THE WIRE GAME

The Wire Game was a widespread scam in Gilded Age America.  Horse race betting parlors dotted urban centers across the country.  But before a national broadcast ground relay had been installed, race results were telegraphed to the far-flung betting parlors from the tracks.  This delay allowed swindlers to bribe telegraph operators to get a timing edge to place sure bets.  While this was well documented, an even bigger derivative scam eventually dwarfed it.  The con men would convince a mark that they had an inside line with a conspirator at Western Union.  The mark would be convinced to make several small bets to win his confidence.  Of course, the hustle required that they create a completely fake betting parlor in order to own the house.  When the time came for a major swing, the con men would ‘miscommunicate’ the tip.  As soon as the sucker bet was pocketed, the whole ‘betting shop’ would vanish, and the cons would disappear into the night. 

Of course, that’s the plot of the 1973 Best Picture winner The Sting.  The movie takes place in Depression-era Chicago – at least a decade after the real con game had disappeared.  Robert Redford and Paul Newman are the con men attempting to loot a gangster played by a sinister Robert Shaw.  The comedy largely revolves around the cons assembling a complete cast and décor for the counterfeit betting parlor.  Everyone from the bookies and cashiers to the cigarette girl is a shill for the illusion.  But the best deception is the ‘live’ radio broadcast of the races whereby the announcer and sound stage create a verisimilitude of the horse race.  All the details must be perfect to allay the suspicions of the mob boss Shaw.  The stagecraft was so convincing that many in the original audience believed that the FBI ‘bust’ in the finale of the movie was real.

As in Hollywood, the corporate world sometimes resorts to stagecraft.  Corporate executives understand that well-executed delusions can manufacture their own reality.  In 1997 Enron needed to buy out a joint venture partner.  But doing so would have required Enron to consolidate $600 million in debt. (Note: this was a time when $600 million was considered material.)  Enron CFO Andrew Fastow created a special purpose vehicle which would borrow the purchase price on the back of an Enron guarantee.  After the transaction, Enron would be able to recognize 100% of the venture’s earnings and avoid booking the debt as a liability.  That worked until an accountant at Arthur Andersen noticed some irregularities in the structure.  The Andersen employee informed her legal department.  On November 8, 2001, Enron announced that it would fully consolidate the subsidiary.  That adjustment led to an earnings restatement of $591 million and an increase in debt of $658 million.  Less than a month later, Enron was gone.  The troubling lesson is that the financial sleight of hand lasted for four years under a veil.  It was revealed only because the fraudsters hadn’t covered their tracks.

Today, investors must come to terms with veils everywhere they turn.  On September 24, 2026, Bloomberg reported that Oracle Corporation had issued a force majeure notice to the developer of a one-gigawatt data center in southern New Mexico intended for Oracle’s use.  It sounded very odd.  Parties to a contract invoke force majeure when they cannot perform due to circumstances outside their control.  It seemed that the problem went in the other direction – that the developer should be notifying Oracle of its inability to deliver the data center on schedule.  According to media reports, it turns out that Oracle is allegedly obligated to make payments to the developer in lieu of lease payments even if the developer does not deliver the project.  This is where the veils matter.  The developer obviously negotiated that nice term and in turn got more favorable financing.  Not in a million years would I have guessed that a tenant could be forced to pay up for a non-functioning asset.

I go into this onerous detail because it illustrates a salient point.  The veils surrounding the special purpose vehicles are masking details that would be relevant to any financial investor.  Just as Enron’s creditors were oblivious to the machinations underneath its off-balance-sheet entities, Oracle’s lenders are forced to find out the terms of agreements only when disclosed.  Special purpose vehicles effectively parcel out a set of risks (construction, power, regulatory).  But they do not make the risks disappear.  Since the Bloomberg report, Oracle’s credit default swaps (the cost to insure against a default) have escalated from 195 basis points to 263 basis points.

The supporters of the AI capex extravaganza espouse a view I call the ‘real money’ thesis.  Their argument is that the companies leading the AI capex wave are substantial businesses spending their own money to meet measurable demand.  The proliferation of special purpose vehicles suggests the opposite.  The sponsors of AI capex are actually doing everything they can to insulate themselves from the downside.  SPVs have a common epithet – ‘bankruptcy remote entities.’  That means that the creditors to the SPV entities can only rely on the assets and contracts of the SPV for recovery.  Res ipsa loquitur.[1] 

Recently, we have also observed the widespread use of credit enhancement.  On August 10, Nvidia announced a $500 billion compute financing platform with major financial institutions to provide financing for Nvidia chips to hyperscalers, large language model developers and enterprises.  The crux of the offering is Nvidia’s commitment to provide $125 billion of residual value support for lenders.  In plain terms, if the lenders need to foreclose on the assets, Nvidia will make creditors whole in a recovery up to a deficit of 25%.  The press release contains this gem:

“Under these strategic partnerships, NVIDIA will work with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create dedicated pools of capital at significant scale at attractive rates for NVIDIA customers.”

Attractive rates for the borrowers mean unattractive rates for the providers of capital, right?  They said the quiet part out loud.  I am sure the bankers winced a little reading that part.  Probably their limited partners winced a little more.

Not to be outdone, Broadcom (according to Reuters) agreed to lend Anthropic $42 billion in the form of convertible bonds tied to Anthropic’s commitment to lease Broadcom’s chips over five years with an aggregate value of $125 billion.  Broadcom is also ‘partially’ guaranteeing another $60 billion of syndicated loans to Anthropic.

What is taking place is credit transformation on an epic scale.  The people extracting the most value are enhancing the credit of those who otherwise might have trouble obtaining alternate financing.  In the press release Nvidia eulogized: “These financing platforms will help customers access scarce compute at scale and build the AI factories that will power every industry and country in the age of AI.”  Only Jensen Huang can make vendor financing sound epochal.

I would love to believe – as Jensen does – that the ‘age of AI’ will usher in a mother lode of wealth for society.  But I am compelled by a simple calculus.  The amounts of money that are being laid out require ever more heroic assumptions about the potential payoffs.  Goldman Sachs analysts calculated that hyperscaler AI revenues would have to reach $417 billion by 2028 to earn a 10% return on invested capital JUST based on 2026–2027 combined capital expenditures.  But of course, compute is an intermediate good.  Someone must buy it and resell some value-added service to an end user.  Helpfully, Goldman calculated that those resellers would have to achieve sales in 2028 of $1.05 trillion. (That’s 2.8% of U.S. GDP.  I know, it’s a global market, but I do not believe that China is going to exactly love importing our software.) Let’s say the year-end 2026 combined revenues of the application layer are $150 billion.  It’s nowhere near that, but I want to be generous.  That implies a three-year compound annual growth rate of 90%.  To be fair, it’s possible.  Google in its heyday from 2002–2007 compounded its revenues 106% annually.  On the other hand, Google’s 2007 sales represented 1/10 of 1% of U.S. GDP.

There is an additional wrinkle.  The preceding assumes the AI capex stops in 2027.  But the currently announced data center builds stretch out to 2032 and beyond.  So, all those calculations just keep ballooning the longer we build.  Bain Capital recently estimated that by 2031, the global AI-services market would need to reach $6 trillion to cover the planned data center build.

Here is my dilemma.  We are led to believe that the ‘age of AI’ will transform the economy’s productive capacity.  We are led to believe that the technology is revolutionizing work and scientific discovery.  But the people at the core of the revolution are increasingly mitigating their own exposure by transferring the risk and keeping optionality.  I feel like a mark inside a modern wire game.  All around me, everything seems legitimate.  Real people betting real money on real horse races.  But the arithmetic tells me otherwise.

[1] Latin for “the thing speaks for itself.”

—T. Brad Conger, CFA
Chief Investment Officer

---

On a quarterly basis, Hirtle & Co. publishes our perspective on the current market. If you would like to be added to our distribution list and receive the full version of our latest Investment Perspective piece, please contact us.

To download a PDF of the excerpt, click here: Investment Perspective 3Q 2026 Excerpt.