Who's Actually Funding the AI Buildout?
In a Nutshell
Magnetar Capital, a $22B asset manager, funds AI compute buildout through innovative DDTL/SPV debt structures backed by take-or-pay contracts from hyperscalers like Microsoft and Meta, collateralizing GPUs and cash flows for 4-5 year amortization, enabling CoreWeave's scale from Ethereum mining to AI training/inference without equity dilution. AI infrastructure faces $660-690B capex in 2026 amid bottlenecks in power distribution/storage, transformers, and skilled labor, shifting from chip constraints to turning chips into reliable revenue assets, with inference demanding distributed, low-latency clusters. Capital rotates from SaaS to capex-heavy infrastructure, labs, and physical AI (robotics/defense), fueled by insatiable demand, sovereign national security builds, and efficiency gains like 90-100x from Hopper to Blackwell.
These notes were generated by AI and may contain inaccuracies.
Neil Tuari of Magnetar Capital, a $22 billion alternative asset manager, discusses their role in the AI compute buildout, financial innovation, GPU depreciation, and future of AI compute.
Magnetar, in its 20th year, has three primary strategies: private credit, venture, and systematic/quantitative public strategy. They specialize in building capital-intensive businesses using creative financing structures to optimize balance sheets.
Magnetar first invested in CoreWeave in 2021, when it transitioned from Ethereum mining to high-performance compute for visual effects (e.g., Marvel movies). This predated the AI boom. They added optionality for various GPU applications, including machine learning and AI training, doubling down as use cases expanded.
Magnetar had prior experience in real estate, energy, power, land—key data center elements—but was new to compute.
By late 2022, AI discussion intensified; in 2023, CoreWeave trained models for OpenAI, driving massive growth due to unprecedented LLM training compute needs.
CoreWeave founders' energy asset management background enabled power/energy access. They focused on scale (capital, energy, power, data centers) and reliability (managing GPU fleets at 99.9% uptime, handling failures/software challenges). Started building tech stack in 2017-2018.
Capex for AI compute/infrastructure from hyperscalers projected at $660-690 billion in 2026, scaling to trillions over years. Challenges: access to energy/power/data centers, plus capital structure. Equity alone causes massive dilution.
DDTL structures/SPV debt: SPV holds capex (collateral: GPUs + contracts). Primary collateral: contracted cash flows from investment-grade counterparties (e.g., Microsoft, Meta)—take-or-pay contracts, ~5 years.
GPUs are secondary/tertiary collateral. Media mischaracterized as risky (depreciating assets like used cars), missing guaranteed offtake.
Debt amortizes over 4-5 years (full payoff, no balloon); capex payback ~2-3 years. Depreciation irrelevant as debt paid via contracts; residual value/upside to cloud provider (e.g., CoreWeave), enabling redeployment without new debt.
Now blending investment-grade (hyperscalers) with non-investment-grade (AI model labs, startups, software companies) in portfolios, balancing risk as operators gain experience.
2023-2024: Chip-constrained. Now (to 2026): chips more available, but bottlenecks in people/power/infrastructure to turn chips into revenue-generating assets. Latest-gen chips (e.g., H100/H200, Blackwell) scarce at scale; demand from high-end players/startups for efficiency.
Dylan Patel (SemiAnalysis) data: Hopper to Blackwell = 90-100x inference efficiency (vs. claimed 30x).
"The more you buy, the more you save."
Demand is insatiable; no dark GPUs (unlike 2000s fiber overbuild). Enterprise AI TAM: $37 billion last year, growing. Tools deliver positive ROI.
Circularity concern only for speculative builds/vendor financing; here, financing supports ROI-positive use cases with hyperscalers as ultimate buyers (positive unit economics).
Growing from training to inference as applications develop. Inference complexities:
- Latency, fungeibility, cost optimization, variable demand (vs. training's 100% utilization).
- Memory throughput problem (prefill/decode phases).
- Distribution: decentralized clusters (e.g., 4-5 MW in separate data centers vs. training's 50-150 MW monolithic).
B10 optimizes distributed inference at scale; software critical.
Started with training for investment-grade hyperscalers. Now: finance distributed inference for inference clouds/application layers. Their COGS highest item is compute; they buy spot/unused capacity (layered margins), pushing to own infrastructure for margins/control.
Inference clouds (e.g., Base) must deliver traditional cloud reliability; current distributed ops lack it. Spot pricing data (SiliconData, Kerman Lee): identical compute varies wildly in reliability/cost/speed.
Nvidia's AI factories: Dedicated clusters for corporates/Fortune 500/AI companies (on-prem-like, workload-specific), alongside hyperscalers/neo-clouds. Power limits scale; smaller units easier.
Inference demand ensures offtake for usable compute.
Over the last year, significant time has been spent in power and energy markets exploring solutions to scale power for the gap. There is a power problem, more nuanced than reporting suggests. It's not just that we can't generate power; there's quite a bit of stranded power across the grid. Utilities focus on peak power with natural gas peakers for high-demand moments, only a few days a year. Generating assets exist but are stranded.
The power problem involves multiple aspects: making existing grid power usable through flexibility and storage. Focus on energy storage and distribution to store unused capacity, peak demand shave, and distribute when needed. Investment in Taurus, building a distributed utility layer like mesh infrastructure to store excess capacity from various sources and distribute when needed. This is a critical layer.
Longer term, there's a generation problem, but shorter term (next 6-12 months), it's distribution and storage. True bottleneck is structural steel, finding electricians, substations, transformers, air chillers—specific power infrastructure to build a powered shell on land. Bottlenecks are people and equipment.
On generation, regulatory challenges. Bring your own capacity: sites starting with 10 MW grid interconnect but scaling to 50 MW by adding solar, natural gas turbines, and other technologies to make sites usable. Much focus on bring your own capacity short term.
Crusoe and flare gas origin story exemplifies lots of energy available that can be made more consumable.
Sovereigns view compute and AI as national security, funding clusters with government capital differently from private clusters. News from India last week, Middle East, Southeast Asia. Interesting aspects: partners to build capacity and cybersecurity implications. Sovereigns need players to rapidly scale GPU compute in their countries, often lacking local expertise to build sovereign ecosystems. Much work needed on cyber side for scaling sovereign AI.
Physical AI is capex intensive. From 2010 to early 2020s, capital asset light mode with SaaS; no Magnetar in SaaS. Since 2021, compute became asset heavy, drawing infrastructure investors. Physical AI extends this.
Scars from 2010s hardware companies that were hard to scale due to software difficulties after heavy hardware spend. Now, general-purpose AI software makes hardware easier to scale by interacting with more hardware.
Compute required flexible capital: not just equity, but debt and project finance. Same need in physical AI due to capital intensity. CoreWeave needed billions to scale cloud. Robotics, manufacturing, drones, defense are capital intensive; AI helps scale faster, but capital intensity remains. Optimize balance sheets for physical AI growth.
Early AI compute contracts structured to support investment-grade buyers with debt-raiseable contracts. Robotics was a way to lose money; now on boards of two robotics companies. Products will support investment-grade buyers in home or industrial settings for bad human jobs or labor shortages.
Capital rotating out of software to infrastructure, labs, AI natives. Every day, industries tank: wealth advisors, consulting, real estate, payments. Towards tail end of 2025 into 2026, step up in usable AI performance (e.g., Anthropic's Claude). AI can disrupt non-AI native industries; market reaction exaggerated.
Two factors: SaaS valuations at lowest in years on free cash flow basis. Free cash flow margins increased significantly over last 4-5 years; revenue multiples same or down. Exaggeration due to individual names vs. sectors—some will maximize AI value, others won't. Hammer hitting all names indiscriminately.
Applications sounding interesting (AI rebuilding Slack, Salesforce) overlook enterprise integration across services/systems, harder to replicate than markets react. Fundamental question: does anybody want to rebuild/own it? Within software, some structurally protected, others at risk—need to select.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
