The DeepSeek lesson

During the past two weeks, DeepSeek unraveled Silicon Valley’s comfortable narrative about generative artificial intelligence by introducing dramatically more efficient ways to scale large language models (LLMs). Without billions in venture capital to spend on Nvidia GPUs, the DeepSeek team had to be more resourceful and learned how to “activate only the most relevant portions of their model for each query,” as Reflexivity president Giuseppe Sette notes.

It didn’t take long for everyone to start interpreting DeepSeek’s feat through the lens of their own biases. Closed-model vendors cried foul over theft of training data (given how much of their own training data was lifted from others, the irony police were out in full force), while open sourcerors saw DeepSeek as a natural fulfillment of open source superiority (despite the fact that there is no correlation between being open and winning in tech).

Lost in all this confirmation bias were two big developments, one positive and the other quite negative. First, AI need no longer be dominated by a billionaire’s club. DeepSeek didn’t democratize AI, exactly, but it has shown that AI entry costs needn’t require seed rounds in the hundreds of billions. Second, although there’s no reason to think open approaches to AI will win, there’s every reason to think that OpenAI’s hyper-closed approach will most definitely lose because it’s customer-unobsessed. Winning in AI won’t be about open versus closed, but rather about customer trust.

‘Techno-feudalism on steroids’

I don’t have anything to add to the financial implications of DeepSeek’s approach. As DeepLearning.AI founder Andrew Ng points out, “LLM token prices have been falling rapidly, and open weights have contributed to this trend and given developers more choice.” DeepSeek, by optimizing how it handles compute and memory, takes this to the next level: “OpenAI’s o1 costs $60 per million output tokens; DeepSeek R1 costs $2.19.” As he concludes, the expectation is that “humanity [and developers] will use more intelligence…as it gets cheaper.”

But who will build the tools to access that AI-driven intelligence? Here’s where things get interesting.

Although it’s fun to eviscerate OpenAI and others for finger-pointing over stolen training data, given these LLM vendors’ propensity to “borrow” copious quantities of others’ data to train their own models, there’s something far more troubling at play. As Me & Qi cofounder Arnaud Bertrand argues, “The far more worrying aspect here is that OpenAI is suggesting that there are some cases in which they own the output of their model.” This is “techno-feudalism on steroids,” he warns: a world in which LLM owners can claim ownership of “every piece of content touched by AI.”

This isn’t open source versus closed source. Closed source software doesn’t try to take ownership of the data it touches. This is something more. OpenAI, for example, is clear(ish) that users own the outputs of their prompts, but that users can’t use those outputs to train a competing model. That would violate OpenAI’s terms and conditions. This isn’t really different from Meta’s Llama being open to use—unless you’re competing at scale.

And yet, it is different. OpenAI seems to be suggesting that its input (training) data should be open and unfettered, but the data others use (including data that competitive LLMs have recycled from OpenAI) can be closed. This is muddy, murky new ground, and it doesn’t bode well for adoption if enterprise customers have to worry—even a little bit—about their output data being owned by the model vendors. The heart of the issue is trust and customer control, not open source versus closed source.

Exacerbating enterprise mistrust

RedMonk cofounder Steve O’Grady nicely sums up enterprise concern with AI: “Enterprises recognize that to maximize the benefit from AI, they need to be able to grant access to their own internal data.” However, they’ve been “unwilling to do this at scale” because they don’t trust the LLM vendors with their data. OpenAI has exacerbated this mistrust. The vendors that will end up winning will be those that earn customers’ trust. Open source can help with this, but ultimately enterprises don’t care about the license; they care about how the vendor deals with their data. This is just one of the reasons AWS and Microsoft were first to build booming cloud businesses. Enterprises trusted them to take care of their sensitive data.

In this early rush for gold in the AI market, we’ve become so fixated on the foundational models that we’ve forgotten that the real, biggest market has yet to emerge, and trust will be central to winning it. Tim O’Reilly is, as ever, spot on when he calls out the “AI company leaders and their investors” for being “too fixed on the pursuit or preservation of monopoly power and the outsized returns that come with it.” They forget that “most great companies actually come after a period of experimentation and market expansion, not though lock-in at the beginning.” The AI companies are trying to optimize for profit too soon in the market’s evolution. Efforts to control model output will tend to constrain customer adoption, not expand it.

In sum, AI vendors that want to win need to think carefully about how they can establish trust in a market that has moved too quickly for enterprise buyers to feel secure. Grasping statements, such as OpenAI’s, about model output data don’t help.