DeepFreak

We are not unbiased.

We are large-scale consumers of AI inference and are not in the business of training large models. We are neither long nor short the stock of any individual company in the AI or semiconductor space₁, but we do pay OpenAI and Anthropic lots of money to perform AI inference for us. Back of envelope, the open source version of DeepSeek-R1 that has been plastered across the news will allow us to devote a few months of our current API spend as capex to replicate a year’s worth of productivity at an often higher level of fidelity with significant capacity left for any number of other productivity tasks. It’s a game-changer and we are very excited about it.

We are also Americans. We like it when Americans win. We also rather enjoy it when the commies lose. Sorry not sorry.

But we cannot help it. We are students of Narrative. And the narrative battle around the emergence of DeepSeek R1 is one of the most compelling watches in recent memory.

In case you are not involved in financial markets, do not read the news at all, and avoid social networks like the plague₂, DeepSeek-R1 is a new large language model (LLM). Like the others you probably know, such as OpenAI’s ChatGPT and Anthropic’s Claude, DeepSeek-R1 can be accessed via a web-based chatbot or through an API. Unlike those competitors, DeepSeek-R1 is “open-source₃“. It can also be accessed at varying levels of distillation and quantization₄ and with mixture of experts methodologies that allows individual inference users (i.e. asking a language model questions) to tailor the accuracy, quality, size, and speed of the model to their available hardware. It also reportedly cost a mere fraction of the resources being devoted by the other big AI players to training bigger, newer, more powerful models and works just about as well as all of them. Y’know, so long as your inference tasks aren’t excessively interested in Tiananmen Square, the sovereignty of Taiwan, or the Great Leap Forward.

When I say that it (reportedly) cost a fraction of the resources, what that really means is that it (reportedly) did not require the developer to acquire nearly as many ultra-powerful GPUs to train it as comparable models. And by “a fraction of the resources” I mean something like 3% (reportedly). While LLMs represent a narrower field of human activity than most, there are not many historical analogs to such a change in production costs in one product generation. We might grant the Bessemer process credit for something like a 70-80% reduction in the cost of producing steel back in the 1850s. A few decades later, the Hall-Héroult process transformed aluminum from precious metal to industrial, contributing to a greater than 97% price reduction over a relatively brief period. Unlike those innovations, what allowed DeepSeek to (reportedly) achieve these efficiencies will probably not convey to the development of all future LLMs. But for the time being, it is a Big Deal.

There is no company for which it is a bigger deal than Nvidia.

Sure, OpenAI, Anthropic, and the handful of megacap technology companies who have poured investor capital into similar products have a new competitor happy to undercut them on pricing. But the entire story of Nvidia is the Common Knowledge that the world would have a nearly limitless appetite for GPUs. If your stock trades at 30 times sales because everybody knows that everybody knows that billions will continue to finance massive data centers with hundreds of thousands of ungodly expensive Nvidia H100 GPUs, the release of DeepSeek-R1 was a narrative-breaking event.

Maybe.

The thing about GPUs is that they are needed (or at leastvery useful) not only for the massive-scale training of LLMs, but for so-called inference tasks. Inference is when you ask Claude if it makes sense to add a little baking powder to the dredge you use to batter fried chicken, or to help you understand the math question your 7th grader just asked so you don’t look like a dope. It is also when companies ask language models to tell them things about their massive internal datasets, among a million other use cases. Is it possible that DeepSeek-R1 and further developments made possible by its methods will accelerate the adoption, use, and implementation of large-scale inference tasks? Is it possible that the narrative of Infinite GPUs for Mass-Scale Training simply transforms into the narrative of Infinite GPUs for Mass-Scale Inference?

The narrative possibilities don’t end there. Remember, DeepSeek-R1 is also a Chinese product. In reality world, that means that most people will access it through non-open-source versions that are probably (i.e. definitely) sending all sorts of prompt data back to a data center in Hangzhou. It also means that even most local installs₅ will include some measure of the censorious predispositions of Beijing instead of the censorious predispositions of Silicon Valley. I mean the sensibilities it had before Big Tech flipped on a dime for the new boss in DC, of course. But those are reality world implications; their echoes in narrative space are at once much larger and more multi-faceted. Could DeepSeek have been lying about how easily they trained R1 to spook US markets and steer its competitors in the wrong direction? Could all of this be a psy-op designed to give China an edge as the world’s AI leader? Could it be a CCP ploy to extract data from private citizens in the west while our lawmakers are focused on finding the most politically connected tech oligarch to get a sweetheart deal for TikTok.

In all, I think there are at least eight distinct narratives being used today to frame the DeepSeek release.

All but one of these narratives have been present since the first day of material DeepSeek-R1 coverage.

Each of these narratives has a distinct and measurable semantic signature – our term for the linguistic characteristics which convey a particular framing of a news event. If you want to know more about what we mean by semantic signatures, this essay from Ben is a good place to start. Each such signature represents our measurement of attempts to establish a narrative – of efforts to direct the reader to a certain interpretation of what DeepSeek-R1 is really about:

The Eight DeepSeek Narratives

• DeepSeek-R1 is a game-changer for AI users
• DeepSeek-R1 is a threat to the business models of AI competitors
• DeekSeek-R1 is a threat to the business models of semiconductor manufacturers
• DeepSeek-R1 is forcing AI competitors to discount and bundle more services
• DeepSeek-R1 is a threat to the AI-driven boom in US mega-cap stocks
• DeepSeek is lying about the ease with which it was trained
• DeepSeek-R1 is a national security threat
• DeepSeek-R1 should be banned

So how has the structure of these narratives evolved? And how is it still evolving? More to the point, if you had constructed a daily dataset of 25 million or so news articles, high volume blogs and Substacks, press releases, transcripts, and other unstructured text content, what would you want to know in order to assess the evolution of the structure of narratives about DeepSeek?