Contracts, Not Chips, Are the Collateral: How the AI Compute Buildout Gets Financed

Neil Tiwari describes Magnetar as a 20-year-old alternative asset manager with three strategies: private credit, venture and a systematic public strategy. Its edge in the AI buildout, he says, comes from a long record of financing capital-intensive businesses with creative structures. The firm already had experience in real estate and energy, which meant it understood much of what a data center is made of, even if compute was new to it.

Magnetar met CoreWeave in 2021, when the company was moving from mining Ethereum to high-performance computing, with visual effects for films as the first use case. Magnetar invested before the AI trade began, adding optionality for GPUs in many kinds of workloads. When CoreWeave began training models for OpenAI in 2023, demand took off, and Tiwari credits the company's founders' energy-asset background and its focus on scale and reliability for winning early.

Scale depends on access to capital, power and data center space, he explains, while reliability means running a huge GPU fleet at around 99.9% uptime despite hardware failures and software problems. CoreWeave had begun working on that around 2017 and 2018. Sarah Guo agrees from her own portfolio companies, saying CoreWeave has a reputation for reliability that not everyone has matched.

Tiwari puts 2026 hyperscaler capex for AI compute and infrastructure at $660-690 billion, scaling to trillions over the next several years. Using only equity would cause enormous dilution, so the less discussed problem is how capital is structured. Guo says she reached the same conclusion when she decided her own portfolio shouldn't take the dilution for a cluster.

His first example is the delayed-draw term loan or SPV structure, in which the collateral is the GPUs plus the contracts. He stresses that the real asset is the contracted cash flow from investment-grade counterparties such as Microsoft or Meta, on take-or-pay contracts of around five years. That is what the press missed when it likened the debt to lending against a used car, since the chips are only secondary or tertiary collateral.

The debt amortizes in full over four to five years, he says, while the payback period on the capex is roughly two to three, so there is no balloon payment and the depreciation question mostly falls away. Any residual value sits with the cloud operator, who could redeploy the hardware without new debt. The structures are also evolving: early SPVs held only investment-grade customers, but newer ones blend those with AI-native companies such as model labs and software startups to balance the risk.

Guo notes that her companies describe compute as supply constrained. Tiwari says that in 2023 and 2024 the constraint was chips, and many predicted an overbuild. By 2026 chips are more available, he says, but building and operating data centers needs people, power and infrastructure, and it is not clear the latest generation is available at scale.

Newer chips matter because of efficiency, not just raw power. Citing a SemiAnalysis article by Dylan Patel, he says Nvidia's claim that Blackwell is 30 times more efficient than Hopper for inference was beaten by analysis showing 90 to 100 times. Guo jokes that her favorite Jensenism is "the more you buy, the more you save," and Tiwari says that on price-performance it is actually true.

On the circular-financing critique, Tiwari says the circularity concern applies to speculative capacity or pure vendor financing, which is not what Magnetar sees. In his view demand is insatiable, unlike the dark fiber of the early 2000s, because he says there are no dark GPUs. He puts enterprise AI at about $37 billion last year and growing fast, and says unit economics are positive for the hyperscalers that are the ultimate buyers.

Tiwari says the split between training and inference keeps shifting toward inference as useful, ROI-positive applications arrive. Inference is more complex than assumed: it brings latency, fungibility and cost-optimization issues, and demand is peaky, unlike a training cluster that runs flat out. He also calls it a memory-throughput problem, with prefill and decode phases that must be optimized across a fleet.

Distribution is the third difference. Training clusters tend to be centralized at 50 to 150 megawatts under one roof, while inference could use four- or five-megawatt sites stitched together, where software matters much more. Guo adds that inference clouds must deliver traditional-cloud reliability, and the two discuss how identical-looking compute can perform very differently in reliability, cost and speed.

The key question for Magnetar now, he says, is how to finance this distributed inference build. For application-layer companies, compute is the largest cost-of-goods line, and inference clouds often resell capacity from other clouds, producing layered margins. That is pushing them to want to own infrastructure, and he also points to Nvidia's "AI factories," dedicated on-premises clusters for large corporations, as a further sign of where demand is going.

Over the past year Tiwari has spent a lot of time in power and energy markets. He agrees there is a power problem but says it is more nuanced than reports suggest: utilities are built around peak demand, leaving generating assets such as gas peakers idle most of the year. Making that power usable depends on flexibility, storage and distribution, and he mentions an investment in a company building a distributed utility layer that stores excess capacity and releases it when needed.

The true bottleneck over the next 6 to 12 months, he says, is almost mundane: structural steel, electricians, substation transformers and air chillers, all needed just to build a powered shell on a piece of land. On generation, regulation is a major hurdle, so developers use "bring your own capacity," such as a site that starts with 10 megawatts of grid interconnect and grows to 50 with solar or gas turbines. Guo points to Crusoe and flare gas as an example of making more energy consumable.

On sovereign buildouts, Tiwari says governments in places such as India, the Middle East and Southeast Asia, and the United States too, treat compute as a national security matter and can fund themselves with government capital. The nuances he sees are finding partners that can rapidly build and scale GPU compute in a country that lacks them, and cybersecurity, where he says much work remains.

On physical AI, he contrasts the asset-light SaaS era with the asset-heavy phase that began around 2021. Hardware was hard to scale in the 2010s because software was an afterthought, but general-purpose AI software can make hardware easier to scale. Robotics, manufacturing, drones and defense remain capital intensive, so he expects them to need debt and project finance like compute did, and Guo agrees that investment-grade buyers will sign contracts that debt can be raised against.

Asked about the capital rotation away from software, Tiwari says he has watched a different industry fall each day, citing wealth advisors, consulting and real estate. He believes usable AI took a big step up in late 2025 and into 2026, and credits Anthropic and Claude, so the disruption fear is not baseless. But he calls the reaction "a bit much" and says it hits whole sectors instead of individual names.

His reasons are twofold. On free-cash-flow terms, SaaS valuations are at their lowest in years, since margins have risen while revenue multiples stayed flat or fell. And while AI might look able to rebuild Slack or Salesforce on paper, those products are integrated across many enterprise systems, which is much harder to replicate. Guo adds that it is also a question of whether anyone wants to rebuild and own such software, and that some companies are structurally more protected than others, so investors must select carefully.