Mark-to-Model

In November 2025 the world's most valuable company did something unusual. Nvidia sent analysts a memo pushing back on comparisons with Enron. It was not hiding debt in special purpose vehicles, the memo said, according to TechSpot; Barron's and The Verge had authenticated it.

The trigger was a viral blog post and the investor Michael Burry, who accuses the big cloud companies of flattering their profits through depreciation. Burry later made clear that he was not comparing Nvidia to Enron at all, but to Cisco in the late 1990s (TechCrunch). A supplier, in other words, that sold real products to real customers during a build-out that ran ahead of demand.

The distinction matters. Enron faked profits and hid losses. None of that has been shown for the AI industry. The chips exist, they run, and they get paid for.

But whether someone is cheating is not the only question Enron raises. The better one is this: which mechanisms carried the company so far that fraud eventually became necessary? And are they back?

To find out, I read the quarterly filings available as of late October 2026 from Nvidia, Oracle, Microsoft, Alphabet, Amazon, Meta and AMD, along with the 2002 investigation into Enron. Four mechanisms come back. None of them is fraud.

What really killed Enron

In the early 1990s Enron ran pipelines out of Houston. Then it turned natural gas into a financial product: long-term supply contracts, traded like securities. The idea was good, and it worked. In 2000 Enron reported about $100 billion in revenue and employed some 20,600 people (ABA Journal).

The lever was an accounting rule. In January 1992 the Securities and Exchange Commission allowed Enron's gas trading business to carry contracts at market value, mark-to-market (Cato Institute). Sign a ten-year contract and you book today's value of all the profit you expect from it, straight away. For a commodity with a daily exchange price, that is sensible. Enron later applied the method to almost everything it did, including deals for which no market existed. There the value came not from a market but from the company's own model. In finance, that is called mark-to-model.

The best-known example is a video-on-demand deal with Blockbuster, signed in the summer of 2000. The service never got beyond trials in a handful of cities, and Blockbuster walked away in March 2001. By then Enron had booked $111 million in profit from the venture, $53 million in 2000 and $58 million in the first quarter of 2001. According to the SEC's complaint, this ran through a joint venture called Braveheart that lacked economic substance (SEC).

This is the mechanism the whole piece is about. Book the profit of a long contract in its first quarter and that contract gives you nothing the next quarter. To keep growing you need a new deal, and a bigger one. Every quarter Enron had to find a larger deal than the quarter before. Not out of malice. Out of accounting.

When the deals stopped being enough, the special purpose entities arrived. The most important were called the Raptors. Enron used them to hedge investments whose value swung around. But the Raptors were funded mainly with Enron's own stock, as the board's investigation, the Powers Report, sets out. As long as the share price rose, the hedge held. When it fell, the hedge fell with it, at exactly the moment it was meant for. Enron never escaped the risk of loss, the report says. The deals made earnings from the third quarter of 2000 to the third quarter of 2001 look almost $1 billion higher than they should have been.

On 16 October 2001 Enron reported a third-quarter loss of $618 million (Enron). Shareholders' equity had to be cut by $1.2 billion, which is also in the Powers Report. Seven weeks later, on 2 December, it filed for bankruptcy. The stock, once worth about $90, stood at 26 cents (ABA Journal). Thousands of employees lost their savings along with their jobs, because those savings were in Enron shares.

The reckoning took years. Arthur Andersen, Enron's auditor, was convicted of obstruction of justice in 2002 and had to close. The Supreme Court overturned the verdict in 2005 because the jury had been wrongly instructed (Supreme Court). By then the firm barely existed. Jeff Skilling, chief executive until August 2001, was sentenced in 2006 to more than 24 years, and resentenced in 2013 to 14 (US Department of Justice). Ken Lay died before he could be sentenced.

The fraud was the consequence, not the cause. The cause was a business that used up more future every quarter than it created.

Training out the no

On top of that came a culture that made dissent expensive. Twice a year Enron ranked its people against one another, and those at the bottom had to go. Inside the company it was called “rank and yank”, as Bethany McLean and Peter Elkind describe in their book “The Smartest Guys in the Room”. In a system like that, your own risk rises with your colleagues' success. What got rewarded was booked profit. And booked profit was whatever the model said.

The warnings came anyway, only from outside or from below. In March 2001 McLean asked in Fortune how exactly Enron made its money. In April a fund manager on the earnings call asked for a balance sheet. Skilling insulted him in front of everyone listening (Fox News). In August the executive Sherron Watkins wrote to Ken Lay that she was “incredibly nervous” Enron would implode “in a wave of accounting scandals” (US House of Representatives).

None of these voices got through in time. Not because they argued badly. Because the company had trained itself out of the places where a no could land.

Money that goes round in circles

Now to AI. The first mechanism is the hedge whose value depends on the thing it is supposed to protect. In the AI industry the equivalent is called circular financing. The Bank for International Settlements described it in its annual report in June 2026: chipmakers and cloud companies take stakes in AI labs or new cloud providers, which in turn commit to buying chips or computing power for years (BIS). The terms are “poorly disclosed”, it warns, with a risk of “the same asset being pledged multiple times”.

Nvidia shows it most clearly. Early in 2026 it put $30 billion into OpenAI, one of its biggest customers. The $100 billion from a letter of intent in September 2025 was “probably not in the cards”, chief executive Jensen Huang said in March (Bloomberg). Nvidia is investing up to $10 billion in Anthropic, Microsoft up to $5 billion. In the same deal Anthropic committed to buying $30 billion of computing power on Microsoft's Azure cloud, running on Nvidia systems (Microsoft).

How fast this is growing shows up in Nvidia's quarterly filings. The carrying value of its stakes in private companies was $3.8 billion at the end of July 2025 and $8.2 billion at the end of October 2025 (Nvidia). By the end of July 2026 it was $47.9 billion (Nvidia). That is 12.6 times as much in twelve months. On top of that come $42.8 billion of shares in listed companies and commitments for another $25 billion, to AI model makers, infrastructure financiers and other private companies, the filing says. Nvidia does not say who exactly they are.

The money also flows back by other routes. Nvidia has committed to buying $36 billion of computing time from cloud providers that also buy its chips. They may resell that capacity, and Nvidia's obligation then shrinks. In August 2026 Nvidia also took on guarantees of up to $105 billion for lease and power payments at a data centre in Ohio to be used by an affiliate of OpenAI. They only fall due if the tenant defaults. Both are in the same quarterly filing.

AMD's arrangement is the plainest. In October 2025 AMD and OpenAI agreed to build 6 gigawatts of computing power on AMD chips. In return OpenAI received a warrant for up to 160 million AMD shares (AMD) at one cent each. Tranches vest as OpenAI takes delivery of chips and as AMD's share price reaches set targets. The first stage is due to start in the second half of 2026; the filings so far do not say whether the first tranche has vested. In February 2026 Meta got a warrant built the same way, also for up to 160 million shares at one cent. The final tranche requires an AMD share price of $600 (AMD).

It is worth reading slowly. AMD pays customers in its own shares for buying AMD chips. What that payment is worth depends on AMD's share price rising. And the price rises because customers buy AMD chips. This is not fraud; it is in a mandatory filing with the regulator. It is simply reflexive: the collateral is made of the thing it is meant to secure. With Anthropic, in July 2026, AMD went the other way round. It committed to an investment of up to $5 billion, and Anthropic plans to deploy up to 2 gigawatts of AMD chips (AMD).

Who is paying whom? The honest answer: everyone is paying everyone. While end-customer demand grows, a loop like this lets investment happen faster. If it stalls, seller, investor and customer come under pressure at the same time.

And off the balance sheet?

Enron's special purpose entities kept risks off the group's balance sheet. That too has a legal, disclosed equivalent today. Meta is building its largest data centre, Hyperion in Louisiana, inside a joint venture 80 per cent owned by funds managed by Blue Owl, at a development cost of around $27 billion (Data Center Frontier). Meta leases the site and guarantees a residual value. At the end of September 2026 Meta put its residual value guarantees for its data-centre ventures at up to $41.5 billion, and its total maximum exposure at $68.4 billion (Meta). No liability is recorded, because payments are “not probable”.

This is not Enron. Meta discloses all of it, and the auditors have seen it. But it is the question that was asked too late in 2001: which obligation sits not on the balance sheet but in the notes?

The future, counted today

The second mechanism is the booked future. Today there is a cleanly regulated tool for it with a sober name: remaining performance obligations, or RPO. It means contracted revenue for work not yet delivered.

The difference from Enron matters. Enron booked expected profits as profit. RPO is neither profit nor revenue but a figure in the notes. The parallel lies elsewhere: the market treats orders as if they were already money.

At the end of August 2026 Oracle reported a backlog of $664 billion (Oracle). In the fiscal year to May 2026 it had taken in $67.4 billion in revenue (Verdict). The backlog is almost ten times a year's revenue. According to the quarterly filing, about 13 per cent of it turns into revenue in the next twelve months, 37 per cent in the two years after that, 34 per cent in years three to five, and the rest later (Oracle).

A large share depends on one customer. According to the Wall Street Journal, OpenAI has ordered around $300 billion of computing power from Oracle over roughly five years (heise). At Microsoft, OpenAI committed in October 2025 to buying an additional $250 billion of Azure services (Microsoft). At the end of 2025, Microsoft said, about 45 per cent of its $625 billion commercial backlog came from OpenAI (The Register). By the end of June 2026 that backlog had grown to $678 billion (Microsoft).

Set against that are OpenAI's own numbers. In March 2026 it raised $122 billion and reported revenue of about $2 billion a month (Yahoo Finance). For 2026 to 2030 it plans to spend around $856 billion on computing and infrastructure, according to the Financial Times, and expects to spend $278 billion more than it brings in over that period (Bloomberg). The computing figure has moved several times in a year. Sam Altman cited commitments of $1.4 trillion in 2025, over a longer period. In February 2026 OpenAI told investors around $600 billion through 2030 (Reuters); in July the Wall Street Journal reported around $750 billion (Quartz).

In April 2026 the Wall Street Journal reported that OpenAI's chief financial officer, Sarah Friar, had privately doubted whether the company could pay for future computing contracts if revenue did not grow fast enough. Oracle fell 3.4 per cent the next morning, SoftBank almost 10 per cent in Tokyo (Reuters). The Reuters report carried no comment from OpenAI.

Oracle itself is funding the build-out with debt and new shares. Its free cash flow for fiscal 2026 was minus $23.7 billion. In the first quarter of the new fiscal year it sold $20 billion of stock, and free cash flow was minus $5 billion. Customers now carry part of the risk: on large contracts they have prepaid for GPUs or bought and supplied them themselves, $75 billion in all.

Enron needed a bigger deal every quarter. The AI industry needs a bigger backlog every quarter. The difference is that at Oracle each dollar only becomes revenue once the service is delivered. Whether the customers can pay by then is not in any footnote.

How long does a chip live?

The third mechanism is the least conspicuous. A server is depreciated over its assumed useful life. The longer the life, the smaller the annual charge and the bigger the profit. It is the simplest dial on a balance sheet, and it is legal as long as the assumption is reasonable.

Microsoft extended the useful life of its servers and network equipment from four to six years in July 2022. In the first fiscal year that added about $3.7 billion to operating income (Microsoft). Alphabet also moved its servers from four to six years at the start of 2023, cutting depreciation by $3.9 billion in the first year (Alphabet). Meta stretched the life of most of its servers and network assets to 5.5 years from January 2025, up from four to five (Meta). Oracle uses six years for servers and network equipment.

Amazon went the other way. From January 2024 it had extended its server lives from five to six years. From January 2025 it cut a subset back to five, citing an “increased pace of technology development, particularly in the area of artificial intelligence and machine learning” (Amazon). By its own estimate, that cost about $0.7 billion of operating income in 2025.

Four companies buy similar chips from the same makers for similar purposes. Three say the equipment lasts longer than they thought. One says it lasts less. Nobody knows today who is right, and that is the point. Useful life is a management estimate. What gets audited is whether it is reasonable, not whether it is true.

Michael Burry calculates that the big providers will together under-depreciate by about $176 billion from 2026 to 2028 (CNBC). That is his estimate; CNBC could not confirm it, and it should be read with care. The disagreement between the companies, though, is in their own filings. And the direction holds: from fiscal 2027 Microsoft is also extending the assumed life of its data centres and office buildings, from 15 to 25 years (CFO Dive).

The BIS estimates that the five largest cloud companies will invest more than a trillion dollars in AI assets across 2025 and 2026. At that scale, every year of useful life added or removed shifts profits by billions.

A number that fitted too well

That leaves the fourth mechanism, culture. It is the hardest to measure, and I am not innocent here myself.

In July 2025 the research group METR published a randomised trial with 16 experienced open-source developers working on their own projects. With AI tools, their tasks took 19 per cent longer. Afterwards the same developers believed they had been 20 per cent faster (METR).

The number became the sceptics' favourite. I had it in an early draft of this piece, as a current finding. Then I checked. In February 2026 METR wrote that its follow-up study now gave only an “unreliable signal” (METR). Between 30 and 50 per cent of developers said they were holding back tasks they did not want to do without AI. Pay had dropped from $150 to $50 an hour. The new estimates point towards a speed-up, with wide uncertainty. METR thinks it likely that developers now gain more from AI than in early 2025.

METR has not withdrawn the 2025 study. But it describes a moment that has passed. I would have booked a number that suited my argument without checking whether it still held. In a piece that faults Enron for booking numbers that suited its argument.

What remains of the study is not the percentage. It is the gap between measured and felt. The developers were far off about their own speed, and off in the direction they liked.

Inside companies it looks similar. In March 2026 Goldman Sachs counted: 70 per cent of S&P 500 management teams talked about AI on their earnings calls, 54 per cent framed it as a productivity question. 10 per cent quantified the effect for a specific use case. 1 per cent quantified what AI did for earnings (Fortune). Across the economy as a whole, Goldman economist Ronnie Walker wrote, “we still do not find a meaningful relationship between productivity and AI adoption”.

Seventy per cent talk. One in a hundred counts. The rest is mark-to-model in the head.

What is different

Fairness is needed at this point, or the comparison is worthless.

Enron sold deals that did not exist in the form it claimed. Nvidia sells chips that compute. In the quarter to the end of July 2026 it took in $96.2 billion, at a gross margin of about 75 per cent (Nvidia). Oracle says demand for AI computing is growing faster than supply. Anthropic reached an annualised revenue run rate of $65 billion at the end of July 2026, according to reports (Axios). These are invoices that get paid, not models.

Enron hid its structures. Today's stakes, warrants, guarantees and backlogs sit in mandatory filings and quarterly reports that anyone can read. That is how this piece was put together. Accounting fraud has not been shown at any of the companies named, and no one is accused of it here.

The danger is a different one. Enron was a company rotting from the inside. The AI industry is a web in which many healthy companies hang off one another. A failure would not be an accounting scandal but a chain reaction: weaker demand, falling share prices, warrants and stakes worth less, customers who can no longer pay their contracts. That is what Burry's Cisco comparison means. Cisco had a real product and real customers. Its shares fell anyway, because they had priced in a future that arrived later than expected.

Who can still say no

Enron did not lack controls. Its board approved finance chief Andrew Fastow running, at the same time, the partnerships Enron was doing business with. Fastow made at least $30 million from them, the Powers Report found. The auditor, Andersen, billed $5.7 million for advice on those deals alone, and approved them. The controls were there. They had simply stopped saying no.

So the question for the AI industry is not whether someone is cheating. It is: where does the no sit today? With the auditor who signs off on a useful life because it is reasonable? With the analyst whose bank earns fees on the next listing? With the finance chief whose doubts end up in the newspaper rather than on the balance sheet?

The worth of an organisation does not show in its best idea. It shows in whether someone inside it can still say no without paying for it.

Sources and method

All company figures come from SEC filings (10-K, 10-Q, 8-K) or earnings releases, as of 26 October 2026. Alphabet, Amazon and Microsoft report the quarter to September only at the end of October and are not included. Meta's quarterly filing of 22 October 2026 is included.

Nvidia: stakes in private companies from the roll-forward table in the note “Investments in non-affiliated entities”: $3,799 million on 27 July 2025, $8,187 million on 26 October 2025, $22,251 million on 25 January 2026, $42,336 million on 26 April 2026, $47,898 million on 26 July 2026. The factor of 12.6 is 47,898 divided by 3,799. The balance-sheet line “Non-marketable securities” is higher, at $51,157 million, because it also includes equity-method investments. Listed holdings: $42,783 million. Commitments to further equity investments: $25 billion. Computing commitments to cloud providers (“AI cloud agreements”): $36 billion. Guarantees for the Ohio data centre: up to $105 billion, entered into in August 2026, for an affiliate of OpenAI according to the filing. Gross margin is my own calculation, 72,142 over 96,221 million dollars of revenue.

Oracle: the $664 billion backlog and the shares of 13, 37 and 34 per cent come from the 10-Q for 31 August 2026, where they are given as “approximately”. The dollar amounts of the segments in the chart are my own calculation (share times 664, rounded), as is the ratio to annual revenue (664 over 67.4). Annual revenue, free cash flow and the $75 billion of prepaid or customer-supplied hardware come from the fiscal 2026 results release, as reported by Verdict. The $20 billion share sale, free cash flow of minus $5 billion and the guidance of at least $90 billion revenue for fiscal 2027 come from the release of 10 September 2026. Neither Oracle nor OpenAI has put a figure on their contract; the roughly $300 billion comes from the Wall Street Journal.

OpenAI: the funding round and monthly revenue are company statements as reported in the press. Plans to 2030 come from reports by the Financial Times (via Bloomberg), the Wall Street Journal (via Quartz) and Reuters. They are drawn from internal documents and have not been confirmed by OpenAI. They partly cover different periods and scopes, so they are shown side by side and not combined.

AMD: warrants for OpenAI (8-K of 6 October 2025) and Meta (8-K of 23 February 2026), investment commitment to Anthropic (release of 22 July 2026). Microsoft: OpenAI's $250 billion commitment from the 10-Q for 30 September 2025; OpenAI's share of the backlog from the call on the quarter to December 2025, as reported by The Register; Microsoft has not given a newer share. Meta: residual value guarantees and maximum exposure from the 10-Q for 30 September 2026; ownership split and development cost of Hyperion from the October 2025 announcement, as reported by Data Center Frontier.

Useful lives: Microsoft 10-K for fiscal 2023, Alphabet fourth-quarter 2023 results, Meta 10-K for 2025, Amazon 10-K for 2024, Oracle 10-Q for 31 August 2026. For Meta, the chart uses the midpoint of the earlier range of four to five years. The $176 billion is Michael Burry's estimate, posted on X and reported by CNBC, and has not been verified.

Enron: report of the special investigative committee of the board (Powers Report, 1 February 2002, filed with the SEC on Form 8-K) for the Raptors, the earnings overstated by almost $1 billion, the $1.2 billion of equity, Fastow's at least $30 million and Andersen's $5.7 million in fees. Blockbuster and Braveheart from the SEC's 2003 complaint, which puts the figure at $111 million; the trials in a handful of cities from McLean and Elkind. Approval of mark-to-market from the Cato Institute, which relies on the SEC's letter of 30 January 1992. “Rank and yank” and the earnings call of 17 April 2001 from McLean and Elkind and contemporary reports. Bethany McLean, Peter Elkind: “The Smartest Guys in the Room. The Amazing Rise and Scandalous Fall of Enron”, Portfolio 2003.

Other sources: BIS Annual Economic Report, June 2026, chapter I. METR studies of July 2025 and February 2026. Goldman Sachs Research, March 2026, via Fortune. Anthropic's run rate from Bloomberg and Axios; a run rate scales a short period up to a year and is not revenue actually earned over a year.

Text and research: Jonas Höttler. As of October 2026. The cover image and chapter images are concept images, AI-generated with Bria.