AI hardware costs are soaring, but LinkedIn is holding its data center investment steady through 2026. By optimizing GPU usage and resource allocation, the company aims to launch new generative AI features without expanding its compute footprint.
While tech giants race to expand their AI infrastructure, LinkedIn is taking a different approach this fiscal year. The professional networking platform is keeping its AI data center spending flat, despite rising hardware prices and growing demand for compute power. According to executives, LinkedIn has managed to double the efficiency of its existing GPUs over the past six months, allowing it to avoid the aggressive expansion seen at companies like OpenAI, Meta, and Google.
Erran Berger, LinkedIn’s chief technology officer for engineering, told WIRED that the company’s goal is to maintain a steady compute and storage footprint while still rolling out more compute-intensive AI features. “That’s a pretty bold statement to make in today's world,” Berger said. Raghu Hiremagalur, CTO for infrastructure, added that the new constraints are pushing engineering teams to find creative solutions as LinkedIn prepares to launch a wave of generative AI tools. The company’s fiscal year began last month and runs through next June.
LinkedIn’s strategy stands out as many competitors scramble to secure the latest chips and data center capacity, often forming unexpected partnerships to keep up with demand. Labor and parts shortages have delayed projects across the industry, and some businesses have even limited customer access to AI tools. The sustainability of relentless AI investment is increasingly questioned. Songyee Yoon, managing partner at Principal Venture Partners and HP board member, noted that LinkedIn’s move signals a shift from experimentation to production discipline in AI. “The companies that win will not simply be the ones that spend the most on infrastructure,” Yoon said.
Efficiency Over Expansion
After Microsoft acquired LinkedIn in 2016, the company experimented with moving to Azure cloud services. However, the economics didn’t work for LinkedIn’s scale, so by 2022, it committed to operating its own data centers in Oregon, Texas, and Virginia. This gave LinkedIn full control over its infrastructure, which proved crucial as the company developed AI-powered assistants for messaging, job search, and recruiting. The cost of each user query has risen over time, and the volume of stored data was doubling annually, according to Hiremagalur.
To address these challenges, LinkedIn optimized every stage of its AI pipeline, from model training to serving user queries. The infrastructure team built tools to track compute and storage usage by individual teams and implemented a system to allocate projects more efficiently, reducing idle time. Hiremagalur reported GPU utilization rates above 95% for training workloads. The company also adopted model distillation, training smaller models from larger ones to cut costs without sacrificing quality. For example, its job recommendation system now uses a compact model that learns from two larger models, improving both relevance and click prediction.
LinkedIn’s newsfeed ranking model, initially expensive to operate, was also streamlined. Berger said dozens of improvements were made, including more efficient model training, reusing previous recommendations, and balancing workloads between CPUs and GPUs. The company even modified foundational Nvidia software to handle larger tasks and shifted some workloads from GPUs to CPUs, which are less costly and easier to source.
Cost Savings and Industry Impact
These efficiency efforts have saved LinkedIn an estimated $24 million over the past year-the equivalent of running about 1,100 GPUs nonstop for twelve months. While this is a small fraction of LinkedIn’s $18 billion in annual revenue, Hiremagalur emphasized the value of agility and resource flexibility. By freeing up capacity, engineers can accelerate new projects and integrate more AI features without expanding the company’s data center footprint.
Despite holding spending flat, LinkedIn continues to upgrade its hardware, replacing aging servers as needed. The company secured some cost savings by purchasing equipment in advance, but Hiremagalur noted that server prices have tripled in recent months. This focus on efficiency reflects a broader industry trend known as “tokenomics,” where enterprises scrutinize the cost of generative AI usage. Chirag Dekate of Gartner observed that businesses are shifting from a “buy more” mentality to a “do more” approach, seeking to maximize output from existing infrastructure. Some smaller firms have responded by cutting unused software, reducing staff, and turning to alternative cloud providers.
However, Dekate cautioned that strict financial constraints could eventually force companies to compromise on either their AI ambitions or their IT spending mandates. LinkedIn is not ruling out future data center growth but is now allocating resources quarter by quarter, rather than relying on rough annual estimates. Berger said the company is “embracing the chaos” while keeping a close eye on long-term return on investment. The duration of this spending freeze remains uncertain.
Shifting AI Infrastructure Strategies
LinkedIn’s decision to prioritize efficiency over expansion comes as the industry debates the sustainability of massive AI investments. The company’s approach contrasts with the ongoing building boom among rivals, as highlighted in recent coverage of Google’s integration of Top Stories into AI Overviews for mobile users, which underscores the competitive pressure to deliver new AI-powered features at scale.
LinkedIn, founded in 2002 and acquired by Microsoft in 2016, now serves more than 1.3 billion users worldwide. The platform reported $18 billion in annual sales and operates major data centers in Oregon, Texas, and Virginia. Its engineering and infrastructure teams are led by Erran Berger and Raghu Hiremagalur, who have overseen the company’s shift toward maximizing GPU utilization and optimizing AI model deployment. As the company continues to roll out new generative AI features, its ability to maintain efficiency without sacrificing performance will be closely watched across the tech sector.
