The release of the Chinese AI company model called DeepSeek has upended the established AI industry. That a model can be developed without access to the most advanced GPU chips and trained at a reported cost of $6 million dollars is anathema to the prevailing Silicon Valley mantra that “more parameters, more GPUs, and more data” is the pathway to better and better models and eventually, Artificial General Intelligence (AGI). DeepSeek’s unveiling has upended this ideology.
The result has been market cap carnage with the value of chip makers, energy companies, data center component companies, and data center operators collectively losing $1T in market cap in a single day. Digital Realty sold off by 9% and Equinix dropped 7%. So, does this spell doom for the data center sector? While it may cause some short-term pain and delay or postpone some technology investments, DeepSeek should be a good thing for the sector over the long term.
The data center sector has been one of the primary beneficiaries of the AI boom sparked by the release of ChatGPT in October 2022. In order to train and operate an LLM you need GPUs, and GPUs require significant amounts of data center capacity. The last two years have seen a tsunami of data center demand that has resulted in record leasing across the sector, a significant increase in pricing, and demand higher than supply. All good things if you are a data center operator.
Two years ago, most folks outside our industry did not even know what a data center was. Today it’s hard not to hear the word mentioned every hour on a business channel or read about it daily in a national newspaper. What has been the result? Everyone wants to be in the data center business. Every week now it seems anyone with a piece of dirt within 2 miles of a transmission line believes their property is a perfect data center location. Energy companies are spinning up divisions to focus or partner with data center developers. Real estate companies whose core business was industrial or commercial are starting data center divisions. Speculators who look to mimic the “Flip this House” tactic of residential homes are scouring for powered land in every corner of the country. All are supported by the idea that massive amounts of power and billions or trillions of data center clusters will be needed to build the yellow brick road to the magic palace that is AGI.
Given the timeline to develop larger-scale data center projects, most of this supply had ready-for-service dates of 2027/28 and beyond. DeepSeek and the concept that AI can be deployed without the gargantuan power envelope that it currently requires is like a cold bucket of water on this wave of speculative development. The risk of oversupply should dramatically decrease as a result of this development, and oversupply was probably the greatest long-term risk to the data center sector. While contradictory at first glance, this development is beneficial to the sustainability of the data center sector, which relies fundamentally on balanced supply and demand over a prolonged period of time.
How to generate and distribute the massive amount of power required for AI has been another challenge and risk for the data center sector. DeepSeek’s success intimates that wide-scale adoption of AI may be doable without the 2x, 3x, or 4x increase in the amount of power required as has been prognosticated by consulting firms and Wall Street. If true, the costs to develop all this new generation and transmission will dramatically decrease, easing the strain on utility ratepayers and investors alike. Many states were already looking into passing legislation to push those costs onto data center companies. A lower power profile for AI will reduce those impulses. This lower power footprint will also be good from an environmental perspective as AI demand was beginning to drive demand for short-term onsite generation via natural gas and recommissioning of coal power plants. These trends conflict with widely shared industry goals for achieving carbon neutrality. A more gradual power ramp should allow for the continued adoption of sustainable energy along with a more moderate use of fossil fuel-based generation. More efficient AI will benefit the planet, taxpayers, and data center operators in the long term.
Finally, DeepSeek should accelerate the rate of AI adoption in the long term as it seems to be a more cost-effective approach to building AI (Of course, this all must be validated in the months ahead). Driving down the costs of AI adoption will allow developers to deploy it everywhere in the software stack, increasing use and consumption. DeepSeek shows that there are multiple paths to AGI that can be more resource efficient than current approaches which rely heavily on “throwing more compute” at the solution. If anything, DeepSeek shows that the AI arms race will not be won just through more chips and more data, but that innovation will be a core competitive factor to success.
Fundamentally, the data center sector is a long-term business. The sector was doing just fine before ChatGPT and would have continued to do so since data center demand is fundamentally indexed to technology adoption, and humans show no indication of reducing their consumption of technology. Regardless of the power footprint, AI and LLM adoption is a “net new workload”, something that is additive to what already exists in data centers today. And there is no doubt that AI will be adopted and impact every facet of our lives. New technology trends like GenAI, which has only just started, along with public cloud growth and digital transformation will continue to drive data center growth for many years to come.
