Broadcom’s $60B AI Debt Financing Is a Red Flag
When Broadcom lends Anthropic up to $42 billion, AI revenue stops meaning what you think it means. This boom is becoming a debt-financing experiment.
Broadcom is reportedly lining up $60 billion in fresh debt financing to support AI-chip infrastructure for Anthropic and others. That is not merely a large deal. It is the moment the AI boom stopped looking like a standard technology cycle and started looking like Wall Street building scaffolding underneath demand. ([au.finance.yahoo.com](https://au.finance.yahoo.com/news/broadcom-starts-amassing-60-billion-041233805.html?utm_source=openai))
Let’s be clear: this does not mean Anthropic is a bad business, Broadcom is reckless, or AI is a fraud. It means the numbers need to be read like an adult reads them.
A supplier financing its customer can be perfectly rational. Aircraft makers do it. Machinery vendors do it. Banks have made a career out of it. But when the supplier is also providing the hardware, arranging the capital, and potentially receiving equity exposure in the buyer, you have to stop pretending every dollar of future demand is independent market validation.
That is the uncomfortable bit. And it matters because the next phase of AI will not be won by whoever puts the biggest number in a press release. It will be won by the firms that can turn eye-watering infrastructure commitments into durable cash flow.
The $60 billion headline is only half the story
Bloomberg reported on October 2 that Broadcom’s Wall Street syndicate was gathering $60 billion in new financing: a proposed $42 billion senior-secured tranche and an $18 billion junior-debt tranche led by Blackstone, which was expected to commit $9 billion itself. Broadcom declined to comment to Bloomberg. ([au.finance.yahoo.com](https://au.finance.yahoo.com/news/broadcom-starts-amassing-60-billion-041233805.html?utm_source=openai))
This comes after Reuters reported, based on Anthropic’s IPO prospectus, that Broadcom had agreed to lend Anthropic up to $42 billion for infrastructure spending. The prospectus reportedly lays out an arrangement where Broadcom is intertwined with Anthropic as a compute supplier, equipment-leasing partner and financing source. Anthropic also warned investors that the relationship created potential conflicts around hardware pricing and access to capacity. ([fidelity.com](https://www.fidelity.com/news/article/default/202610010606RTRSNEWSCOMBINED_KBN3VH3ZP-OUSBS_1?utm_source=openai))
That is a lot of roles for one counterparty.
Broadcom is not just selling shovels during a gold rush. It is helping finance the miner buying the shovels. In a normal market, that should make you ask one question before you get excited about revenue projections:
Would the customer still buy at this scale if the seller were not helping fund the purchase?
Sometimes the answer will be yes. Frontier AI labs genuinely need extraordinary computing capacity. The models are expensive to train, expensive to serve and increasingly embedded in real enterprise workflows. Anthropic’s Claude is not some novelty app with a cartoon logo and no customers.
But that does not make the financing structure irrelevant. It makes scrutiny more important.
This is what the AI capital cycle now looks like
For years, the clean version of the AI trade was simple: Nvidia sells chips, hyperscalers buy them, startups rent the compute, and customers eventually pay for useful software.
That version is gone.
Now the money loops through chip designers, cloud providers, specialist infrastructure firms, private-credit funds, data-centre developers and AI labs. Each party has a logical reason to participate. Broadcom sells custom AI silicon and networking. Anthropic secures more compute. Blackstone earns credit exposure. Banks arrange debt. The market gets a new pool of assets to finance.
Individually, none of that is outrageous.
Collectively, it creates a system where the same belief — that demand for AI compute will remain enormous for years — is doing a hell of a lot of work. If that belief is right, everyone looks clever. If it weakens, the pain does not stay neatly inside one startup or one chip company.
Broadcom’s own strategy has been explicit. In June, it announced a platform with Apollo and Blackstone intended to support more than 20 gigawatts of global AI deployments through 2028, beginning with a $35 billion tranche tied to Anthropic’s capacity expansion. ([investors.broadcom.com](https://investors.broadcom.com/node/64396/pdf?utm_source=openai))
And in April, Broadcom said Anthropic would access roughly 3.5 gigawatts of next-generation TPU-based AI compute from 2027 through its expanded collaboration with Broadcom and Google. ([investors.broadcom.com](https://investors.broadcom.com/static-files/c906d370-921b-4bc2-bb7b-57877dfcf1ae?utm_source=openai))
These are not ordinary SaaS contracts. Gigawatts are power-station language. Once you are discussing debt packages, custom silicon, leased capacity and multi-gigawatt deployments, you are no longer analysing a software company in the old sense. You are analysing industrial infrastructure with software economics layered on top.
That is both the opportunity and the risk.
The overlooked angle: the smart money is moving assets off balance sheets
Broadcom is not alone in discovering that AI hardware is too expensive to casually carry.
Reuters reported on October 2 that Amazon was in talks to move roughly $8 billion of advanced Nvidia chips into a special-purpose vehicle, then lease the chips back. The proposed structure would put thousands of Grace Blackwell chips, spread across more than a dozen US data centres, into a vehicle financed by outside investors through debt. Amazon and Nvidia did not immediately comment to Reuters. ([streetinsider.com](https://www.streetinsider.com/Reuters/Amazon%2Bseeks%2Bto%2Boffload%2B%248%2Bbillion%2Bof%2BNvidia%2Bchips%2Bto%2Binvestors%2C%2BFT%2Breports/27137844.html?utm_source=openai))
Do not confuse that with Amazon losing faith in AI. It is more likely the opposite: Amazon wants the compute, but it also wants someone else to fund more of the asset base.
That is a rational operator’s move. If you can preserve capital, keep building capacity and retain access to the hardware through leases, why would you voluntarily let every chip sit on your own balance sheet?
But it also tells you something important. The biggest companies in the world are not treating AI infrastructure as a normal capex programme anymore. They are looking for financial engineering because the spending is becoming too enormous to ignore, even on their balance sheets.
That should sober up founders who think the answer to every hard business question is “we’ll just use more AI.” Compute has a cost. Serving users has a cost. Agentic products can create wildly variable costs. And borrowing, leasing or raising equity does not make a lousy unit economy good. It just delays the day you have to admit it.
Why this does not automatically mean ‘bubble’
There is a lazy take doing the rounds: supplier financing equals bubble, therefore sell everything and hide under the doona.
Rubbish.
Big infrastructure booms always involve financing innovation. Railways, telecoms, energy, commercial aviation and cloud computing all required capital structures that looked aggressive before the assets produced mature cash flows. The issue is not whether financing exists. The issue is whether the underlying asset earns enough, for long enough, to service it.
AI could absolutely justify huge infrastructure investment. Models are getting more capable. Enterprises are paying for automation, coding assistance, customer support, research, security and workflow tools. The prize is real.
But investors should separate two things that are currently being bundled together:
1. AI is useful. 2. Every AI infrastructure commitment will earn a brilliant return.
The first can be true while the second is very much not.
The AI market will produce enormous winners. It will also produce overbuilt data centres, obsolete hardware, squeezed cloud margins and vendors who mistook financed demand for permanent demand. That is not cynicism. That is what happens whenever capital races ahead of proven utilisation.
The real risk is concentration, not technology
The part I find more interesting than the circular-financing headlines is concentration.
Broadcom is building a deeper commercial relationship with a tiny number of frontier-model companies. Anthropic gets greater access to compute and financing. Broadcom gets a major customer and a powerful position in the custom-chip stack. That can be a fantastic partnership.
It can also leave both parties exposed if one model lab loses its technical edge, hits a regulatory wall, cannot monetise fast enough, or simply discovers that its customers will not pay enough to cover premium inference costs.
The same goes for the lenders. Private credit is brilliant until it gets paid for ignoring correlations. If the collateral is AI hardware and the borrower’s revenues depend on an AI boom continuing, those risks are not independent. A downturn can pressure chip values, utilisation rates and borrower cash flow at the same time.
That is the bit nobody should wave away with a PowerPoint slide full of hockey sticks.
What this means for you
If you are a founder, stop using AI spend as a badge of seriousness. Track gross margin after model and inference costs every week. Not quarterly. Every week. If usage doubles, know exactly whether your margin improves, stays flat or gets smashed.
If you are building on someone else’s model, negotiate for flexibility. Use a multi-model architecture where it makes sense. Avoid designing your product so tightly around one provider that a price rise, capacity constraint or product change can kneecap you overnight.
If you run an established business, do not ask, “How do we get an AI strategy?” Ask, “Which three expensive workflows can we automate with a measurable payback inside 12 months?” That question produces a business case. The first one produces a committee.
If you invest, separate the picks-and-shovels story from the financing story. Revenue backed by independent customer demand is worth more than revenue made possible by a vendor-funded capital loop. Read the commitments, the leases, the guarantees and the counterparty exposure. The boring footnotes are where the actual risk lives.
And if you are tempted to call this all a bubble or all a revolution, resist both lazy labels.
The better verdict is this: AI is becoming real infrastructure. Real infrastructure creates fortunes, but it also creates debt, concentration and spectacular mistakes when people confuse a massive buildout with a guaranteed return.
The winners will not be the loudest AI companies. They will be the ones that can still make money after the financing party ends.