DeepSeek's "Ingenuous" Marketing Strategy
What is DeepSeek's strategy, and how everything might play out
On January 20th, DeepSeek made a splash in AI news by releasing an “open source” reasoning model called R1 that’s about 27 times cheaper than OpenAI’s O1 (when used via API - see the pricing comparison in the next section).
In this post, we will analyze DeepSeek’s “shock and awe” price leader strategy and the competitive ramifications, while debunking some common tech blogosphere narratives about DeepSeek’s models.
Also, there are inescapable parallels between DeepSeek and TikTok, not only due to their connections to China, but their mastery of marketing and serious technical prowess. Similarly, DeepSeek’s success will inevitably flare up national security concerns, especially as its models integrate deeply into Western consumer and enterprise apps. I end with some predictions about how existing players (OpenAI, Meta, etc) may respond to DeepSeek’s challenge.
In a Nutshell / TLDR
Evidently, DeepSeek’s short-term strategy is to rapidly capture a vocal developer (and startup ecosystem) fanbase as a dirt-cheap alternative to OpenAI and a better model versus Llama.
Leveraging a vocal fan base and an “us versus them” narrative (in this case, “open source” versus “closed source” model) to enter the U.S. market has been a brilliant move by DeepSeek. But it also echoes TikTok’s PR playbook in that the company recruits grassroots support to grow mindshare. I foresee Mark Zuckerberg and Meta to fast follow by releasing a state-of-the-art reasoning model, lest it loses the narrative as the leading open source LLM provider.
I use the term “open source” here with a big caveat, for reasons I will explain in the coming section (hint: strictly speaking, the models are “open weight”).
But DeepSeek’s real prize is to establish a direct-to-consumer presence in the U.S. (with a ChatGPT like app, or by powering other B2C apps), via leveraging startup ecosystem traction. This also pressures the U.S. tech industry (e.g. AWS, Google Cloud) to incorporate DeepSeek into its platforms (Bedrock, Vertex, etc), establishing a beachhead in the U.S. enterprise - something Alibaba Cloud couldn’t manage. Presence in both consumer and enterprise markets gives DeepSeek more variety in LLM training data, which it (and China by extension) desperately needs.
This aggressive price leader strategy, which feels similar to BYD or Temu’s go-to-market strategy, forces OpenAI and Anthropic to accelerate their product releases, and reveal their “hand” sooner, while pressuring Meta to spend more to defend its open source leadership. A tertiary effect will be that OpenAI and Anthropic will remain private for longer, and require further help from our own public sector to allow price matching to DeepSeek.
All DeepSeek needs to win market share is to match OpenAI’s performance via “fast-follow”, and convince companies it can do so. If this strategy works, we will eventually have the TikTok debate but with DeepSeek, which should catch enterprise AI space off-guard.
On one hand, it’s great that OpenAI and Anthropic have real competition, which fosters innovation outside of the Big 2. On the other hand, the nature of the competition and potentially distortive pricing strategy can be problematic, and cause an allergic reaction for some akin to that re: TikTok. Now, let’s dive a bit deeper into DeepSeek and its strategy along with national security ramifications.
What is DeepSeek and DeepSeek’s Strategy
DeepSeek is an AI research lab founded in 2023 by a China-based quantitative hedge fund called High-Flyer. Since the release of ChatGPT, many Chinese labs and companies played catch-up to the SOTA (state of the art) set by OpenAI, but in late 2024 two Chinese labs (DeepSeek and Alibaba’s QWEN group) finally managed to release models close OpenAI and Anthropic models, at least on benchmark performance (which always should be taken with a grain of salt).
The cornerstone of DeepSeek (and Qwen)’s strategy is that they position themselves as “open source model providers”, and pursue aggressive growth strategy via selling the API access far below cost (most likely), which also takes a page out of Silicon Valley’s blitz-scaling playbook:
Since OpenAI and Anthropic got a lot of flak by developers and venture capital community for being “closed source”, this allows the Chinese labs to convert anyone on the fence about using proprietary model APIs, and the cost-sensitive startups. It also allows ecosystems and applications around DeepSeek to grow quicker, especially around use cases forbidden by OpenAI (e.g. NSFW content).
Given the API version of DeepSeek R1 is 27x cheaper than OpenAI’s O1, we are looking at Restoration Hardware versus Temu levels of price gap. Except - the performance on LLMs are much more objective than merchandise, and most likely indistinguishable on most tasks. This is a strong incentive for any company to at least test out DeepSeek to instantly make more money.
Is it Truly Open Source, Though?
That said, DeepSeek’s use of “open source” is a misnomer in the strictest sense, given that they didn’t open source its training code nor training data, but just the model weights. DeepSeek did release a paper describing the training process, though the training code itself wasn’t open sourced, either. The paper itself was narrative heavy and short.
This prevents anyone from fully understanding how the model got created at a deep level, or modifying the model significantly. This is understandable given DeepSeek’s models have significant pro-China censorship baked in them, and presumably they won’t appreciate anyone “uncensoring” these models, or reverse engineer the censorship process.
To be fair, this fuzzy notion of “open source” has been used by American and European companies as well (e.g. Llama, Falcon, etc), and they have stricter EULAs (end user license agreements) with restrictions on who gets to use, etc. It has been an exception, rather than the rule, that “open source” models release their training data or training code.
To give DeepSeek credit, the model weights were released under an unrestricted commercial license (MIT), which also means companies can use it without having to pay royalties beyond a certain usage. Llama and Falcon, etc, came with some restrictions. That said, the MIT license itself doesn’t mean anything for 99%+ of startups since Llama(2), etc, only restricted commercial use if you had hundreds of millions of users.
Another nuance is that most companies and developers do actually not self-host open source LLMs, but use services like together.ai to consume them. Thus, practically speaking, most developers consume open source models in the same fashion as they consume OpenAI or Anthropic models - via third party APIs. Also, most large models like DeepSeek R1 are too tricky to self-host technically speaking, and beyond most developers’ scope of responsibilities. Where open source actually means something is when you need extensive model finetuning, which something most people don’t do.
What does matter - in practice - is price and performance, since switching from OpenAI to DeepSeek can now slash LLM bills (which is a sizable COGs item) by 90%+. While DeepSeek’s hosted API version will use your data for their training purposes (and be sent to China most likely), there are no regulations forbidding that for U.S. companies, which means many companies will switch over.
Thus, all in all, “open source” is more of a marketing strategy, if we realize that most companies will be using DeepSeek APIs just like they are using any closed source LLM API (outside of LLMs used on personal devices e.g. Ollama or LM Studio).
The Bigger Picture - DeepSeek’s Deeper Strategy
But there’s an even bigger picture to DeepSeek’s go to market strategy. As a disclaimer, these forecasts are not based on any insider knowledge or conversations with DeepSeek, and we are just speculating from DeepSeek’s public actions and the typical trajectory of Chinese startups selling to U.S. consumers (Tiktok, Shein, etc). The “big picture” is probably this:
Own the “open source” framing: essentially, take a page out of Meta’s Llama playbook to create an “us versus them” against OpenAI and Anthropic, allowing them to capture developer mindshare, especially those who prefer to self-host LLMs. Since reasoning models didn’t have a strong “open source” version, DeepSeek effectively supplanted Llama in this category. Meta is at risk of losing the narrative on open source LLM leadership.
Gain global developer good will by offering APIs at low cost: many startups are always looking to increase margins, and have little loyalty to one provider assuming performance won’t suffer, regardless of the labs’ nationality. In some sense, DeepSeek is footing startups’ LLM bill, and many will love that. Expect many startups to do POCs soon.
Paint a positive image of DeepSeek (and Chinese Labs) as the savior to “closed source AI”: This is an underrated benefit of the “shock and awe” arrival of DeepSeek - even though it’s not fully open source, it is now seen as the “open source” + “good performance” option. This puts a halo effect around models from Chinese labs. Also, it puts OpenAI and Anthropic in an awkward position where they must demonstrate and maintain a much larger performance advantage over DeepSeek to convince customers to stay. OpenAI will now be forced to release MORE stuff, MORE frequently, even when it wants to slow play its hand.
Gain market share in the consumer and developer markets: once DeepSeek takes control of the Internet narrative, then it can drive more traffic to the DeepSeek chatbot apps, which has been squarely OpenAI’s turf. Increased awareness allows DeepSeek to directly serve U.S. consumers, like Tiktok, while gaining valuable data. Also, DeepSeek APIs will be demanded inside coding IDEs like Cursor, which opens up access to more tokens in coding realm.
Gain entry into the U.S. enterprise market: As developer momentum builds, within 3 months - 6 months time, Google and AWS will need to decide whether to make DeepSeek available on Vertex AI or Bedrock, which opens up the doors to enterprise SaaS.
If these steps were to actualize, then DeepSeek will have an enormous ecosystem of deployments, all serving DeepSeek models (obviously). That’s another wedge for China in defeating the U.S. in the AI reace.
Also, DeepSeek does not need to be better than OpenAI for this strategy to work! It clearly can offer GPT-4o level performance at dirt cheap prices, and that’s good enough for most AI applications (presumably with state subsidies, breakthroughs in inference technology, or both). OpenAI can hold on the reasoning model leadership, but only a small % of AI workloads require the best LLM currently (but expected to grow).
The National Security Angle
It is not a secret that DeepSeek’s API and hosted version may use your data for training purposes (here’s their privacy policy, where they clearly state servers are located in China).
Since the lab is located in China, you can assume your entire prompts, PII data, API tokens, codebase, and other proprietary data will go to China as well, especially when interacting with the DeepSeek app. Of course, developers can self host the LLMs, but DeepSeek’s largest and best models are difficult to self-host.
If DeepSeek is successful at growing in the U.S. (and western market), it will have acquired a source of training data that it controls, as opposed to relying on scraping or purchasing.
From consumer apps, it will gain conversational data, work task data, access to documents (PDF Files, Word Files, etc) that are attached to threads, images, etc.
From coding apps, it will gain access to codebases, configuration files, developer documentation, etc.
From enterprise apps, it will gain access to proprietary workflow data, prompts, internal policies, any data that is RAG’ed into the context, etc.
Needless to say, this happens also when you are using any AI app, U.S. or not. Not all apps provide opt-out, either.
Gaining this data is very important for any AI research lab, especially since ChatGPT has a chokehold on the consumer AI market (which I don’t see it relinquishing for a while, if not ever). More data that’s inaccessible to other AI labs is crucial for the company to maintain competitiveness.
Thus, as more users flock to DeepSeek’s consumer apps, and more applications use DeepSeek APIs, more data will be at risk - except this data has more value than the data that’s been leaking through TikTok.
Given that security experts and the Supreme Court unanimously agree that TikTok leaking data is a threat, it’s hard to view DeepSeek not becoming a security problem especially if it becomes successful. What’d be interesting is whether and how developers would react if there’s any talks of a DeepSeek ban, since it’s a different demographic from the average TikTok consumer.
Note, many people don’t care if their data is sent to LLMs for training purposes, or are simply unaware of that fact. American research labs also train on your data, but at least you can opt out in some cases. Thus, again, we have the Meta versus Tiktok debate - if OpenAI trains on your data, why can’t DeepSeek do the same? How you feel about this is mainly a function of your views on the importance of U.S. national security.
DeepSeek’s Future
When Llama2 was first released in mid 2023, the AI community got the first viable alternative to GPT3.5, and that was a huge deal from “hype” perspective. Over time, however, it turned out OpenAI was just fine and kept increasing market share, while Llama class of models (and its variants) were embraced modestly at best. The majority of developers elected not to self-host, but use some serverless API version of Llama (through together.ai, etc).
But in the last 1.5 years, a lot has happened, and companies have already deployed many AI apps. Adoption will be quicker than that of Llama’s, barring any regulatory or compliance related guidance. Companies are looking to increase margins, so I expect many companies to use DeepSeek in mostly hosted mode. They will be expecting their favorite cloud provider to price match DeepSeek, which puts AWS and GCP in awkward positions as well. Those who elect to self-host DeepSeek will mostly be those who are already self-hosting. I expecting few companies taking on self-hosting just for the sake of using DeepSeek.
That being said, end users also have a right to know what models apps are using. DeepSeek or not, companies should be clear about what models are being used and where their data is going. Ideally this information shouldn’t hide in privacy policies, but be exposed in the user interface like how ChatGPT does it.