0:00
/

Who Pays for the GPUs?

E19 – Ankur Patel (Ares) on how GPU financings are structured and what role ABF is playing in financing the AI revolution

A one-gigawatt data centre costs $15 to $20 billion of physical infrastructure. The GPUs and servers inside it cost two to three times that again. The largest line item in the financing is the asset with the shortest life.

That is residual-value risk. Ankur Patel, a partner in the Ares Credit Group, is not trying to underwrite it as an asset call.

Thanks for reading! Subscribe for free to receive new posts and support my work.

The loan tenor is three to six years and sits inside the offtake, not alongside the economic life of the chip. Amortisation is set so the remaining principal at maturity is meant to meet whatever the GPU is still worth. If a gap remains, a creditworthy name writes a residual-value guarantee over it.

Tenor, amortisation and the RVG are there so you are not taking leftover chip risk. Ankur’s line on the recording is the underwriting conclusion: he is not relying on the assets to get out. The credit is the customer contracted to pay for the compute. The hardware is secondary collateral.

“One of the main goals from a lender perspective is to make sure you have as little residual value at the end. Residual value is the value of the GPU at the end of three years, four years, five years.”
— Ankur Patel, 44:07

If a GPU financing lands on the desk, these are the credit questions. How much residual is left at year three, four, five? Is the offtake longer than the loan, and what are the termination rights? Who writes the residual-value guarantee, and does that name still pay if the offtaker is gone? If the sector overbuilds, do you still have contracted cash flow, or a depreciating box with no bid?

We spent the hour on how that is documented: what asset-based finance actually funds in a data centre, why power is the binding constraint, and how ABF, term loans and high yield take different slices of GPU risk. Ankur sits in Ares’ alternative credit business — just under $60 billion across 40 to 50 asset classes — with a concentration in heavy assets, including digital and AI.


What is in it

  • What ABF actually funds in a data centre. Powered shells and turnkey construction, then substations, turbines, engines and battery systems, then the contractor fleet of excavators, cranes and dump trucks on site — and increasingly the GPUs themselves.

  • Tenor sits inside the offtake, not alongside it. Three to six years, deliberately shorter than the contract paying for the compute, because any portion of the loan that outlives that contract is exposure to a resale market that may not exist.

  • Amortisation is set against residual value. The balance is walked down to meet whatever the chip is still worth at maturity, and where a gap remains a creditworthy name writes a residual-value guarantee over it.

  • The underwriting order puts the chips last. Offtake rating first, then the operator and its uptime SLAs, because a terminated contract leaves you with full racks and no payer. Then the guarantee and who stands behind it. Chip generation — B300 and Rubin over H100 — is a preference, never the exit.

  • Power is the binding constraint. Data centres take five to six percent of US grid power today and go north of ten percent within three to five years, against a 30 to 35 gigawatt shortfall — which is why developers are buying options on land near gas and building generation on site while they wait for an interconnect.

  • Pre-lease behind-the-meter power is the newest ABF trade. That spend used to be equity funded. ABF now bridges the twelve to twenty-four months to a signed lease with an SPV, a pledge of the power collateral and a dynamic borrowing base, before a bank takeout recycles the capital into the next project.

  • The overbuild answer is supply-side. Ankur is not especially worried, and his three reasons are power availability, equipment supply chains and zoning approvals — brakes on the pace of building rather than a view on whether the compute gets used.


Chapters

00:00 Financing the AI data centre build-out

02:34 What ABF actually finances: shells, power, GPUs, site equipment

06:26 How ABF differs from bank lending and working capital

08:22 Asset-based finance in plain English: the three buckets

11:24 Lending against assets versus owning them

12:24 Where ABF sits: investment grade to opportunistic

17:04 Insurance balance sheet versus Pathfinder funds

19:00 Sourcing edge and a 100-person investment team

20:30 The asset classes inside infrastructure finance

23:17 Collateral and covenants versus contracted cash flows

24:38 Why ABF underwrites portfolios, not single assets

27:29 Digital versus non-digital infrastructure

31:54 Power as the binding constraint on data centre build

35:42 Behind-the-meter power as an ABF opportunity

39:17 Where ABF capital enters the build phase

40:39 The GPU financing market and who borrows

42:29 How GPU financings are structured: SPV, tenor, offtake

46:07 Segments of GPU financing: high yield, term loans, ABF

47:39 Underwriting GPU risk and lender protections

50:22 The data centre overbuild bear case

52:01 Structural protections against overbuild

53:07 Red flags and lessons

54:57 Three-year outlook: how big ABF gets and who funds it


About the guest

Ankur Patel is a Partner in the Ares Credit Group, working in the firm’s alternative credit business — the asset-based finance platform, just under $60 billion of AUM across 40 to 50 asset classes, with over half of that in the Pathfinder family of funds. He is a generalist on a roughly 100-person investment team, with a concentration in heavy assets: real estate, infrastructure, digital and AI. His career runs close to thirty years — a neuroscience degree in the 1990s, then strategy consulting, then investment banking, then more than fifteen years in asset-based investing, first at Fortress and for the past eight years at Ares.

Ankur on LinkedIn · Ares Management · In the Gaps


Support material

F+f 19 Ankur Patel Transcript
116KB ∙ PDF file
Download
Download


Fixed + Floating

Fixed + Floating is a credit podcast hosted by Portfolio Manager Josef Pschorn. Long-form conversations with credit market practitioners - portfolio managers, analysts, restructuring advisers and academics - on private credit, high yield, distressed debt, CLOs, liability management and credit policy.

Subscribe on YouTube · Spotify · Apple Podcasts

Follow on LinkedIn · X · Fixed + Floating on LinkedIn

If there is a guest or a topic you want covered, reply to this post or message me on LinkedIn.

Fixed + Floating is for informational purposes only. Not investment, legal or tax advice. Guest views are their own and do not represent their employer.

Discussion about this video

User's avatar

Ready for more?