Apollo and Blackstone Shopping $36 Billion Debt Deal to Supply Anthropic with Google TPUs

Apollo Global Management and Blackstone are working to bring additional investors into a roughly $36 billion debt financing deal that would supply Anthropic with Google's custom tensor processing units.

That's $36 billion. With a B. For context, that's larger than the GDP of over 80 countries.

The deal structure is notable: the debt will be used to purchase Google's TPUs, which Anthropic will then lease, according to Bloomberg. It's essentially a massive equipment financing arrangement, the kind of thing you'd see for aircraft or shipping containers, except the equipment here is AI accelerators.

What does this tell us about AI compute economics?

The scale here is striking, but it's the structure that caught my attention. Anthropic isn't buying these chips outright. They're not even taking on the debt directly. Instead, Apollo and Blackstone are effectively becoming AI infrastructure landlords, purchasing hardware that Anthropic will lease.

From my time in hardware, I've seen this model before in industrial automation. When you need expensive equipment but want to preserve capital, you lease. But $36 billion worth of leased compute suggests something about Anthropic's cash position, or at least their preference for keeping that cash liquid for other purposes (like, say, actually training models).

The involvement of private equity giants also signals that traditional finance sees AI infrastructure as a bankable asset class. TPUs depreciate, sure, but apparently not fast enough to spook Apollo and Blackstone.

Why Google TPUs instead of Nvidia?

This is where it gets interesting. Anthropic has been a Google Cloud customer for years, and Google has made significant investments in the company. Using Google's custom TPUs rather than Nvidia's GPUs isn't surprising given that relationship, but the scale of this commitment is.

TPUs are designed specifically for machine learning workloads. They're not general-purpose chips. That means Anthropic is betting heavily on a single vendor's silicon roadmap. If Google's next-generation TPUs underperform, or if there's a supply issue, Anthropic doesn't have the flexibility to pivot to alternative hardware without, well, renegotiating a $36 billion deal.

Look, that's a calculated risk. Google has been iterating on TPUs since 2016 and they're now competitive with Nvidia's best offerings for training large language models. But it's still a risk.

The numbers in context

Let's put $36 billion in perspective:

• Anthropic's last reported valuation was around $18 billion (as of early 2024)
• The entire global semiconductor equipment market was roughly $100 billion in 2023
• Nvidia's data center revenue for fiscal 2024 was about $47 billion

So this single debt deal is worth twice Anthropic's last known valuation and represents a meaningful fraction of global chip spending. That's an ambitious number, even by AI industry standards.

It remains unclear how the economics will work over the lease term. We don't know the interest rate, the lease duration, or what happens to the TPUs at the end of the agreement. Bloomberg's reporting doesn't include those details, and neither Apollo nor Anthropic have disclosed them publicly.

What this means for the AI infrastructure race

The AI industry has a compute problem. Training frontier models requires massive amounts of specialized hardware, and there's basically not enough of it. Companies are scrambling to secure supply through any means necessary: direct purchases, cloud contracts, and now, apparently, $36 billion private equity financing arrangements.

This deal suggests that the compute arms race is entering a new phase. When private equity firms start treating AI chips like infrastructure assets, sort of like toll roads or power plants, you're seeing a maturation of the market. Whether that's good or bad for innovation is another question.

For Anthropic specifically, securing this much compute capacity would be significant. The company has been competing with OpenAI and Google DeepMind for frontier model development, and training runs are only getting more expensive. Having guaranteed access to tens of billions of dollars worth of TPUs would help close any compute gap.

But the real test is whether this deal actually closes at the reported size. Apollo and Blackstone are still shopping it to additional investors, which suggests they want to spread the risk. Finding partners for a $36 billion debt deal in any sector isn't trivial. In a sector where the underlying technology is evolving as rapidly as AI, it's even harder.

I've seen enough spec sheets to know that announced deals and closed deals are different things. We'll see where this lands.