LLMs Are Over

Hey y’all — we just had the most significant shift in AI since the launch of ChatGPT.

That sounds like hyperbole because you hear it from big AI newsletters every time a new foundation model comes out. The difference is that, this time, it’s actually true.

The last two years have been defined by companies with proprietary LLMs. That’s where value has accrued. They were so far ahead that everyone else flocked to them.

That world is now gone.

No, we haven’t taken it out back and buried it. But there’s the first undeniable sign that LLMs, while incredibly disruptive, aren’t immune to traditional market forces.

And value is going to accrue elsewhere.

We actually have a bit of a roadmap for what will happen next.

Let me explain…

LLMs Are Over

ICYMI

Quick summary (skip ahead if you’ve been following the news):

• Last weekend a Chinese company called DeepSeek released an open source foundation model that nearly matches OpenAI’s best model on most metrics

• They did it for, allegedly, only ~$5 million (ironically US sanctions on NVIDIA’s hardware being sold into China created an environment where Chinese companies were forced to massively innovate cost reduction in order to compete)

• As a result, using DeepSeek is 100x cheaper than using OpenAI

• It caused a massive stock selloff in the US of NVIDIA and other companies

• There are unconfirmed reports floating around that say DeepSeek’s models were explicitly trained on outputs from OpenAI’s models

So What?

You may remember last year when Sam Altman was in India and said, in response to an audience question, that it was pointless to compete with OpenAI on building foundation models:

At the moment, that looks a bit silly.

Go back to the launch of ChatGPT for a second.

It felt like such a large step forward in what was possible that everyone sort of collectively lost their minds.

This was going to be a new world, that we were all curious about, and that would rewrite the entire rule book.

We pretended to forget that value accrues at the data and application layers more than at the infrastructure layer.

So while it may be a brave new world for software within markets, it isn’t true for the markets themselves.

An Example

In the 90s Oracle, IBM, and Microsoft charged premium prices for relational database management systems.

Eventually, free open-source alternatives like PostgreSQL and MySQL became viable competitors and completely commoditized the market.

As the market evolved, the cost of running software built on top of relational databases went to near-zero and value accrued where differentiation remained — which was in the data itself (the data layer) and how that data was used (the application layer) rather than the database itself (the infra layer).

This is exactly what’s happening again today.

OpenAI, Anthopic, Google, and others are trying to keep their foundational models proprietary — they’re fighting a battle they’re going to lose.

Zuck saw this coming and didn’t even bother building a business around Llama in the same way — he had Meta release it as open sourced tech.

DeepSeek is simply the new kid on the block taking it once step further. In a few months, it’ll be someone else.

What Happens Next

We’re about to see the real winners of the AI boom start to emerge.

In the databases example, the winners were companies like Google and Salesforce.

Even though Google search runs on a relational database, it’s been extremely valuable for 25 years because it has the best index of the interest (data) and it presents and leverages that data in valuable ways (applications).

Don’t get me wrong — OpenAI will almost certainly be just fine (Microsoft made it out of the database era just fine, after all). The fact that ChatGPT (their application layer product) loses money per subscriber right now isn’t great, but they have other paths.

In fact, the news this week (post-DeepSeek) that they’re raising $40 billion and the recent news that they’re investing $19 billion into a physical infrastructure project called Stargate are, I believe, directly related.

Despite the clip I shared at the start of this issue, Sam is smart and had to have known this was coming. So, he’s going towards the areas that are more defensible.

But any company just building foundation models that don’t have a strong product, distribution partnership, or physical infrastructure play may run into issues no matter how large they are.

How to Win

I see three main ways to win going forward:

1. Own the customer

2. Own proprietary data

3. Sell physical infrastructure or access to it

You’ll notice that in all three of those categories, big companies have a huge advantage over startups.

The startups that become resilient unicorns will need to find unique advantages in one of those three areas and move incredibly quickly.

Since setting up physical infrastructure is very expensive, and unique datasets will likely become harder to build, the easiest path to a unicorn is what it’s always been — make something people want.

How We Can Help

Advertise in my newsletter to get in front of 70,000+ founders.