LinkedIn Skips Data Center Expansion, Doubles GPU Efficiency Instead

LinkedIn, owned by Microsoft since 2016, has decided to hold its data center spending flat for the fiscal year beginning last month and ending next June. Unlike the major cloud and AI platforms racing to build and finance massive new data centers, LinkedIn's executives announced to WIRED that the company will keep its GPU investment steady and its compute and storage footprint flat. This decision became possible after LinkedIn's engineering teams found ways to use their existing GPUs twice as efficiently over the past six months.

The path to this efficiency gain began years earlier, when LinkedIn moved to build its own data centers in Oregon, Texas, and Virginia in 2022. At the time, the company had attempted to use Microsoft Azure, but the fit was poor. "Microsoft Azure was growing like crazy, the level of customer demand was through the roof, and at the same time we saw skyrocketing growth on the LinkedIn side," said Raghu Hiremagalur, LinkedIn's chief technology officer for infrastructure. Owning its own data centers gave LinkedIn control over every detail of its infrastructure, setting the stage for the optimization work that followed. Around the same time, LinkedIn began developing AI-based assistants to help users write messages, find jobs, and recruit candidates—work that came with escalating costs. "Every query that's coming to our site has increased in cost over time," Hiremagalur explained, noting that LinkedIn's stored data was doubling annually. "That is not a sustainable place to be."

LinkedIn's response was systematic. Hiremagalur's team developed measurement tools to understand how much compute and storage individual teams were using, then set up an allocation system to distribute projects across data center computers more efficiently, reducing idle time. The results were striking: LinkedIn achieved GPU utilization on the training side at north of 95 percent, a level Hiremagalur described as "the best that I have seen." The company also adopted model distillation, training smaller AI models from larger ones. For job recommendations, a single smaller model learned from two larger models to both identify relevant job openings and predict which users would click on them. Although smaller, the model sacrificed no quality—users were discovering jobs they had not found before because the model understood their preferences better.

LinkedIn made dozens of other improvements. The company streamlined model training, reused information from earlier recommendations, and better balanced workloads between CPUs and GPUs. It even reworked foundational software on Nvidia processors to handle tasks larger than they were designed for, and rerouted some work to run on CPUs instead of Nvidia GPUs, which are more expensive, harder to procure, and consume more electricity. Together, these efforts saved about $24 million（約38億円） over the past 12 months—equivalent to roughly 1,100 GPUs running around the clock for a year.

Erran Berger, LinkedIn's chief technology officer for engineering, acknowledged that these savings are modest for a company with $18 billion（約2.9兆円） in annual sales, but emphasized the deeper value. "One of the goals we've set is to try to basically keep our compute footprint flat or as close to flat as possible while shipping more compute-hungry things to production," he said. "That's a pretty bold statement to make in today's world." Berger believes efficiency gains could compound over time, enabling LinkedIn to get more value from future data center expansions. Both executives stressed that the constraints will push engineering teams to be more creative and that freed-up resources allow engineers to move faster on new projects. "I really want to double underscore that for a company of our scale, to say a full year we're going to do this with no incremental storage and compute is no small feat, but it's taken a ton of work to get there," Hiremagalur said.

LinkedIn's move reflects a broader trend in the industry toward "tokenomics"—deeper analysis of the cost of using generative AI. Songyee Yoon, managing partner of Principal Venture Partners and board member at server maker HP, called the LinkedIn approach encouraging: "It suggests AI is beginning to move from experimentation into production discipline. The companies that win will not simply be the ones that spend the most on infrastructure." Gartner's Chirag Dekate added that enterprises are evolving from a "buy-more era" to a "do-more era," and that even smaller businesses without LinkedIn's infrastructure control are finding ways to cut costs—purging unused software, moving to cheaper neoclouds, and adopting the lowest-cost AI models for each project.

Despite the flat spending, LinkedIn's data centers are not stale. The company will continue buying new servers to replace aging machines as they break down or age out over the coming months. However, the company locked in some savings by purchasing ahead; Hiremagalur noted that server costs have surged dramatically, with some jumping three-fold in price in recent months. LinkedIn is not ruling out future data center growth—Berger said the company has chosen to "embrace the chaos" and take things quarter by quarter while keeping an eye on long-term return on investment. But Dekate and LinkedIn's own executives warned that the flat-spending strategy could face limits: the hardware demands of AI are shifting rapidly, and at some point, either AI ambitions or spending mandates may have to give.

LinkedIn's decision to hold its data center spending flat stands out sharply against the broader tech industry's infrastructure race. While OpenAI, Meta, and Google are scrambling to secure funding and form partnerships to build massive new data centers, LinkedIn is pushing back against what it calls unsustainable spending. The company's move is grounded in a specific technical achievement: over six months, LinkedIn's engineering teams doubled the efficiency of their existing GPU fleet, effectively getting twice as much work from the same hardware.

This efficiency gain did not happen by accident. LinkedIn spent years building its own data centers in Oregon, Texas, and Virginia after finding that Microsoft Azure (its parent company) did not make economic sense for the scale of the professional network. Once in control of its own infrastructure, the company invested heavily in measurement, optimization, and redesign. Teams developed tools to track how much compute and storage individual projects used, built allocation systems to reduce idle hardware, and deployed smaller models trained through distillation from larger ones. The result was not just cost savings—the company estimates about $24 million（約38億円） over 12 months—but also maintained or improved product quality, such as job recommendation tools that help users discover openings they would not have found before.

LinkedIn's approach reflects a broader industry shift sometimes called "tokenomics," where enterprises move from a "buy more to save more" mindset to a "do more with what you have" mindset. Gartner's Chirag Dekate notes that even smaller businesses without LinkedIn's infrastructure control are finding their own cost-reduction strategies. However, both LinkedIn's executives and industry observers caution that the flat-spending model faces limits. As Dekate warns, "at some point, something has to give"—either AI ambitions will be compromised, or spending mandates will have to shift. LinkedIn's executives have publicly stated that hardware prices are rising sharply (some servers have jumped three-fold in recent months), and the relentless growth in AI compute demands may eventually force the company to increase its budgets again.