How does DeepSeek compare to ChatGPT and what does its GRPO methodology mean for AI competitive dynamics?
DeepSeek achieves competitive performance at lower training cost via GRPO, accelerating AI capability commoditization and shifting competitive advantage toward proprietary data and fine-tuning.
- DeepSeek achieved GPT-4 class performance at a fraction of the training cost, challenging assumptions about AI infrastructure investment requirements - Group Relative Policy Optimization replaces traditional RLHF with a more computationally efficient reinforcement learning approach - Falling model training costs accelerate the commoditization of AI capabilities at the foundation model layer - Durable competitive advantage in AI is shifting toward proprietary data, workflow integration, and domain-specific fine-tuning - Technology investors must reevaluate AI infrastructure plays in light of DeepSeek's capital efficiency benchmark
The rapid evolution of Large Language Models has revolutionized artificial intelligence, enabling machines to perform tasks once thought to be the exclusive domain of human intelligence. Among the many LLMs that have emerged, DeepSeek and ChatGPT stand out as two of the most advanced. While both models generate human-like text and perform complex reasoning tasks, they differ significantly in their underlying architectures and training methodologies.
ChatGPT: The Established Standard
ChatGPT, developed by OpenAI, is built on the GPT architecture and trained using Reinforcement Learning from Human Feedback (RLHF). This approach uses human evaluators to rank model outputs, training the model to maximize human preference scores. The result is a model that excels at natural conversation, creative writing, and general-purpose reasoning. The training process requires substantial computational resources and human annotation at scale.
DeepSeek: A Different Approach
DeepSeek, developed by a Chinese AI research team, differentiates itself through its use of Group Relative Policy Optimization (GRPO). Unlike RLHF, which relies on human preference rankings, GRPO trains the model by generating multiple candidate responses to the same prompt and using the relative quality differences within that group as the training signal. This approach reduces dependence on expensive human annotation while maintaining competitive performance on reasoning and coding benchmarks.
Group Relative Policy Optimization Explained
GRPO is a reinforcement learning technique that evaluates model outputs relative to each other rather than against an absolute human standard. For a given prompt, the model generates a group of candidate responses. These responses are ranked by a reward function, and the model is trained to increase the probability of higher-ranked responses. The key advantage is efficiency: by using the model's own outputs as the training signal, GRPO reduces the annotation bottleneck that constrains RLHF-based training.
Investment and Valuation Implications
DeepSeek's emergence raises important questions about the economics of frontier AI development. If competitive performance can be achieved at significantly lower training cost through techniques like GRPO, the capital intensity of AI development may be lower than current frontier lab valuations imply. This challenges the premise that scale and capital alone determine competitive position in AI, and suggests that architectural innovation remains a significant driver of competitive advantage.
DeepSeek's emergence as a competitive frontier LLM developed at a fraction of OpenAI's training cost represents the most significant challenge to US AI incumbency since the GPT era began. The underlying methodology, Group Relative Policy Optimization, demonstrates that reinforcement learning from human feedback can be replaced by a more computationally efficient alternative without material performance degradation. For technology investors and acquirers, this signals that AI model training costs will continue to fall and that the durable competitive moat in AI is shifting from compute infrastructure to proprietary data, workflow integration, and domain-specific fine-tuning.
When to speak with Chatsworth
You may benefit from an advisory conversation if your board is evaluating timing, valuation expectations, buyer universe quality, or diligence readiness. Chatsworth provides senior-led perspective on process design and execution risk independently of whether a mandate results.
Speak with the team →Marcus Magarian, Managing Director. M&A advisory, strategic advisory, cross-border transactions, and technology-sector advisory. chatsworthgroup.com/our-team/marcus-magarian
LinkedIn Profile →Strategic perspectives on M&A, capital formation, and cross-border transactions, delivered directly.
Subscribe