Apollo and Blackstone Raise $36B to Lease Google TPUs for Anthropic in Largest Chip Rental Deal Ever

$36 Billion for Compute: The New Economics of AI Training

In a deal that underscores the staggering capital demands of large-scale AI model training, private equity giants Apollo Global Management and Blackstone have jointly raised approximately $36 billion to lease Google's custom Tensor Processing Units (TPUs) exclusively for Anthropic, according to a report from AIbase. The arrangement—described as the largest chip rental transaction in history—provides Anthropic with access to a massive fleet of TPU accelerators without the upfront capital expenditure of purchasing hardware outright.

This structure is unusual: rather than Anthropic buying chips directly from Google or NVIDIA, the compute capacity is financed by external investment firms that then lease the infrastructure to the AI company. Apollo and Blackstone are effectively betting on Anthropic's long-term need for compute power while assuming the asset risk. The deal signals a deepening intersection between traditional finance and the rapidly scaling AI infrastructure sector.

The Parties and Their Roles

Anthropic, the developer of the Claude series of large language models, has long relied on Google Cloud infrastructure and TPU hardware after the search giant invested $300 million in the startup in early 2023 (later expanded). Unlike many AI labs that favor NVIDIA GPUs, Anthropic has leaned heavily on Google's custom chips—a strategic choice that reduces dependency on NVIDIA while tying Anthropic closer to Google's ecosystem.

Apollo Global Management and Blackstone are two of the world's largest alternative asset managers, with combined assets under management exceeding $1.5 trillion. Their involvement in AI hardware leasing is a notable expansion beyond traditional real estate and private equity deals. According to the report, the $36 billion will be used to procure and deploy TPU pods in data centers, likely operated by Google Cloud but financed through a special-purpose vehicle that separates ownership from usage.

While the exact timeline and terms of the lease have not been disclosed, industry insiders estimate that the funding could secure Anthropic enough compute to train multiple frontier models over several years. Each TPU v5e pod can deliver up to 400 petaflops of mixed-precision performance; a $36 billion allocation would enable capacity on an unprecedented scale.

Why This Deal Matters for the AI Industry

The transaction represents a fundamental shift in how AI companies acquire compute. Traditionally, hyperscalers (Google, Microsoft, Amazon) and a few well-funded startups have purchased chips directly from manufacturers or cloud providers under long-term contracts. By shifting to third-party lease financing, Anthropic insulates its balance sheet from hardware depreciation and can scale compute up or down more flexibly—though the lease terms likely include minimum commit levels.

More importantly, this deal may accelerate the commoditization of AI training infrastructure. If Apollo and Blackstone prove that chip leasing is a viable asset class, other financiers may enter the market, reducing the capital barrier for startups that cannot afford billion-dollar hardware purchases. In turn, that could fuel more competition at the frontier AI level, as smaller labs gain access to clusters they could not otherwise afford.

The structure also carries risks. Should Anthropic's need for compute diminish—due to a shift in model architecture, a slowdown in funding, or a pivot to inference rather than training—the leased TPUs could become stranded assets. Apollo and Blackstone are presumably protected by long-term contracts and strong covenants, but the market will watch closely whether such deals become a standard financing vehicle or a one-off experiment.

Impact on Google and NVIDIA

Google stands to benefit directly: tens of billions of dollars worth of TPU sales (via lease arrangements) without having to finance them internally. This effectively turns Google Cloud into a platform that offers third-party financing, similar to how AWS partners with infrastructure funds. For Google, locking in Anthropic as a massive TPU customer also strengthens its position against NVIDIA's dominance in AI accelerators. TPU market share, while still small compared to NVIDIA's GPUs, could see a significant boost from this deal.

NVIDIA, meanwhile, may face heightened competition. While its H100 and B200 GPUs remain the gold standard for AI training, the emergence of large-scale financing for alternative chips (TPU, AMD MI300X, or custom ASICs) could erode its near-monopoly in the long run. If investors see TPU leasing as a viable asset class, they may extend similar deals for other non-NVIDIA hardware, creating a more diversified compute market.

However, it is important to note that this deal does not directly reduce NVIDIA's momentum. Anthropic has always been a Google Cloud customer; the lease is an extension of an existing relationship. Many other AI labs are still doubling down on NVIDIA infrastructure. But the signal is clear: the cost and scarcity of AI chips is forcing creative financing solutions that could reshape hardware procurement.

Forward-Looking Analysis

The Apollo-Blackstone-Anthropic deal is a watershed moment for AI infrastructure financing. It validates the thesis that large-scale compute can be treated as a leased asset class, much like aircraft or data center real estate. Over the next 12-18 months, we can expect other private equity firms to explore similar arrangements with major AI labs, potentially involving other chip vendors.

For the AI community, this means two things. First, frontier model development will continue to require enormous capital, but the barriers are shifting from technical capability to financial engineering. Anthropic did not have to raise equity or debt specifically for hardware; instead, it turned to lease financing. Second, the dependency on specific hardware vendors may become more liquid as financiers seek to diversify chip portfolios to manage risk. Labs could soon have access to multi-vendor compute pools financed by Wall Street.

Watch for future disclosures from Apollo and Blackstone on the performance of this asset class. If the returns meet expectations, expect a wave of copycat deals—possibly involving NVIDIA GPUs—that could further accelerate AI model training capacity worldwide. For now, Anthropic has secured the compute it needs to chase AGI without tying up billions in silicon.