LinkedIn says it won't be spending big on AI hardware this year — but it has a good reason why

LinkedIn has announced it will not increase its AI data center investments or expand its compute and storage footprint in the upcoming fiscal year, instead opting to maintain GPU investment flat. The company attributes this decision to its recent success in improving GPU efficiency by roughly 100% over six months, achieved through optimization techniques including utilization enhancement, model distillation, and workload allocation.

Despite being part of Microsoft since its acquisition in 2016, LinkedIn's approach diverges from that of its parent company, which is aggressively expanding its AI infrastructure. This strategy allows LinkedIn to focus on adding new features rather than purchasing additional hardware, making it a unique case among large corporations navigating AI development.

LinkedIn's journey to this point was not without challenges, as it initially considered migrating to Azure in 2019 but ultimately decided to build its data centers in Oregon, Texas, and Virginia, citing difficulties with transferring its in-house tooling to Azure. However, this decision has paid off, as LinkedIn's CTO for infrastructure, Raghu Hiremagalur, emphasizes that owning the full stack enables the company to instrument every layer and view efficiency as a continuous investment.

The cost of serving each query has been steadily increasing as stored data doubles annually, a situation LinkedIn views as unsustainable. While LinkedIn acknowledges that it has not solved any fundamental problems, its transparency about its compute constraint stands out in an industry where most companies focus on hardware growth.

The company's approach challenges the prevailing assumption that AI product ambition necessitates proportional hardware growth, suggesting that with continued efficiency gains, consumer platforms can achieve significant AI features without expanding their hardware footprint.