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:

• Nvidia-Backed IPO’s Cratering Demand Sends Warning on AI Funding

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. (…)

EARNINGS WATCH

The Q3 2026 earnings season will kick off next week. Most of the hyperscalers are scheduled to report during the last week of October.

Goldman Sachs:

• Consensus expects S&P 500 year/year EPS growth of 27% in Q3. This compares with realized growth of 33% in Q2, excluding some accounting distortions. Analyst estimates show the beneficiaries of AI infrastructure spending accounting for over 50% of S&P 500 EPS growth this quarter alongside hyperscaler capex growth of 116%.

• We expect most companies will once again surpass consensus earnings estimates this quarter. For the median S&P 500 stock, consensus estimates show year/year EPS growth slowing from 14% in Q2 to 9% in Q3. Input cost pressures will likely limit the magnitude of sequential margin expansion in Q3, and the dollar should provide a modestly smaller tailwind to S&P 500 revenue growth this quarter than last quarter. However, recent economic growth data have been strong and S&P 500 earnings revision breadth has remained positive.

• Last quarter, hyperscaler cloud revenue growth accelerated to 48% alongside surging revenue backlogs, boosting investor confidence in the likely return on AI investments. Our equity analysts anticipate a further acceleration to 55% in Q3. Productivity gains from AI adoption remained in early stages based on company commentary last quarter, but we expect both commentary and income statements will increasingly reflect AI productivity gains in coming quarters.

• Analyst estimates show wide dispersion at the sector level, with Info Tech and Energy powering aggregate S&P 500 earnings growth. The two sectors combined are expected to contribute nearly 80% to this quarter’s earnings growth, and the median stock in each of those sectors is expected to grow EPS by more than 30% year/year. In contrast, bottom-up estimates are particularly pessimistic in the consumer sectors, where estimates show no EPS growth at the aggregate sector level.

• AI infrastructure stocks are expected to drive more than half of S&P 500 EPS growth in Q3. The top 10 contributing stocks are expected to account for over two-thirds of aggregate S&P 500 earnings growth this quarter, with Micron (MU) and Nvidia (NVDA) together accounting for more than 1/3 of index growth.

 

 

Factset:

Overall, 116 S&P 500 companies have issued quarterly EPS guidance for the third quarter. Of these companies, 44 have issued negative EPS guidance and 72 have issued positive EPS guidance.

The number of companies issuing negative EPS guidance is below the 5-year average of 62 and below the 10-year average of 57. This quarter marks the lowest number of S&P 500 companies issuing negative EPS guidance for a quarter since Q3 2021 (39).

On the other hand, the number of companies issuing positive EPS guidance for the third quarter is above the 5-year average of 42 and above the 10-year average of 40.

In fact, this quarter marks the highest number of S&P 500 companies issuing positive EPS guidance for a quarter since FactSet began tracking this metric in 2006 The previous record was 65, which occurred in Q2 2021.

As a result, the percentage of companies issuing positive EPS guidance is 62% (72 out of 116), which is above the 5-year average of 40% and above the 10-year average of 41%. This quarter marks the highest percentage of S&P 500 companies issuing positive EPS guidance for a quarter since Q2 2021 (71%).

At the sector level, the Information Technology sector has the highest number of companies issuing positive EPS guidance of all 11 sectors at 44. This number is well above the 5-year average of 23.6 and well above the 10-year average of 20.4 for the sector.

Note that 44 of the 116 S&P 500 companies that have issued a positive guidance for Q3 are IT companies. Note also that only 6 sectors are beating their 5-Y average. Five are not.

John Authers:

Smaller companies have now lost almost all their outperformance versus mega-caps since the start of the year (as measured by the Russell Top 50 and 2000 indexes), while value stocks (bought because they were cheap) have reached a fresh nadir for the decade compared to growth, according to S&P 500 indexes:

Bear in mind that orthodox market theory recognizes both a “size” and a “value” anomaly – over the long term, small companies and cheap stocks are expected to outperform. So this is very unusual. Putting these trends together, the gap between large growth stocks and small-cap value that has opened since the launch of ChatGPT is now yawning:

Earnings, of course, justify this. Growth companies should be expected to grow their profits by definition, but the way their earnings outstripped those of value companies in the second quarter was unprecedented in recent history.

This is not a world where small is beautiful. In fact, small is tough nowadays:

• You have to be big to navigate ever changing tariffs and various supply chain bottlenecks.
• You have to be big to quickly adjust your supply chains around the world.
• You have to be big to negotiate volumes and prices with stranded suppliers.
• You have to be big to absorb all the costs of the above.
• You have to be big to spend on AI and build software to remain competitive.

As a case in point, it’s no coincidence that sales of “nonstore retailers” (largely AMZN) have increasingly outpaced total sales since ChatGPT and, even more so, since tariffs and the US/Israel war with Iran.

Did you miss The AI Boost to S&P 500 Profitability?

AI CORNER

CoreWeave, a major data center developer, held its first Fully Connected Conference September 29-October 1. Frpm Goldman Sachs’ account (my emphasis):

• Industry conversations continue to point to a healthy AI demand backdrop, with both training and inference workloads expanding. Post training is one of the faster growing areas of spend, while robotics remains an emerging driver given the significant compute and data preparation requirements needed for simulation and model development.

• Participants also noted that cloud selection is often driven by access to available compute rather than meaningful platform differentiation, with customers sourcing incremental capacity wherever it can be secured.

• (…) demand for AI capacity continues to exceed available supply. Industry participants pointed to roughly 14-18 months from permitting to data center delivery, limiting how quickly new capacity can come online. (…) power availability, permitting, and site development remain key bottlenecks.

• CoreWeave specifically continues to see demand materially above available capacity, with qualified pipeline demand multiples larger than available infrastructure, growing backlog, and continued geographic expansion driven by customer demand. It reiterated 4.2GW of contracted power today and progress toward its 8GW 2030 target. [+17.5% CAGR]

• Commentary suggested pricing conditions remain favorable. Management reiterated ~5-10pts of contribution margin improvement in 2Q, a ~25% SKU price increase in July disclosed an additional ~10% increase since July; noting signed 3Q short duration contracts (3-6 months) are pricing around ~$40mn/MW. Roughly 70% of 2Q deals included customer prepayments.

• On unit economics, our industry discussions support the view that older generation GPUs may have longer useful lives than initially expected. While some of this demand may reflect ongoing industry capacity constraints, it is also tied to workload continuity, software optimization, and the benefits of staying on proven architectures that are already embedded within research and development workflows. The key debate is how much of the observed demand reflects genuine longevity of older architectures versus customers taking whatever capacity is available. In our view, the answer is likely somewhere in between: capacity scarcity may be extending the economic life of older GPUs, but these renewals also suggest that for certain workloads, the performance benefits of the latest generation may not outweigh the costs and disruption associated with migrating established infrastructure and workflows. CoreWeave customer examples include:

  • A100 (launched in 2020): an AI-native scientific discovery company signed a renewal with extending into 2029. The 3yr contract was signed at pricing in line with typical 1yr terms. This customer is leveraging A100s to produce and refine data sets and fine-tune models to accelerate scientific discovery.

  • H200 (launched in 2023): an early stage research lab renewed a 3yr contract at a premium to the original contract. This customer is leveraging H200s to support development and deployment of smaller families of models build for a specific architecture.

• Management highlighted growing traction across financial services, industrials, healthcare, sovereign deployments, and robotics.
• Security, compliance, and data governance remain the largest friction points for enterprise AI deployments.
• Regulated industries are an expanding customer base, with financial services emerging as a meaningful vertical (~10% of backlog came from financial services as of September). (…) quantitative trading firms such as Hudson River Trading and Jane Street cited as examples.
• Physical AI emerging as a meaningful demand vector, with several customers discussing how advances in multimodal AI are changing development economics. Brandon Hootman (lead of construction autonomy at Caterpillar) noted that development cycles historically tied to new machine launches and measured in years are now measured in months through direct field iteration. He highlighted multimodal systems that combine video, LiDAR, physics engines, and real-world operational data, noting that one hour of real machine data can now generate 100-1,000 hours of simulation data, significantly accelerating development and training workflows.

GS also updated its data center outlooK

• (…) we raise our year-end 2026 US data center capacity forecast by 5 GW to 64 GW, while reducing our year-end 2027 forecast by 5 GW to 90 GW.

  • We now expect US data center power demand to grow 38% (12 GW) in 2026 and 38% (17 GW) in 2027 (all assessed on a December vs December basis). This is assuming a utilization rate consistent with industry reports, as well as bottom-up regional models – where each market carries its own utilization assumption – that converge to a weighted average utilization rate of 70%.

• Effective peak spare capacity tightened across every key region this summer, in part due to record-breaking peak-summer demand.

Also from GS:

Consensus estimates show hyperscaler capex growing by 116% year/year in Q3 2026. This compares with 87% growth in Q2. We expect that hyperscaler capex spending will grow by more than 50% in 2027, surpassing current consensus estimates for spending of roughly $1.1 trillion next year, but also expect that the rate of both capex growth and upside surprises will diminish relative to recent quarters.

 

Last quarter, the hyperscalers reported accelerating cloud revenue growth and large revenue backlogs that signaled the monetization of their capex investments. Explicit management commentary discussing those returns complemented the strong results. Our equity analysts expect the trend of accelerating cloud revenue growth will continue this quarter, with year/year growth increasing from 48% in Q2 to 55% in Q3. Continued signs of AI capex monetization will be important both for the performance of the hyperscalers and for the outlook for capex growth.

 

David Rosenberg:

It may pay to note that there is only one sector at a new high, and it is tech. Increasingly, the market is running on one engine: a familiar cast of tech companies that are building out (and benefiting from) the AI revolution. (…)

But just about everything else is going down. Shares of healthcare firms, banks, and consumer staple companies are declining. So are small-cap stocks (down -0.6% yesterday and off -8% in the past two months). And blue chips, like the Dow industrials.

Over the past month, eight of the eleven sectors of the S&P 500 have dropped, weighed down by the rise in interest rates, while just three — the technology sector, energy, and communication services companies like Meta and Alphabet — have risen. (…)

Fewer than half of the stocks in the S&P 500 closed above their 200-day trendline.

Meanwhile:

The Atlanta Fed’s GDPNow real GDP estimate for 2026: Q3: 3.7%, thanks to AI.

• GS: Trade Deficit Widens by More Than Expected in August; Lowering Q3 GDP Tracking to +3.1%

Plague in Russia?

Last week, the Moscow Times (a non-state-controlled news organization) reported that a 28-year-old civilian anti-plague lab worker in Siberia died on October 1 after being exposed to the plague. Nearly 200 of her contacts are quarantined. Beyond that, the story gets very murky. Independent verification is not readily available in Russia, and there are all sorts of conflicts of interest here.

Plague? Like the Black Death? Kind of. Plague is a rare but serious disease caused by the bacterium Yersinia pestis, the same bacterium behind the Black Death in the 1300s. It lives in wild rodents and their fleas. Thankfully, it’s much less of a problem today because of improved hygiene and living conditions. There are three forms: bubonic (the most common, which causes painful swollen lymph nodes), septicemic (in the bloodstream), and pneumonic (in the lungs). Pneumonic plague is the most dangerous and the only form that spreads person to person.

If early accounts are correct, this story is about the pneumonic plague. It’s concerning, as any lab accident/leak is, but the risk of this becoming a Covid-19 pandemic is very, very low for a few reasons:

1. Pneumonic plague is not traditionally efficient at spreading. While it does spread person to person through respiratory droplets, in previous outbreaks, one infected person typically infected one to three other people. Transmission occurs in close, sustained, face-to-face contact (like household caregiving, crowded sleeping quarters, or unprotected clinical workers). This isn’t like measles, which can linger in the air for hours and hours.

2. We have antibiotics. They work really well. In fact, a few cases of pneumonic plague pop up randomly in the U.S. The last case was in July 2025 in Arizona after a man was exposed to sick cats. Forty-six of his contacts received antibiotic prophylaxis, and none became sick. Outbreaks with person-to-person spread still happen in places where plague is endemic, like Madagascar. In 2021 in Madagascar, for example, 22 cases were confirmed, and eight people died. But antibiotics stopped the outbreak.

3. The response seems very rapid. Nearly 200 of the deceased Russian woman’s contacts were in quarantine and under close medical monitoring. Some reports say five hospitals are also under quarantine.

4. This didn’t happen in a crowded megacity. The institute is in a small town, not Moscow, which means less contact between people and less potential for spread.

Dr. Marisa Donnelly, epidemiologist, and Dr. Liz Marnik, immunologist via Your Local Epidemiologist:

On a scale from 0 (not worried) to 10 (super worried), our worry that this will turn into something big is a 1. This will change if there are more cases or we find out if the pathogen was altered in the lab. (…)

We don’t have lice, fleas, and rodents everywhere in our living quarters anymore. And our medical systems are better. But plague is still around. It’s endemic in some places, like Madagascar, and sporadic in others, including the U.S. An estimated 1,000 to 2,000 people are infected worldwide each year. (…)

If this woman died from the plague, the risk of a pandemic is very, very small.

But no worries: Trump says he will speak with Vladimir Putin about pneumonic plague

We know that Putin always speaks the truth, particularly when speaking with Trump: “”President Putin says it’s not Russia. I don’t see any reason why it would be.”

Remember how George W. Bush, after his first meeting with Putin in 2001, responded to a reporter asking him if he could trust Putin: “I looked the man in the eye. I found him to be very straightforward and trustworthy… I was able to get a sense of his soul”.

Joe Biden recounted that during a 2011 meeting with Putin in Moscow, he explicitly referenced Bush’s comment, telling Putin: “I looked in your eyes, and I don’t think you have a soul.”

According to Biden, Putin smiled and responded, “We understand one another.”