LONDON, June 19 (Reuters Breakingviews) – In his book “The Great Crash 1929”, John Kenneth Galbraith coined the term “bezzle”. The economist defined the term as “an inventory of undiscovered embezzlement in – or more precisely not in – the country’s businesses and banks.” According to Galbraith, the bezzle grows “in good times [when] people are relaxed, trusting, and money is plentiful”. We are living through such times, once again. Enthusiasm for anything related to AI has made investors too relaxed and trusting about the financial activities of the leading companies involved in the technology revolution.
The bezzle comes in various forms. It is found when companies overstate their earnings and understate their expenses. Accounting shenanigans are a common feature. During the British railway mania of the 1840s, for instance, the “Railway King” George Hudson paid dividends out of capital and arranged dubious transactions between companies under his control. The bezzle often involves manufactured profits, such as the gains on stock trades by Japanese corporations in the late 1980s. During the tech boom at the turn of the century, telecom equipment providers, including Nortel and Lucent, boosted their earnings by providing loans to less creditworthy customers. Some telecom companies at the time engaged in sham swaps of network capacity to boost reported revenue. Liabilities are concealed. Sweden’s “Match King” Ivar Kreuger in the early 1930s and Enron CEO Jeff Skilling in the late 1990s both created complex financial structures that kept debts off their audited accounts.
The vendor financing trend of the dot-com bubble has been superseded by vendor-investing. Microsoft committed to large investments in OpenAI, which in turn agreed to use the software giant’s cloud services. Nvidia owns more than 10% of the neocloud firm, CoreWeave which is a large customer for the $5 trillion chip giant’s graphics processing units (GPUs). When OpenAI secured commitments for of new funding in February, $50 billion came from Amazon.com and $30 billion from Nvidia. Naturally, the ChatGPT-maker committed to using Amazon Web Services and Nvidia’s GPUs.
That’s not all. Unrealised gains on AI investments are already inflating the earnings of Big Tech companies. More than half the profits (reported as “other income” in the earnings statement) of Alphabet and Amazon in the first quarter came from the increased valuation of their stakes in the unlisted companies, principally OpenAI and Anthropic.
Related-party transactions in the AI ecosystem are the norm. Prior to its successful May IPO, chipmaker Cerebras Systems earned most of its revenues in Abu Dhabi, whose G42 AI venture was previously a shareholder. In February, Musk’s SpaceX acquired his xAI business at a valuation of $250 billion. Prior to SpaceX’s IPO, Google committed to pay the company nearly $1 billion a month to rent computing capacity. Google’s parent Alphabet was an early investor in Musk’s space-to-AI venture.
Financial reporting of AI costs is another area deserving scrutiny. In recent years, big tech cloud computing firms – known as hyperscalers — have extended the depreciation schedules for their expensive AI chips from around 3 years to between 5 and 6 years. The investor Michael Burry claims that this has served to inflate profits. Others, however, argue that the change is justified because, while GPUs have a short useful life training AI models, they can be profitably deployed elsewhere for longer periods.
Data centres take up to three years to build. Hyperscalers that stockpile GPUs are not required to depreciate the chips until they are used. In the meantime, they are lodged on the balance sheet as “construction in progress”. At the end of last year, Alphabet, Amazon, Meta Platforms and Oracle reported nearly $220 billion worth of such assets.
Balance sheets in the AI world don’t include many of the liabilities incurred by the leading players. Broadcom a $35 billion special purpose vehicle, supported by buyout giants Apollo Global Management and Blackstone, with the semiconductor giant providing computing capacity for Anthropic and OpenAI as well as a financial backstop. Prior to the financial crisis, another artefact of financial engineering, the Variable Interest Entity (VIE), was used to conceal banks’ subprime exposures. Meta’s planned $27 billion is funded by a VIE operated by private credit operator, Blue Owl Since the Facebook owner only holds a minority stake in the platform, Hyperion’s liabilities will not be consolidated in its accounts.
Analysts at Morgan Stanley estimate that a total of around $1.8 trillion worth of liabilities in the AI domain have been kept off hyperscalers’ balance sheets. This sum includes around $1 trillion worth of purchase commitments that don’t have to be reported unless the company expects a loss. Lease commitments, which amount to around $800 billion, also don’t appear on the balance sheet until the lease commences. These off-balance sheet risks have grown so large that the International Monetary Fund notes them as a concern.
What’s clear is that businesses at the core of the AI system are bleeding cash. In the first quarter, xAI’s costs were around three times sales. Earlier this year, OpenAI projected that it would burn through $25 billion this year, Though Anthropic expects to report its first quarter of positive adjusted operating profit, it is also likely to burn cash in the foreseeable future. By the end of this year, the hyperscalers’ once mighty free cash flow will have evaporated thanks to their continued capital spending, according to Gerard Minack of Minack Advisors. Alphabet has been raising vast amounts from creditors and equity investors. Nvidia this week is launching a $25 billion bond, despite reporting $72 billion of cash and marketable securities at the end of the last quarter.
During the boom period, the bezzle increases rapidly. Few care to pay attention. “In depression all this is reversed,” writes Galbraith. “Money is watched with a narrow, suspicious eye… Audits are penetrating and meticulous.” The AI revolution has yet to meet its moment of reckoning. One day it will.
Editing by Peter Thal Larsen; Production by Streisand Neto.
