At the start of AI project development, companies generally paid expenses in cash. Microsoft, Google, Amazon, and Meta raised the funds needed to build their data centers from operating profits, and investors were usually satisfied with the results. It now seems that this stage is coming to an end; the news of this week resembles the inbox of a loan officer, with reports of a $40 billion debt agreement for Nvidia chips, the cancellation of a $5 billion IPO, and an estimate that the industry will have to get $1.5 trillion from outside sources by 2028.
Since I use AI tools daily, I think the question of who pays for them deserves more thought than it currently gets. When you run a business that relies on such tools, you should watch your vendors’ balance sheets as closely as you do their product launches.
The SpaceX deal in plain numbers
As investingLive summarizes in a Financial Times report, SpaceX aims to raise about $40 billion to buy Nvidia AI chips. The proposal involves about $30 billion in investment-grade debt and $10 billion in bank loans, with Apollo expected to lead and Pimco among a small group of lenders in discussions. The transaction is expected to close in 2027, and the chips are expected to power AI computing for the Grok model and the Colossus data centers.
There are several reservations to note. Unnamed sources provided the figures; the original article is behind a paywall, and neither SpaceX, Apollo, nor Nvidia has commented. As a result, the exact lender group and pricing remain unconfirmed. That said, the market reaction was telling: SpaceX’s shares are said to have fallen about 1% after hours, while Nvidia’s rose about 0.5%, suggesting investors already anticipate this kind of demand.
Why $1.5 trillion matters
Morgan Stanley estimates that AI infrastructure will need about $1.5 trillion in external financing by 2028. That is money raised from bond buyers, banks, and private credit funds rather than generated by the companies themselves, and it comes on top of what the largest technology firms already spend from their own cash flow. In August, Nvidia reportedly partnered with several major financial firms on platforms aimed at mobilizing more than $500 billion for AI infrastructure, and this week’s Tech Startups roundup also notes large financing deals involving Broadcom and Oracle.
Why debt changes the math
Cash-funded investment is forgiving. If a data center takes longer to fill, the owner absorbs the delay out of profits. Debt-funded investment has a schedule, because interest is due whether the demand shows up or not, and I see three pressure points in that arrangement.
The first is utilization. A chip purchase pays off only if the hardware runs at high utilization, and idle capacity still costs interest. The second is depreciation. AI chips age quickly as newer generations arrive, so a loan that runs longer than a chip’s useful economic life leaves a gap. The third is refinancing. If lenders get nervous, the next round costs more or doesn’t happen at all, and Morgan Stanley reportedly warns that lenders are growing more cautious, making investor appetite for a $30 billion offering an open question.
None of this means a bust is coming. It means the margin for error is thinner than it was when everyone paid in cash.
The warning signs are showing up.
Not every deal this week landed well. Firmus, an Australian company backed by Nvidia and Blackstone, canceled a roughly $5 billion listing that would have valued it above $30 billion. Reporting cited in the Tech Startups roundup says only 42 MW of its planned 1 GW capacity is currently operational, meaning investors were asked to price the full vision while less than 5% of it was built.
OpenAI’s revenue figure was also revised. Axios reported annualized revenue of about $50 billion rather than the $70 billion that had been circulating, and the reported gap stems from how cloud partner sales are counted rather than from a confirmed drop. Even so, it shows how much weight the market places on self-defined revenue figures. Flock Safety, an AI surveillance company last valued around $7.5 billion, reportedly plans to cut about 270 jobs, or 18% of its staff, amid privacy opposition.
What I would watch if you run a business
You don’t need to own a data center to feel this, and my advice is to treat your AI vendors like any supplier you depend on for survival. Start with cloud pricing. Providers with debt need steady margins, so I wouldn’t assume compute gets cheaper on a straight line.
The next consideration is vendor risk. If your AI vendor depends on a financing round to keep expanding capacity, it makes sense to ask what happens if that round closes late, and whether you could move your workflow elsewhere without rebuilding it from scratch. Contract flexibility matters for a similar reason, so I would avoid multi-year commitments to a single model provider unless there is a reasonable way to exit. Finally, when a vendor quotes growth, ask how that figure is defined, since the OpenAI revision is a reminder that definitions matter.
The takeaway
In my opinion, demand isn’t fake, and spending hasn’t stopped; what has changed is the subject of the debate. A year ago, the discussion centered on whose model performed better on benchmarks; now it centers on who can find the money to keep a half-built data center running, and I expect the next twelve months to sort companies by their loan terms almost as much as by their engineering.
Watch the bond market, not just the product launches. The first signs of strain will probably show up in credit spreads, canceled offerings, and revised revenue numbers, well before they show up in an earnings call. If you rely on an AI vendor, this is a good week to read their funding news.
