Barclays: Global AI Computing Power to Support 15-220 Billion Agents by 2025

Barclays has released a research report indicating that the current global AI computing power is sufficient to support between 15 to 220 billion AI agents by 2025. This capacity is deemed adequate to meet the majority of the demand in the United States and the European Union. The report suggests that the existing AI computing power is likely to satisfy the needs of these regions, which could potentially dampen the demand for more efficient AI models and computing power. This analysis comes on the heels of a similar report by TD Cowen, which also discussed the computational requirements for AI.

The implications of this report are significant for the AI industry. If the current computing power is indeed sufficient, it could lead to a reduction in the urgency to develop more advanced and efficient AI models. This could impact the investment and research efforts in the AI sector, as companies and institutions may reassess their priorities and resource allocations. However, it is important to note that the demand for AI computing power is not static and can evolve rapidly with technological advancements and new applications.

The report also highlights the importance of open-source models in reducing costs and increasing accessibility. Open-source models can democratize AI technology, making it more available to a broader range of users and organizations. This could foster innovation and drive the development of new AI applications across various industries. The role of open-source models in the AI ecosystem cannot be overstated, as they provide a foundation for collaboration and advancement in the field.

Barclays' research points out that the AI industry needs to shift from meaningless benchmark tests to the deployment of practical Agent products. The report emphasizes that low inference costs are crucial for profitability. Open-source models are expected to lower these costs, making AI technology more accessible and affordable. Despite the apparent sufficiency of current computing power, there is still a gap in the specialized computing power required for efficient, low-cost Agent products.

Barclays' findings reveal several key points about the supply and demand of AI computing power. By 2025, there will be approximately 15.7 million AI accelerators (GPUs, TPUs, ASICs, etc.) online globally, with 40% (about 6.3 million) dedicated to inference. Half of these inference capabilities, around 3.1 million, will be specifically used for Agent/Chatbot services. The industry is also seeing a shift towards more cost-effective open-source models, such as Salesforce's Agentforce, which uses the Mistral open-source model with parameters ranging from 7B to 141B, instead of the most expensive proprietary models.

The report also notes the rapid growth in downloads of open-source models like DeepSeek, Llama, and Mistral, as reported by Hugging Face. This trend is expected to accelerate as the industry transitions from chatbots to more advanced Agent technologies. However, despite the apparent abundance of computing power, there are structural challenges. If Agent products become widely adopted and prove to be highly useful for consumers and enterprise users, there may be a need for more affordable, smaller, yet high-performance base models, additional inference chips, and possibly repurposing installed training GPUs for inference.

This indicates that while overall computing power may seem sufficient, there is still a significant gap in the specialized computing power required for efficient, low-cost Agent products. Companies focusing on high-efficiency inference cost structures and developing small, efficient models may have a competitive advantage in the AI Agent race. Conversely, companies relying solely on large models without considering unit economics may face greater challenges.

Barclays also highlights the economic challenges posed by the inference costs of AI Agents. Compared to traditional chatbots, Agent products generate approximately 10,000 tokens per query, which is 25 times more than the 400 tokens generated by chatbots. This significantly increases the inference costs. The economic benefits of different models vary greatly; for example, an Agent product based on the OpenAI o1 model costs $2,400 per year, while one based on the DeepSeekR1 model costs only $88 per year, providing 15 times more user capacity. The demand for "super Agents" is also on the rise, with OpenAI planning to launch products that consume up to 35.6 million tokens per month and handle 44 queries per day, far exceeding the 2.6 queries per day of ordinary Agents.

From an economic perspective, the pricing model based on tokens will determine the market competitiveness of different models. As Barclays' research points out, the low inference cost is crucial. Due to their autonomous nature, Agent AI products consume tokens at a much higher rate than chatbots. Additionally, while "super Agents" have potential, their high inference costs may limit their widespread application. Investors evaluating such products should carefully consider their economic viability.