Is the First-Mover Advantage Over for AI?

DeepSeek is offering a pricing and investment advantage previously unseen in the AI market. Here's a closer look at the model and the current landscape of AI.

The Chinese model, DeepSeek, stormed onto the world stage in late January,  claiming a significantly lower cost than its rival, OpenAI’s o1 model. At an estimated 1/20^th the cost of OpenAI’s model, it demonstrates a significant leap in the accessibility of AI, taking it from a medium-to-enterprise product to one that many more businesses will look at seriously.

DeepSeek also claims that it can use less memory with a caching trick, making its day-to-day runtime easier in addition to its lower cost of entry. These factors combine to have DeepSeek Chat (their lower-tier offering) model compete head-on with Gemini Flash and OpenAI’s GPT-4o mini:

More importantly, its Reasoning R1 model competes with Claude 3 Haiku, a model designed to perform like GPT-4o mini or Gemini Flash. In essence, DeepSeek offers a more advanced reasoning model for the price of a Chat model from other vendors.

This novel positioning provides DeepSeek with a pricing and investment advantage previously unseen in this market: serious price competition.

DeepSeek’s Challenges

DeepSeek’s location in China makes it vulnerable to regulatory action, as concerns about its privacy mount, especially given its obligation to share data with the Chinese government upon request. This has renewed the discussion around the banning of Chinese firms that collect U.S. data and send it outside the U.S. (in this case, to China), reminiscent of the recent discussion around TikTok. Similar actions that target DeepSeek have been proposed in the U.S. Senate. It has been banned in Italy and is under investigation in France and Ireland.

DeepSeek also faces stiff competition from inside China, with rival Tencent developing an AI model which claims similar performance but at higher speeds. With the revelation that high-end chips banned for export to China are not necessary for performant Generative AI, more competition should be expected.

A Brave New World: The New AI Market

Note: This section contains educated guesses based on typical market behavior and recent historical regulatory actions, and is speculative in nature. For investment or regulatory advice, consult financial or legal experts. All opinions are solely those of the author.

First-mover OpenAI and its cluster of rivals, such as Anthropic, invested heavily in the development of these models, often placing heavy investments into very large hardware clusters, including heavy investment into GPUs (OpenAI alone has over 20,000). DeepSeek represents the first competitor to successfully innovate a method that is 1) cheaper and 2) not dependent on the cutting-edge chips from NVIDIA that are banned for export to China (this point is contested).

Given that this method is being used by DeepSeek’s rivals, it seems likely that a market shift toward cheaper, more affordable models has begun. Even if a particular model is banned, other competitors will have been paying attention and most likely will develop their own cheaper LLM, let alone the possibility of cut-out corporations for existing players. One way or another, it seems the genie is out of the bottle. Even without circumvention, it seems unlikely that the proliferation of cheaper models will be stymied so easily.

This should not be taken to mean that the existing cutting-edge players are out; they will likely release their own optimized models to compete on the same field as DeepSeek and its cohort. However, it is hard to imagine that they would abandon their head start in research into more advanced models. That said, this split focus may slow them down as they divert resources to such defensive development, so it would not be strange to see the breakneck pace of reasoning model releases slow somewhat.

Overall, these market changes will likely develop along three core precepts:

1. The increasing accessibility of these models will likely drive adoption as the value proposition improves for companies and users previously shut out due to cost or uncertainty about it.
2. Some regulatory turmoil may occur over individual models or countries, but these are unlikely to significantly slow the market shift since this type of work has now been demonstrated as possible.
3. Cutting-edge model releases may slow somewhat as core vendors develop models to compete with DeepSeek and its followers, but are unlikely to stop.

Getting the Most for Your Money: Living in the Maturing AI Market

As price competition heats up, enterprises would be well-served to remain flexible. Avoiding vendor lock-in, which typically has some benefits, will provide additional competitive advantages for firms as model pricing fluctuates and new players emerge on the scene. Incorporating model-agnostic architecture remains key to this success and plays a crucial role in your IT strategy going forward.

Developing a model-agnostic architecture will revolve around three pillars:

1. ‍An up-to-date tracking of available models: To take advantage of per-token pricing, loss-leader positioning, and other market shifts, a periodic review of the market will provide the ability to detect and take advantage of these opportunities.‍
2. A robust A/B testing framework: While saving some money on models is nice, if a model is hurting your KPIs, chasing the cheapest model is counterproductive. Using best-in-class analytics and experimentation helps you ensure that the latest model is also the greatest model.‍
3. A model capture software development kit (SDK): Where the rubber meets the road for your applications, a uniform format and set of code standards will allow your enterprise applications to seamlessly swap and test models with minimal overhead.

Are you looking to make sure your enterprise is model-agnostic? Contact Concord!

‍