Image: METAL
Summary
- A $15 billion fundraising for Anthropic, led by Morgan Stanley, is moving forward.
- Google will secure a 20% stake in a Texas data center and supply its own AI chips under the arrangement.
- Under the guarantee structure, if Anthropic fails to meet its obligations, Google will cover the rent and power costs instead.
Google has agreed to provide a debt guarantee for Anthropic's $15 billion fundraising. In this Morgan Stanley–led deal, Google will secure a 20% stake in a Texas data center in exchange for the guarantee, and will supply the AI chips to go inside it — its own TPUs.
This isn't lending money — it's guaranteeing the borrower's debt. Anthropic gets to raise funds on terms it likely couldn't secure on its own credit alone, while Google locks in a long-term buyer for its chips without spending a cent upfront.
One deal, three threads tied together
| Thread | What Google gives | What Google gets |
|---|---|---|
| Finance | Debt guarantee on $15 billion fundraising | Deal enablement · lower effective borrowing cost |
| Real estate | Guarantee risk exposure | 20% stake in Texas data center |
| Semiconductors | TPU supply volume | Long-term demand anchor for its own chips |
The guarantee's terms are specific. If Anthropic fails to meet its obligations, Google will pay the data center's rent and power costs in its place. The fact that power costs are written alongside rent in the contract signals how large a share electricity now takes up in AI data center cost structures.
This fundraising is separate from the $35 billion chip purchase agreement signed earlier this year. Combined, the scale of Google's exposure to Anthropic's infrastructure grows considerably larger. Around the same time, a cost-sharing arrangement involving Broadcom for chips was also reported.
Why this structure emerged
As AI infrastructure investment scales up, fundraising itself becomes the bottleneck. A single data center can cost trillions of won, with returns not materializing for years. Model companies are generally unprofitable, making it hard for them to pass traditional loan underwriting.
When a chip company steps in as guarantor, three problems get solved at once.
- The model company raises funds on terms it couldn't get on its own credit.
- The chip company locks in demand for its products through contract, without spending cash.
- Lenders can treat the debt as backed by a major corporation, lowering perceived risk.
But the risk doesn't disappear — it concentrates in one place. If Anthropic falters, Google is left covering rent and electricity bills. Taking a 20% stake alongside the guarantee looks like a way of securing compensation for that risk in advance.
What's happening outside NVIDIA's orbit
The AI infrastructure narrative has largely been framed as a race for GPU supply. This deal shows a different axis: a company with its own accelerator using finance as leverage to manufacture demand for its own chips. Instead of simply selling chips, Google is becoming the financier behind the building the chips will sit in.
TPU is Google's own custom-designed AI accelerator. It has historically been used more within Google's own cloud than sold externally, but that changes once a major model company takes on volume through a long-term contract. It's a structure that uses contracts to compensate for the criticism that the ecosystem around it is weaker than general-purpose GPUs.
Degree of vertical integration, in table form
| Layer | Integration tool | Appearance in this deal |
|---|---|---|
| Semiconductors | Proprietary accelerator | TPU supply |
| Power/real estate | Data center equity stake | 20% in Texas |
| Finance | Debt guarantee | $15 billion |
| Demand | Long-term purchase agreement | Separate $35 billion |
It's unusual for a single company's name to appear across all four layers.
What this means for practitioners in Korea
First, dependency structures are deepening. When adopting an AI service, it's no longer enough to look only at the model provider. Which cloud and which accelerator the model runs on, and what conditions bind that contract, now directly affect pricing and availability. It's worth checking contracts for notification clauses covering changes in model, region, or accelerator.
Second, power is the bottleneck. It's no coincidence that power costs appear in the guarantee terms. For anyone evaluating a data center project domestically, grid connection timing and rate structures now determine roughly half of a project's viability.
Third, memory demand will move in tandem. Anthropic CEO Dario Amodei recently noted that Korean memory manufacturers have already made substantial investments. As this line of business grows, the volume flowing into Korea's semiconductor and data center industries will move with it.
What remains unconfirmed
Whether the fundraising has formally closed, the composition of the lending group, and the exact trigger conditions for the guarantee all fall outside what's been publicly reported. Even the size of the data center is described differently across reports. This article covers only what has been confirmed.
Source: Compiled from Maeil Business Newspaper and other domestic reports.
Comments