Okay I drank too much coffee this morning, so fine let’s talk about it. These past few weeks have been some of the most consequential in the world of artificial intelligence that I can remember. Quick deep dive here we go:

US AI Supremacy Is Over

You’ve probably seen some of the big headlines this week about Chinese AI Startup DeepSeek skyrocketing it’s way to #1 in the App Store, surpassing ChatGPT. For those who haven’t:

• DeepSeek (open-source AI chat model controlled by the Chinese gov) just hit #1 in the US App Store after announcing their new V3 model

• The model is a huge deal because it performs on par or better than anything we’ve ever seen, and DeepSeek trained it on a shoestring $5.6M budget (like 45x cheaper than OpenAI, Google, etc.). Didn’t seem possible until now.

• Nvidia stock fell 17% on the news, wiping out $600B, the biggest 1-day drop in market HISTORY. Nvidia sells GPUs, used to train AI models. 45x more efficient models means a lot less GPUs need for training than the market thought

• The question: how did DeepSeek do it? It kinda looks like they just reverse-engineered ChatGPT with some breakthrough cost-cutting training innovations along the way (more on that below).

• Now OpenAI and Microsoft are accusing DeepSeek of violating their terms of service and stealing their IP. Seems the US Government may get involved. It all raises some questions about AI ethics / data ownership.

• Not to mention Alibaba just released Qwen 2.5 Max AI, it’s most powerful model ever. It can write code, search the web, generate amazing video (for free)

Long-story short, Chinese company managed to create the world’s best AI model for 50x less money by copycatting the work OpenAI and friends. What was once believed to be a vast competitive advantage for US companies has turned out to be a commodity.

I think many could’ve seen this coming — China is famous for taking US-originated innovation and going hard on it until they achieve superiority (same happened with electric vehicles, mobile payment, and 5G cellular).

I just don’t think anyone imagined it happening so fast, all at once like this. My question is, what does this mean for what we thought about the AI landscape in the first place?

How DeepSeek Made 95% Cheaper AI

First — okay seriously how tf did they do it?

It’s actually pretty interesting. A few really smart folks have written about this at length, I am not claiming to be one of them. I think Jared Friedman’s summarization of Jeffrey Emanuel’s interpretation sums up the 3 or 4 technological innovations DeepSeek made on the training side that enabled them to achieve a 45x cost reduction. From Jared’s Tweet:

• [DeepSeek] trained on 8bit instead of 32bit floating point numbers, which gives massive memory savings

• Compress the key-value indices which eat up much of the VRAM; they get 93% compression ratios

• Do multi-token prediction instead of single-token prediction which effectively doubles inference speed

• Mixture of Experts model decomposes a big model into small models that can run on consumer-grade GPUs

Each of these are pretty massive breakthroughs, but wouldn’t have been possible without the work of US companies to lay the groundwork. Essentially they took the rubric that OpenAI / Anthropic / Meta / Google had outlined for training AI models that scored at PhD levels across various performance benchmarks, and reverse engineered it with a few massive breakthroughs in efficiency. Another excellent overview here.

I suppose like any feat that was once assumed impossible (sending rockets to orbit, running a sub-4 minutes mile, creating AGI), once you know the destination is reachable you can begin to explore shortcuts. The Bannister Effect.

When AI Models Are a Commodity

Once you realize that A) China now rivals the US in LLM performance and B) They actually found out how to do it for 95% cheaper in short time, the big hairy question is — What happens now? Here’s my 2-cents:

First, over the short term it shows how fragile the art of differentiated innovation is becoming as AI advances.

UI became a commodity over the past 18-months as AI tools like Cursor and Lovable made it possible for anyone to design the perfect, custom website interface without writing any code. I wrote about this at length here, about how AI has killed SaaS. I’ve also proven this point here in a series of videos about building a Coleslaw Maps website by having AI code it for me.

Now it appears that the underlying AI models themselves are commodities. Which was kind of unthinkable two or three years ago. So basically AI has made both UI and the models themselves commodities. So what’s left?

In theory, The Data. Salesforce’s Marc Benioff says it perfectly:

I mostly agree with Marc. Every few months it feels like another piece of the puzzle for web-based startups gets nerfed by ChatGPT and friends (and now China). Where a great user interface or a robust model used to be a competitive advantage, now there will be none.

When I play it out into the future I anticipate that only a few things will be left in 5-10 years: AI-native companies, web/AI infrastructure companies, service companies, and data companies. All else will be forced into one of those buckets or completely extirpated.

For me this is why I’ve pivoted to going all-in on a data company over the past 6 months. I’ve written a lot about that in a few previous blog posts, happy to write more if you’re curious. The hypothesis continues to be proven correct.

Second, over the long-term this very well could mark the start of a bonafide AGI Cold War between China and the USA. It’s hard to imagine either side stopping until society either achieves true Artificial General Intelligence, or we reach a technological limit.

We may realize that it’s impossible to achieve an artificial super intelligence no matter how much compute we have or how good the models are. Finding enough data may be the real limiting factor. Part of me hopes that is all true.

Or, we may unlock some portal as a society into a realm beyond our imagination where machines are truly smarter than human comprehension. A singularity event from which there’s no going back.

Anyways, just some thoughts, don’t mean to be doom-and-gloom (though it is raining here in Chicago). Genuinely just fascinated by all this and trying to predict where it leads us. Would love to hear your thoughts.

Happy Friday,
Peace,
Ramsey