A $13.7 billion compute contract is a strange object. It reads like a number, behaves like a debt instrument, and settles in megawatt-hours. So when I saw the headline that Rumble — a video platform with a business model closer to a cable channel than to a hyperscaler — had secured a $13.7 billion compute agreement with Anthropic out of a facility in Maysville, Kentucky, the first thing I did was not open the press release. I opened a spreadsheet and tried to convert the dollar figure into physical units. Almost nothing about the result compiled cleanly.
The contract, as reported, is roughly two orders of magnitude larger than anything Rumble has historically booked in annual revenue. That mismatch is the hook, but it is not the story. The story is the unit conversion. A compute contract is denominated in dollars, but it is delivered in floating-point operations and megawatt-hours, and those two currencies drift apart at different rates over time. When you sign a five-year dollar-denominated agreement on a three-year hardware asset financed by short-duration capital, you have not bought compute. You have written a duration bet, and nobody has priced the mismatch.
The most interesting part of this deal is not that it exists. It is that we are watching compute stop being an operating expense and start becoming a financial instrument, without any of the disclosure infrastructure that financial instruments normally carry.
Context: How Compute Became a Contract Asset
To understand why a $13.7 billion figure is even legible, you have to understand what changed in the AI infrastructure market between 2023 and 2026. For most of the cloud era, compute was rented. AWS, Azure, and GCP sold GPU-hours on demand, at spot, with a markup that reflected the cost of overprovisioning, the cost of capital, and the convenience premium. The customer carried almost no risk. The hyperscaler carried utilization risk, hardware obsolescence risk, and the financing risk of building capacity before demand was proven.
That model broke when frontier model training runs grew larger than any single cloud region could accommodate on short notice. Anthropic, OpenAI, and Google DeepMind began signing multi-year reservations to guarantee priority access to capacity that did not yet exist. CoreWeave built its entire business on this — long-dated take-or-pay contracts backed by GPU collateral and debt. Lambda Labs, Crusoe, and a scattering of regional operators followed. By 2025, the compute reservation market had quietly become a capital markets category of its own, complete with structured financing, collateralized GPU debt, and secondary trading of contract positions.
The Maysville facility fits this template but pushes it further than usual. It is a single-site deliverable from an operator with no track record in hyperscale AI infrastructure, against a counterparty that is one of a handful of labs capable of consuming that much capacity. The deal is not a cloud contract. It is a bespoke infrastructure project with a single customer, structured as a services agreement. That distinction matters because a bespoke project has no spot market to fall back on. If the customer walks away, the asset does not redeploy at the same price. It redeploys at whatever the secondary GPU market will bear, which, as we will see, is a number that moves fast in the wrong direction.
It is also worth noting what Rumble brings to this table. The company operates content delivery infrastructure, which means it understands bandwidth, edge routing, and the operational discipline of keeping latency-sensitive workloads alive at scale. That is genuinely adjacent to AI serving. It is not, however, the same as operating a 100-plus megawatt training campus. The gap between a CDN and a hyperscale training facility is not a gap in sophistication. It is a gap in physics.
Core: The Arithmetic That Has to Hold
Let me walk through the numbers the way I would walk through a circuit audit. I want to know whether the contract's dollar figure is physically deliverable, and under what assumptions.
What $13.7 Billion Actually Buys
Assume a five-year term, which is the standard structure for deals of this size. That gives $2.74 billion per year, or roughly $228 million per month. Divide by 730 hours per month and you get approximately $312,000 per hour of contracted revenue.
Now convert to silicon. H100-class rental rates in the wholesale market have hovered in the $2 to $4 per GPU-hour range depending on commitment length, interconnect quality, and cooling. Take $3 per GPU-hour as a mid-case. That implies roughly 104,000 GPUs running at full utilization, or about 139,000 GPUs at a realistic 75% utilization.
Power that. An H100 SXM draws around 700 watts at load. Add host CPU, memory, networking, and a power usage effectiveness of roughly 1.25 for a modern liquid-cooled facility, and you land near 900 watts per GPU of total facility load. At 139,000 GPUs, you are looking at approximately 125 megawatts of IT load and roughly 150 megawatts of total site draw.
So the physical translation of $13.7 billion over five years is: a 150-megawatt training campus with roughly 140,000 flagship accelerators, running at three-quarters utilization, for five consecutive years. That is not a small facility. It is a top-tier site, and it is the kind of asset that takes two to four years of construction and interconnection work before the first GPU powers on.
The Revenue-per-Megawatt Sanity Check
Traditional colocation economics run in the range of $2 million to $4 million of annual revenue per megawatt of critical IT load. A hyperscale AI lease can push that to $8 million or $10 million because the tenant is paying for GPU density, not just floor space.
Now do the division on this contract. At 150 megawatts and $2.74 billion per year, you get approximately $18.3 million of annual revenue per megawatt. That is roughly double even the aggressive AI-dense benchmark and four to nine times a conventional colo.
That number is not automatically wrong. It is possible if the operator is selling GPU-hours directly rather than leasing shells — because then you are monetizing silicon, not real estate. But it tells you something important: this contract is not a landlord-tenant arrangement. It is a compute resale arrangement, and the operator is carrying the hardware risk. That is a very different exposure than a data center REIT would accept, and it is the reason the financing structure matters more than the real estate.
Where Maysville Fits
Maysville, Kentucky is not an accident. It sits in a region with industrial power priced well below coastal markets, meaningful available capacity on the regional grid, and access to the kind of land and water resources that large cooling systems require. Kentucky has actively courted data center investment, and the economics of siting compute in the Ohio River valley rather than Northern Virginia are real.
But siting is a solved problem. The unsolved problem is interconnection. A 150-megawatt load does not appear on a grid by request. It enters a queue, and that queue in most U.S. regions is now measured in years. The binding constraint on this contract is not capital and not silicon. It is the queue position, and the queue position is not something a press release can accelerate.
If the Maysville site already has interconnection rights secured for a large fraction of that load, then the deal is more credible than the headline suggests. If it does not, then the first two years of the contract are effectively a construction financing agreement, and the deliverable slips.
The Depreciation Cliff
Here is where the arithmetic gets uncomfortable, and where I think the real story lives.
A five-year take-or-pay contract on GPU compute assumes the underlying hardware maintains economic relevance for five years. It does not. NVIDIA's architectural cadence has compressed to roughly annual major releases, and each generation delivers a step-function improvement in performance-per-watt and performance-per-dollar. Hopper gave way to Blackwell. Blackwell is giving way to what follows. The secondary market for previous-generation accelerators does not appreciate. It collapses.
I watched this dynamic from the inside in 2024, when I spent six weeks optimizing a ZK-Rollup prover circuit and hitting a wall that had nothing to do with cryptography. The wall was memory bandwidth. Every gate reduction I made was bounded by how fast the proveing hardware could move data, and the hardware I was targeting was already being displaced by the next generation mid-project. Tracing the gas leak in the untested edge case is easy when the edge case is a bug. It is much harder when the edge case is the hardware roadmap.
Apply that to Maysville. Suppose the deal signs today with Hopper-class or early Blackwell silicon. By year three, the same dollar buys two to three times the effective FLOPs on the open market. By year five, the gap is wider still. That means the effective price that Anthropic pays per unit of useful compute rises over the life of the contract relative to the spot benchmark, even though the nominal dollar price is flat.
Now flip the lens. From Rumble's side, the revenue is locked in dollars. The cost of delivering that revenue — in depreciation, in power, in replacement hardware to stay competitive — is not locked. If the contract requires maintaining a certain level of effective performance, and the operator cannot do that on the original silicon, then the operator has to eat the upgrade. If the contract only requires delivering GPU-hours on the original silicon, then the customer is buying depreciating compute at a fixed nominal price, which is a bad trade in a market where compute-per-dollar is deflating fast.
One of the two parties is mispricing duration, and the contract structure determines which one.
The most likely resolution, and the one I would bet on, is that the contract is not a fixed GPU-hour commitment at all. It is more plausibly a capacity reservation with a dollar ceiling, where the operator has flexibility to refresh hardware and the customer gets priority access rather than a fixed unit count. That structure makes the deal survivable. It also means the $13.7 billion number is a maximum, not a promised revenue stream — which is a very different thing to put in a headline.
The Financing Layer
A $13.7 billion delivery obligation implies capital expenditure in the range of $5 billion to $7 billion for the silicon, the shell, the power infrastructure, and the interconnection. That capital does not come from operating cash flow. It comes from debt, equity, or a combination, and the cost of that capital determines whether the project clears its hurdle rate.
Collateralized GPU debt has become a real asset class, but it carries an uncomfortable feature: the collateral depreciates faster than the loan amortizes. Lenders have responded by shortening tenors and demanding higher coverage ratios. A five-year contract paired with a four-year loan on three-year-life collateral is a structurally fragile stack. The code is a hypothesis waiting to break, and so is a capital structure built on a depreciation schedule that assumes hardware ages gracefully.
If the Maysville project is financed with long-dated debt matching the contract term, the operator has a real refinancing wall in year three or four, at the exact moment when the underlying asset is losing market value fastest. That is not a crisis today. It is a crisis if compute prices normalize, and normalization is the base case in any hardware market with this much capacity being announced.
The Revenue Recognition Question
There is a subtler issue. If this is structured as a take-or-pay contract, Rumble may be able to recognize the contract value over time in a way that makes its income statement look dramatically different from its cash flow. That is not fraud. It is accounting. But it is the kind of accounting that has historically preceded uncomfortable restatements when the underlying delivery milestones slip.
I spent part of 2025 auditing optimistic verification modules on a cross-chain bridge, and the failure mode there was structurally identical: a system that looked solvent on paper because the accounting assumed completion of a step that had not yet been reached. Bridges fail when the message-passing assumption breaks. Compute contracts fail when the delivery assumption breaks. The mechanism differs; the shape is the same.
Contrarian: The Security Blind Spot Is the Contract, Not the Data Center
Everyone watching this deal is asking the wrong question. They are asking whether the data center can be built. The more important question is what the contract becomes once it exists.
A $13.7 billion compute agreement is not a procurement document. It is a tradable claim on future compute output. And once you have a tradable claim, you have the raw material for derivatives. Compute futures already trade informally. Compute forwards, compute options, and compute-indexed structured products are the natural next step. Once those exist, you get tranching, and once you get tranching, you get the same failure mode that produced 2008: assets whose risk was priced at the instrument level, not the underlying level.
Here is the blind spot. The smart contracts that will formalize these instruments will be audited. They will pass. They will be formally correct. And they will still fail, because the thing they are referencing — the delivered compute — is a physical asset whose value is decaying on a schedule that no on-chain oracle can observe.
I made this argument in 2026 when I audited the zero-knowledge proof system behind an AI-agent identity protocol. The cryptography was elegant. The proof aggregation logic had a subtle soundness error that would have permitted Sybil admission, and I flagged it. But the deeper issue was that the protocol was pricing identity as a static credential when identity is a dynamic property that decays with the behavior it describes. Modularity is an entropy constraint. Every layer you add to abstract away the physical world also adds a surface that can silently drift out of sync with it.
Compute contracts have exactly this problem. The abstraction layer — the dollar figure, the contract term, the promised capacity — is clean and legible. The physical layer — the GPUs, the power, the cooling, the depreciation curve — drifts. And the further you get from the physical layer, the longer the drift goes undetected, because the abstraction is designed to look stable.
The specific risk I would watch is not default. It is silent repricing through hardware substitution. If the operator, under pressure to maintain margin, quietly runs the workload on lower-tier or older hardware while reporting the contract as delivered, the abstraction stays intact and the underlying degrades. That is the kind of failure that no audit catches, because nothing on the document is wrong. Only the physics are wrong.
There is also a concentration risk that gets almost no attention. A single-site, single-customer contract of this size creates a shared fate between two firms with very different balance sheets. If the customer reduces its training cadence — and capex discipline at AI labs is not guaranteed forever — the operator has no second tenant to absorb the shortfall, and the asset has no liquid market to clear into. Liquidity mining taught us this lesson already: the number on the dashboard reflects the commitment, not the demand. A $13.7 billion contract is a claim about future appetite. Appetite is the one input that has never been reliably forecastable.
Takeaway
The Maysville deal is legible as a demand signal and dangerous as a duration bet. The $13.7 billion figure will be quoted in every summary, and the number that actually matters — the ratio between the contract term and the economic life of the hardware delivering it — will not be quoted anywhere.
The forward-looking question is not whether Rumble can build the site. It is whether the contract was written with a hardware refresh clause, how the financing amortizes against a decaying collateral base, and whether the customer's commitment is a floor or a ceiling. Those three details determine whether this is a transformative infrastructure deal or the first visible crack in the compute-as-asset-class thesis.
I will be watching the loan documents before I watch the GPUs. Latency is the tax we pay for decentralization, but duration is the tax we pay for optimism, and it is collected all at once at the end.