Lecture 6 talked about the paradox of the application layer. One key judgment was: open-source models are flattening the moat of the application layer.
This piece puts that judgment into concrete terms — one look at the API price list is enough to understand how much pressure application-layer companies are under right now.
Going back to early 2024, GPT-4 was the undisputed king. API input was $30 per million tokens, output was $60. Every company building AI applications was calling GPT-4, and there was a generational gap between user experience and open-source models. At the time, even if your product was just a thin wrapper, as long as it was plugged into GPT-4, you could charge users money.
Two years later, things have changed.
Looking at current API prices, GPT-4o is $2.5 per million input tokens and $10 output. DeepSeek V3 is $0.27 input and $1.10 output. Qwen 3.5 is $0.39 input and $2.34 output. Mistral Small is $0.10 input. DeepSeek R1, benchmarked against reasoning models like OpenAI o1, costs only 5% of the latter.
On average, open-source models cost only one-tenth to one-twentieth as much as top closed-source models.
Price collapse is not the key point; comparable capability is.
DeepSeek V3 is already close to GPT-4o on several benchmarks. DeepSeek R1 is on par with OpenAI o1 in math and code. Llama and Qwen are both catching up to GPT-4 in general capability. This means that for the vast majority of application scenarios — customer service, writing, analysis, code completion — the output quality of open-source models is already “good enough.”
Good enough plus ten times cheaper is a terminal combination.
What does this mean for application-layer companies?
Their old business model was: call GPT-4, wrap it in an interface, and charge users a subscription fee. The spread was the profit. Now that open-source models have pushed down the underlying cost, the margin space in the application layer has been squeezed very thin. Users will say: your product is so expensive, but the backend may just be using an open-source model — why should I pay? Competitors will say: I’m using DeepSeek too, and I’ll charge you one-third the price for the same features. The moat is gone.
What’s even more troublesome is that the deployment barrier for open-source models is falling fast. Llama, Qwen, and DeepSeek can all be deployed locally, and large enterprises can pull an open-source model into their own intranet without even using an API. That completely cuts off the link where the application layer calls model APIs. Enterprise customers used to be the fattest piece of meat for the application layer; now they are increasingly inclined to build in-house.
The core judgment in the course — that the best companies in the application layer have not gone public, while the publicly listed companies in the application layer have very thin moats under the impact of open source — is being verified by reality layer by layer just like this.
It’s not that you can’t make money in the application layer; it’s that the secondary market can hardly buy the batch that can truly make money. The ones that can be bought are, for the most part, being squeezed from both sides by open-source models and by the “application-ization” of large model companies.
Looked at from the other side, the victory of open-source models is precisely the victory of the infrastructure layer. DeepSeek training needs GPUs, Llama deployment needs GPUs, and Qwen inference needs GPUs. The more widely open-source models spread, the greater the demand for inference compute becomes, and that inference compute ultimately still has to flow down to the chain of Nvidia, HBM, cloud, and power.
That is why “bullish on open source” and “bullish on infrastructure” are really just two sides of the same judgment.