Mining Power Rental Service Provider Baseten is planning to finalize a $1.5 billion fundraise, pushing its dual-layer valuation to $13 billion.

According to DataBeat's Beating monitoring, AI inference service provider Baseten is finalizing a new $15 billion funding round. Compared to a $10 billion funding plan based on a $110 billion valuation discussed three months ago, the final funding size has been increased by $5 billion. The valuation has also been adjusted to a two-tier structure, with some investors subscribing at a $110 billion valuation and another part co-investing at a $130 billion valuation. This round is led by Altimeter Capital, Conviction, Spark Capital, Sands Capital, and Wellington Management, with Wellington Management making its first investment in the AI inference track.

Baseten builds a software layer on top of the computing power of 20 cloud service providers to help enterprises fine-tune and run open-source models. Current customers include the coding assistant Cursor, Mercor, and OpenEvidence. As open-source model performance from DeepSeek, Kimi, and NVIDIA Nemotron approaches that of closed-source cutting-edge models, enterprises are shifting towards a hybrid architecture to control costs. Baseten CEO Tuhin Srivastava stated that switching to open-source models for specific tasks costs only 30% of what closed-source models would cost.

The surge in open-source models has driven strong demand for underlying computing power, leading to a substantial increase in Baseten's revenue. By the end of the first quarter of 2026, annualized recurring revenue had grown from $2 billion at the beginning of the quarter to $6 billion, a 20-fold increase from the same period last year. However, Baseten relies on leasing GPUs from cloud giants and reselling them, facing the risk of squeezed profit margins. Despite facing profit concerns, traditionally cautious institutional investor Wellington Management still entered the game, completing its first investment in the AI inference market, seen in the industry as a signal of the enduring vitality of the computing intermediary track.