big tech

Nvidia has recruited Wall Street to bankroll its own customers.

ADVERTISEMENT

ADVERTISEMENT

The US chipmaker said last week it had signed memorandums of understanding with Wall Street’s largest asset managers, including Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs and KKR to raise upwards of half a trillion dollars for AI companies to borrow against, money that will buy its chips and build the servers that run them.

The six firms will set up what Nvidia calls “compute financing platforms,” drawing on institutional money, insurance funds and private credit. Borrowers can use the proceeds for the chips as well as servers, networking equipment, buildings and power supply.

Nvidia has the option to guarantee up to a quarter of any given deal, which lowers the interest rate its customers pay while leaving most of the credit risk with the lenders.

CEO Jensen Huang said he approached only these six companies and none refused.

Keeping that spending off their own books is precisely the point, and the fact that such a structure is needed at all tells investors a great deal about where the constraints in the AI boom now lie.

The financial engineering rests on a single reclassification. Graphics processing units (GPUs) have always been treated as equipment that loses value quickly, superseded whenever a faster generation arrives.

Nvidia is effectively asking lenders to treat them instead as long-lived infrastructure, closer to a toll road or a power plant, that can be borrowed against for years.

“These are revenue-generating assets now,” Huang said, describing them as productive, long-lived and transferable between customers.

Why the money had to come from somewhere else

The timing reflects a squeeze that has been building all year.

Microsoft, Amazon, Alphabet, Meta and other hyperscalers whose cloud platforms host most of the world’s AI workloads have together guided roughly $720 billion (€624bn) to $745 billion (€646bn) of capital spending in 2026, an increase of about 77% on last year.

What analysts expect the hyperscalers to spend in 2027 alone has more than doubled in the space of a year, from a consensus of $480 billion (€416bn) in August 2025 to $1.08 trillion (€943bn) this month, a rise of about 127%, according to Bank of America.

The pattern has repeated at every stage.

Analysts who already considered last year’s investment unsustainable then watched the hyperscalers guide higher at the start of 2026, revise those figures upward again through the year, and pencil in larger sums still for next year and 2028.

Moody’s has warned that spending on this scale is eating into free cash flow and pushing tech groups into heavier borrowing. Alphabet recorded negative free cash flow of $5.9 billion (€5.1bn) in a quarter when it spent $44.9 billion (€38.9bn) on projects.

That is the pressure the structure of Nvidia’s Wall Street deal relieves.

Debt raised through these “compute financing platforms” sits with the financing vehicles rather than on a hyperscaler’s own accounts and also has Nvidia’s backing, which protects credit ratings and leaves room for conventional borrowing elsewhere.

For smaller operators the effect is larger still as companies such as CoreWeave and Nebius, which lack investment-grade ratings and pay dearly for credit, gain access to capital on terms previously reserved for the giants.

What the market actually read into it

The reaction was more ambivalent than the headline number suggests, and came weeks after a July selloff driven by doubts over whether AI spending will pay for itself.

Essentially, equity investors saw a bottleneck being cleared while credit investors saw something else: the cost of insuring Nvidia’s own debt against default rose after the news and has roughly doubled since late May.

Their doubt concentrates on the reclassification previously mentioned.

“Chips depreciate fast and lose value the moment a newer generation arrives,” warned Nigel Green of financial advisory firm deVere Group, noting that lending against them only works if the collateral holds its value.

Critics also point out that Nvidia is helping finance purchases of its own products, deepening the circularity that already worries the sector.

Goldman Sachs CEO David Solomon called it “a pivotal moment of a historic AI investment cycle.”

Whether it proves pivotal in the direction Solomon means depends on a question nobody can yet answer: what will the value of a current GPU be in five years?

In its Annual Economic Report, published on Sunday, the Bank for International Settlements (BIS), known as the central bank for central banks, warned that the enormous spending on AI is accumulating financial vulnerabilities that could amplify any future shock and spread from markets into the wider economy.

ADVERTISEMENT

ADVERTISEMENT

Presenting the findings, BIS general manager Pablo Hernández de Cos said the message was one of “urgency”, with policymakers urged to act before any reversal makes the eventual adjustment more painful.

At the core of the warning is the scale of the spending, despite massive investment having supported global growth over the past year.

The five largest “hyperscalers”, the technology giants racing to build AI infrastructure, are on track to commit more than $1 trillion (€878bn) to AI-related investment across 2025 and 2026, a pace that is outstripping their earnings and free cash flow and pushing some to borrow heavily to keep up.

The BIS suggests this race is fuelled by a belief that only a handful of dominant players will ultimately prevail, encouraging firms to pour money into projects whose returns remain deeply uncertain.

Echoes of past manias

The report sets today’s AI boom against a long historical lineage, from the canal mania of the 1830s and Britain’s railway mania of the 1840s to the electrification of the 1920s and the dotcom bubble.

Each began with a genuine technological breakthrough that attracted more capital than commercial returns could justify, the BIS notes, with each episode ending “with an eventual reversal in investment, inducing economy-wide recessions”.

Compounding the danger are stretched share prices and opaque financing.

The BIS highlights the spread of “circular financing”, in which chipmakers and cloud giants take equity stakes in AI labs that then commit to buying their chips and computing power, effectively recycling money back to the original investors as revenue.

Much of the funding now flows through hedge funds and private credit vehicles that face lighter scrutiny than banks.

According to Zhang Tao, the BIS chief representative for Asia and the Pacific, that reliance on non-bank channels means an AI downturn could unwind into a sharper, faster crash than a traditional banking crisis.

The hidden costs of data centres

Beyond financial markets, critics argue the true cost of the AI build-out is being obscured in plain sight.

A central concern, examined by the Wall Street Journal, is how the technology giants account for their data centres.

By assuming the expensive equipment inside them will stay useful for longer, firms can spread its cost over more years, lowering the depreciation charged against profits in any given period and making earnings look healthier than the underlying cash burn implies.

However, the specialist chips at the heart of these facilities may become obsolete far faster than those extended schedules assume, leaving a gap between reported profits and economic reality, as well as a balance sheet more exposed than it appears should demand disappoint or a sizable need to replace hardware arise.

The physical scale is staggering.

Columbia University economist Stijn Van Nieuwerburgh estimates the build-out could cost in the region of $8 trillion (€7tn) over the next six years, financed in part through the kind of off-balance-sheet arrangements the BIS flagged.

The costs are also no longer confined to corporate accounts.

Some economists now warn of a so-called “third wave” of inflation, after the pandemic and tariffs, driven this time by the AI build-out. As chip manufacturers prioritise high-margin parts for AI servers, the resulting squeeze on memory and storage has rippled out to consumer electronics.

For example, Apple raised prices on its MacBooks, iPads and other devices last week, citing an “extraordinary surge in demand for memory and storage” and saying it had “never seen a component price increase this much, this quickly”.

The company’s shares fell around 6%, their worst day in over a year, as Microsoft, Nintendo and Sony have also made similar moves.

Beyond hidden costs and inflationary pressures, where the strain may spread furthest is raw power.

Goldman Sachs expects data centres to account for nearly half of the growth in US electricity demand by 2030, with consumer power prices forecast to rise around 6% a year through 2026 and 2027.

The BIS itself notes that the build-out’s hunger for electricity is already pressuring prices and input costs, with potential spillovers to inflation, though it stresses, as do many economists, that AI could yet prove disinflationary if its promised productivity gains eventually arrive.