LinkedIn Swaps OpenAI Models; Its Own Ran on 2 GPUs vs 152

The Information reported on 3 November that LinkedIn swapped out OpenAI models to keep its server costs flat. In the report’s public opening, LinkedIn says it has “flattened the cost of running its AI-powered Hiring Assistant”, its recruiting agent, even as usage of the tool has steadily increased in the past year.

The public part of the report does not name the models that replaced OpenAI’s. LinkedIn’s own engineering blog shows the economics of one alternative. In December 2024 it said its EON-8B model, built on Llama 3.1-8B, was “75x and 6x cost effective in comparison to GPT-4 and GPT-4o respectively”.

LinkedIn’s chart puts that in GPUs. To serve one query per second of an interactive AI Premium experience, it counted 152 A100 GPUs for GPT-4, 12 for GPT-4o and 2 for EON-8B. By our arithmetic, that is 98.7% fewer GPUs than GPT-4.

What The Information Reported

The report, by Aaron Holmes in The Information’s Applied AI newsletter, was published at 18:31 UTC under the headline “LinkedIn Swapped Out OpenAI Models to Keep Its Server Costs Flat”.

It frames the problem this way: “The more people use AI features in apps, the more the companies selling those apps have to spend on servers to run them, potentially squeezing their profit margins.”

It describes Hiring Assistant as a tool that “uses large language models to automate candidate search and outreach for recruiters”. The rest of the article is behind The Information’s paywall, and we have not read it.

What LinkedIn’s Engineers Published

LinkedIn’s engineering post of 18 December 2024 describes EON, a family of domain-adapted foundation models. LinkedIn says it measured Llama, Mixtral and Mistral as base models before building them.

The post says LinkedIn launched Hiring Assistant in October 2024 for select charter customers. A key step is an evaluation agent that assesses candidate suitability against AI-extracted job requirements.

On that step, LinkedIn says EON-8B outperformed OpenAI’s GPT-4o mini and Llama-3-8B-instruct on candidate-job-requirements matching accuracy “by absolute 4% and 30% respectively, without further fine-tuning on the matching task”.

The GPU figures come from the post’s Figure 4: GPT-4 (March 2023) 152, GPT-4 Turbo (November 2023) 112, GPT-4o (May 2024) 12 and EON-8B (June 2024) 2. LinkedIn says the GPT counts were calculated from its “production deployment of GPT instances on Azure” and the EON count from on-premise deployment.

What Microsoft Said on 28 October

Microsoft owns LinkedIn. In his prepared remarks for the September quarter, Satya Nadella said “25,000 customers are using our agentic hiring solutions”.

He added that recruiters using Hiring Assistant “are finding interview-quality applicants 33 percent faster, while reviewing 83 percent fewer profiles to find a qualified match”.

Nadella also set out a first-party model goal: “delivering capabilities comparable to leading models within three to six months of their release at a fraction of the training cost”. His example: “MAI Cyber now powers over 40 percent of the scans in our agentic MDASH code scanner, at 18 percent lower cost per scan.”

OpenAI remains a large partner. Microsoft’s 10-Q says it holds an approximate 24% interest in OpenAI and recorded $10.0 billion of revenue from commercial arrangements with OpenAI in the quarter, revenue-sharing payments included.

The same 10-Q no longer shows LinkedIn as its own revenue line. LinkedIn Talent Solutions and Sales Solutions now sit in Industry solutions, and Marketing Solutions and Premium Subscriptions in Search and advertising. The last annual figure, in the fiscal 2026 10-K, was $19.8 billion, up 11%.

The Business Lever

As we read it, the lever is gross margin on an AI product whose use is growing. The Information’s report says LinkedIn held the cost of running Hiring Assistant flat while usage rose.

As we read LinkedIn’s 2024 chart, model choice sets much of that cost: by our arithmetic, the GPUs GPT-4o needed for one query per second could serve six with EON-8B.

Connecting the Coverage

Our related pieces:

• Microsoft Stops Reporting LinkedIn’s $19.8B Revenue Line: where LinkedIn now sits in Microsoft’s accounts.
• Microsoft Azure Grows 48% as Free Cash Flow Falls to $18.6B: Microsoft’s September quarter, including OpenAI.
• Microsoft Calls Its AI Stack ‘Model-Forward, Agent-First’: Microsoft’s own model strategy.
• GLM Agent Passes Anthropic’s Fable 5 at 30x Lower Task Cost: inference cost gaps between models.
• Match Group Filing: AI Tokens, Staff Lift Product Costs 14%: AI serving costs in another app company’s filing.
• Palantir: Tuned Nvidia Model Tops GPT-6 Astra at 1/200 Price: another tuned open model set against a frontier one.

The Structural Read

As we read LinkedIn’s 2024 chart, GPT-4o had already cut the GPU count from 152 to 12, and LinkedIn’s own EON-8B cut it to 2 for the same AI Premium experience.

As we read Microsoft’s remarks, it applies the same approach to its own products: Nadella described MAI Cyber running over 40 percent of scans in its MDASH scanner “at 18 percent lower cost per scan”.

As we read the 10-Q, OpenAI is still a large customer and partner for Microsoft, with $10.0 billion of commercial revenue in the quarter; LinkedIn’s choice is about one product’s serving cost.

Three Implications

FOR RECRUITERS Microsoft says recruiters using Hiring Assistant find interview-quality applicants 33 percent faster while reviewing 83 percent fewer profiles.

FOR MICROSOFT LinkedIn’s revenue no longer appears as its own line in the 10-Q, so the cost effect is not visible in Microsoft’s reported numbers.

FOR OPENAI The Information’s report describes a Microsoft-owned customer moving work off OpenAI models; Microsoft’s 10-Q shows $10.0 billion of quarterly revenue from OpenAI arrangements.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework argues: “The new job of finance is to manage the ratio between the two — work produced per token burned. That ratio is the gross-margin equation of the agentic era.”

As we read it, LinkedIn’s reported swap is that ratio managed from the cost side: the work Hiring Assistant does grew, and LinkedIn says the cost of running it did not. The framework is about AI economics in general; it does not discuss LinkedIn.

What Is Not Established

The public part of The Information’s report does not say which models replaced OpenAI’s in Hiring Assistant, when the swap happened, or what the costs and usage figures are. We have not read the paywalled text, so we do not know whether it does.

LinkedIn’s GPU comparison dates from 2024 and measures an AI Premium experience, not Hiring Assistant. It compares models available then, not current OpenAI models. Microsoft’s filings do not break out LinkedIn’s AI serving costs. We did not contact LinkedIn or Microsoft.

The Bottom Line

The Information reports that LinkedIn swapped out OpenAI models and kept Hiring Assistant’s running cost flat as usage grew. LinkedIn’s own 2024 figures show why model choice moves that cost: 152 A100 GPUs per query per second for GPT-4, 12 for GPT-4o and 2 for its Llama-based EON-8B. Microsoft says 25,000 customers now use its agentic hiring solutions.