AINEWS
LinkedIn's two CTOs say the 1.3 billion user platform will run flat on compute and storage from July 2026 to June 2027, fueled by 2x GPU efficiency.
LinkedIn told its engineering teams they would not get a single new GPU for the fiscal year that began in July 2026. Six months later, those teams had made the existing chips twice as efficient. The company now plans to hold compute and storage flat through June 2027 while continuing to ship generative AI features, according to its two chief technology officers.
In a WIRED report published July 29, Erran Berger said the hardware demands of AI are shifting rapidly, and the company is choosing to use what it has rather than chase the buildout arms race. Raghu Hiremagalur called the flat-spend posture a non-trivial engineering commitment. "No incremental storage and compute for a full year is no small feat," he told WIRED, describing the constraint as a forcing function for kernel-level and inference-stack work.
The stance makes LinkedIn one of the largest platforms to publicly diverge from the rest of big tech. OpenAI, Meta, and Google are all racing to fund, partner, and staff massive AI data centers, WIRED reports, while coping with labor and parts shortages that are already pushing customer-usage limits. LinkedIn says it has found another way through: a 2x improvement in how hard its existing GPUs work over the past six months. The platform counts more than 1.3 billion users, which gives the company a usage base smaller platforms cannot match.
The mechanism sits in code, not hardware. LinkedIn Engineering has open-sourced Liger-Kernel, a Triton-based kernel stack for efficient LLM training. Triton kernels are small, hand-tuned GPU programs that squeeze more work out of each chip without waiting for new silicon. The kernel work, combined with inference-side efficiency gains, is what made the 2x figure real, Berger said.
The financial backdrop is favorable. Microsoft, LinkedIn's parent, reported Q4 FY2026 earnings on July 29, 2026, with LinkedIn revenue rising and the Microsoft CEO calling out site-level agentic AI as a growth lever. Agentic AI, where the model takes multi-step actions on a user's behalf, is typically more compute-intensive per interaction than ranking or feed generation, which is why a parent-company call-out is the kind of commitment that would normally pull a platform toward buying more hardware. The flat-compute posture means LinkedIn has to ship that category of feature on the same GPUs it already has. Staffing Industry Analysts covered the same earnings call, noting the agentic-AI emphasis.
The constraint is also time-bounded. The fiscal year ends in June 2027, and Berger acknowledged that the engineering calculus could change if hardware demand patterns shift. If the AI capex arms race accelerates again, LinkedIn's discipline is reversible on a calendar, not a doctrine.
Skeptics in the industry are watching, not cheering. Songyee Yoon called LinkedIn's stance "encouraging for the industry" in the WIRED report, the kind of qualified praise insiders give a credible but unproven bet. Whether the 2x efficiency figure can be replicated across more demanding model classes, or whether the flat-spend pledge will hold once a high-value AI feature requires new hardware, remains to be tested.
The watch item is June 2027. If LinkedIn enters fiscal year 2028 still running flat compute, the constraint will have been vindicated. If it caves by Q2, the experiment will have run out of road.