Oracle, Broadcom and SpaceX Seek Blockbuster Debt Deals to Pay for AI Chips Apollo, Blackstone and Goldman Sachs among lenders in talks to finance megadeals worth tens of billions of dollars apiece
Big players in artificial intelligence are lining up a series of blockbuster financing deals to pay for computing hardware, part of a rush for capital as data-center build-outs race forward.
In recent weeks, Broadcom has been working to arrange more than $50 billion in financing for OpenAI’s custom artificial intelligence chip, which the firms are developing together, according to people familiar with the discussions.
Apollo and Blackstone are among the lenders Broadcom has talked to about participating in the deal, people close to the situation said. Talks are early and the size of the deal could change.
Separately, Oracle is in talks with Apollo and Goldman Sachs to arrange money for a big purchase of chips, people familiar with the matter said. And SpaceX has talked to lenders in recent days about a $40 billion chip financing for Nvidia NVDA chips, according to a person familiar with the discussions. The Financial Times earlier reported on the SpaceX talks.
The wave of deals reflects the mounting cost of building AI infrastructure. Cloud providers such as Amazon Web Services and Oracle have traditionally financed computing hardware through their own cash flows. For their AI build-outs, the companies issued hundreds of billions of dollars of bonds, pushing the public debt market to its limits. Now, some buyers are turning to Wall Street investment firms to help fund purchases totaling tens of billions of dollars per deal.
There is also a new group of chip buyers, including OpenAI and Anthropic, who don’t have the financial firepower to purchase their own hardware. Leading AI labs historically rented the bulk of their computing capacity from cloud providers, but they now want to own more of their own infrastructure to help lower costs and reduce their reliance on other firms. (…)
The well-known bottlenecks to AI growth are power, equipment, memory/chip capacity and specialized labor.
Add capital.
I highlighted parts of the WSJ article that characterize the AI infrastructure issues:
- The “race forward”: model companies are in a race for leadership/supremacy. The best models will win big, but this is a never ending race.
- The race necessarily brings “mounting costs”, aggravated by significant geopolitical issues.
- Owning their “own infrastructure to help lower costs” necessitates huge amount of capital …
- … right when governments across the world also need financing, “pushing the public debt market to its limits”.
- Hence the “rush for [unconventional] capital” which, in the current context, should read the “rush for affordable capital”.
So far, AI racers have been willing to pay the rising costs to participate. For some of them (e.g. Anthropic, OpenAI) this is an existential race.
They are also willing to pay the increasing cost of capital but, much like for the physical issues, capital is not infinite, particularly when its cost keeps rising. Lenders will eventually balk when they start questioning the safety of their capital given the ever rising borrowers’ liabilities.
The chart plots various market yields all indexed to 100 on January 2026. Financing costs are up 20-25% even for the better credits (treasuries are up 26%!). They are up 36% for the worst.
If capital gets scarce at the top of the chain, the whole chain slows down.
Demand for capital is booming along with AI-related capital spending. Hyperscaler capex is expected to reach $750-$800 billion this year and $1.2 trillion next. Total US corporate bond issuance over the past 12 months through August was a record $3.0 trillion, including $1.4 trillion and $1.6 trillion issued by nonfinancial and financial corporations, respectively. At the same time, US Treasury borrowing totaled $2.1 trillion over the past 12 months through September, including $1.3 trillion in notes and bonds. National savings faces demographic headwinds as retiring Baby Boomers stop saving and draw down their net worth. (Ed Yardeni)
Maybe we should read something from these facts:
- Blackstone stock is down 40% in the last year, 23% in the last 6 weeks. Its forward PE dropped from 32 to 18.
- Apollo stock is down 24% in the last year, 18% in the last 6 weeks. Its forward PE dropped from 20 to 12.
- KKR stock is down 40% in the last year, 23% in the last 6 weeks. Its forward PE dropped from 20 to 12.
- Goldman Sachs stock is down 16% in the last 6 weeks. Its forward PE dropped from 17 to 14 (and from 27 one year ago).
In effect, everybody’s cost of capital is rising fast.
Well, not everybody’s:
The sudden plunge in demand for an Nvidia-backed data center company’s initial public offering is revealing fresh cracks in the AI funding boom.
The planned $5.5 billion listing by Australia’s Firmus Grid Ltd. has become shrouded in uncertainty after the deal failed to attract adequate support for the A$11 marketed share price, according to people familiar with the matter.
Some investors turned cautious just days after the company said it received indications of interest well above the offer size, putting it on track for a $30 billion valuation, the people said. While Firmus closed order books on Thursday, it has so far given no clear indication of the price or the deal structure, an unusual communication gap that’s fueling speculation the price may be cut or the IPO scrapped altogether.
The deal underscores growing concern over how much capital AI infrastructure companies are demanding from public markets at a time when borrowing costs are rising. Much of Firmus’ valuation is based on the company successfully building a pipeline of data centers across Asia serving customers such as Meta Platforms Inc. and OpenAI. Currently it operates two data centers. The IPO proceeds were needed to help fund construction of the broader network.
“Investors still believe in AI,” said Maxence Visseau, Dubai-based chief investment officer at Arkevium Capital, a multi-strategy investment firm. “What they won’t do is pay any price for companies that spend huge amounts on data centers, depend on a few big customers, and promise profits years from now.” (…)
UniSuper, one of Australia’s biggest pension funds, was among institutional investors not taking part in the IPO.
“We think that Firmus indeed has a compelling story. It just doesn’t have a compelling valuation,” Chief Investment Officer John Pearce said in an investor update published Thursday. “So much has to go right to justify the valuation.” The fund was also concerned that Firmus would have to continue to raise debt and equity to fund its expansion plans, he said.
“Investors are increasingly on edge,” said Phil Wool, head of portfolio management at Rayliant Global Advisors. “Firmus was going to be one of the biggest Australian IPOs ever, so from that perspective, it registers as a historical fail.” (…)
Firmus was valued at $10.5 billion in early August after a fundraising round which included Jane Street and Blackstone Inc., meaning it was looking to nearly triple its valuation in two months. The Australian company, which had revenue of $51 million in the 2026 financial year, plans to build data centers it calls AI factories using hardware from backer Nvidia. It has a pipeline of 912 megawatts, of which only 46MW has been built, according to investor documents seen by Bloomberg. (…)
Some AI cloud companies are turning to risky debt to raise capital. At the same time as JPMorgan Chase & Co. was joint lead manager on the Firmus listing — along with Bank of America Corp., Morgan Stanley and Morgans Financial Ltd. — it was also pitching a yield of about 11% on a $5 billion leveraged-loan sale on behalf of Volta Infrastructure Holdings Ltd. to finance a data center complex in Norway. (…)
But with KKR & Co. estimating $8 trillion is needed to complete the global AI buildout, pressure will only intensify for companies to raise capital. (…)
BTW, Bloomberg’s Chris Bryant’s column today: The AI Giants Are Facing a Severe Case of Financial Indigestion
Today’s FT:
China races to build data centres in bid for AI supremacy Beijing is rolling out computing infrastructure at breakneck speed in Inner Mongolia
(…) While China is struggling to secure enough advanced chips to satisfy soaring AI demand, it has been able to mobilise the land, electricity and construction capacity needed to build the data centres that house them.
China already has 24 gigawatts of operational data centre computing capacity, more than the rest of Asia combined but less than half the 56GW in the US, according to SemiAnalysis, the chip-focused research firm. A further 50GW is under construction or has been announced. (…)
Inner Mongolia had 117GW of installed wind capacity by June this year, the largest fleet in China and nearly four times the entire capacity of the UK. It also has roughly 130GW of fossil-fuel power capacity, mostly made up of coal power.
David Fishman, energy analyst at consultancy The Lantau Group, said Inner Mongolia likely had the “largest electricity local oversupply” of any administrative region in the world. “There is an immense amount of electricity that operates at very low capacity,” he said. “Data centres help soak up this excess power.” (…)
But China’s advantage extends beyond access to power. Developers in Ulanqab are also building data centres more cheaply and quickly than is typical in the US.
Contractors from across the country have won orders for the projects, with construction costs in the region about 20 per cent lower than in larger cities, according to Goldman Sachs.
Thousands of workers have been brought in and housed in temporary accommodation beside the sites, many of which are scheduled to be completed within 12 to 18 months. US counterparts typically take between 18 and 24 months.
China has pioneered the use of prefabricated modules — uniform, shipping-container-like units housing computing racks that can be assembled rapidly. SemiAnalysis calls the approach “Lego data centres” with the design also adopted by US hyperscalers. (…)
Beijing is also using incentives to steer the enormous build-out towards its goal of reducing China’s reliance on foreign technology. Data centres that use domestic processors rather than Nvidia’s AI processors receive better tax benefits and discounts on electricity and water bills, according to people familiar with the policies. (…)
Lee said demand for computing power was accelerating as Chinese technology companies expanded their use of AI, particularly AI agents, contributing to rising prices for rented chip capacity.
“There is no risk of overbuilding,” he said. “Token consumption is going through the roof.”
Instead, the extraordinary speed with which China can build and power data centres is exposing the problem at the heart of its AI infrastructure push: securing enough advanced processors. (…)