The AI revolution has a dirty secret: it runs on money as much as it runs on math. While the world marvels at the latest reasoning models and agentic breakthroughs, a quieter but equally consequential revolution is unfolding in the plumbing of finance — one that determines whether the trillion-dollar AI infrastructure buildout actually happens or stalls out.
At the center of this financial engineering sits Magnetar Capital, a $22 billion alternative asset manager that has become one of the most consequential — and least discussed — players in the AI compute ecosystem. On a recent episode of the No Priors podcast, host Sarah Guo sat down with Neil Tiwari, the managing director leading AI infrastructure at Magnetar, to pull back the curtain on how capital is being structured, deployed, and innovated to power the physical backbone of artificial intelligence.
The conversation ranged from the mechanics of GPU-backed debt instruments to the surprising bottleneck of structural steel, painting a picture of an industry where financial creativity is just as important as technological innovation.
From Ethereum Mining to the AI Gold Rush
Magnetar's journey into AI infrastructure didn't begin with a grand thesis about large language models. It started, improbably, with cryptocurrency.
"We actually stumbled across the compute problem before it was compute. We met CoreWeave back in 2021, and that was when they were actually transitioning from mining Ethereum into high-performance compute."
At the time, CoreWeave was repurposing the same GPUs that had been mining crypto for visual effects work — rendering Marvel movies and similar compute-intensive media. Tiwari and his team made their first investment before the AI trade had even started, drawn instead by the optionality they saw in GPU-based computing.
"I wouldn't say that we could foresee everything that happened. We just happened to be at the right place at the right time, and we continued to double down as the company progressed."
What made Magnetar an unlikely but natural fit for this space was its diverse background across asset classes. While the firm had no prior data center investing footprint, it had deep experience in real estate, energy, and property investing — precisely the building blocks of modern data centers.
"A lot of the elements for what constitutes a data center — power, energy, land, real estate — we had a lot of the background in those spaces. I think we were new to compute. That was a new sector for us."
The real inflection came in early 2023, when CoreWeave began training models for OpenAI. The sheer volume of compute required to train large language models was unprecedented, and the company that had quietly built reliability and scale over years suddenly found itself at the epicenter of the most important technology wave in a generation.
The Art and Science of Financing GPUs
With projected AI infrastructure CapEx from hyperscalers alone reaching $660 to $690 billion in 2026 — and scaling to trillions over subsequent years — the question of how to finance this buildout becomes existential. As Tiwari bluntly put it, equity alone won't cut it.
"This is billions to trillions of dollars of CapEx. And just using equity dollars alone is not an efficient way to scale this. That's obviously massive dilution."
The solution Magnetar helped pioneer involves structured debt instruments built around special purpose vehicles (SPVs). Inside each SPV sits the capital expenditure — the GPUs and the contracts associated with them. But here's the crucial insight that Tiwari says the media consistently got wrong: the GPUs themselves are not the primary collateral.
"What was oftentimes characterized in the media was these debt structures had GPUs as collateral, and that's like putting a used car as collateral, which is obviously just gonna depreciate incredibly fast. What got missed was the GPUs themselves were actually like the second or tertiary level of collateral in those instruments. The primary collateral was the contracted cash flows from investment-grade counterparties."
Those counterparties — Microsoft, Meta, and similar hyperscalers — sign take-or-pay contracts typically spanning five years, creating a reliable revenue stream that debt holders can lend against with confidence. The debt itself is structured to fully amortize over the contract period, meaning there's no balloon payment at the end and the depreciation of the underlying GPUs becomes largely irrelevant to the risk profile.
This is a fundamentally different proposition than what the skeptics assumed. It's closer to infrastructure project finance than speculative tech lending.
The Evolution of Debt Structures
As the market has matured, so have the financial instruments. In the early days, only investment-grade counterparties could anchor these structures — the space was too nascent and the operators too unproven for anything else. Now, Tiwari explains, the portfolio construction has become more sophisticated.
"What you're starting to see is a blend of investment grade and non-investment grade... You're seeing those companies get mixed in alongside the IG companies to build a portfolio. 'Cause now you have the history that you can do this."
This evolution is critically important for the broader AI ecosystem. It means that AI labs and startups — not just hyperscalers — can increasingly access debt-financed compute, potentially democratizing access to the infrastructure needed to train and run frontier models.
The Circularity Question
One of the most persistent criticisms of the AI infrastructure boom is that it resembles circular financing — companies buying chips from NVIDIA to build clouds that serve customers who are themselves funded by the same investment ecosystem. Tiwari pushed back firmly on this characterization.
"I've... you don't see any dark GPUs, any GPUs unused. And then number two, you're starting to see actual economic value. Last year, enterprise AI had about 37 billion of total TAM, and it's continued to grow like crazy."
The comparison to the dot-com era's dark fiber — massive overbuilds of telecommunications infrastructure that went unused — doesn't hold, he argued. Every GPU deployed is being utilized. The tokenomics of inference are delivering positive ROI. The demand is real, the value creation is measurable, and the ultimate buyers are the largest, most creditworthy companies in the world.
From Training to Inference: A New Frontier
As the AI industry matures, the shift from training to inference workloads is reshaping how infrastructure needs to be built, operated, and financed. Tiwari described inference as far more complex than he initially expected, with challenges around latency, demand variability, and memory throughput that make it fundamentally different from training.
"Training cluster might be 50, 100, 150 megawatts in one kind of four walls. I think you're starting to see distributed inference which could be four or five megawatts and five separate data centers and stitching them together in different areas."
This architectural shift toward distributed inference clusters creates entirely new financing challenges. Training infrastructure is centralized and serves a small number of massive customers; inference infrastructure is distributed, serves a broader customer base, and requires different approaches to reliability and optimization.
Tiwari highlighted NVIDIA's concept of "AI factories" as a glimpse of where things are heading — dedicated compute environments built for specific companies or workloads, potentially sitting on-premises rather than in shared cloud environments. For Fortune 500 companies consuming massive amounts of inference, owning their own AI factory may become as natural as owning their own data infrastructure was a generation ago.
The Surprisingly Mundane Bottlenecks
Perhaps the most striking revelation from the conversation was what's actually slowing down the buildout in 2026. It's not chips — though the latest generation remains scarce — and it's not even power generation in the traditional sense.
"The true bottleneck, at least in the short term, the next 6 to 12 months, is... things like structural steel. It's finding electricians that can build out the power infrastructure — substations, transformers, air chillers."
Steel. Electricians. Transformers. The most advanced technology buildout in human history is being constrained by the same factors that limit any large-scale construction project. This reality check serves as a reminder that AI infrastructure, for all its digital sophistication, is ultimately a physical enterprise.
On the power side, Tiwari offered a more nuanced view than the prevailing "we can't generate enough electricity" narrative. He pointed to significant stranded power capacity across the grid — utilities built for peak demand that sits idle most of the year — and argued that the near-term challenge is more about distribution and storage than raw generation.
"There's actually quite a bit of stranded power across the grid, across the country... I look at the power problem as being kind of multiple-fold. The first one is, how can you take the power we have on the grid and actually make it usable?"
Magnetar has invested in companies like Torus, which is building a distributed utility layer to store excess capacity and distribute it when needed — essentially creating a mesh network for power that mirrors the distributed nature of emerging inference infrastructure.
Physical AI: The Next Capital-Intensive Wave
Looking ahead, Tiwari sees the same capital intensity dynamics that defined the compute buildout extending into physical AI — robotics, manufacturing, drones, and defense. The key difference from the hardware failures of the 2010s, he argued, is that general-purpose AI software now makes hardware much easier to scale.
"Part of the scars was, it was so difficult to scale hardware companies because the software was so difficult to build. What you're seeing now is now that you have more general purpose software via AI, it can make the hardware easier to scale."
The implication is clear: the project finance playbook that Magnetar developed for GPU clouds — structured debt backed by contracted cash flows from creditworthy buyers — will need to be adapted for robotics companies, drone manufacturers, and other physical AI ventures. The capital intensity isn't going away; if anything, it's spreading.
The Software Rotation: Overblown or Overdue?
On the topic of public markets rotating out of traditional software stocks in favor of AI infrastructure and AI-native companies, Tiwari struck a measured tone. He acknowledged the step-up in AI capability that's driving the rotation but cautioned against painting with too broad a brush.
"Free cash flow margins have steadily increased significantly for SaaS as a whole over the last four or five years, and revenue multiples have stayed the same or gone down. And so, to me, that's a bit of an exaggeration because it really has to do with individual names versus sectors."
The nuance matters: some software companies are structurally protected by deep enterprise integrations that AI can't easily replicate, while others are genuinely at risk. The market, Tiwari suggested, is hitting all names with the same hammer rather than making the more difficult but more accurate company-by-company assessment.
The Capital Layer as Competitive Advantage
What emerges from this conversation is a portrait of an AI ecosystem where financial innovation is as critical as technical innovation. The ability to structure creative financing — blending equity and debt, building portfolios of investment-grade and non-investment-grade counterparties, engineering amortization schedules that match asset depreciation — is not just a back-office function. It's a competitive moat that determines who can build at scale and who gets priced out.
The AI buildout is often framed as a story about chips, models, and energy. But underneath all of it runs a layer of capital infrastructure that is just as complex, just as innovative, and arguably just as consequential. As the industry scales from hundreds of billions to trillions in cumulative CapEx — and as the same dynamics extend into inference, sovereign AI, and physical AI — the companies and investors who master this financial engineering will shape the physical reality of the AI age as much as any model architecture or chip design.
In an era where the conversation is dominated by parameters and benchmarks, it's worth remembering that every GPU, every data center, and every AI factory begins not with a line of code but with a line of credit.