CoreWeave’s Contract Machine: How a Neocloud Turned GPUs into a Utility
When a Side Hustle Becomes a $55 Billion Contract Stack
CoreWeave was never intended to resemble a utility. It started as a scrappy GPU shop born out of crypto, riding cycles of Ethereum mining and cheap silicon. Five years later, it is sitting on roughly $55.6 billion of revenue backlog, about 96–98 percent of which is locked into multi‑year, take‑or‑pay contracts, supported by roughly 2.9 gigawatts of contracted power across more than 40 sites.[1]
Most start‑ups spend their first decade chasing logos, figuring out pricing, and praying for product‑market fit. CoreWeave did that in fast-forward, then quietly built something stranger: a GPU-denominated, contract-heavy cloud platform that behaves more like a vertically integrated utility than a conventional software company.
The story of CoreWeave is not just “AI demand is big” or “NVIDIA is powerful.” It is about how a small specialist provider hacked the power‑compute stack—from silicon access to customer contracts to power procurement—and turned GPU scarcity into a balance sheet, a moat, and a set of very specific risks.
This piece walks through that story in five parts: where CoreWeave came from, how the business model actually works, what lies underneath it in terms of power and infrastructure, how the capital stack is structured, and what that combination means for the next cycle.
1. Origins: From Crypto Rigs to Contracted GPU Utility
The origin story is familiar to anyone who has followed the first wave of “neocloud” providers. CoreWeave began its life as a crypto-mining operation, buying GPUs when the rest of the market still viewed them as gaming cards and hash-rate machines. When the Ethereum party ended, the founders faced a binary choice: wind down or repurpose the silicon for something with more durable demand.
They picked the harder path. Instead of chasing generic cloud workloads, CoreWeave focused on high‑performance, parallelizable jobs—first VFX and rendering, then machine learning, and eventually the full spectrum of AI training and inference. The bet was simple: if you could consistently offer better performance‑per‑dollar than hyperscalers on these workloads, and you had the discipline to wrap that in long‑dated contracts, you could ignore most of the broader cloud market and still build a large business.
That bet intersected with three macro forces:
NVIDIA’s stranglehold on AI silicon. The company secured early and preferential access to H100, then B200 and Blackwell‑class hardware, in part by leaning into NVIDIA’s partner program and in part by being willing to scale fast on new architectures.[2]
The GPU supply crunch of 2023–2025. When demand from OpenAI, Meta, and others outstripped hyperscaler capacity, customers became willing to sign multi‑year, take‑or‑pay deals with specialists if it meant earlier access to hardware.
The rise of workload‑specific clouds. In Powering the AI Edge: Equinix in the Decade of Density, I argued that infrastructure would bifurcate into general‑purpose clouds and a handful of highly optimized, workload‑specific platforms.[3] CoreWeave is what that thesis looks like in practice: an AI-first, GPU-dense, contract-heavy cloud.
The result is a company that, by late 2025, looks nothing like a boutique hosting shop. It is an index on frontier‑model demand with a customer list that includes OpenAI, Meta, and Microsoft; a multi‑billion‑dollar capacity backstop from NVIDIA; and a contract stack that locks in years of cash flow—if the infrastructure shows up on time.
2. The Product: Selling GPU‑Hours, Not Servers
CoreWeave’s business model begins with a deceptively simple product definition: it sells GPU hours, not machines, not racks, and not generic “instances.”
Most of its revenue comes from three interlocking pieces:
Training capacity. Customers buy access to large, tightly networked clusters of GPUs—tens of thousands of H100s or newer parts—either as:
GPU‑hour blocks, where they pay a fixed price per GPU‑hour; or
Reserved clusters, where they contract for a slice of a cluster over a fixed period.
Inference capacity. Here the product looks more like a utility tariff. Customers pay for throughput and latency—measured in tokens processed per second, under strict service-level objectives—rather than raw GPU hours. Some take dedicated capacity; others ride shared pools.
Storage and networking. As the platform scaled, AI object storage and multi‑cloud networking stopped being bolt‑ons and started to matter. Storage alone is now north of $100 million in annual recurring revenue, and network features (like direct paths into other clouds for inference) help keep workloads sticky.[4]
The key design choice is contractual. By 2024–2025, roughly 96–98 percent of CoreWeave’s revenue came from multi‑year, take‑or‑pay agreements, typically two to five years long.[5] On‑demand and spot usage exist but are deliberately kept small.
These contracts share some common features:
Fixed or formula‑based GPU‑hour pricing. Customers agree to a dollar-per-GPU-hour (or equivalent) rate that may fluctuate within narrow bands but does not vary with day-to-day market conditions.
Minimum commitments. Even if utilisation drops, customers owe a base payment. That creates annuity‑like revenue for CoreWeave and gives lenders comfort.
Prepayments and term loans. Some anchor deals, particularly with OpenAI and Meta, are accompanied by structured financing arrangements or prepayments that help fund the hardware.[6]
This is not a traditional SaaS model. It is more like a hybrid between capacity contracts in power markets and reserved‑instance deals in cloud. The advantage is visibility: a $55.6 billion backlog provides management and investors with multiple years of contracted revenue coverage.[1] The trade‑off is rigidity: pricing cannot be repriced overnight if hardware costs fall, and utilisation risk sits squarely on CoreWeave’s shoulders.
In "The Price of a GPU Hour," I laid out a rough willingness-to-pay band for high-end AI GPUs, with Blackwell-class hardware clearing around $9–$ 12/GPU-hour at the top end, depending on configuration and region.[7] Neocloud providers like CoreWeave have already seen H100 pricing compress into the $2–3/GPU‑hour range in some cases, even as hyperscalers still post higher list prices.[8]
CoreWeave’s hedge against that compression is utilisation and performance. It is betting that:
Its clusters can deliver more useful tokens per GPU hour than its peers.
Its contract structure keeps racks busy most of the time.
Its storage and networking revenue can grow faster than headline GPU‑hour prices fall.
If those three things hold, GPU‑hour price compression is painful but survivable. If they do not, a contract stack built in a tight market could become a margin headwind in a looser one.
3. The Engine Room: Performance as a Business Model
Most cloud providers talk about performance. CoreWeave tries to monetize performance differentials directly.
SemiAnalysis’ ClusterMAX benchmark, which ranks large GPU clusters on “goodput” (effective throughput), currently gives CoreWeave the only Platinum rating, estimating around 96 percent goodput on certain large‑scale training workloads versus roughly 90 percent for typical hyperscaler configurations.[9]
Under the hood, that advantage comes from:
Dense NVLink fabrics with non‑blocking topologies and up to 3.2 Tbps per node.
Quantum‑2 InfiniBand and BlueField‑3 DPUs, offloading network and storage tasks.
A control plane—Mission Control, Slurm on Kubernetes, and internal schedulers—that is topology‑aware, aware of failure domains, and optimized for long‑running training jobs.[10]
Why does this matter for the business model?
Because a 10 percent uplift in effective throughput at the same price is equivalent, for a buyer, to a 10 percent price cut with no change in invoice. That is a real edge when large customers are explicitly tracking dollars per million tokens and comparing providers on that basis.
CoreWeave leans into this in three ways:
Cluster‑scale sales. Instead of selling small instances and letting customers cobble together clusters, it sells access to pre‑built, well‑characterized clusters—tens of thousands of GPUs wired in ways that benchmarks can validate.
Inference SLAs. For inference, it is willing to write contracts around throughput and latency rather than raw hardware. That shifts the conversation to outcomes and makes performance a differentiator, not just a marketing line.
Lifecycle management. Older fleets (A100s, then H100s as Blackwell ramps) do not get dumped. They are re‑contracted for one‑ to three‑year inference deals. In at least one case, a 10,000‑plus H100 cluster was re‑upped two quarters before expiry at pricing within five percent of the original deal.[11]
This is where the story connects back to Powering the AI Edge. The unit that matters is no longer megawatts of IT load; it is tokens per watt, per rack, and per dollar of contracted spend.[3] CoreWeave’s bet is that if it can stay on the right side of that ratio, its contract machine will still look sensible when GPU supply finally catches up.
4. The Infrastructure Stack: Power, Shells, and the Tyranny of Energization
To understand CoreWeave’s business model, you have to look at what sits underneath the contracts: power and real estate.
By late 2025, CoreWeave had around 590 MW of active (energized) power and about 2.9 GW of contracted power, spread across roughly 41 data centers.[1],[12] More than 1 GW of that remained “open to sell” but was expected to energize over the next 12–24 months.[12] Management guided to >850 MW of active power by year‑end.
That infrastructure is built on three legs:
Third‑party powered shells. CoreWeave leases capacity from specialist data‑center developers. No single provider accounts for more than about 20 percent of contracted power, and the fleet spans more than 30 sites.[13]
Self‑developed sites. In places like Kenilworth, New Jersey and Lancaster, Pennsylvania, CoreWeave is building its own shells, with roughly mid‑three‑hundreds of megawatts of potential capacity. This gives it more control over schedule and design.[14]
Anchored campuses. In ERCOT, for example, CoreWeave has a large allocation—on the order of several hundred megawatts—at Galaxy Digital’s Helios campus, effectively utilising existing transmission and generation infrastructure rather than developing a new site.[15]
The single most important operational risk in this stack is time-to-energisation.
In the third quarter of 2025, delays from one shell provider were sufficient to push a meaningful portion of expected fourth-quarter revenue into the first quarter of 2026, even though the total contract value was preserved through extensions and remedies.[2],[16] Construction‑in‑progress swelled, asset turnover sagged into the mid‑30s per cent, and the company’s liquidity became more sensitive to whether sites hit their energisation dates.
This is exactly the dynamic I described in The Tyranny of Time: How “Bridge Power” Became a Feature, Not a Bug: when the cost of waiting is high enough, developers will accept higher operating costs, complex hybrids, or non‑ideal sites if it means energising faster.[17]
CoreWeave’s model bakes this into the business logic:
It is willing to pay up or structure contracts creatively to get power and shells on time.
It treats energy risk as a first-order variable, like GPU availability, not as a footnote.
It is open to BYOP‑style solutions—nuclear PPAs, gas‑plus‑battery hybrids, behind‑the‑meter projects—where they reduce schedule risk.[18]
For investors, the key point is that CoreWeave’s “product” is not just GPU‑hours. It is GPU-hours that arrive on a specific timetable, tied to contracts that start billing regardless of whether a site is ready. That turns construction and interconnection timing from an execution detail into a core part of the business model.
5. The Capital Stack: Growth Story, Credit Story
All of this lives on a very busy balance sheet.
In 2025, CoreWeave guided to $5.05–5.15 billion in revenue and $12–14 billion in capital expenditure.[19] By the third quarter, it had already raised more than $14 billion of debt and equity year‑to‑date, including $1.75 billion in senior notes and a $2.6 billion GPU‑secured term loan tied to the OpenAI contract.[20] Total capital raised since early 2024 is in the ballpark of $25 billion, with first major maturities around 2028.[20]
Some other markers:
Q3 2025 revenue of about $1.365 billion.[19]
Liquidity of roughly $3 billion in cash and marketable securities.[19]
Q3 interest expense of roughly $311 million, with guidance for $350–390 million per quarter going forward.[21]
If this feels familiar, it should. In Racks, Real Estate, Risk – A Deep Dive on Digital Realty, I argued that large data‑center platforms are fundamentally capital‑structure stories, not just growth stories.[22] CoreWeave is following a similar playbook, compressed into a shorter timeline and wrapped around GPUs instead of generic racks.
The twist is NVIDIA. The chipmaker not only supplies critical hardware; it:
Holds roughly 7 percent of CoreWeave’s equity; and
Has agreed to a roughly $6.3 billion capacity backstop through 2032, effectively acting as a buyer of last resort for unused capacity.[6],[23]
That backstop reduces utilization risk—if demand wobbles, NVIDIA will still pay for capacity—but it does not eliminate pricing risk (renewal prices could fall) or refinancing risk (debt still has to be rolled or repaid).
So the business model, at the capital level, looks like this:
Sign long‑dated, take‑or‑pay GPU contracts with high‑quality counterparties.
Use those contracts to raise large amounts of debt and structured capital.
Deploy that capital into GPUs, power, and shells as fast as practical.
Rely on utilization, performance, and software‑layer revenues to keep margins healthy as hardware cycles and pricing evolve.
It is a powerful model when the world wants more AI, more GPUs, and more tokens. It will be stress-tested to see if any of those axes flatten.
6. Strategic Position: Where CoreWeave Fits in the Power–Compute Map
By now, CoreWeave is firmly in the “neocloud” bucket alongside peers like Lambda and Crusoe: GPU‑dense, performance-optimised, much smaller than hyperscalers but laser‑focused on a specific workload set.
Its strategic position is defined by three relationships:
With customers. CoreWeave is both a partner and a risk for hyperscalers and frontier‑model labs. OpenAI, Meta, and Microsoft all buy from multiple clouds, and they are perfectly capable of building more of their own infrastructure. CoreWeave earns its place by offering better economics, earlier access, or specific performance advantages on certain workloads.
With suppliers. NVIDIA is a supplier, an equity holder, and a quasi‑offtaker via the capacity backstop. That alignment is a moat today. It could become a constraint if NVIDIA’s own priorities shift or if custom silicon from hyperscalers begins to significantly impact the market.
With the power system. As explored in META’s Power Play: A Field Guide to Energy Procurement in the Power‑Broker Era and Power Brokers of AI – Vistra, Calpine, Talen, the next phase of competition will be about who can secure, shape, and pay for power most intelligently, not just who can rack the most GPUs.[12],[18]
CoreWeave is positioning itself as:
A contracted‑power specialist, comfortable with large anchored campuses and structured power deals.
A technology‑forward orchestration layer that can sit on top of different physical campuses and power structures.
A credit‑aware developer, building infrastructure only when there is contracted demand to support it.
That story is still being written. The next few years will show whether CoreWeave becomes:
A GPU‑denominated utility with high utilization, strong contracts, and durable margins; or
A highly levered call option on NVIDIA’s roadmap and frontier‑model demand, vulnerable to cycles in GPU supply, customer bargaining power, and capital markets.
Either way, the business model is no longer a mystery. CoreWeave is what happens when you treat compute like capacity, contracts like fuel, and time-to-energisation as a resource every bit as scarce as GPUs.
Endnotes
[1] CoreWeave backlog, revenue mix, and contracted power figures drawn from company disclosures and broker analysis as of the third quarter of 2025, including Compass Point and Mizuho coverage.
[2] NVIDIA alignment and early access to new GPU generations based on broker research and NVIDIA partner program disclosures.
[3] Powering the AI Edge: Equinix in the Decade of Density – Part 1–5, Renewables Investor, 2024–2025.
[4] Storage and networking revenue contributions drawn from Mizuho and RBC coverage highlighting AI object storage surpassing $100 million in annual recurring revenue.
[5] Contracted revenue mix and take‑or‑pay structure derived from Loop Capital and company commentary, which estimate 96–98 percent of revenue coming from multi‑year committed agreements in 2024–2025.
[6] OpenAI, Meta, and NVIDIA anchor‑deal values taken from Compass Point, HSBC, and company disclosures, including a $22.4 billion OpenAI contract, $14.2 billion Meta contract, and $6.3 billion NVIDIA capacity backstop running through 2032.
[7] The Price of a GPU Hour, Renewables Investor, 2025.
[8] Neocloud versus hyperscaler H100 pricing ranges, including $2–3/GPU‑hour for some neocloud offerings and $5–8/GPU‑hour list prices at major clouds, drawn from expert calls and RBC pricing surveys in late 2025.
[9] SemiAnalysis ClusterMAX Platinum ranking and 96 percent goodput estimate for CoreWeave versus ~90 percent for hyperscalers from Compass Point’s November 2025 initiation on CoreWeave.
[10] NVLink, InfiniBand, and DPU‑enabled architecture specifics—3.2 Tbps fabrics, 100k‑plus GPU clusters, Quantum‑2 and BlueField‑3—drawn from HSBC and Needham technical descriptions of CoreWeave’s network and control‑plane stack.
[11] Example of a >10,000 H100 cluster re‑contracted two quarters before expiry at pricing within 5 percent of the original agreement from CoreWeave’s Q2 and Q3 2025 earnings calls.
[12] Active and contracted power figures—~590 MW energized, 2.9 GW contracted, >1 GW open to sell—based on CoreWeave’s Q3 2025 earnings call and Needham coverage.
[13] Portfolio diversification—no single third‑party shell provider above ~20 percent of contracted power—and parallelization across 30‑plus sites from company commentary and Needham research.
[14] Self‑build powered‑shell capacity estimates in Kenilworth, NJ, and Lancaster, PA based on Northland Securities’ November 2025 note on CoreWeave’s development pipeline.
[15] Galaxy Digital’s Helios campus allocation and broader ERCOT power‑planning context from RBC’s AI and hyperscale infrastructure research.
[16] Contract extension mechanics and SLA remedies (including rent discounts) related to powered‑shell delays from Needham’s post‑Q3 2025 analysis and company earnings commentary.
[17] The Tyranny of Time: How “Bridge Power” Became a Feature, Not a Bug, Renewables Investor, 2025.
[18] Power Brokers of AI – Vistra, Calpine, Talen, Renewables Investor, 2025, which covers AWS’s 1.9 GW nuclear PPA with Talen, gas‑plus‑battery hybrids, and the rise of structured BYOP models for data‑center loads.
[19] FY25 revenue guidance ($5.05–5.15 billion), capex guidance ($12–14 billion), and Q3 2025 revenue ($1.365 billion) from Mizuho and RBC summaries of CoreWeave’s guidance.
[20] Debt and equity raises—$1.75 billion senior notes in July 2025, $2.6 billion GPU term loan, >$25 billion total capital raised since early 2024—and maturity profile from company earnings calls and Evercore credit commentary.
[21] Q3 interest expense (~$311 million) and forward guidance ($350–390 million per quarter) from Mizuho’s November 2025 CoreWeave note.
[22] Racks, Real Estate, Risk – A Deep Dive on Digital Realty, LinkedIn essay, 2025.
[23] NVIDIA’s ~7 percent equity stake in CoreWeave and the $6.3 billion capacity backstop’s treatment in backlog versus remaining performance obligations from Evercore, HSBC, and company disclosures.