LinkedIn has replaced some of its OpenAI models as part of an effort to stop AI server costs from climbing, according to The Information. The Microsoft-owned network says it has flattened the cost of its AI features, even though more usage normally means a bigger compute bill. The publicly visible part of the report doesn’t say which models LinkedIn dropped or what replaced them, so those details remain unclear.
The swap also fits a broader plan LinkedIn has already discussed in public. In coverage from this summer, including a WIRED interview, the company said it wouldn’t expand its AI data centers for about a year.
⚡ The quick version
- What changed: LinkedIn swapped out some OpenAI models, The Information reports.
- The goal: Keep server costs flat while AI features keep running and growing.
- The hardware plan: Keep GPU investment and overall compute and storage roughly flat for about a year.
- The efficiency gain: LinkedIn says it roughly doubled the output of its existing GPUs in about six months.
- The catch: The plan isn’t a permanent commitment. LinkedIn’s Erran Berger said the company will take it quarter by quarter.
💸 Why this matters
AI features have an awkward cost structure. Traditional software gets cheaper to serve as it scales, but every AI request burns fresh compute. As The Information frames it, the more people use AI inside an app, the more that app’s maker spends on servers, and that can squeeze margins.
What stands out here is who’s making the move. Microsoft is one of OpenAI’s closest partners, and LinkedIn still decided that some of its workloads didn’t need to stay on OpenAI models. If a company with that kind of relationship is willing to swap, nobody should treat their model choices as locked in.
It’s also a contrast with the wider industry. Commentary on LinkedIn’s plan points to the largest US hyperscalers committing somewhere around $600 billion to $700 billion in capital spending this year. LinkedIn is making the opposite bet: get more from what it already owns.
🔧 How LinkedIn got there
LinkedIn has said there wasn’t one big breakthrough behind the GPU gains. It credits a combination of smaller improvements:
- Better utilization: Keeping expensive GPUs busier instead of letting capacity sit idle.
- Model distillation: Training a smaller, cheaper model to imitate a larger one, which keeps much of the quality at a lower running cost.
- Smarter workload allocation: Sending each job to the hardware that makes the most sense for it.
The company plans to spend that headroom on new features rather than more hardware. That’s the part worth noticing: the savings aren’t just going to the bottom line, they’re funding the product roadmap.
🧭 What practitioners should take from it
We don’t know the specifics of LinkedIn’s model swap, but the general playbook is clear enough for any team shipping AI features:
- Audit your workloads by task. Many production jobs may not need a top-tier model. Find out which ones do.
- Measure cost per feature. A single monthly API total won’t show you where the money goes.
- Build for switching. If changing models means rewriting half your app, you’ve lost your negotiating power and your ability to save.
- Budget for the migration. A new model can mean new evaluations, prompt changes and more testing. Cheaper tokens don’t make that work free.
There’s a limit to how far LinkedIn’s example travels, though. It runs its own data centers in Oregon, Texas and Virginia, so GPU utilization gains flow straight back to the company. Smaller teams renting everything through an API can still route tasks to cheaper models, but they won’t have the same control over the hardware.
🔮 What comes next
Expect more AI buyers to treat models as interchangeable parts, with the expensive ones saved for the work that really needs them. That’s not great news for any lab counting on customers to stay put as usage grows.
The test for LinkedIn is whether costs stay flat as it adds new features. Berger’s quarter-by-quarter caveat leaves room for a change of course, so watch whether the freeze actually lasts the full year. More details on the model swap are available in The Information’s original report.