We need to be alert to a scenario where the "deceleration theory" of US AI giants may constitute substantial strangulation of open-source models.

In August 2024, on the JRE #2190 episode, venture capital guru Peter Thiel and Joe Rogan, the well-known conspiracy theorist, began a serious discussion about what an alien society that has mastered curvature / faster-than-light travel would look like.

Thiel reasoned very rigorously:

It's not that they might be demons or angels, it's that they must be demons or angels, if you have faster than light travel.

Aliens that have mastered faster-than-light travel do not "might" become demons or angels, but they "must simultaneously" be demons or angels.

This tech investor who has invested in both Anthropic and OpenAI put it this way:

Faster-than-light travel is an extremely powerful weapon that can strike you before you notice it, leaving no way to defend against it. A rival can occupy the entire universe before the other side even reacts.

Therefore, for a civilization that possesses such technology, something extreme must happen at the social level, and there are only two completely mutually exclusive paths:

1. Demon: Totalitarian hive mind with merged consciousness, so that no individual can launch a curvature weapon on their own;

2. Angel: Completely altruistic with no self-interest, and will never launch weapons to occupy others.

In the eyes of this highly influential figure in the US tech industry, there is no such thing as an intermediate state for a civilization that masters indefensible weapons: you must be both a demon and an angel at the same time to use this unpredictable weapon properly.

This is more than likely a reflection of today's AI deceleration narrative.

01 What Amodei Is Actually Talking About When He Calls for AI Slowdowns

Amodei argues that AI is too dangerous so its development should be slowed down, and Sam Altman from OpenAI agrees with this view.

The real "demon" part is hidden in the details of Dario and Sam's proposals:

Amodei puts forward three points: 1. Third-party on-site assessment; 2. Limit the growth rate of AI capabilities within the sphere of influence of the United States, with global coordination; 3. Maintain chip restrictions on China.

Sam agrees, adding that regulation should only target frontier models, with auditable training pipelines, release permits, alignment processes, and product liability rules.

Only the top 2-5 labs that have sufficient capital, legal teams, safety teams, and the ability to lobby Capitol Hill to set audit rules can afford such costs. (This forms a kind of club made up of OpenAI, Anthropic, plus Google, xAI, and Meta).

No wonder Elon Musk also agreed with this on X. This is a major event, and Musk is not confused about it.

This restriction has the same damaging effect on Chinese open-source labs and new US labs:

Even with sufficient capital, it is almost impossible to achieve overtaking through curves such as "cheaper training methods, open-weight distribution, and vertical post-training", because the track has been defined by the duopoly as that you have to dance in shackles following the security format set by OpenAI.

Once the duopoly (or pentopoly) at the model level gets legislation passed, it will further set rules for what computing power should look like:

Which clusters count as frontier facilities, which chips can be used for training, which cloud providers can sell compliant inference services, and which open-source models are "safe enough".

In the end, all standards and ecosystems will still revolve around the United States, since the licensing rights are held by the US, and the computing power monopoly is also in the US.

The narrative that "AI is dangerous so it needs to be slowed down" is an angelic story with half-truths, which does not need to be elaborated on here.

But very few people have pointed out that what Amodei and Sam really want is the demonic side of the deal:

By limiting the scope of competitors and actively defining the regulatory boundary, they want to permanently consolidate the leading position of Anthropic and OpenAI in the form of legislation, and completely eliminate potential competition from open-source models.

They are fully aware that legislating to ban open-source models will not stop the capability catch-up of open-source models, but only make both sides bear the corresponding cost of this comparison.

Once the legislation is passed:

1. U.S. computing power operators cannot load "unsafe" open-source models

2. U.S. AI users must also use safe closed-source models.

There is a very clear economic logic behind this:

Hock Tan gave the clearest public statement to date at Goldman Sachs Communacopia (September 8, 2026):

At present, the global inference infrastructure costs about 200 billion US dollars per year, while the model-related revenue is only about 150 billion US dollars;

Among them, the token volume of closed-source and open-weight models is roughly half and half, and open-source models even have a slightly larger share, but the revenue of closed-source models is about 120 billion US dollars, and that of open-source models is only about 30 billion US dollars. It is obvious that the average pricing of open-source models is nearly 25% of that of closed-source models.

In other words, the closed-source sector spends about 100 billion to earn 120 billion US dollars; the open-source sector spends about 100 billion to earn only 30 billion US dollars.

It is worth noting that this 30 billion US dollars of revenue does not go into the financial statements of Chinese open-source developers such as Qwen, Zhipu AI, and Kimi.

The entire open-weight industry chain is getting paid, including self-operated APIs, Bedrock/Azure hosting premiums, and intermediaries like Fireworks/Together. The cash received by Chinese labs is far less than that figure.

This is the underlying logic why U.S. computing power operators and NVIDIA highly praise Chinese open-source models:

The tokens are trained in China, but a large amount of inference revenue goes into the wallets of AWS, Azure, xAI cloud, and NVIDIA.

For xAI and AWS, as long as they provide secure access, compliance support, billing services, and VPC, they can start charging. The stronger the open-source ecosystem is, the more profit the computing power centers make.

The more commoditized models become, the more profits go to the computing power layer. The fundamental reason why AWS abandoned training frontier models and turned to hosting open-source models is not hard to figure out:

Computing power centers are essentially commoditizing frontier models by leveraging open-source models, so that they can collect most of the open-weight revenue without bearing the training costs.

02 The Open-Source Camp: Another Side of Angel and Demon

In addition to the economic reasons why chipmakers oppose the ban on open-source, David Sacks also fiercely criticized Amodei on X:

As the AI Czar, David Sacks sees the same thing as Jensen Huang:

NVIDIA favors open-source, because free weights can drive the downstream GPU load to a sufficiently large scale.

As the supplier of 80% of the world's computing power, the United States promotes open-source because Sacks clearly sees that restricting open-weight models will not stop China from launching the next Kimi, but will only eliminate US companies like Harvey that build vertical models on top of open-source, with the cost borne by US AI users.

In Sacks' view, the U.S. dominance in AI is a full-spectrum dominance over a 5-layer cake: the higher the load running on U.S. chips, the more standards, toolchains, security certifications, and enterprise procurement lists will revolve around the United States.

If commercial use of open-source models on U.S. clouds is banned, the business will not disappear, but will flow to overseas data centers with looser regulations;

The political cost of eliminating open-source will be borne by U.S. developers, while overseas regions will not follow the ban on open-source.

Therefore, the core logic of the open-source camp is not to benefit China, but to keep global intelligent traffic locked on U.S. computing power.

03 Both Sides Are Fighting for the Same Cake

Let's summarize the above points.

The open-source camp: Expand the U.S. computing power load with open weights, so that standards and ecosystems can grow upward from chips/clouds, which is conducive to the hard infrastructure of U.S. computing power layer and chip hegemony.

The deceleration camp: Narrow the model layer through frontier licensing, so that standards and ecosystems can be pushed downward from models to the supply chain. This helps the model duopoly to further capture the remaining 25% of open-source ARR and completely eliminate Chinese open-source competition.

The two sides have different methodologies, but their end goals are isomorphic:

Both are committed to consolidating the U.S. AI multi-layer stack — chips, clouds, models, applications, and rules. Sacks is worried about regulatory capture by closed-source labs; Amodei is worried that open weights will widen the capability gap; both of them take maintaining a leading position over China as a must, and both take "the United States must hold the steering wheel" as a premise.

04 Conclusion

The most paradoxical situation today is that the people who propose that "AI may get out of control and deserves the level of caution for human civilization" also claim that "this technology can only be controlled by the sovereign states/companies/chip stacks that are capable of restraining it".

On the surface, the former is the angel and the latter is the demon, but just as Thiel said, you must be both an angel and a demon at the same time.

Amodei is a typical representative of this:

On one hand, Amodei tirelessly writes in his blog about out-of-control risks, RSI, agents taking over the network, and that "we (Anthropic) owe humanity a slowdown" — no AI developer could be more angelic than Amodei.

On the other hand, Amodei loudly demands on-site assessments, legislation covering all US model companies, continued chip restrictions against China, while his own lab keeps running on the most expensive clusters, eliminating the possibility for computing power centers to access open-source models, completing the harvesting of AI possibilities in a very 1996 Microsoft-style way.

Machiavelli once said that you have to be a demon to achieve the goals declared by angels.

In Thiel's curvature weapon theory, the intermediate state is eliminated by physics; in the frontier model scenario, the intermediate state is eliminated by politics.

Pure angels have long left the game table, and pure demons cannot deceive human beings into imposing self-restrictions.

That exact position is where Amodei chooses to be the demon.

This article is compiled based on public materials, for information exchange only, and does not constitute any investment advice

This article is from the WeChat official account "Jinduan" (ID: jinduan006), written by Mu Zhi, and authorized for release by 36Kr.