Welcome to this week’s edition.

So let me get this straight.

We spent trillions building AI infrastructure on the assumption that a handful of companies would keep buying the compute.

Then cheaper Chinese and open-weight models started catching up.

And now some of the biggest AI companies are calling for stricter safety regulations. 😁

Maybe they are right about the risks.

But I think we should pay attention to who gets to write the rules and who can afford to follow them.

Let's get into it ~

Microsoft President Brad Smith recently called for stronger global AI safety regulations.

His argument is that even milk and washing machines face more safety requirements than AI.

Microsoft wants:

And I actually agree with much of this.

If your AI agent breaks into somebody else's infrastructure, "it behaved unexpectedly" should not be an acceptable excuse.

But there is another side to this story.

On October 14, Anthropic's international chief argued that open models carry greater safety risks than closed alternatives.

There is a legitimate argument here.

Once model weights are released, developers cannot easily recall them, enforce centralized safeguards or monitor how everyone uses them.

But open-weight models also give organizations more control, less vendor dependence and the ability to run AI locally.

Importantly, Anthropic's CEO has explicitly said the company does not support blanket bans on open-weight models. Instead, it supports mandatory safety testing for sufficiently capable models, open and closed.

Question:

What happens when safety requirements become so expensive that only the largest AI labs can afford compliance?

That is where safety regulation can also become a competitive advantage.

According to Jefferies research reported by The Register:

Around half of the $2.3 trillion contracted cloud backlog across Microsoft, Alphabet, Amazon and Oracle is attributed to OpenAI and Anthropic.

Meanwhile, cheaper Chinese and open-weight models are pressuring API prices.

And Bain estimates that AI infrastructure investment could require approximately $6 trillion in annual AI revenue by 2031 to remain economically justified.

Existing applications might generate $1.8 trillion, leaving a projected $4.2 trillion gap to fill.

These are forecasts obviously.

Cheaper models could also drive enormous growth in usage.

But you see the problem.

Cloud providers are making huge infrastructure commitments while their biggest AI customers face increasing competition and pricing pressure.

That creates some interesting incentives around regulation.

I support enforceable AI safety requirements.

Especially for autonomous agents that can access credentials, execute code, and affect real infrastructure.

I don’t like the excuse “Yeah my agent did it, not my fault”. There should be consequences on the agent owner or delegator.

But a small open-weight research project should not face the same compliance burden as a frontier model with demonstrated dangerous capabilities.

Regulation should be:

Risk-based. Independently evaluated. Transparent.

And most importantly, the same capability-based standards should apply to closed models too.

Otherwise, we risk creating a market where the biggest AI companies can afford to follow the rules they helped shape, while everyone else struggles to compete.

But the direction seems like going into banning the opensource models to not lose to the competition.

I like cheaper models.

Threat actors like cheaper models too.

And the largest AI labs have actual reasons to be concerned about both security and competition, but should not abuse their power to get a monopoly.

For enterprise architects, my advice stays the same:

Keep testing alternative models. Avoid unnecessary vendor lock-in. Design your AI architecture so you can change providers.

Because tomorrow's biggest AI security risk might not be your model.

It might be your complete dependence on the company selling it.

That's it for this week.

What do you think? Are we seeing genuine AI safety regulation or the beginning of regulatory capture?

See you in the next one.

Reuters: Microsoft calls for global AI safety rules

Bloomberg: Anthropic executive warns about open-source AI

Anthropic: Official position on open-weight models

The Register: AI infrastructure investment and competition