LinkedIn is taking a different path from the aggressive AI infrastructure buildout seen across much of big tech. The professional social network says it does not plan to expand its AI data centers this fiscal year, even as it prepares to ship more generative AI features.
Executives told WIRED that LinkedIn intends to keep GPU investment steady while holding its compute and storage footprint flat. The plan covers LinkedIn’s fiscal year that began last month and ends next June.
A Bet on Efficiency Over Expansion
The core claim behind LinkedIn’s plan is simple: the company believes it can do more with the hardware it already has. LinkedIn says it found ways to use its existing GPUs twice as efficiently over the past six months.
That matters because AI features can quickly raise the cost of serving each user interaction. LinkedIn has been building AI-based assistants that help users write messages, find jobs, and recruit candidates. According to Raghu Hiremagalur, LinkedIn’s chief technology officer for infrastructure, the amount of data LinkedIn stored was doubling annually, and the cost of queries had risen over time.
“That is not a sustainable place to be.”
Erran Berger, LinkedIn’s chief technology officer for engineering, framed the goal as keeping the company’s infrastructure from growing while product demands continue to rise.
“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,” says Erran Berger, LinkedIn’s chief technology officer for engineering. “That’s a pretty bold statement to make in today's world.”
Why LinkedIn Has More Control Than Many Companies
LinkedIn’s infrastructure strategy changed after Microsoft acquired the company in 2016. A few years later, LinkedIn tried moving to Microsoft Azure, but the company concluded that moving a giant social network into general-purpose data centers did not make economic sense.
In 2022, LinkedIn went all-in on its own data centers in Oregon, Texas, and Virginia. That gave the company more direct control over the systems running its products. It also positioned LinkedIn to tune its infrastructure for its own AI workload rather than depending entirely on general cloud capacity.
That control is central to the current plan. LinkedIn has more than 1.3 billion users, and its decision to avoid a near-term data center expansion stands out at a time when companies such as OpenAI, Meta, and Google are spending heavily on AI infrastructure.
The broader industry context is difficult. The source article notes labor and parts shortages, limits on customer usage of some AI tools, and rising questions about whether constant AI investment is sustainable. LinkedIn’s approach does not reject AI expansion, but it does challenge the idea that growth must always begin with more hardware.
How LinkedIn Is Stretching Its AI Hardware
LinkedIn’s efficiency work spans the AI pipeline, from training models to serving them in response to user queries. Hiremagalur’s team built measurement tools to see how much compute and storage individual teams were using. It also created a system to assign projects to machines more efficiently so that computers sat idle less often.
Hiremagalur described GPU allocation and utilization on the training side as north of 95 percent. That kind of utilization is one reason LinkedIn believes it can hold the line on capacity while continuing to deploy AI products.
The company has also used distillation, a technique in which smaller AI models are trained from larger ones. For job recommendations, LinkedIn used a single model that learned from two larger models. The model was designed to identify relevant openings and predict which users were likely to click on them.
Berger says the smaller model is cheaper to operate without sacrificing quality. He said people are finding jobs they had not found before because the model better understands their desires.
LinkedIn also reduced the cost of the model that decides which posts appear in users’ newsfeeds. Berger said the company made dozens of changes, including streamlining training, reusing information from earlier recommendations, and balancing work more effectively between CPUs and GPUs.
- Measure compute and storage use by team.
- Allocate projects so machines are idle less often.
- Use smaller models where they can preserve quality.
- Move some tasks from Nvidia GPUs to CPUs.
- Reuse recommendation data when possible.
The Savings Are Real, but the Stakes Are Larger
LinkedIn says its efficiency work has saved about $24 million over the past 12 months. The company compares that to roughly 1,100 GPUs running around the clock for a year.
Those savings are modest next to LinkedIn’s $18 billion in annual sales, and executives acknowledge that. But Hiremagalur argues that engineering craft and agility matter too. When teams free up compute, they can start new projects sooner and add more AI capability without increasing the overall computing footprint.
The company has not frozen its hardware entirely. LinkedIn has committed to purchasing new servers so it can keep upgrading machines as they age out or break down over the coming months. It also bought ahead to lock in some savings as hardware costs rose.
“The cost of all of this hardware has just gone through the roof,” Hiremagalur says, noting some servers have jumped three-fold in price in the past few months. “It's just nuts.”
What This Signals for AI Spending
LinkedIn’s plan could still change. The source article notes that AI hardware demands are shifting quickly, even though executives say they have already considered surging prices for memory chips.
Still, the move reflects a broader focus on what the article calls “tokenomics,” or deeper analysis of the cost of using generative AI tools. Chirag Dekate, who helps businesses think through AI cloud strategies for Gartner, described a shift from buying more infrastructure toward doing more with what companies already have.
“Enterprises are evolving from a buy-more era to a do-more era,” says Chirag Dekate, who helps businesses think through their AI cloud strategies for the consultancy Gartner. “Until now, the mantra was, buy more to save more. But buy more only increases costs.”
Songyee Yoon, managing partner of Principal Venture Partners and a board member at the server maker HP, sees LinkedIn’s move as a sign that AI is moving from experimentation into production discipline.
That is the central lesson of LinkedIn’s decision. The company is not stepping away from generative AI. It is trying to prove that better utilization, model design, and software engineering can create room for more AI without immediately adding more data centers.