Ema Unlimited is pleased to share this blog from Everest Group authors, Vaibhav Bansal, Vershita Srivastava, and Yusuf Ahsan.

Agentic AI and its reliance on foundation models

The AI and automation space has been evolving at an unprecedented pace. A couple of years ago, Generative AI (gen AI) reshaped our understanding of AI’s potential. Just as the world was beginning to adapt to gen AI, agentic AI then emerged as the new kid on the block and captivated everyone’s attention. While gen AI was about knowledge assistance, agentic AI takes it to the next level through goal-based action, autonomy and adaptability

Though agentic AI is highly disruptive, it is not an isolated new technology. It builds on the capabilities of earlier innovations, especially that of gen AI and foundation models. The foundation models essentially act as the brain for agentic AI, facilitating input analysis, intent recognition, goal setting, advanced reasoning, workflow planning and decision making. This continuous loop of perceive-reason-decide enables agents to perform a range of tasks in varying environments.

DeepSeek has shaken the foundation model ecosystem

Since the launch of ChatGPT in Nov 2022, Open AI has been the industry leader for foundation models. There have also been challenges to that dominance from competitors such as Anthropic, Google, and Meta. All this while certain norms were established by these players around the model ecosystem, such as the need for advanced chipsets, high training costs, significant usage costs and the effectiveness of larger models.

DeepSeek’s entry has challenged these norms like never before. Early tests indicate its superior performance vis-a-vis competitors as depicted in Exhibit 1.

Moreover, this comes at a fraction of the cost of the existing models as outlined in Exhibit 2.

While a few claims may be debatable, DeepSeek has presented a strong proof of concept and disrupted the industry. The question now though is how such limited resources have delivered such a superior offering?

Let’s dive a bit ‘deep’ to ‘seek’ the answers:

• Mixture of experts model: DeepSeek’s R1 utilizes multiple expert networks, with only a few activated at a time. For any given query, DeepSeek R1 utilizes only 37 billion parameters out of the total 671 parameters, resulting in a reduced computation cost
• Optimizing expert deployment: It operates on a combination of shared experts with broad capabilities and specific experts with narrow focus areas. A built-in load balancing mechanism ensures even distribution of tasks enhancing efficiency and resilience.
• Training mechanism: It utilizes a multi-stage training pipeline which combines reinforcement learning and supervised fine-tuning as opposed to relying only on supervised fine-tuning which inflates cost.
• Multi-token prediction: Traditional models generate responses one token at a time, with each prediction depending on the previous one. DeepSeek predicts multiple tokens in parallel improving processing speed without major accuracy trade-offs

With these techniques, DeepSeek has challenged the dominance of the incumbent model providers who were winning owing to their compute capacity and scalability. So, how will these advances impact the broader foundation model ecosystem?

• Plummeting usage costs: With DeepSeek’s entry, other model providers will be forced to reduce their usage cost. These have been coming down gradually since ChatGPT was launched but expect that to accelerate going forward. Even if a direct cost cut doesn’t happen, more intelligent models will be available at the same price.
• Rationalized hardware demand: DeepSeek utilized H800 chips by Nvidia instead of the latest H100 GPUs, which the majority of its competitors utilize. Despite the overall demand for chips continuing to climb, the product mix will lean away from the higher end GPUs, eroding manufacturer’s profit margins and reducing hardware costs for end users in the process.
• Increasing commoditization: DeepSeek’s approach could be a differentiator in the short run, but will cease to be a moat in the longer run. As other model providers catch up, the techniques adopted will get standardized with foundational models becoming commoditized.
• Open-sourcing: Open-source models provide the opportunity to build, modify and integrate the offering as per requirements. While open-source models have found good adoption earlier, DeepSeek is accelerating the creation of an open-source community.
• Surge in explainable models: With Explainable AI, DeepSeek prioritizes transparency and insights on how it arrives at the conclusion. This fosters a trust in DeepSeek and will push others to open their models similarly.

Expected impact on Agentic AI adoption

There has been a lot of excitement around Agentic AI’s potential to transform businesses and enable the shift towards a combined digital-human workforce. However, there have also been headwinds. The most common is the concern around cost, contributed by the high usage cost of proprietary models. Also, with usage-based pricing becoming the norm, enterprises are increasingly unprepared for uncertainty around foundation model bills.

Another challenge is the Agentic AI infrastructure. Most organizations are unable to invest in their own infrastructure to host models. Hence, they consume these models through cloud services, which subjects them to data security risks and becomes detrimental in regulated industries.

Concerns around the lack of trust in the outputs generated by an AI agent, pushes back the whole narrative around autonomy.

The disruption caused by DeepSeek is expected to solve these challenges to a good extent. Exhibit 3 demonstrates the expected short-term impact of foundational model commoditization on players involved in an agentic value chain.

While there might be variation in the extent of impact on individual players in the short term, the overall agentic AI market is expected to get a boost in the longer term. Let’s understand how:

• Democratization of agentic AI: Cost reduction, availability of open-source models and greater explainability will lead to a surge in adoption. As Jevons paradox kicks in, enterprises will be able to infuse more processes with agentic AI, without compromising their bottom line leading to increased overall spending. Agentic AI will get more democratized, expanding its reach to a range of enterprises, beyond those with significant financial resources.
• Acceleration in agent development: While foundation models have received all the attention, the real value for enterprises is at the application layer. With commoditization, the development of domain-specific AI agents will accelerate and with the supply-side becoming more competitive, ultimately this will offer more choice and better quality to enterprises.
• Innovations in model ecosystem: DeepSeek has shown us the ‘art of possible’ and might mean reduced short to mid-term profitability for foundation model players. We can expect this to be a trigger for further model engineering, optimization and innovation with advances on domain-specific and multi-modal models. Further breakthroughs in this field are going to benefit the entire agentic ecosystem boosting adoption
• Acceleration in model fusion: Enterprises that continue to rely on just one foundation model for all their needs are going to be at a continuous disadvantage going forward. The enterprise demands today are so complex and diverse that a single model can’t fulfil all of them, while being accurate as well as cost-effective. With new innovations like DeepSeek taking place every few weeks and the goalpost for the best or the most superior model constantly changing, enterprises can’t keep on switching their entire agentic ecosystem from one model to another. Therefore, model fusion technologies which dynamically combine or leverage different foundation models, are expected to become more prevalent. These technologies will not only optimize model cost and accuracy for enterprises, but will also make their agentic ecosystem future-proof by providing the flexibility to incorporate new innovative models seamlessly.

Capitalizing on the perfect opportunity

DeepSeek has provided the answer to the most prominent concerns plaguing enterprises in their agentic journey. However, these are by no means completely addressed. Enterprises, especially in regulated or data sensitive industries, are grappling with data privacy and security concerns, which DeepSeek may not have done much to alleviate. In fact, the exposure of DeepSeek’s data including chat history and sensitive information has aggravated this concern. DeepSeek’s evasive nature when encountering questions related to controversial national issues has put into question its reliability in the face of censorship.

Nevertheless, DeepSeek has expanded the horizon through innovation. With a slew of advances expected to follow, models getting more and more commoditized and model fusion becoming a norm, this is the perfect opportunity for enterprises to embark on their agentic journey. The barriers to entry have been brought down significantly allowing enterprises to sail ahead on their agentic journey without a heavy financial luggage.