DeepSeek jolts AI industry: Why AI’s next leap may not come from more data, but more compute at inference

The AI ​​landscape continues to develop at a rapid pace, and recent developments have challenged established paradigms. In early 2025, DeepSeek, a Chinese AI lab, launched a new model that sent Shockwaves Through the AI ​​industry And leads to 17% Nvidia stocks fell, as well as Other stocks related to demand for AI data centers. This market reaction was reportedly stemming from DeepSeek’s ability to deliver high-performance models with a small percentage of U.S. competitor costs, triggering people’s attention Impact on AI data centers.

To relate the destruction of DeepSeek to context, we believe that a broader shift in AI landscape that takes into account the scarcity of other training data. Since major AI labs have now trained models on many public data available on the internet, data scarcity is Further improvements in slowing down pre-training. As a result, model providers are seeking “test time calculation” (TTC), where inference models (such as the “O” model series that open AI) “think” and then answer questions while inference as an alternative to improving overall model performance. The current idea is that TTC may show improvements with the scaling laws that have been driven by pre-training and potentially achieve the next step of transformative AI advancement.

These developments suggest two significant shifts: First, laboratories running on smaller (reported) budgets are now able to release state-of-the-art models. The focus of the second class is TTC as the next potential driving force for AI progress. Below, we unravel both trends and the potential impact on the competitive landscape and the broader AI market.

Impact on the AI ​​industry

We believe that the shift to TTC and the increased competition among inference models may have a broader impact on the Artificial Intelligence Landscape Cross hardware, cloud platform, basic models and enterprise software.

1. Hardware (GPU, dedicated chips and computing infrastructure)

• From large-scale training clusters to on-demand “test time” spikes: In our opinion, the shift to TTC may have an impact on the type of hardware resources required by AI companies and how to manage them. Instead of investing in growing GPU clusters, AI companies can increase investment in reasoning capabilities to support growing TTC demand. While AI companies may still need a large number of GPUs to handle inference workloads, the difference between Training workload Inference workloads may affect how these chips are configured and used. Specifically, because the inference workload tends to be more Dynamic (and “Spikey”)capacity planning may be more complex than batch-oriented training workload.
• The rise of inference optimization hardware: We believe that turning attention to TTC may increase the chances of alternative AI hardware specializing in low-latency inference time computing. For example, we may see more demands for GPU alternatives, such as application-specific integrated circuits (ASIC) Reasoning. As access to TTC is more important than training ability, General purpose GPU, For training and reasoning, it may decline. This shift can benefit professional reasoning chip providers.

2. Cloud Platform: High Standards (AWS, Azure, GCP) and Cloud Computing

• Quality of Service (QoS) becomes the key difference: In addition to involving model accuracy, one problem that can prevent enterprises from adopting AI is the unreliability of the reasoning API. Issues related to unreliable API inference include Response time fluctuations,,,,, Rate Limit and difficulties Process concurrent requests and Adapt to API endpoint changes. The increase in TTC may further exacerbate these problems. In this case, we believe that cloud providers that can provide QoS guarantees for these challenges have a great advantage.
• Despite increased efficiency, cloud spending has increased: Rather than reducing the need for AI hardware, it is better to adopt more efficient large language model (LLM) training and reasoning methods to follow Jevons Paradox, a historical observation that increasing efficiency can drive higher overall consumption. In this case, an effective inference model may encourage more AI developers to leverage inference models, thereby increasing the need for computing. We believe that recent model advances may lead to an increased demand for cloud AI computing for model reasoning and smaller professional model training.

3. Basic Model Providers (OpenAI, Human, Cohere, Deepseek, Mistral)

• Impact on pretrained models: If a new player like DeepSeek can do it with Frontier AI Lab In a small fraction of the reported cost, proprietary pre-training models may become less defensive. We can also expect TTC to innovate further in the transformer model, which, as DeepSeek shows, may come from sources outside of more established AI labs.

4. Enterprise AI Adoption and SaaS (Application Layer)

• Security and privacy issues: Given the origins of DeepSeek in China, it may continue Review From a security and privacy perspective, the company’s products. In particular, the company’s Chinese API and chatbot products are unlikely to be widely used by corporate AI customers in the United States, Canada or other Western countries. It is reported that many companies are Move to block DeepSeek websites and applications use. We want the DeepSeek model to be even by Third party In the U.S. and other Western data centers, companies may be restricted from adopting models. Researchers have pointed out examples of security issues captivity,,,,, Prejudiced and harmful content generation. Give it Consumers’ concernsWe may see experiments and evaluations of the DeepSeek model in enterprises, but due to these concerns, it is unlikely that business buyers will leave existing people.
• Vertical specialization gains appeal: In the past, vertical applications using basic models focused primarily on creating workflows designed for specific business needs. Technologies such as retrieval function generation (RAG), model routing, function calls and guardrails play an important role in generalized models adapting to these dedicated use cases. Despite the significant success of these strategies, there has been concern that the underlying model can significantly improve can make these applications obsolete. As Sam Altman warned, a major breakthrough in model capabilities could be “Steamroll” Application Layer Innovation These were built as wrapping paper around the foundation model.

However, if the advancement in train time calculations is indeed stable, the threat of rapid displacement will be reduced. In a world where the benefits of model performance come from TTC optimization, new opportunities may be open to application layer players. Innovations of post-training training algorithms in specific fields – e.g. Structured and timely optimization,,,,, The reasoning strategy of lurking opinions and effective sampling techniques – may provide significant performance improvements in the target verticals.

Any improvement in performance is particularly important in the context of inference-focused models such as OpenAI’s GPT-4O and DeepSeek-R1, which typically show multi-second response times. In real-time applications, reducing latency and improving the quality of reasoning within a given field can provide a competitive advantage. As a result, application layer companies with domain expertise may play a key role in optimizing inference efficiency and fine-tuning output.

DeepSeek shows that emphasising the increasing amount of pretraining is the only driving force behind model quality. Instead, this development emphasizes the importance of TTC. Although the direct adoption of DeepSeek models in enterprise software applications is still uncertain due to ongoing scrutiny, their impact on driver improvements in other existing models is increasingly obvious.

We believe that DeepSeek’s advances have prompted established AI labs to incorporate similar technologies into their engineering and research processes, thus complementing their existing hardware advantages. As predicted, the resulting model cost reduction appears to help increase model usage and align with the principles of the Jevons paradox.

Pashootan Vaezipoor is the technical director of Georgian.