In the span of a week, Chinese tech startup DeepSeek has upended the global AI industry and left Silicon Valley execs reeling.
DeepSeek's AI capabilities rival those of OpenAI's ChatGPT, but at a fraction of the cost. But what really rocked the market was its efficiency, which reduced the GPU and power needed by multiple factors. Tech experts and pundits are digging into DeepSeek's reasoning logic, hardware used, and training time required to verify the efficiency gains—but this nonetheless appears to be a remarkable breakthrough.
Until this past week, the tech industry believed that to build bigger and better AI systems, companies would need to spend billions on massive data centers to provide the necessary computing power the models require. This led to exploding demand for land, water, and power to enable construction and operations. Data center cost models factored in many variables including improvement in cooling capability, increased chip efficiency, and new reactors—all of which could be reasonably predicted, a key for investors given the long financing timeframes. But new and unpredicted AI model efficiency gains can also generate uncertainty, as they can throw off data center returns, disrupt economic models, and potentially reduce investor yields.
The DeepSeek moment should cause industry participants to consider the possibility that different GenAI software architectures can run 10 to 20 times more efficiently and fundamentally impact the supply and demand curve of AI compute power. And it is only a question of when the next disruptor comes to further bend the curve. This higher efficiency and the resulting lower costs will eventually drive greater adoption, vastly accelerate demand, and unlock new regions for AI infrastructure development. However, in the short-term, AI compute demand could fall due to this step change.
We believe DeepSeek's breakthrough will alter the following data center industry fundamentals:
Redefining data center capacity for AI scale
While current AI infrastructure investments focus on facilities that need 300 to 1,000+ megawatts (MW) of power capacity, companies may now be able to construct facilities that require only 50 to 200 MW and still meet market demand. This enables smaller industry players to build and operate their own data centers, rather than rely on hyperscaler infrastructure. Supply chain of the HW and scale is however a legitimate consideration for smaller data center players or new entrants. Advanced GPUs are often on allocation basis plus network equipment, cables, interconnects, cooling technologies etc all play role, and advanced planning is critical to meet customer demands and ROI.
Edge data centers for AI inference are also now viable given the decrease in required compute power. These smaller data centers are positioned closer to end users, thus delivering faster services with minimal latency.
Unlocking AI infrastructure investments in new regions
This opens immense possibilities for where and how companies can build and power new data centers. Outside of Texas and Washington, few U.S. states possess the power generation or energy grid necessary to bring 300 to 1,000 MW of capacity regularly online. Renewables aren't able to meet this demand; nuclear and natural gas are the only sources that can deliver this type of power.
But now, far more locations across the U.S. and Canada become viable options for new "Mixture-of-experts" AI training and inferencing data centers (as opposed to foundational AI models), and those that have ambition to develop in this area have a lower energy and financing threshold to cross. There is a case to harvest excess power generated to support smaller facilities that can run on 50 or fewer MW, providing a new opportunity for industry leaders to expand their AI infrastructure using power that would have otherwise gone to waste. Similarly, countries that previously could not afford to operate their own AI models will now be able to enter the game, such as those in Europe and the Global South.
Financing and securitization challenges
Current hyperscale data centers are financed and securitized based on future revenue. As such, AI server capacity and utilization levels must be forecasted conservatively over the life of the data center to ensure proper value.
With new disruptors like DeepSeek, the cost curve for all AI workflows will decrease. Over time, continued downward pressure will make it increasingly difficult for investors to justify the ROI at significantly lower levels of revenue driven by the lower cost of compute. Investments that seemed prudent at the time may cause data centers to operate at a loss as unpredicted efficiency gains shock the revenue model. Financing of the builds will need to be reevaluated as construction and capacity planning adjusts. While compute demand may recover with time, data centers may still underdeliver financially if new cost models fail to factor in new efficiency standards.
We will soon release a playbook on how we believe the different players will need to act in this evolving space. Disruption will continue to permeate across the AI landscape, and while we can't predict the future, it will certainly be a rollercoaster.
