it sit, and who controls it when the financing question is finally settled.
Frame it that way and the capex stack stops looking like a bet and starts looking like a deposit.
Consider the three outcomes the Goldman–SemiAnalysis argument actually spans. Enterprise returns arrive: the cloud providers’ valuations recover, the semiconductor concentration gives back some margin, the United States holds the intelligence layer. Enterprise returns do not arrive and the providers cut capex: their cash flow improves, the semiconductor complex takes a severe correction, the investors who funded the build are impaired—and the data centers, the fabricated silicon, the power interconnections, and the trained engineering networks remain, in American geography. The status quo persists: the build continues, infrastructure keeps accumulating.
In all three outcomes the distribution of pain among enterprises, institutions, and households shifts violently. In all three the question of where the physical capital settles, and toward which jurisdiction it accumulates, does not move.
This structure has a precedent, and it is not the one usually offered. The fiber-optic boom of the late 1990s is the analogy SemiAnalysis and others reach for: pioneers take the arrows, settlers take the land. It is too gentle. The closer precedent is the American railway debt crises of 1873 and 1893.
In both, vast railway construction was financed substantially by foreign capital—much of it British bondholders buying into American railroad debt. In both, the financing structure collapsed: the Panic of 1873 began with the failure of Jay Cooke and Company over Northern Pacific financing; the Panic of 1893 took down the Philadelphia and Reading, the Northern Pacific, the Union Pacific, and the Atchison in succession. British bondholders were impaired on a continental scale. And in both, the rails remained. The physical network—graded, tracked, bridged—stayed in American ground and became the circulatory system of twentieth-century American industrial supremacy.
The British investor did not win. The British investor was, at the time, devastated. The loss was the physical fuel of a hegemony the investor did not share in.
This is the structure the railway and the chartered company have in common, and it is the structure of the AI capex stack. The financial outcome is independent of the sovereignty outcome. Whoever loses the money, the infrastructure and the control over it settle in a determinate place. The $1.5 trillion in projected debt is denominated in dollars, subscribed substantially by American institutions, and converted into data centers on American land, silicon from a supply chain under American protection, and power drawn from the American grid. Win or lose at the financial layer, the physical layer accrues to one side.
One actor sits at the center of this and is consistently described in terms too small for its position. SemiAnalysis calls Nvidia the central bank of the AI ecosystem, and uses the phrase to discuss pricing restraint. The phrase deserves to be taken to its conclusion.
A central bank is not defined by pricing restraint. It is defined by three things: a monopoly on the issuance of the unit everyone must use, a lender-of-last-resort function that makes it unbypassable in a crisis, and a position no participant can route around. Nvidia approaches all three. Its CUDA software ecosystem functions as an issuance monopoly—the switching cost away from it is a near-total loss of accumulated tooling. In any contraction, a customer’s real alternative to buying its accelerators is exit from frontier AI altogether, which makes it the bottleneck no one can route around. And the boundary conditions it sets—on price, supply, and architecture—determine the gross margins available to every model lab, every cloud, every downstream agent business.
Read at the power layer, Nvidia plus TSMC plus the American grid is a distributed central bank for machine computation. Its monetary policy is the GPU release cadence. Its reserve currency is compute. Its headquarters, taken as a system, is in the United States and the territory under American security guarantee. This is not a metaphor about pricing. It is a description of where monetary sovereignty over computation physically resides.
Imperium in imperio
Assemble the two halves.
At the payment, protocol, and pipe layers, a private consortium exercises sovereign functions over a territory the formal state has not reached, at a speed that outruns the state’s capacity to follow. At the compute, capital, and power layers, the physical substrate of machine intelligence accumulates toward a single jurisdiction regardless of who is financially harmed in the building of it.
Both halves are the chartered-company form. A private entity, granted or seizing the functions of sovereignty, operating in a territory ahead of the formal state, with a financial fate disconnected from the sovereign fate of the territory it builds.
The word for this in the older literature is imperium in imperio—a sovereignty lodged inside a sovereignty. The chartered companies were exactly that, and the tension was understood at the time to be unstable: a private power performing public functions will either be absorbed by the state or will hollow it out. The Hudson’s Bay Company was absorbed. The East India Company was absorbed, after the Crown decided in 1858 that a private corporation could no longer be permitted to govern a subcontinent.
What is emerging in 2026 is not the decentralization its early promoters described, and it is not a simple extension of the existing state. It is the chartered-company form, revived at machine speed, with a topology its predecessors did not have: partial embedding and partial escape. The currency layer is embedded in the sovereign completely. The protocol layer has escaped it completely. The physical layer accrues to it regardless of outcome. These are not contradictions to be resolved. They are the shape of the thing.
The open question
The chartered companies were eventually re-absorbed. The Crown bought back Rupert’s Land. Parliament took India in hand. The railways were brought under regulated-utility law and, later, antitrust. In each case the absorption was possible because the territory had an edge—a physical, mappable boundary the sovereign could eventually reach, govern, and tax.
The question that determines everything downstream is whether the territory of machine commerce has such an edge.
If it does—if the protocol layer can be brought under a registry of authorized title, if the settlement instrument’s embeddedness in the dollar is the leash by which the whole apparatus is eventually reeled in, if the physical concentration of compute proves as governable as a railway network—then the present moment is the chartered-company interval, and it ends the way such intervals have always ended, with the sovereign arriving late and buying back the land.
If it does not—if a territory that completes its transactions faster than any court can convene has no boundary at which the sovereign can stand—then this is the first chartered company whose territory has no edge, and the comparison to Rupert’s Land breaks at exactly the point that matters.
This essay does not predict which. The conditions that would decide it are the things worth watching: whether authorization re-enters the protocol layer as an enforceable standard, whether the dollar-embeddedness of the settlement layer is used as leverage, whether compute concentration triggers the antitrust response that the railway concentration eventually did. Each is a parameter, not an event. Each is observable before it resolves.
The agent economy has not escaped sovereignty. It has revived the oldest method by which private power and sovereign power have always negotiated a frontier: the charter. What remains undecided is whether, this time, the frontier has an end.