1/ DeepSeek originally represented 'lower-cost AI', but it might have pushed American tech giants into a more aggressive debt race. BofA predicts that, assuming about $200 billion in debt issuance each in 2026 and 2027, Hyperscaler's investment-grade debt could rise from $288 billion to $659 billion…

1/ DeepSeek was originally synonymous with “lower-cost AI,” yet it may have inadvertently pushed U.S. tech giants into a more aggressive debt-raising race.

BofA estimates that, assuming approximately $200 billion in debt issuance each in 2026 and 2027, the investment-grade debt held by Hyperscalers could surge from $288 billion to $659 billion.

Izabella Kaminska argues the turning point was DeepSeek. 👇 https://t.co/iizW6H9euJ

2/ Her reasoning: once DeepSeek open-sourced its weights, model capabilities became easier to replicate and undercut.

As models gradually commoditize, the competition shifts from a “single winner” paradigm to another arena:

Whoever builds the most data centers, secures the most power, and locks up the most computing capacity stands a better chance of establishing a durable moat.

3/ From this perspective, AI infrastructure spending doesn’t need to adhere to traditional economic logic.

Over-investment itself becomes a strategy: continuously raising the cost of competition until rivals can no longer finance their operations.

Even if one incurs massive losses in the process, as long as the end result is a network of data centers, power grids, and chip production capacity, a nation may still take over these assets.

4/ Izabella likens this to the Soviet Union pouring resources into steel factories to压低成本 (drive down costs) through unmatchable scale.

But recent advances from Kimi have disrupted this logic.

If greater capital efficiency can also yield powerful models, a moat built solely on debt and computing stacks risks being circumvented by new technological approaches.

5/ Leopold Aschenbrenner’s Situational Awareness fund offers another real-world cautionary tale.

The fund bet heavily on AI infrastructure using leverage, saw its portfolio value drop 67% in July, and was subsequently forced to sell most of its public equity holdings.

Izabella suggests this could push the market away from “infinite expansion” and toward a greater focus on capital efficiency.

6/ Another possibility is that AI development becomes more secretive and centralized.

The prior conflict between Anthropic and the Pentagon over military applications and safety restrictions already demonstrates that computing power, models, and state power are hard to fully separate.

The next phase may not merely concern which company wins, but who ultimately controls these models and infrastructures.

Source:

1/ A company repeatedly questioned for having “no moat” was acquired by Stripe within roughly three years of its founding.

The acquisition price for OpenRouter hasn’t been disclosed, but it may be one of the fastest acquisitions at its scale.

What did it get right?