DeepSeek R1 is positive for AI training demand

On January 27, 2025, companies across sectors and markets experienced significant share price impact due to the emergence of DeepSeek’s latest AI model (R1). Despite signs of some market rebound since then, many clients have been following up with respect to BCG’s Global Data Center Model projections, which forecast global data center power demand to grow at a 16% compound annual rate from 2023 to 2028—33% faster than the 2020–2023 period—reaching approximately 130 GW by 2028.

Key Takeaway: DeepSeek R1 does not impact BCG power demand projections.

• The demonstrated efficiency improvements in the DeepSeek models are an expected step in the data center power demand growth journey and align with BCG’s proprietary Global Data Center Model assumptions.
• BCG’s Q4 2024 projection of data center power demand incorporates significant algorithmic improvements to training and inference efficiency over the forecast period, as improvements like those made by DeepSeek are aligned with historical and observable trends.
• We have committed to making market updates periodically (independent from the impact of DeepSeek) and plan to make the next update within Q1 2025.

Why not? (in a nutshell): DeepSeek’s latest AI model (R1) represents a notable algorithmic improvement, combining impressive efficiency gains with innovative techniques. However, DeepSeek’s innovation aligns with the broader historical trajectory of AI development rather than signaling a paradigm shift. As learnings from DeepSeek’s approach are applied, power demand growth will persist, and infrastructure remains critical – despite recent market movements.

Below, we explore themes and implications derived from this advancement.

1.     Efficiency gains may seem exponential but are in fact relatively incremental. While this gain is notable in the markets right now – taking a step back, we see this as in line with broader trends in algorithmic innovation.

2.     AI scaling will continue to absorb available compute. Increasing inference demands, expanding AI applications like real-time reasoning and generative media, and self-optimizing efficiency gains will drive sustained and exponential growth in computational requirements, absorbing available compute despite efficiency improvements.

3.     Long-term power demand growth will remain robust. Data center power demand growth remains within expected ranges but is increasingly significant in absolute terms, with sustained growth driven by training and inference workloads that require significant power, while traditional enterprise computing—projected to account for 55% of demand by 2028—continues to account for the majority of demand for computing power through 2028.

4.     Market movements do not reflect underlying value. Market volatility following DeepSeek R1 reflects shifting sentiment but does not diminish the critical role of power infrastructure in sustaining AI-driven demand. As inference scales, reliable power remains critical to support continued growth. We still anticipate a significant gap to firm power to serve demand from data center operations globally.

Please note that this is an evolving set of perspectives, and we anticipate further developments and learnings in the coming days and weeks. 

Detailed Insights

1. Efficiency gains may seem exponential but are in fact relatively incremental

• DeepSeek R1 demonstrates improved computational efficiency: DeepSeek R1 is the first non-US state-of-the-art (SOTA) model, trained on significantly less compute than existing proprietary models while achieving impressive efficiency. Additionally, the open-source nature of DeepSeek R1 allows for broad access, setting it apart from proprietary alternatives.
• DeepSeek R1 is not unique but is an improvement: While DeepSeek R1 is a more efficient model, it is not the only notable improvement in training efficiency in recent months. Google's Gemini 2.0 Flash demonstrated similar model advancements in December 2024, including extended context handling, multimodal capabilities, and autonomous task execution. While Gemini 2.0 Flash achieves ~5% greater accuracy in complex reasoning tasks than DeepSeek R1, it still operates at higher compute requirements and requires more energy per query. Gemini also remains closed source, limiting its accessibility and potential for viral impact.
• Cost narrative is incomplete: DeepSeek's R1 optimized energy consumption by building on top of open-source models leveraging existing compute investments. DeepSeek claims that R1 was developed with an estimated $5.6 million investment, though this excludes the total cost-to-achieve. (e.g., base model design cost, hardware purchase costs, researcher cost, inference infrastructure cost, etc.).
• Efficiency gains are part of a long-term trend: Compute power has improved by multiple orders of magnitude over the last 20 years, with regular algorithmic advances driving exponential growth. Orders of magnitude additional improvements remain possible, which has been the trajectory of machine learning to-date. Events like DeepSeek’s R1 are part of this predictable progression and should be expected to continue.

2. AI scaling will continue to absorb available compute

•  Inference models increase compute needs: While efficiency gains like those in R1 may temper immediate compute needs for model training, long-term compute demand trends remain intact. As training becomes more efficient, more models will be trained and deployed at scale. Inferencing demand will increase as more end-users engage each model. Despite efficiency improvements, training models will persist, and the need to deploy (inference) models will continue to proliferate as use cases grow.
• Compute demand is elastic: As AI innovation scales toward AGI (Artificial General Intelligence) and ASI (Artificial Superintelligence), it will absorb as much compute as is available. Lower costs for model training and inference will drive increased adoption. Emerging applications such as real-time AI reasoning, generative media creation, and large-scale simulations will sustain and amplify compute demand despite efficiency gains.
• Efficiency improvements will continue at pace: As AI systems are deployed to optimize themselves, future improvements in training and inference efficiency are likely to accelerate. This suggests that orders of magnitude additional gains remain possible, reinforcing the trajectory of rapid innovation.

3. Long-term power demand will remain robust

• Growth remains aligned with trends: Power demand from data centers has historically grown at 12% annually. While today’s 16% annual growth rate remains within anticipated ranges, the larger baseline of demand means that this growth appears more significant in absolute terms.
• Continued training and increased inference will drive sustained growth: Training workloads will persist as model efficiency improvements enable increased training runs. Inference workloads, unlike training, will continue to scale as adoption increases and will dominate future demand. As high-demand inference workloads proliferate, the increased compute need will drive increased power demand, despite increased efficiencies.
• Traditional cloud compute will persist: Enterprise workloads such as file storage, transaction processing, and conventional business applications still represent the majority of data center power demand today and will continue to do so, accounting for roughly 55% of demand by 2028. These workloads provide a stable and growing base for data center buildout, demonstrating continued power demand growth even as AI efficiencies evolve.

4. Market movements do not reflect underlying value 

• Market movements reflect sentiment, not value: Market volatility following DeepSeek R1 reflects shifting investor sentiment, recalibrating valuations due to evolving expectations about AI-driven infrastructure needs. However, these fluctuations do not diminish the critical role of power infrastructure in sustaining data centers as AI demand accelerates. Continued model training, growing inference workloads, new AI applications, and sustained traditional cloud compute make reliable power essential for efficiency, innovation, and long-term value, reinforcing the need for expanded capacity.

Conclusion

DeepSeek R1 reinforces the ongoing trajectory of AI efficiency gains but does not alter the fundamental drivers of data center power demand growth. As AI adoption scales, training persists, inference workloads expand, and traditional computing remains a steady demand base, power infrastructure will remain a critical enabler of sustained growth, regardless of short-term market fluctuations.

Authors: Braden Holstege, Clark O’Niell, Vivian Lee, Ross LaFleur, Michael McKissack, and Pattabi Seshadri

We will continue to monitor developments and remain available to discuss specific impacts for your organization or projects. 

Please see BCG’s latest published findings on data center growth here: Breaking Barriers to Data Center Growth.