LinkedIn doubles GPU efficiency and freezes AI hardware spending

LinkedIn plans to keep its GPU investment, compute footprint, and storage capacity roughly flat through its next fiscal year, even as it adds more compute-intensive features. The company says it can do that because it has approximately doubled the efficiency of its existing GPUs in six months.

The improvement did not come from one breakthrough. LinkedIn attributes it to better GPU utilization, model distillation, workload allocation, and changes to how tasks are divided among training, inference, storage, and systems design. The company intends to use the resulting headroom for product development rather than expanding its data centers.

Why LinkedIn is holding hardware spending flat

Speaking to Wired, Erran Berger, LinkedIn’s engineering CTO, said the company expects to keep its compute footprint “flat or close to it” while putting more demanding features into production. Raghu Hiremagalur, LinkedIn’s CTO for infrastructure, described the decision as a major undertaking:

Recommended reading

AWS brings Superblocks vibe coding into private clouds

“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.”

The strategy is also a response to LinkedIn’s costs. Hiremagalur said the cost of serving each query had been rising steadily, while stored data was doubling every year. LinkedIn concluded that the cost curve was unsustainable and focused on bending it through efficiency improvements.

That approach is possible partly because LinkedIn operates its own data centers in Oregon, Texas, and Virginia. The company can instrument and optimize more of the infrastructure stack instead of relying on a cloud migration that did not work as planned.

The Azure decision that shaped the strategy

Microsoft acquired LinkedIn in December 2016 for $26.2 billion. In 2019, LinkedIn announced a plan, code-named Blueshift, to move its infrastructure to Azure. It quietly shelved the project in 2022, citing Azure’s demand pressures and choosing instead to scale its own infrastructure. Later reporting also found that LinkedIn’s internal tools did not transfer cleanly to Azure.

That decision was widely treated as a setback at the time. LinkedIn now presents ownership of its infrastructure as an advantage: controlling the full stack gives its engineers more visibility into utilization, workload placement, and the cost of running models.

The company’s position contrasts sharply with Microsoft’s broader infrastructure push. Microsoft reported $41 billion in capital expenditure in its most recent quarter, added 31 data centers during that quarter and 88 across the year, and expects to spend more than $50 billion in the current quarter. Much of Microsoft’s spending is driven by Azure customer demand, including capacity contracted for OpenAI, rather than LinkedIn’s internal workloads.

An outlier in the AI infrastructure boom

LinkedIn is not claiming to have eliminated the need for more hardware. Instead, it is arguing that a large platform can ship new generative AI features for a year without proportional growth in compute and storage. That makes it an outlier as the four largest US hyperscalers collectively commit roughly $600 billion to $700 billion in capital expenditure for the calendar year.

The result is a useful test of a widely assumed relationship: that more ambitious AI products necessarily require more infrastructure. If LinkedIn’s efficiency gains support its roadmap, that assumption becomes weaker. If they do not, the company’s flat-capacity plan will look less like a durable alternative and more like a temporary effort to contain a difficult cost curve.