For Financial Services AI, DeepSeek Changed Everything and Changed Nothing

What Just Happened?

The pace of change and events this last two weeks has been bewildering. I have below summarized what just happened with links to social posts with deeper insights, and also described how financial services need to react.

DeepSeek launched their LLM to great acclaim. So too did Alibaba with their new Qwen model. However, DeepSeek captured the headlines with how

• They navigated the maths – like a top dollar data scientist
• They delved deep into the hardware, including GPUs to improve compute efficiencies – like a top performing software engineer
• Ensured data management and transfer efficiency, like a top tier data engineer.
• Challenged the “proprietary” business logic, deliberately or otherwise, by open sourcing the model, cutting $billions from Silicon Valley tech stocks

Why DeepSeek? Perhaps because they were Quants, maybe with a Citadel or XTX-like structure, they could solve broad problems and dive deep into many areas of tech.

How has DeepSeek changed and not changed technical and commercial priorities, particularly in financial services?

First, agentic AI. Despite the MIT “Imagination in Action” session at Davos deeming “agents” a victim of the brilliantly descriptive term semantic bleaching, agents are genuine and useful. The bullets below, recreated from here summarize the AI Agent Technology Stack well, with DeepSeek disrupting the Foundation Models Layer.

• User Interface Layer, e.g. Streamlit, React
• Agent Orchestration Layer, e.g., LangGraph, Swarm
• Core Agent Logic Layer, e.g. LangChain, LlamaIndex
• Tool Integration Layer, e.g., Zapier, {Vectorize}, Hugging Face
• Foundation Models Layer, e.g., OpenAI, DeepSeek, Anthropic, Mistral, etc
• Infrastructure Layer, e.g., AWS, SingleStore, Nvidia, Pinecone

Yet the AI Agent stack is more than the LLM, with commercial value in turn deriving from much more than the AI agent in an “Application Layer.” One AI tech investor notes a ton of use cases where agents apply, but he says "the real winners will emerge at the application layer". I agree.

To achieve this, enterprises and organizations combine the foundation model and agent with business logic, as one investor posted on how execs from certain data-driven Decision Intelligence companies had foreseen the future driving business logic with AI, as have others. Such observations are not new, but fundamental to traditional AI going back to its foundations in the last century as well as contemporary GenAI.

Yet this in no way diminishes DeepSeek’s contribution. One Head of AI notes the journey to production remains hard, but the lesser computational costs of the LLM, which DeepSeek transforms, do help.

To conclude, try this heuristic:

Trusted AI

=

Trusted Data + Context + Model (e.g. LLM) + Compute.

As the last two weeks showed, DeepSeek solves well for Compute, and alongside Alibaba provides more LLM model choice to customers, but financial services firms wanting Trusted AI given their needs for transparency, hallucination mitigation, accuracy and compliance must leverage their organizational and enterprise data as trusted, contextualized Data. In other words, DeepSeek disrupts the compute and LLM layer, but it still needs to work with a strong data foundation and a "contextual fabric" to provide value to financial services organizations.

What About European AI?

For those who suggest  Europe has AI struggles given regulatory “barriers,” I find a lot to be positive about. The French Mistral AI, has similar model and open source approaches to DeepSeek. Don't forget too that French tech entrepreneurs are behind Hugging Face, where a senior lead lambasted a recent "protectionist" blog from Bay Area's Anthropic CEO. I too work for a strongly positioned  UK organization dedicated to aiding FinCrime, regulatory and other Decision Intelligence-oriented AI workflows, and until recently I worked for a currently Irish/UK domiciled "vector database" provider which targets the front office and capital markets. 

But yes, It has been a heck of a couple of weeks!