In the fast-moving AI infrastructure race, lower-cost chips may not always be the safest bet. As demand for artificial intelligence computing power continues to soar, financing has become a major challenge for companies trying to secure alternatives to NVIDIA’s widely used GPUs.
Anthropic, one of the leading AI startups, appears to be taking a cautious and diversified approach. Rather than depending on a single type of processor, the company is spreading its compute strategy across multiple chip platforms, including NVIDIA GPUs, Google TPUs, Amazon Trainium chips, and AMD GPUs.
The reason is becoming clearer: in today’s AI market, the cost of silicon is only one part of the equation. Lenders and investors are also paying close attention to resale value, long-term demand, and risk. NVIDIA GPUs remain expensive, but many financing providers still view them as more reliable assets because they are widely adopted and are expected to hold stronger residual value over time.
That concern became especially visible in Broadcom’s large financing arrangement tied to Anthropic’s AI computing needs. The deal reportedly faced complications due to rising interest rates and broader uncertainty surrounding cloud infrastructure providers that offer GPU-as-a-Service and AI inference capacity.
The financing structure was ambitious. Broadcom planned to create a Special Purpose Vehicle, or SPV, to secure around $40 billion in loans from a group of banks. An additional loan from Blackstone would raise the total financing package to roughly $63 billion. The SPV would then purchase AI chips designed by Broadcom and Google and lease them to Anthropic. Anthropic’s lease payments would be used to repay the debt.
However, the deal reportedly encountered delays during negotiations. To help finalize the agreement, Broadcom’s CEO personally stepped in and offered banks more attractive loan terms, including a higher interest rate and guarantees intended to reduce potential losses. Those added assurances helped push the financing package across the finish line.
The episode highlights a key reality in the AI hardware market: cheaper AI chips can come with added financing risk. Custom AI accelerators such as TPUs and other application-specific chips may offer strong performance and cost advantages for certain workloads, but they may not yet carry the same lender confidence as NVIDIA GPUs.
This helps explain why Anthropic is not relying entirely on one supplier or chip architecture. According to its disclosed compute commitments, the company’s planned AI compute mix includes 4.7 gigawatts of NVIDIA GPUs, 4.5 gigawatts of TPUs, 1.2 gigawatts of Amazon Trainium chips, and 0.8 gigawatts of AMD GPUs.
NVIDIA still represents the largest share of Anthropic’s compute footprint, even though its GPUs are generally more expensive than some alternative AI accelerators. That suggests the company values the flexibility, availability, ecosystem support, and financing appeal of NVIDIA hardware alongside the potential cost savings of custom silicon.
For the broader AI industry, Anthropic’s strategy offers an important lesson. Winning the AI race is not just about finding the cheapest chips. It is about securing enough computing power, managing financial risk, and building a resilient infrastructure plan that can support massive model training and inference workloads over the long term.
As AI companies continue to scale, the competition between NVIDIA GPUs, Google TPUs, Amazon Trainium, AMD GPUs, and custom AI chips is likely to intensify. But financing may become just as important as performance. In a market where tens of billions of dollars are needed to build AI infrastructure, the chips that lenders trust most could have a major advantage.
