Baseten Raises $1.5 Billion Series F at $13 Billion Valuation as Inference Demand Explodes

Baseten, a San Francisco company that helps businesses deploy and run AI models in production, has raised a $1.5 billion Series F at a $13 billion valuation — one of the largest infrastructure rounds of the year and a marker of where AI spending is heading next.

The round was backed by a deep bench of growth investors, including Altimeter Capital, Conviction Partners, Spark Capital, Sands Capital and Wellington Management. It topped a week of megadeals that also included a $1 billion Series E for marketing-analytics firm AppsFlyer and $650 million for AI inference chip company Groq.

From training to serving

Baseten's rise captures a structural shift in the AI economy. For three years, the biggest checks flowed to labs training ever-larger models. Now, as those models move into real products, the bottleneck — and the money — is moving to inference: the unglamorous, relentless work of running trained models fast, cheaply and reliably at scale.

That is Baseten's business. It abstracts away the hardware and orchestration needed to serve models in production, letting companies ship AI features without building their own GPU infrastructure. As enterprises push agentic and multimodal features into production, inference volume — and the cost of getting it wrong — is climbing steeply.

The bigger picture

The $13 billion valuation slots Baseten alongside a cohort of inference-focused infrastructure players — from Groq to Together AI to Fireworks — all betting that the next phase of the AI buildout is about deployment economics, not model bragging rights. With Chinese open models like GLM-5.2 and DeepSeek V4 driving inference costs toward zero and volumes toward the stratosphere, the companies that make serving efficient stand to capture a durable slice of the value.

For investors wary of frontier-model capex, inference infrastructure offers a picks-and-shovels bet: whoever wins the model race, someone still has to run the models.