The Governed-Compute Thesis

Start with the category error, because everything downstream depends on getting it right.

The four objects share a name and nothing else. A model token has no issuer and no redemption. It is a measurement artifact — a count that depends on the tokenizer, the model architecture, the input-output mix, context length, caching, batch size, latency target, and quality class. Two workloads that both consume a million tokens can differ in cost by an order of magnitude and in business value by more.

A crypto token has an issuer, a transfer mechanism, and a market price. Its supply schedule is a design choice. Its price responds to speculation as readily as to use.

A tokenized deposit is a claim on a specific bank. The Bank for International Settlements treats two properties as non-negotiable for any monetary arrangement: a common unit of account, and singleness, meaning claims denominated in that unit are redeemable at par with central bank money, with finality. Those properties are what make the instrument money. They are not incidental features a software platform can adopt selectively.

A usage credit is a contractual entitlement. Its issuer owes a service, not a sum. It is not redeemable for cash, and it should not be transferable.

Strategy that treats these as one object produces predictable failures. It markets an entitlement as an investment. It prices a service against a benchmark it cannot control. It promises fungibility across models that are not fungible. And it invites a regulatory question the business was never structured to answer.

The AI compute market is often drawn as a stack with margin migrating upward. That picture is too simple. It is better described as five layers, each selling a different thing, each carrying a different risk, and each measuring in a different unit.

A durable strategy respects these boundaries. Physical units measure capacity. Model units measure processing. Usage credits commercialize governed consumption. Money settles the contract. Every layer can be programmable without becoming the same instrument.

The distinction is not academic. It determines whether a platform can change providers without renegotiating customer contracts, whether a usage credit can be marketed without securities exposure, and whether a customer experience can be organized around outcomes rather than provider-specific price lists.

The clearest public evidence on infrastructure economics is CoreWeave’s first quarter of 2026, and it is routinely misread.

The company reported revenue of $2.078 billion, up approximately 112 percent year over year, against a revenue backlog of $99.4 billion. In the same quarter it reported a GAAP operating loss of $144 million, a negative 7 percent operating margin, adjusted operating income of $21 million at a positive 1 percent margin, $1.147 billion of depreciation and amortization, and $536 million of net interest expense.

Read carelessly, this looks like a price war. It is not. It is capital intensity.

Depreciation alone equals roughly 55 percent of revenue. That is the cost of buying accelerators fast enough to serve contracted demand, recognized over their useful life. Net interest expense of $536 million sits below operating income and reflects how that hardware was financed. These are two distinct pressures on two distinct lines, and collapsing them into one narrative about pricing produces the wrong conclusion.

Revenue growth does not eliminate depreciation, financing cost, utilization risk, or customer concentration. A $99.4 billion backlog is a commitment to deliver, and delivering requires more of exactly the capital that compresses the margin.

This is not a prediction that infrastructure providers fail. It is an observation that owning the metal is a capital business with capital-business returns, and that a control layer sitting above it has a different cost structure and a different risk profile. Published GPU pricing reinforces the point: rates vary widely by provider, form factor, region, commitment, spot availability, network fabric, and service level, as market indices such as the SemiAnalysis GPU Cloud Index track in detail. A price index rising from a trough indicates scarcity in a particular contract segment. It does not establish that normalized cost per unit of capability is rising, that all providers are expanding margin, or that the next hardware generation preserves the same rate.

The most common mistake in AI cost forecasting is picking one direction.

Research from MIT’s FutureTech group, published as The Price of Progress, measures both. Using pricing data from Artificial Analysis and Epoch AI, the authors find that the price of achieving a given level of benchmark performance has fallen roughly five- to tenfold per year. Over the same period, the price of running frontier models has risen between three- and eighteenfold per year, driven by larger models and heavier reasoning demands. Algorithmic efficiency alone contributes roughly a threefold annual improvement.

Efficiency reduces the cost of a fixed capability. If you need last year’s performance, you will pay far less for it this year. But demand does not hold capability fixed. It moves toward longer context, tool use, multimodal processing, and agentic loops that consume more tokens along more expensive execution paths. The buyer who benchmarked a workload eighteen months ago and assumed the bill would fall has watched the workload change underneath the assumption.

The commercial consequence is that no vendor should promise a customer that token prices only fall. Nor should any vendor build a business on the premise that they only rise. The real problem is managing a changing mix of workloads against a contracted service envelope. That is a governance problem, not a pricing prediction.

At GTC 2026, Jensen Huang framed AI factories as power-limited systems. Every facility operates under a hard ceiling in gigawatts, and at a fixed power level, whoever achieves the highest throughput per watt has the lowest production cost. Tokens per watt is therefore the metric that governs an operator’s revenue capacity.

This is correct, and it is an operating metric. It is not a unit of account.

Pricing research increasingly recognizes the limits at both ends of the spectrum. As Bessemer’s pricing work documents, raw token pricing is technically measurable but unintelligible to a nontechnical buyer, who cannot forecast a budget from it. Pure outcome pricing aligns value but transfers uncontrolled cost and scope-definition risk to the supplier. Neither extreme survives an enterprise procurement cycle intact.

The practical settlement is a governed usage unit whose consumption varies by workload class and execution policy. The customer receives a stable contractual framework. The supplier retains responsibility for the underlying cost model, routing, and capacity optimization. That is a division of risk both sides can actually sign.

Institutional tokenization is real, and it is the most frequently misappropriated precedent in this market. Tokenized real-world assets are a measurable and growing category, tracked live by aggregators such as RWA.xyz, and major banks have moved settlement infrastructure into production, with J.P. Morgan’s Kinexys among them.

The BIS describes tokenized deposits as digital representations of commercial bank money on programmable platforms. They confer a direct claim on the issuing bank. They are redeemable at par for central bank money of the same currency, with finality. The BIS position is that tokenization should be built into the existing two-tier system anchored in central bank money, precisely because singleness and elasticity are what keep the system stable under stress.

Those properties are the instrument. Par redemption, bank balance-sheet status, supervision, and settlement liquidity are not decorations on a programmable ledger. They are the reason it functions as money.

The broader lesson is the one worth taking. Programmability becomes more valuable when it is paired with strong identity, governance, legal rights, controlled access, and reliable settlement. That convergence is the relevant precedent. Financial tokenization for its own sake is not.

Decentralized physical infrastructure networks offer a genuine warning, and it is worth stating precisely rather than dismissively.

Render Network’s published burn-mint documentation describes the mechanism plainly. Customers convert fiat to RENDER and the tokens are burned on completion of work. Node operators, however, are compensated through scheduled emissions allocated epoch by epoch: 9,126,804 RENDER in year one and 5,905,580 in year two, on a predefined schedule.

The structure is deliberate and openly documented. It also means supplier compensation and customer payment are two separate flows. When the emission schedule is the larger of the two, reported network activity may not represent sustainable commercial demand. A tradable token additionally creates a second source of demand, speculation, which is independent of compute utilization entirely.

None of this proves that every decentralized compute model is structurally invalid. Render, Akash, Bittensor and io.net differ in workload, supply mechanism, token design, customer base, and revenue disclosure. Emissions, lease spend, market capitalization, provider payouts, and protocol revenue are not interchangeable metrics, and comparisons that mix them are meaningless.

The useful output is a set of questions any compute-token network should be able to answer from public disclosure. A network that cannot answer them has told you something.

Everything above is diagnosis. This is the constructive claim.

The opportunity is not to become another GPU marketplace or another crypto network. It is to become the governed workload plane through which institutional customers convert approved data, approved models, and purchased capacity into accountable work.

The atomic object is the Work Order. A Work Order connects the customer institution, the authorized actor, the research or business purpose, the source manifest, the workflow type, the execution policy, the budget authorization, the estimated reservation, the actual settlement, the output location, the status, and the evidence trail. The usage credit is simply the budgeting and settlement unit attached to that object.

What the customer sees are workloads, projects, grants, datasets, approvals, and outputs, plus execution policy classes, the estimated reservation against actual settled usage, and the evidence trail. What stays behind the boundary is provider selection, routing logic, cache strategy, batching, scheduling, model combinations, and workload-placement algorithms.

For a research university, this is the difference between a token wallet and an institutional system. A Work Order records the grant or project, the purpose, the actor, the data policy, the execution class, the source manifest, and the maximum credit budget. Required approvals (data-use restrictions, IRB, FERPA, HIPAA, CUI, export control, or sponsor conditions) are checked before execution rather than reported afterward. The institution buys accountable research capacity. It does not need to know which provider token, accelerator, or routing path produced each intermediate step.

A stable public contract requires a detailed private cost model. The two are not in tension; the first depends on the second.

The execution-cost layer has to include reserved-capacity cost, hardware depreciation or lease expense, power and cooling, networking, storage, runtime operations, security, model serving, orchestration, observability, support, expected utilization, and the treatment of failed work. Token processing is one component of that cost. It is nowhere near the whole of it, which is exactly why token counts make a poor commercial unit.

This market runs on claims that sound like facts. Here is a test that separates them, and it applies to every vendor in the category — including the one publishing this essay.

Six distinct things get described with the same confident verb tense: a market thesis, an approved architecture, implemented software, a deployed service, a contracted commercial position, and measured production performance. They are not equivalent, and the distance between the first and the last is where most of the disappointment in enterprise AI lives.

For any claim of the form “the platform does X,” ask what evidence would exist if it were true.

The minimum proof package is short enough to request in a single call. An active deployed release and service path. Gateway policy and fail-closed negative tests. Hold, execute, settle, refund and concurrency tests. Tenant-isolation and authorization evidence. A reconciled usage ledger with an exportable audit record. Measured throughput, latency, utilization and fully loaded cost. And executed contracts, or planning assumptions clearly labeled as such.

What the thesis claims: that governance and settlement become more valuable as compute and models fragment; that capital intensity creates real risk for infrastructure owners even during strong demand; that institutional buyers need stable commercial abstractions and auditability; and that routing and workload control can create switching and policy value.

What the thesis does not claim: that all model tokens are fungible; that every NeoCloud fails or every hyperscaler wins; that usage credits are money, deposits, stablecoins or investments; or that any single broker’s throughput proves the ultimate margin structure of the category.

And because a thesis that cannot be wrong is not worth holding, here is what would falsify it.

None of these look likely today. All of them are observable, and each would be visible well before it was decisive. That is the point of writing them down.

The most important shift underway is not from dollars to tokens. It is from ungoverned consumption to governed work. The winning institutional platform will not merely meter compute. It will connect identity, policy, budget, data, execution, output and evidence into one enforceable transaction. Everything else in this market is a supplier to that transaction.