Anthropic’s Report on Lobsters, Pregnant Women, and Banned Chips: A Critical Look at US‑China AI Chip Policy

Today's article is about lobsters, pregnant women, and banned chips.

Could it be that pregnant women eating lobsters while gaming help fetal development? It sounds absurd; these three have nothing to do with each other.

No joke, today I saw a blog post from Anthropic.

These three keywords were mentioned in the blog.

In addition, Anthropic submitted a more detailed report to the US government.

Reading the whole report made me laugh.

First, let's see what's up with the lobster.

This news link was provided in Anthropic's own report.

The news report is dated 2023, but it mentions smuggled 2014‑released Nvidia Quadro K2200.

Indeed, it's a 2014 GPU, considered obsolete in 2023, unlikely to be used for AI training.

What about the pregnant women? I followed the link and found this news—

The news was published in 2022, smuggling a CPU, or a fake pregnant woman.

Seeing this passage in the report, I began to understand.

Using this method, how many “pregnant women” and “lobsters” are needed to build a training cluster?

These examples illustrate the absurdity of the report.

Of course, the report has more interesting points; let me introduce its content.

Summarized in one sentence, the report's purpose is:

“To keep America ahead, we must ban selling chips to China.”

With this stance, the rest of the paper argues around it, main arguments are:

“We at Anthropic are great.”

“Artificial intelligence is powerful.”

AI’s significance for US national security means the government must pay attention.

The first two points are relatively normal; the core argument follows.

Anthropic’s logical chain is roughly:

1. Advanced AI requires massive compute.

2. Chinese chips are inferior to US chips.

Specifically, China’s top chip Huawei Ascend 910C delivers only half the performance per watt of Nvidia’s top chip GB20; Chinese chip yields are insufficient and its software ecosystem is lacking.

It also notes that if China tries to achieve compute by stacking many cards, it would require massive electricity.

In short, Chinese chips are weak, Chinese firms depend on US chips, so restricting export of top chips is needed to curb China’s AI development.

This is not Anthropic’s first China‑targeted report; each is similar, elevating to US national security.

In contrast, other US AI giants have not acted similarly.

Google focuses on models, OpenAI holds frequent launches; they are improving internally and competing on products.

So what prompted Anthropic?

The report mentions DeepSeek 23 times, apparently spurred by DeepSeek V3 and R1.

Compared to Google, OpenAI, and Meta, Anthropic has the thinnest foundation.

Its only product is its model; its moat is the model.

If DeepSeek’s coding ability surpasses Anthropic Sonnet 3.7, it would truly hit Anthropic’s weak spot.

Although Sonnet 3.7’s coding ability is strong now, no one can guarantee it stays that way.

By the way, Google upgraded Gemini 2.5 Pro to version 05‑06; its coding ability now surpasses Claude 3.7 in benchmarks:

The above image shows benchmark data from lmarena.ai.

The most intense reaction came from Nvidia, the biggest AI hardware supplier.

Nvidia’s response was blunt:

Nvidia’s comment, while self‑interested, acknowledges the ban targets its products.

But it also reflects its view of the current landscape:

References:

https://www.anthropic.com/news/securing-america-s-compute-advantage-anthropic-s-position-on-the-diffusion-rule https://www-cdn.anthropic.com/7449887b6715e3a35f362b1301e5b5d8a6b116e5.pd