Deepseek Jolts AI Industry: Why the next AI jump may not come from more data, but more clever in conclusion

The AI ​​landscape continues to develop at a rapid pace, with the latest developments provoking established paradigms. In early 2025 Through the AI ​​industry and lead to a 17% dropping in NVIDIA stocks, together with Other stocks related to searching the AI ​​data center. This market reaction is reported to result from Deepseek’s apparent ability to provide highly efficient models of part of the price of rivals in the United States, provoking a discussion of Consequences for AI data centersS

In order to context the interruption of Deepseek, we believe it is useful to look at a broader displacement in the AI ​​landscape, which is guided by the shortage of additional training data. So delay further improvements in preliminary trainingS As a result, model providers seek to “calculate the test time” (TTC), where models of reasoning (such as “O” series of AI models) “think” before answering a question during the conclusion as an alternative method to improve the overall efficiency of the model. This thinking is that TTC can show improvements to scaling similar to those who have ever driven pre -workouts, potentially allowing the next wave of transformative AI progress.

These developments have shown two significant changes: first, laboratories working on smaller (reported) budgets are now capable of launching state-of-the-art models. The second change is the focus on TTC as the next potential engine of AI progress. Below, we unpack both trends and potential consequences for the competitive landscape and the broader AI market.

Consequences for the AI ​​industry

We believe that the transition to TTC and increased competition among the reasoning models can have a number of consequences for the broader AI landscape In hardware, cloud platforms, foundation models and corporate software.

1. Hardware (graphic processors, specialized chips and computing infrastructure)

• From massive training clusters to tests “Time”: In our opinion, the transition to TTC may have consequences for the type of hardware resources that AI companies require and how they are managed. Instead of investing in more and more GPU clusters dedicated to training, AI companies can instead increase their investments in opportunities for conclusions to support the growing needs of TTC. Although AI companies are likely to still require a large number of graphics processors to cope with the load of conclusions, the differences between workout And the workload of the conclusion can affect the way these chips are configured and used. More concretely, since the load of the conclusions is usually more Dynamic (and “Spikey”)Capacity planning can be more complicated than it is for batches oriented.
• Rise of optimized hardware conclusion: We believe that the focus to TTC is likely to increase the opportunities for alternative AI hardware, which specializes in calculating the low latency conclusion. For example, we may see more search for GPU alternatives as app -specific integrated circuits (Asic) for conclusionS As access to TTC becomes more important than the training capacity, the dominance of general purpose graphic processors, which are used for both training and conclusions can reduce. This change can be beneficial for specialized chip providers for conclusions.

2. Cloud platforms: hyperskalers (AWS, Azure, GCP) and Cloud Compute

• The quality of service (QoS) becomes a key differential: One question that prevents AI from receiving an enterprise, in addition to the concerns about the accuracy of the model, is the unreliability of API of the conclusion. The problems associated with the unreliable conclusion for API includes Fluctuating reaction times., Speed and difficulty Working with simultaneous requests and Adapting to API’s final point changesS Increased TTC can further exacerbate these problems. In these circumstances, the cloud supplier that can provide QOS models to cope with these challenges would, in our opinion, have a significant advantage.
• Increased cloud costs despite the profits of efficiency: Instead of reducing the demand for AI hardware, it is possible that more effective approaches to the training and the conclusion of a large language model (LLM) may follow the Jevons paradox, a historical observation, which leads to more general consumption. In this case, effective models of conclusions can encourage more AI developers to use reasoning models, which in turn increases the demand for calculation. We believe that the recent progress of the model can lead to an increased search for the AI ​​Compute cloud both for the conclusion of the model and for less, specialized training of the model.

3. Suppliers of Basic Models (Openai, Anthropic, Cohere, Deepseek, Mistral)

• Impact on pre -trained models: If new players like Deepseek can compete with Frontier Ai Labs With some of the reported costs, own pre-trained models can become less protective as a moat. We can also expect additional innovations in TTC for transformers models and as Deepseek demonstrates, these innovations can come from sources beyond the more admitted AI laboratories.

4. AI and SAAS (Application layer)

• Concern for security and confidentiality: Given the origin of Deepseek in China, it is likely to continue control from the products of the company in terms of security and confidentiality. In particular, API and Chat offers of the company based in China are unlikely to be widely used by Enterprise AI customers in the US, Canada or other Western countries. Many companies have been reported transition to a block Using the website and Deepseek applications. We expect Deepseek models will face control even when hosting third In the US and other Western data centers that can limit the acceptance of enterprise models. Researchers already indicate examples of security concerns around shutter., addictions and harmful content generationS Given Consumer attentionWe can see experiments and evaluation of Deepseek models in the enterprise, but it is unlikely that businesses of enterprises are moving away from those in force due to these problems.
• The vertical specialization acquires grip: In the past, vertical applications that use basic models focus mainly on creating workflows designed for specific business needs. Techniques such as generation with generation (RAG), model routing, call calling and fuses play an important role in adapting summary models for these specialized cases of use. Although these strategies have led to remarkable success, there is a constant concern that significant improvements to basic models can make these applications outdated. As Sam Altman warned, a great breakthrough in the capabilities of the model could “can” can “Steamroll ”Innovation of the app which are built as packaging around the main models.

However, if the progress in the calculation of train time is indeed paid, the threat of rapid shift decreases. In a world where the profits in the performance of the model come from TTC optimizations, new opportunities can be opened to the player player of the app. Innovation in Domain-specific algorithms after exercise-like Structured optimization of fast., Latency reasoning strategies And efficient sampling techniques – can provide significant performance improvements within target verticals.

Any improvement in productivity would be especially appropriate in the context of models aimed at reasoning, such as Openai’s GPT-4O and Deepseek-R1, which often show many seconds of response. In real -time applications, reducing latency and improving the quality of conclusion in a domain can provide a competitive advantage. As a result, companies for applying applications with expert experience in the domain can play a major role in optimizing the effectiveness of the conclusion and fine results.

Deepseek demonstrates a reducing emphasis on the ever -increasing quantities of pre -workout as the sole engine of model quality. Instead, development emphasizes the growing importance of TTC. Although direct adoption of Deepseek models in corporate software applications remains uncertain due to continued control, their impact on driving improvements in other existing models becomes more clear.

We believe that the progress of Deepseek has caused established AI Labs to include similar techniques in their engineering and research processes, complementing its existing hardware benefits. The resulting reduction in the cost of the model, as provided, seems to contribute to the increased use of the model, aligning the principles of Jevons paradox.

Pashootan Vaezipoor is a technical leading role in Georgian.