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 evolve at a rapid rate, recent developments questioning the established paradigms. At the beginning of 2025, Deepseek at the Chinese AI laboratory unveiled a new model that sent shock waves in the AI ​​industry and resulted in a 17% fall into the Nvidia stock, with Other actions related to the request of the AI ​​data center. This reaction of the market was largely reported from the apparent capacity of Deepseek to provide high performance models to a fraction of the cost of competitors in the United States, causing a discussion on the discussion on the Implications for AI data centers.

To contextualize Deepseek’s disruption, we think it is useful to consider a broader change in the AI ​​landscape being motivated by the scarcity of additional training data. Because the main AI laboratories have already formed their models on most of the public data available on the Internet, the shortage of data is slow down additional pre-training improvements. Consequently, providers of models seek to “calculate the calculation of time” (TTC) where the reasoning models (such as the “O” series of “O” of Open AI) “think” before answering a question at the time of inference, as an alternative method to improve the overall performance of the model. Current thinking is that the TTC can present improvements in the scaling law similar to those which have once propelled prior pre-training, potentially allowing the next wave of AI transformative advances.

These developments indicate two important changes: first, laboratories operating on smaller (reported) budgets are now able to publish advanced models. The second change is the emphasis on TTC as the next potential engine for the progression of AI. Below, we unpack these two trends and the potential implications for the competitive landscape and the wider AI market.

Implications for the AI ​​industry

We believe that the transition to TTC and the increased competition between reasoning models can have a number of implications for the wider AI landscape through the hardware, cloud platforms, foundation models and corporate software.

1. Material (GPU, dedicated chips and calculation infrastructure)

• From massive training bunches to the peaks of “testing time” on demand: In our opinion, the transition to TTC can have implications for the type of material resources that AI companies need and how they are managed. Rather than investing in increasingly important GPU clusters dedicated to workloads, IA companies can rather increase their investment in inference capacities to meet the growing needs of TTC. While IA companies will probably still need a large number of GPUs to manage the workloads in inference, the differences between Work charges And the inference workloads can have an impact on how these chips are configured and used. More specifically, because the workloads in inference tend to be more Dynamics (and “Spikey”)Capacity planning can become more complex than for work -oriented workloads.
• Climb of equipment optimized by inference: We believe that the change in focusing towards the TTC is likely to increase the opportunities for alternative AI equipment specializing in the calculation of the low -latency inference time. For example, we can see more request for GPU alternatives such as integrated circuits specific to the application (Asics) for inference. As access to TTC becomes more important than training capacity, domination of GPU for general use, which are used both for training and inference can decrease. This change could benefit suppliers of specialized inference fleas.

2. Cloud platform: Hyperscalers (AWS, Azure, GCP) and Cloud Compute

• The quality of service (QOS) becomes a key differentiator: A problem preventing the adoption of AI in the company, in addition to the concerns about the precision of the model, is the lack of reliability of the inference APIs. Problems associated with non -reliable API inference include Fluctuating response time,, Limitation of rates and difficulty Manage simultaneous requests And Adapt to API termination point changes. Increased TTC can exacerbate these problems. Under these circumstances, a cloud supplier capable of providing models with QOS insurances that raise these challenges, in our opinion, have a significant advantage.
• Increase in cloud expenditure despite the efficiency gains: Rather than reducing the demand for AI equipment, it is possible that more effective approaches to training and inference of the large language model (LLM) can follow the Jevons paradox, a historical observation where improved efficiency leads to higher overall consumption. In this case, effective inference models can encourage more AI developers to be taken advantage of reasoning models, which, in turn, increases the demand for calculation. We believe that the advances of the recent model can lead to increased demand for calculating the Cloud for the inference of the model and a smaller and specialized specialized model training.

3. Foundation Model Providers (Openai, Anthropic, Cohere, Deepseek, Mistral)

• Impact on pre-formed models: If new players like Deepseek can compete with the AI ​​frontier laboratories to a fraction of the reported costs, pre-formulated models can become less defensible as a moat. We can also expect new innovations in TTC for models of transformers and, as Deepseek has demonstrated it, these innovations can come from sources outside the more established AI laboratories.

4. Adoption of corporate and SaaS (request layer)

• Regarding security and confidentiality: Given the origins of Deepseek in China, there will probably be careful examination Company products from the point of view of security and confidentiality. In particular, API and Chatbot offers based in the company’s China are unlikely to be widely used by corporate AI customers in the United States, Canada or other Western countries. Many companies would have move to block Using the website and Deepseek applications. We expect the Deepseek models to be examined by a meticulous exam even when they are hosted by third party In the United States and other Western data centers, which can limit the adoption of model companies. Researchers are already pointing to examples of safety problems prison breakdown,, bias and generation of harmful content. Given Watch out for consumersWe can see the experimentation and evaluation of Deepseek models in the company, but it is unlikely that business buyers will move away from the holders because of these concerns.
• Vertical specialization is gaining ground: In the past, vertical applications that use foundation models focused mainly on creating workflows designed for specific commercial needs. Techniques such as generation of recovery (RAG), routing of models, function calls and railings have played an important role in adapting generalized models for these specialized use cases. Although these strategies have led to notable successes, there has been a persistent concern that significant improvements to the underlying models could make these applications obsolete. As Sam Altman warned, a major breakthrough of the model’s abilities could “Steamroll application layer innovations » which are built as packaging around foundation models.

However, if the progress in the train calculation is indeed plated, the threat of rapid movement decreases. In a world where model performance gains come from TTC optimizations, new opportunities can open up to application layer players. Innovations in post -training algorithms specific to the field – such as Structured prompt optimization,, Reasoning strategies sensitive to latency and effective sampling techniques – can provide significant performance improvements within targeted verticals.

Any improvement in performance would be particularly relevant in the context of reasoning models such as GPT-4O and Deepseek-R1 of Openai, which often have response times for several seconds. In real -time applications, reducing latency and improving the quality of inference in a given area could provide a competitive advantage. Consequently, appliance companies with an expertise in the field can play a central role in optimizing the inference efficiency and fine results.

Deepseek demonstrates a drop in emphasis on ever increasing quantities of pre-training as the only engine of the quality of the model. Instead, development highlights the growing importance of TTC. Although the direct adoption of Deepseek models in corporate software applications remains uncertain due to the meticulous examination, their impact on improving the improvement of other existing models becomes clearer.

We believe that Deepseek’s progress has prompted the AI ​​laboratories to integrate similar techniques in their engineering and research processes, completing their existing material advantages. The reduction in the costs of the resulting model, as expected, seems to contribute to increased use of the model, aligning the principles of the Jevons paradox.

Pashootan Vaezipoor is a technical head at Georgian.