@FinanceYF5: How Anthropic used less than $9 billion in annualized revenue to secure nearly $50 billion in AI infrastructure. 1/ When Anthropic announced its $50 billion US AI infrastructure plan, its annualized revenue was still under $9 billion. Then, it secured nearly $50 billion in debt financing for over 1GW of TPU and 5 data centers. This…

How Anthropic Leverages Less Than $9 Billion in Annualized Revenue to Unlock Nearly $50 Billion in AI Infrastructure

1/ When Anthropic announced its $50 billion U.S. AI infrastructure plan, its annualized revenue was still less than $9 billion.

Subsequently, it secured nearly $50 billion in debt financing for over 1GW of TPUs and 5 data centers.

This suggests that for frontier AI expansion in the short term, what’s truly scarce may not be money, but long-term contracts and credit support acceptable to financial institutions.

2/ The key to this financing is transforming future computing demand into assets that can be lent against today.

Anthropic signs long-term leases with Fluidstack; SPVs hold the TPUs or data centers; banks, insurers, and private credit provide the funding; the equipment and facilities serve as collateral.

Broadcom and Google aren’t upfront funders but assume partial risk in case of default.

3/ The TPU segment is the largest: Funds managed by Apollo, Blackstone, and several banks have committed $34.5 billion in debt to the AI XPV Platform.

The SPV purchases over 1GW of Google TPUs and leases them to Anthropic for five years.

Anthropic pays rent to service the debt, with Broadcom supporting $30 billion of this, carrying a maximum risk exposure of about $29 billion.

4/ This $34.5 billion is structured into three tranches.

Tranches A1 and A2 total $30 billion, have higher repayment priority, and are supported by Broadcom; Tranche B is $4.5 billion, unsupported, and last in line.

The difference is directly reflected in pricing: Tranche A2 has an interest rate of 5.75%, while Tranche B reaches 8.5%. Credit support clearly makes capital cheaper.

5/ The data center segment uses a similar project finance structure.

Five project companies have issued approximately $15.18 billion in debt to build 1.43GW of capacity across Lake Mariner, Barber Lake, Abernathy, River Bend, and Meridian Arc.

Institutional capital is invested upfront, future rental payments service the debt, and Google provides conditional support for a portion of the rent.

6/ Lake Mariner best illustrates how this structure works.

TeraWulf handles construction, the project company borrows $3.2 billion; Fluidstack begins paying rent after capacity delivery and then provides data center and operational services to Anthropic.

If Fluidstack defaults, Google can step in to pay rent, take over the lease, or pay termination fees.

7/ Epoch AI concludes that, at least in the short term, financing will not be the primary bottleneck for frontier computing growth.

Broadcom, Apollo, and Blackstone also plan to scale this model to over 20GW of deployment before 2028.

However, its sustainability ultimately hinges on one thing: whether AI companies’ long-term revenues can fulfill the massive leases signed today.

Original article: https://epochai.substack.com/p/will-financing-bottleneck-ai-compute?utm_source=substack&publication_id=3755861&post_id=210914311&utm_medium=email&utm_content=share&utm_campaign=email-share&triggerShare=true&isFreemail=true&r=oqv1&triedRedirect=true…

Anthropic Tests Claude’s Ability to Design Novel Protein Binders De Novo

In a human-expert-crafted protein design prompt, Claude autonomously designed binders for 14 out of 15 targets.

Subsequently, Anthropic partnered with Adaptyv Bio and Twist Bioscience to independently build and test the proteins designed by Claude.