· macrofireside.com · August 26, 2026
Ahead of the much-awaited NVIDIA results today after market close, I felt a response to Jensen Huang’s essay of August 10, 2026, might be appropriate. The essay announced the arrival of a massive AI compute financing platform in collaboration with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.
The idea is to raise a substantial amount of capital from a multitude of third-party names to fund NVIDIA’s computing infrastructure which can then be rented to clients of various kinds: hyperscalers, corporations, and others. The formidable distribution networks (a.k.a. reach into capital providers) that these top asset managers have built over decades can come in handy in this context.
In turn, it has led Jensen Huang to claim that his firm’s compute is an asset class unto itself. This essay examines the claim closely and finds that it does not pass muster even at the definition stage, because an asset class represents so much more. For instance, it is an ecosystem characterized by multiple counterparties/issuers risk across the whole spectrum, and not by a single name risk; so much so that no issuer can set the price.
On the other hand, the proposed architecture has some resemblance to the mid-2000s structured credit machine which I am personally familiar with. More directly, when the chips are down, you know where the buck will eventually stop. A public balance sheet is not ostensibly part of the current conversation, and it will be déjà vu all over again.
The claim on the table
“NVIDIA compute is an investable asset class,” so thundered Jensen Huang in his August 10th essay. Between chip engineering and financial engineering, an envi(di)able ecosystem, a compute island for all things AI, can be built. With every CUDA release enriching the chip side coupled with the residual value of hardware underwritten up to a quarter of each project, $125 billion in aggregate, reassuring buyers and expanding its redeployment base, the firm can pull it off. So believes Jensen. Six of the biggest names in global capital are on board to help NVIDIA create a $500 billion third-party financing platform to make it happen. That includes Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.
Let me be clear at the outset that I admire Huang. NVIDIA, the firm he built and patiently nurtured, now dominates the AI revolution. Truly a rare distinction. So is his 10th of August piece which does not shy away from asking tough questions, including funding circularities, and seeks answers in public. But that’s still not the same as answering those. To me, its thesis falls on its own terms, no matter where we may be in the new AI paradigm. One stock does not an asset class make.
What an asset class actually is
Simply put, an asset class is a mélange of assets which share a broadly common valuation framework but can vastly differ among themselves in terms of risk characteristics due to issuer differences. Independent price discovery is therefore a must. Commercial real estate qualifies because a new building in the neighborhood does not make the existing stock go obsolete. Similarly, a new toll road does not cut the value of the roads financed on earlier vintages.
Treasuries, corporate credit, farmland: in each case the investor’s fate is spread across issuers, geographies and technologies, and the secondary market prices the collateral without asking the seller’s permission.
Now hold the proposed structure against that standard. One company designs the collateral. The same company manufactures it, prices it, sets the annual cadence on which it becomes obsolete, invests in several of the largest customers whose rents constitute the collateral’s cash flows, and now proposes to backstop the collateral’s residual value from its own balance sheet. Every exposure inside the $500 billion loads on one factor. In the language of the correlation desk, this is a portfolio whose default correlation to a single name approaches one, which makes it not an asset class but a syndicated single-name exposure in infrastructure clothing, and the tailoring is very good indeed.
The seller holds the obsolescence clock
Real infrastructure earns long-duration financing because nobody controls its decay: a bridge depreciates on physics, while a GPU depreciates on a product roadmap, and the roadmap belongs to the vendor arranging the financing. Hopper gave way to Blackwell, Blackwell is giving way to Rubin, and NVIDIA has committed itself publicly to an annual rhythm. The essay presents the A100, launched in 2020 and still in commercial service, as evidence of a decade-long economic life. The example proves less than it appears to: old silicon runs, but the financing question is whether it competes.
In a power-constrained data center the scarce input is not the chip but the rack slot and the megawatt behind it, and every improvement in performance per watt raises the opportunity cost of feeding power to the prior generation. The metal keeps running long after the marginal economics have quit. Amazon, which knows a great deal about server economics, shortened depreciation on a portion of its fleet last year. Useful-life assumptions are drifting shorter across the hyperscalers just as lenders are being invited to underwrite them longer.
CUDA is real, and so is the software moat. But CUDA extends usefulness, not competitiveness. A software release that improves the whole installed base improves the newest installed base most, and the financing math turns on relative economics, not absolute ones.
There is a second hand on that clock. The largest renters of NVIDIA compute are also the largest funders of its replacements: the hyperscalers each build custom silicon expressly to need less of the collateral, and AMD prices against it from outside. The pressure lands where the residual value lives, since an aging GPU earns out its later years in inference, and inference is the workload where in-house parts compete best. The decade-of-life assumption requires customers to keep renting old NVIDIA silicon at rates that clear the vehicles’ debt service, at the very moment those customers are spending billions to escape it.
Rental prices are not residual values
The essay’s empirical centerpiece is rental pricing: one-year H100 rates up from roughly $1.70 to $2.35 per GPU-hour between October and March, on-demand medians up in similar fashion, Blackwell capacity commanding $5 to $7. Every number in that list is true, and none of them bears on the question. Spot and one-year rental strength in the middle of a shortage tells you the market is short compute today. It tells you nothing about the value of a five-year-old GPU in year six, which is the number the residual-value assumption actually requires. Lenders in 2006 watched home prices rise with just as much precision, and it told them just as little. Observable strength in current cash flows is precisely the condition under which residual assumptions get written most aggressively.
There is also the awkward matter of who sets the rents. NVIDIA has invested in OpenAI, CoreWeave and other large offtakers. In the weeks before this announcement, Bloomberg had reported that the company was in talks to backstop as much as $250 billion of OpenAI’s compute leases at a single Ohio campus, and in discussions to finance $350 billion of chip purchases for the same project. The cash flows that would collateralize this financing are, in part, rents paid by entities the vendor has funded, and the reach keeps extending: on Friday the company took a minority stake in Cloverleaf Infrastructure to develop the power behind the data centers at the earliest stage. Huang answers the circularity charge by pointing to independent underwriting. Independence of the underwriter does not create independence of the cash flows.
The wrap returns
Read the residual-value support mechanism twice, because we have seen it before under another name. In the last structured credit cycle, of which I was a part, the monoline insurers wrapped the senior tranches, and the wrap worked beautifully for as long as nobody needed it. The guarantees failed the moment they were called, because the events that trigger such guarantees are never idiosyncratic. The guarantee now has a stated size, $125 billion against roughly $158 billion of trailing net income, so the question of capacity is no longer hypothetical. If it is ever drawn in size, it will be drawn on everything at once, since the trigger will be either the company’s own next product generation or an industry-wide demand break. In either scenario the guarantor’s earnings, share price and balance-sheet capacity are impaired when the guarantee must pay. On the desk we called this wrong-way risk, and I have rarely seen a purer expression of it: the seller of the put and the underlying of the put are the same name.
We have run this machine before
The pattern is familiar to anyone who worked through the 2000s: origination gets separated from risk, underwriting gets declared independent, and the exposure is tranched, independently rated and distributed, which spreads the risk around without destroying a dollar of it. The senior paper migrates to the balance sheets of pensions, insurers and money funds, which is to say, to the public, and the machine runs until the collateral assumption at its base gives way. The telecom buildout offers the nearer analogy. Lucent and Nortel extended billions in vendor financing to customers who bought their equipment, booked the sales as revenue, and discovered in 2001 that the receivables and the customers were the same trade. NVIDIA’s structure is more sophisticated, the capital is third-party, and the underwriting names are the best in the world. The 2006 vintage of CDOs also had the best names in the world on every page of the offering circular.
Securitization is the reported design here, not a projection of mine. Bloomberg reports that the debt will be collateralized by the compute itself and issued through private placements and bonds from special-purpose entities, tens of billions of dollars at a time, with the vehicles leasing the compute to NVIDIA’s clients and the first deals expected within months. Goldman Sachs, the one bank in the consortium, is positioning to run the books on the public deals and to distribute the paper through an asset-management arm overseeing nearly four trillion dollars.
BlackRock’s chief executive has already made the pitch: high credit quality, and attractive yields for investors “overinvested in equities.” So the offer amounts to paper whose residual-value assumption rests on one equity’s product cycle, marketed as relief from equity concentration. Once that machinery starts, the exposure stops being a private-markets curiosity and becomes a systemic fact, held in size by institutions the state cannot allow to fail. The sell side has already begun to capitalize the machine as vendor revenue: Morgan Stanley estimates the usage-linked revenue share embedded in the platforms could bring NVIDIA as much as $51 billion a year at a gross margin near 100 percent. A toll on the financing of one’s own product is a remarkable business, and it deepens the circularity rather than resolving it.
Uncle Sam holds the last tranche
Which brings us to the part of the movie we have also seen. Concentrate a systemic quantum of risk on a single name, distribute it through the institutions that hold the nation’s retirement savings, declare the underlying technology essential to national security, and the ending writes itself. If the structure breaks, it will be rescued in the interest of saving the system, and the public sector balance sheet will absorb what the private structures cannot. AI has already been elevated to strategic infrastructure in Washington’s vocabulary, which strengthens the bailout case in advance and lowers the cost of capital today. That is the quiet subsidy inside the $500 billion. The carry is private and the tail is pre-socialized, and everyone at the table is too sophisticated not to know it.
What it means for the book
None of this is a short signal on the tape today, and the distinction matters. To be clear, I have long exposure to NVIDIA stock. Financing innovations extend cycles before they end them. The structured credit machine ran for two full years after the first cracks appeared in 2005, and the money lost by impatient shorts was donated to patient ones. Near term, the announcement is bullish for NVIDIA’s demand visibility, since it unlocks buyers who could not previously fund their appetite. The market’s reading has been cooler, and it now has two weeks of tape behind it. The stock fell close to three percent when the news leaked on August 10, recovered to an intraday high near $228 on the 17th, and then fell for seven straight sessions, its longest losing run since 2022, bottoming near $206 on Monday before Tuesday’s bounce back to $213.
As of Friday’s count it was up 21 percent for the year against 63 for the semiconductor index, so the market pays the vendor’s ecosystem three times what it pays the vendor. Some of that is the rates backdrop, with tens near 4.7 percent leaning on every long-duration growth multiple, and honesty requires the caveat. But half a trillion dollars of announced financing that cannot hold the share price up for a fortnight is a market that has not yet decided what it heard, the demand unlock or the financing tell.
What the announcement changes is the character of the cycle. When a capex boom migrates from cash-funded to balance-sheet-funded to vendor-assisted to structured third-party capital, the marginal buyer is telling you it needs progressively more help. That progression has marked the late innings of every buildout I know of, from the railroads to telecom to housing, and it is no longer a forecast either: last week Broadcom was reported in talks to raise $60 billion of debt, and Alphabet raised roughly $50 billion of equity after printing its first negative free cash flow quarter as capex outran operating cash.
To be fair to the demand side before any watch list: the four largest hyperscalers ended the second quarter with a combined $2.3 trillion of contracted backlog, and the company guides to a $91 billion quarter with China excluded, so nothing in the near-term numbers argues for a break. The first hard data point after this essay arrives tonight after the close, when the July quarter prints, with the options market pricing a move of about five and a half percent on it.
The watch list from here: the terms and spreads on the first platform deals when the memoranda become signed agreements; any disclosure of residual-value support utilization; hyperscaler depreciation schedules, which are the honest accounting of useful life; and the shape of the GPU rental curve, since long-term commitments pricing below spot would be the first sign the shortage premium is rolling over.
Once the paper exists, so will the derivatives on it. The debt, structured or straight, and the CDS written on top will become the market’s running proxy for the health of the AI economy. That is not an unmixed blessing, as those of us who lived through the Lehman negative basis trade can vouch. It is the market’s job to handicap asset prices continuously, and once there is a traded price for AI credit, the outlook for the buildout will gyrate with the spread, and the spread will move for reasons that have nothing to do with the volume of intelligence produced or rented.
Until then, own convexity and prize liquidity, and respect the possibility that this runs further than any of us find reasonable. The definitional point stands regardless of the tape. One stock does not an asset class make, and $500 billion does not change the arithmetic. It only raises the stakes on being right about the difference.
Let’s see what Jensen Huang has to say after market close today, Wednesday, August 26, 2026.
Disclaimer
The Macro Fireside is published for informational and educational purposes only. Nothing in this publication constitutes investment advice, a solicitation, or an offer to buy or sell any security, financial instrument, or investment product of any kind. The views expressed are solely those of the author and do not represent the views of any employer, affiliated entity, or counterparty.
All analysis reflects the author’s independent judgment as of the date of publication. Market conditions, data, and regulatory environments change rapidly; no representation is made that any information herein remains current or accurate after the publication date. Past performance of any asset class, strategy, or instrument referenced herein is not indicative of future results.
Readers should conduct their own independent research and due diligence, and consult a qualified financial adviser, attorney, or tax professional before making any investment decision. The author may hold positions in securities or instruments discussed in this publication. Such holdings are subject to change at any time without notice and without obligation to update this publication.
This publication is not directed at, and should not be relied upon by, any person in any jurisdiction where its distribution or use would be contrary to applicable law or regulation. By reading this publication, you acknowledge and agree that the author and The Macro Fireside bear no liability for any investment decisions made in reliance on its contents.
Sources & References
NVIDIA Corporation, press materials, Jensen Huang, LinkedIn commentary, CNBC, Financial Times, Forbes, Amazon.com, Inc., investor disclosures, Bloomberg and Reuters.