Kuroishi wants to raise another $36 billion to bet on Anthropic! The “securitization” of AI computing power is booming. Who is paying for potential credit defaults?

The Zhitong Finance App learned that Wall Street private equity and alternative asset management giant Blackstone Group (Blackstone) has begun preliminary discussions with investors to test the market's interest in the second huge AI-related debt financing scheme. The plan will be used to fund AI application leader Anthropic to use TPU computing power chips from Google, a subsidiary of Alphabet.

Some media reported, citing information revealed by people familiar with the matter, that one of the preliminary plans is to raise at least 36 billion US dollars in debt funds. People familiar with the matter said that specific details, including the size of the financing, transaction structure, and even whether the financing will eventually be led by Kuroishi, are still being discussed and may change. As they were not authorized to comment publicly, these individuals requested anonymity.

These large-scale AI financings first strengthened order certainty in the computing power industry chain, followed by the bursting of the AI bubble and liquidity risk. Blackstone's second round of financing of at least $36 billion, combined with the $35 billion AI XPV initial transaction that Apollo and Blackstone have already completed, essentially translates Anthropic's revenue expectations for the next few years into current Google TPU AI computing power cluster procurement, optical communication/optical interconnect, data center CPUs, high-performance Ethernet switching infrastructure, data center HBM/DRAM/NAND storage components, and data center power chains.

The initial financing of the AI XPV platform corresponds to more than 1 GW of computing power, and plans to support more than 20 GW of deployment by 2028; this means that chip orders are no longer entirely dependent on AI Labs' current cash flow, but can be “capitalized” through SPV, long-term leases, and supplier credit support. As a result, in the short term, the visibility of orders from Broadcom's AI computing power chain, Google's TPU ecosystem, network equipment, power, and data center core hardware vendors will be significantly extended.

The risk is that financial engineering has not eliminated repayment risks; it has only transferred them from AI Labs' balance sheets to SPV creditors, chip vendors, cloud computing platforms, and their guarantors. This AI boom is not bound to break down, but its pricing core is shifting from technology penetration to whether the credit chain can be captured by real cash flows. Castle Securities anticipates that AI chip financing debt may exceed 500 billion US dollars by 2028, making “credit leverage risk poised” a reasonable investment warning phrase for analysts.

AI financing competition breaks another record: Blackstone plans to raise at least 36 billion US dollars, Anthropic computing power layout bets on Google TPU chips

If the deal were to be completed on a scale close to that, it would exceed the $35 billion debt financing arrangement between Apollo Global Management (Apollo Global Management) and Blackstone about two months ago for Anthropic to rent a customized AI chip made exclusively by Google. The financing was one of the largest credit transactions in the history of the private credit market.

Representatives for Blackstone, Apollo, Anthropic, and Google all declined to comment.

As Silicon Valley competes to build artificial intelligence infrastructure, leading companies have mutually agreed on complex and often circular financing transactions to ensure access to computing power resources. Google was one of Anthropic's earliest investors. It has purchased its equity assets many times, and is now increasingly providing guarantee support for financing the large-scale construction of this AI startup's data center.

This potential new round of funding comes on the heels of Anthropic's secret filing for an initial public offering in the US stock market. The company is trying to get to the open market ahead of rival OpenAI. As the developer of Claude's AI large model series of products, Anthropic plans to rent high-performance computing chips at five data centers with help provided by Google, with the support of previous AI debt financing arrangements.

Global technology companies are using every sector of the credit market to meet the unprecedented capital needs of artificial intelligence, which has forced Wall Street asset management giants to collectively participate in designing new debt structures to keep up with the pace of industry expansion. Due to market concerns that investment in artificial intelligence may not achieve the expected positive return on investment, some companies have also recently been forced to pay high benchmark returns when issuing new debts.

Broadcom, Apollo and Blackstone launched a collaborative platform called the “AI XPV Platform” this year to help leading artificial intelligence technology development companies, including Anthropic, finance AI computing power infrastructure. The $35 billion AI-related debt financing, which was completed about two months ago, is the first financing provided by the platform.

In this deal, Broadcom provided underwriting support for the payment of the largest portion of senior debt. Earlier, there were media reports that Morgan Stanley acted as an advisor to Broadcom and helped arrange the deal. Broadcom representatives did not respond to related requests for comment.

Wall Street begins the “securitized computing power” process: Anthropic and Meta realized some future demand ahead of schedule, and the AI bull market entered the credit leverage stage

The $35 billion AI XPV financing that has already been implemented is used to support Anthropic's expansion of computing power by more than 1 GW, while the initial scale of the second round of financing under discussion by Blackstone is at least 36 billion US dollars. Blackstone's proposed new round of financing of about $36 billion, combined with the $35 billion AI XPV initial funding already in place, is transforming part of Anthropic's strong future revenue expectations into current TPU and data center orders; the first phase of the platform supports more than 1 GW of computing power and aims to promote the deployment of more than 20 GW of AI infrastructure around the world.

All of this means that chip orders are no longer entirely dependent on AI Labs' current cash flow, but can be “capitalized” by Wall Street asset management giants through SPV, long-term leases, and supplier credit support. As a result, in the short term, the visibility of orders from Broadcom's AI computing power chain, Google's TPU ecosystem, network equipment, power, and data center core hardware vendors will be significantly extended.

The risk is that financial engineering did not eliminate repayment risk; it only transferred it from the AI Lab balance sheet to SPV creditors, chip vendors, cloud platforms, and their guarantors. Anthropic's financing structure involved SPV buying Google TPU and leasing it to Anthropic. The largest portion of the senior debt relied on Broadcom to provide differential or residual value support, so the $25 billion core debt was able to reduce the yield to about 5.75%; the yield on secondary debt without equal guarantees was about 8.5%. This forms a typical “closed loop of supplier financing”: chip manufacturers facilitate customer purchases through credit endorsements, and purchase orders are converted into the chip manufacturer's revenue and valuation basis. Once Anthropic's revenue growth rate, computing power utilization rate, or IPO financing falls short of expectations, risk will be transmitted in reverse along the “lease payment — SPV repayment — chip residual value — supplier guarantee” chain.

Facebook's parent company Meta shows that this off-balance sheet capital commitment has been upgraded from a project-level arrangement to a core financing model for giant technology companies. By the end of June 2026, Meta's commitment to leasing, and therefore not included in balance sheet lease liabilities, had reached US$278.99 billion, a significant increase from the previous quarter; in July, it also signed data center leases of about US$68 billion for a period of 18 to 20 years, and also had irrevocable contractual commitments of US$349.31 billion and up to US$14.72 billion or cloud capacity procurement obligations.

Meanwhile, Meta has raised its 2026 capital expenditure guidance to $130 billion to $145 billion. Strictly speaking, these large-scale AI financing and debt projects are not “hidden debts” in the traditional sense, but their economic essence is long-term fixed payment obligations: when the facility is put into use, rental debt will gradually enter the statement, and cash flow pressure has already been locked in through contracts.

Castle Securities predicts that by 2028, the issuance of new bonds used only for AI chip procurement may exceed 500 billion US dollars, which is equivalent to more than 5% of the US investment-grade bond index at that time. Of these, the amount issued in a single year in 2028 may exceed 250 billion US dollars. Most debt terms are only 3 to 5 years to match the short economic life of chips. The real danger is not just “too much debt,” but a maturity mismatch between short-term debt, rapidly depreciating assets, and long-term, highly uncertain AI revenue: if interest rates remain high, model inference prices fall, the residual value of old devices plummets due to iterations of chip architecture updates, or the credit market requires higher issuance premiums, a large number of projects may simultaneously face refinancing walls around 2028, thereby squeezing technology, media, and communication bond allocations, causing credit spreads to expand and private credit liquidity to shrink.

Therefore, according to some senior Wall Street analysts, investment strategies should value “funded orders” separately from “terminal cash requirements”: prioritize customer diversification, sufficient net cash, orders with prepayments or irrevocable promises, and chip and infrastructure leaders that can continuously generate free cash flow; for projects that rely on SPV, supplier guarantees, revolving investment, and rent repayment of a single AI laboratory, it is necessary to increase risk and discount, and focus on monitoring credit spreads, CDS, lease-adjusted leverage ratio, chip residual value, and computing power utilization. Overall, this AI boom is not inevitable, but its pricing core is shifting from technology penetration to whether the credit chain can be captured by real cash flows.