AI credit has been le sujet du jour for the last half of 2026, and this month Wall Street answered it with a $500 billion consortium built on the same architecture that produced the 2008 financial crisis. Nvidia announced its $500 billion financing consortium with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on 10 August, structured to make graphics-processing-unit (GPU) compute "an investable asset class" [1]. On the 12th, Reuters reported Oracle's five-year credit default swap at roughly 200 basis points against about 40 a year ago, with Société Générale's implied cumulative default probability now above 16% [2]. And Nikkei Asia had reported on 21 July that Alphabet, Microsoft, Amazon, Meta, and Oracle now carry approximately $1.65 trillion of off-balance-sheet obligations tied to GPU purchase commitments and long-term data-center leases — roughly eightfold growth in about four years, and now more than the $1.35 trillion those five firms disclose on balance sheet [3]. The market buying the new consortium is paying five times more than a year ago to hedge the same names underneath it. Half of Big Tech's economic leverage sits in footnotes.

This is a systemic-credit story dressed up as a chip story, and the architecture that would fix it is neither exotic nor hypothetical. It is a public-blockchain infrastructure. It is running today, at institutional scale, for tokenized treasuries and stablecoin settlement, with roughly $17.2 billion of tokenized real-world assets on Ethereum alone [4]. Wall Street's response to the fragility has been to build a bigger platform on top of the plumbing that produced the fragility. Can we ever learn from the past?

Innovation Is Not a Container

A special-purpose vehicle (SPV) slices cash flows and sets the queue for who gets paid when things go wrong. Wall Street has been building them for forty years. Calling one an innovation because it holds GPUs instead of mortgages is like calling a shipping container innovation because it now holds semiconductors.

Read the CoreWeave arc. On 30 March, CoreWeave closed an $8.5 billion delayed-draw term loan (the DDTL 4.0 Facility), rated A3 by Moody's and A (low) by DBRS, marketed as the first investment-grade financing secured by GPU-related high-performance-computing infrastructure and an associated customer contract. Non-recourse to the parent. Matures March 2032. Draws through June 2027. Prices at the Secured Overnight Financing Rate (SOFR) plus 2.25% floating and roughly 5.9% fixed [5][6]. The release does not disclose the GPU count, the chip generation, the utilization assumption, or the residual-value curve underwriters used. The anchor customer is Meta, but only through Bloomberg reporting. The contract itself, and its termination and step-in provisions, sit outside the public record [7]. That is the state of disclosure on the first investment-grade GPU-backed financing in history.

Ten weeks later, on 18 May, the same borrower group closed the $3.1 billion DDTL 5.0 Facility. Ba2 from Moody's, BB+ from Fitch. Investment grade to high yield, in ten weeks, on materially the same Nvidia hardware [8]. The variable that changed was the customer's credit — DDTL 5.0 supports two "large non-investment-grade customer contracts" rather than Meta. Two months after that, in August, DDTL 5.5 priced at SOFR + 5.50%, with the borrower and its arrangers deliberately building in a maturity mismatch: a five-year loan against underlying customer contracts averaging around three years [9]. That last one, at least, prints what it is. The other two do not.

The wrappers are elaborate, and they are also completely orthogonal to the thing that matters, which is what a lender or a supervisor can see about the GPUs behind the paper. Wall Street's answer to that gap this month was a $500 billion consortium. Ours is different.

The Depreciation Problem

Chip generations turn on roughly an annual cadence now. Performance-per-watt gains are front-loaded within each generation. The resale market for used AI accelerators is thin, correlated across borrowers, and closed to almost everyone who is not already a GPU cloud. Imagine you underwrote a five-year GPU loan against a fleet of H100s in Q1 2024, which was fine because Blackwell had not yet shipped. It is now late 2026. The H100 rental price has moved considerably. Your loan still matures in 2029. The tail here is not the straight-line depreciation on the borrower's financials. It is the regime change no past agreement is measuring.

The hyperscalers themselves cannot agree on how long this equipment lasts. In its 2024 10-K, Amazon disclosed a shortening of a subset of its servers and networking equipment from six years to five effective January 2025, plus a $920 million accelerated-depreciation charge in Q4 2024 on equipment retired early. Amazon's stated reason, verbatim: an increased pace of technology development, particularly in artificial intelligence and machine learning. Cost to 2025 operating income: roughly $0.7 billion [10]. In the same 10-K filing cycle, Meta went the other way. Effective January 2025, most of its servers and network assets extended from five years to five and a half, worth about $2.9 billion in reduced depreciation for 2025 [11]. Two of the largest operators of AI-capable hardware in the world, disclosing opposite views of that hardware's useful life at the same time, under the same accounting standard.

Private lenders writing GPU loans inherit that argument. AI-related sectors accounted for 34% of private-credit deal activity in 2025, up from an average of 17% over the prior five years [12]. Private-credit borrower leverage now sits at 5–6x debt to earnings before interest, taxes, depreciation, and amortization (EBITDA), and closer to 7x once EBITDA adjustments are unwound, versus roughly 4x in leveraged loans [12]. On top of that base, someone is now writing GPU-backed paper that assumes a residual-value curve Amazon and Meta cannot agree on. Whoever that paper ends up in — the private-credit fund, the insurer, the retail annuity, the pension plan downstream — is bearing a bet on Jensen Huang's next product cycle without necessarily knowing they are.

How Risk Becomes Systemic

1) The exposure surface

The AI trade is not a chip-stock story or a neocloud story. Gavekal Research put it plainly this month: an AI-capex rollover would weigh on corporate profits throughout the economy, erase much of the wealth effect supporting consumption, and remove a significant support for bond yields [13]. Most macro desks in 2026 already work off that assumption, which is also why a repricing of AI compute would not stay in the AI complex.

Start with the balance sheets people can see. Bond sales by Amazon, Alphabet, Meta, and Oracle reached roughly $195 billion in the first six months of 2026, about 80% more than the $108 billion those four raised across all of 2025 [2]. Then look at the ones people can't. Nikkei Asia's July study estimated $1.65 trillion of off-balance-sheet obligations across Alphabet, Microsoft, Amazon, Meta, and Oracle: GPU purchase commitments, long-term data-center leases, and joint ventures that do not count as debt under accounting rules until the facilities go live. Meta alone accounts for roughly $420 billion, nearly triple its reported debt [3]. Roughly half of Big Tech's economic leverage lives in footnotes, and it is growing.

The migration path is short and it reaches into everyone's book. Bank capital funds the private-credit funds, which fund the neoclouds, which buy the GPUs. Life insurers also fund the private-credit funds, and retail annuities and pension plans fund the insurers. US life insurers backed by private-equity owners hold close to $900 billion in insurance liabilities, up from $67 billion in 2012, and wrote 35% of new US annuity sales in 2023 [12]. And now, as of 10 August, Nvidia's $500 billion consortium with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR adds another institutional-capital channel directly on top of that stack [1]. Larry Fink called compute "a critical asset class driving the next leg of global economic growth" at the launch. Compute is also credit exposure to a hardware asset with an annual product cycle, funded through a plumbing layer supervisors already struggle to see through. Both things are true. Only one of them was on the press release.

2) The pricing wedge

The market and the risk specialists are already trying to price this. It is doing so with the crudest instrument in the toolbox. Oracle's CDS at 200 basis points against a DDTL 4.0 print at SOFR + 2.25% is the same underlying AI-credit exposure being priced in two directions in the same market. Origination prints tight because the collateral has been reclassified as accepted. The CDS market widens violently because the risk itself has not changed — only its disclosure has. Between those two prices, actors on the losing side of the mispricing have no way to seize a hedge. The information they would need — collateral state, cash flows, live leverage — is not visible. Transparent credit collapses that gap. Spreads still move; they move continuously, and every participant acts on the same picture. Not a feature crypto is trying to bolt on. The default state of any market where the collateral is visible.

3) 2008 was fundamentally an opacity problem

Blockchain infrastructure would not have prevented the mortgage crisis. The subprime losses were going to arrive; that is not the interesting part of the story. The interesting part is that between HSBC's first writedown in February 2007 and Lehman's collapse in September 2008, no one could answer the basic questions in real time: who owned the exposure, how big it was, and against what pledge. Nine months of that. Counterparties pulled funding on suspicion because measurement was not available, and the suspicions turned out to be broadly right, which made them righter. A discovery lag of that length turns a survivable loss into a systemic event. It is where contagion runs.

With enumerable leverage, observable collateral, and programmable covenants that fire before impairment — cash sweeps, distribution lock-ups, reserve calls — the same losses would still arrive, but they would arrive to a market that could see them and act on them. That is the difference between a bad quarter and a crisis. The industry knows this. The 2008 postmortems are on every risk officer's shelf. Choosing to build the next $3 trillion of credit exposure on the same architecture that produced 2008's blindness is not a technical decision. It is a choice.

4) Opacity as subsidy

There is a stronger version of the transparency argument, and it deserves stating plainly. Opacity is not a neutral feature of a market. It is a benefit that accrues to whoever profits from concealment. Originators price paper tighter than the risk deserves because the buyer downstream cannot see the risk. Fund managers charge fees on mark-to-model NAVs that would look different under mark-to-observable-cash-flow. Rating agencies get paid to bless what cannot be independently verified. The retail annuity buyer at the end of the chain pays for all of it, twice— once in the fee that flowed to each layer above them, and once again when the concealment ends.

When the borrower's collateral state and  compliance are visible in real time to every downstream holder, the incentive to write mispriced paper collapses at the underwriting desk. Not because the regulator caught it. Because the buyer would not have bought it at that price. Junk credit can still exist, and it just gets sized correctly rather than wrapped up as investment grade. DDTL 5.5 at Ba2/BB+, SOFR + 5.50%, with its openly disclosed maturity mismatch, is what pricing meeting risk looks like [9]. The unhealthy version is the same exposure wrapped so it prints tighter. Every intermediary in that wrapping chain makes money on the wrapping.

The problem is not leverage. Uncounted leverage is the problem, and the reason it stays uncounted is that concealment pays. Fix the visibility and you fix the incentive. Bad paper stops getting written because the buyer can finally see it. That is what transparent credit does. It aligns the incentives of the people writing the paper with the people who will end up holding it — which, historically, has been the harder part.

What Blockchain Does, and What It Doesn't

Blockchain-based credit infrastructure will not prevent losses, fix underwriting, or make risky assets safe. It changes what lenders and supervisors can see about a live credit, and how quickly. The whole claim, and it is bigger than the industry gives it credit for because it is the specific claim that makes a 2008-shaped discovery lag structurally harder to reproduce.

In practice, that means the following. Confidentiality remains fine. Customer identities, contract prices, and revenue details stay private. What can be shared with authorized parties — lenders, trustees, auditors, regulators, each with role-appropriate access — is a tamper-evident record of what was reported, when, by whom, and under which rules. Cryptographic attestations prove things about the collateral without exposing the underlying documents. Uptime, cash sweeps, lien priority, that sort of thing. Provenance is not the same as economic truth; on-chain data is only as good as the sensors, attestors, and dispute machinery feeding it. But the current alternative delivers neither provenance nor truth. It delivers quarterly certificates.

Applied to GPU credit specifically, the digital-twin data set is not exotic. Hardware identity and lien status. Uptime and energy usage. Contracted revenue and customer concentration. Debt balance and reserve accounts. Most of it is already collected somewhere by the borrower, the servicer, or the site operator. Moving it onto a shared permissioned record is an engineering exercise, not a research problem. The Financial Stability Board is currently trying to reconstruct this picture through surveys and vendor data feeds, and by its own admission the market's growth is outrunning the exercise [12]. A permissioned lien and leverage register synchronized with off-chain filings does the same work in something closer to real time, which is the actual point.

The Bank for International Settlements has already sketched the blueprint from the top down: a tokenized unified ledger combining central-bank reserves, commercial-bank deposits, and government bonds on a programmable platform that integrates messaging, reconciliation, and settlement [14]. Applied to credit specifically, blockchain adds programmable behaviour that pairs with the covenant grid. A coverage ratio drops below 1.4x: monthly reporting escalates, uptime attestations get required, the lender committee reviews demand. Anchor-customer downgrade: holdco distribution lock-up triggers, new-debt basket suspends, underwriting resets. Attested oracle failure: draws freeze, independent audit runs, lender group decides. None of it is autonomous liquidation. It is meant to buy creditors time and options before losses crystallise, with humans in the loop for anything material. A far superior system.

The constraints are definitely there: legal enforceability stays off-chain; a tamper-evident lien record is not a substitute for a properly filed security interest, and courts will not defer to a smart contract that conflicts with statute. Oracle failure is a genuine risk, and each facility's documentation has to specify who signs each attestation, what their legal liability is if the data are wrong, and what happens during a dispute-window draw freeze. Tokenization does not manufacture secondary-market depth also. Private consortium infrastructure may also capture much of the tokenization benefit without ever touching a public-chain token — the BIS itself has stated it prefers a system anchored in central-bank money [14], although we believe that public permission-less is far superior and resilient.  All of that is already available and has been tested. None of it is a reason to build the next $3 trillion of AI credit on infrastructure that already failed once. Crypto rails make it harder for risky assets to be misclassified as safe, and this could be the difference between having a crisis and avoiding one from the get go. 

Ethereum, and Where We Land

If AI credit is going to be issued, serviced, and settled on programmable rails at institutional scale, Ethereum is today the only public infrastructure equipped to carry it. Not because ether is going up. Because everything else that would need to exist for the architecture to work — regulated stablecoin liquidity, institutional custody, deep tokenized-asset activity, a settlement environment with security institutions can price — already exists there.

RWA.xyz identifies approximately $17.2 billion of tokenized real-world assets on Ethereum across 2,087 instruments as of late August, more than every other public chain below it combined [4]. The share is compressing, from around 52.85% in early June to roughly 45–46% by early August as the total on-chain RWA market grew to $38 billion and competing chains absorbed the growth [15]. The absolute footprint keeps rising while the market share erodes. Both facts are true, and we are watching the second closely. Ethereum's proof-of-stake consensus is the largest economically-secured settlement environment institutions can currently price; finality requires two-thirds of staked ether to justify, and an attacker attempting to finalize conflicting checkpoints would need at least one-third of the total stake to violate protocol safety and be slashed [16]. Every proof-of-stake chain has a security budget. Ethereum's is currently the biggest. And the base layer is neutral in a way a single issuer's proprietary chain is not — no bank owns it, no consortium can rewrite history when the disputes get bad.

Wall Street's answer to the AI-credit fragility problem this month was a bigger financing platform on top of the same architecture that produced the fragility. Another SPV is not innovation — it is a container that has been on the shelf since the 1980s, and a $500 billion consortium built on top of it is doing at scale what CoreWeave did in miniature between March and August: writing bigger paper against an asset whose economics keep sliding out from under the disclosure. The alternative is public-blockchain infrastructure with $17 billion of tokenized real-world assets on Ethereum today, stablecoin flows measured in the hundreds of billions, and regulated tokenized-treasury products with names on them investors already know. It is not experimental. It is not hypothetical. The industry building the paper against $3 trillion of hyperscaler and neocloud obligations has spent 2026 talking about GPUs as an asset class and not talking about the settlement layer underneath. The reason is not that the alternative does not work. It is that the current infrastructure pays too many people too well.

That is the $3 trillion question. It has an answer. Wall Street has chosen not to use it.

Disclaimer

This material is provided for information purposes only and does not constitute investment advice, a recommendation, or an offer or solicitation to buy or sell any financial product or service in any jurisdiction.

The views expressed are those of the author as at the date of publication and are subject to change without notice. While reasonable care has been taken, no representation or warranty, express or implied, is made as to the accuracy, completeness, or reliability of the information contained herein, and no reliance should be placed upon it.

This material has not been reviewed or approved by the Dubai Financial Services Authority, the British Virgin Islands Financial Services Commission, or the Cayman Islands Monetary Authority, or any other regulatory authority.

This material is not intended for distribution to, or use by, any person in any jurisdiction where such distribution or use would be contrary to local law or regulation. In particular, this material does not constitute a public offer of securities in the British Virgin Islands or the Cayman Islands.

Any investment or service referred to herein will be available only to persons who meet the relevant eligibility criteria under applicable laws and regulations, including, where applicable, Professional Clients (as defined by the Dubai Financial Services Authority) or other equivalent categories of sophisticated or professional investors under applicable law.

Investments involve risk, including the possible loss of capital. Past performance is not a reliable indicator of future results.

“Axys” is a trading name used by a network of independent businesses. Axys does not operate as, and is not, a separate legal entity in any jurisdiction and has no distinct legal personality. Each business operating under the Axys name is independently owned and operated.

Axys Capital  Ltd is authorised and regulated by the Dubai Financial Services Authority. Axys Investment Management Ltd is authorised and regulated by the British Virgin Islands Financial Services Commission.

References

1. NVIDIA Corporation, "NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital," 10 August 2026. Available at: [Nvidia — $500 billion AI compute financing partnership](https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital).

2. Reuters, "Soaring AI hyperscaler default hedges aren't what they seem," 12 August 2026. Available at: Reuters on hyperscaler CDS.

3. Kohei Yamada / Nikkei Asia, "Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding," 21 July 2026. Available at: Nikkei Asia on hidden AI debt.

4. RWA.xyz, "Analytics on Tokenized Real-World Assets — Networks and RWA League Table," accessed 31 August 2026. Available at: RWA.xyz networks table.

5. CoreWeave, Inc., "CoreWeave Closes Landmark $8.5 Billion Financing Facility, Achieving First Investment-Grade Rated GPU-backed Financing," 31 March 2026. Available at: CoreWeave investor release, 31 March 2026.

6. CoreWeave, Inc., Form 8-K disclosing the DDTL 4.0 Facility, filed with the US Securities and Exchange Commission, 30 March 2026. Available at: CoreWeave Form 8-K on DDTL 4.0 Facility.

7. Bloomberg News, "CoreWeave Inks $14 Billion Meta Deal in Latest Sign of AI Demand," 30 September 2025 (subsequently expanded to approximately $21 billion through 2032). Available at: Bloomberg on CoreWeave–Meta contract.

8. CoreWeave, Inc., "CoreWeave Closes $3.1 Billion Loan Facility, Expanding Access to Public Markets for GPU-Backed Financing," 18 May 2026. Available at: CoreWeave investor release, 18 May 2026.

9. CoreWeave, Inc., "CoreWeave Closes $2.6 Billion Loan Facility, Expanding Financing Flexibility for AI Infrastructure," 10 August 2026. Available at: CoreWeave investor release, 10 August 2026.

10. Amazon.com, Inc., "Form 10-K for the Fiscal Year Ended December 31, 2024," change in useful lives of servers and networking equipment and Q4 2024 accelerated-depreciation charge. Available at: Amazon 2024 Annual Report (10-K).

11. Meta Platforms, Inc., "Form 10-K for the Fiscal Year Ended December 31, 2024," Change in Accounting Estimate note on server and network asset useful lives. Available at: Meta Platforms 2024 Annual Report (10-K).

12. Financial Stability Board, "Report on Vulnerabilities in Private Credit," 6 May 2026. Available at: Report on Vulnerabilities in Private Credit (PDF).

13. Gavekal Research, "How Sustainable Is The AI Capex Boom?", 19 August 2026 (subscription). Available at: Gavekal — How Sustainable Is The AI Capex Boom?.

14. Bank for International Settlements, "Next-generation monetary and financial system takes shape, based on a tokenised unified ledger," press release accompanying the BIS Annual Economic Report 2025, 24 June 2025. Available at: BIS press release, 24 June 2025.

15. Crypto Briefing, "Ethereum dominates tokenized RWA market with 52% share," 7 August 2026 (citing rwa.xyz data as of early June 2026). Available at: Crypto Briefing on Ethereum RWA share.

16. Ethereum Foundation, "Proof-of-stake (PoS)," official developer documentation on consensus, finality, and economic security, accessed August 2026. Available at: Ethereum Foundation, Proof-of-stake documentation.