TL;DR

Baseten's $1.5B Series F and Together AI's $800M Series C confirm the serving layer is where AI infrastructure capital now concentrates, and inference is a margin-structure bet.

Two rounds inside ten days confirmed that the serving layer, not training, is where AI infrastructure capital is now concentrating. On June 22, Baseten announced a $1.5 billion Series F for its AI inference platform, led by Altimeter Capital, Conviction and Spark Capital with Sands Capital and Wellington Management as co-leads, reportedly structured across two tranches at $13 billion and $11 billion valuations. On July 1, Together AI closed an $800 million Series C at an $8.3 billion post-money valuation for its cloud platform for training, fine-tuning and running open-source models.

The common driver is the shift from experimentation to production agents: enterprise spend is migrating toward latency, reliability and cost-per-token at scale. The investor composition is equally telling, Wellington and Sands are public-market allocators, making Baseten's round function effectively as a pre-IPO raise.

Together's timing carries an extra edge. June's 18-day export-control suspension of Anthropic's frontier models gave every enterprise buyer a live demonstration of single-vendor dependency risk, and self-hosted open-weight deployments, Together's core pitch, are the cleanest hedge available. That control argument may now matter more than the cost argument, which closed-model price cuts keep eroding.

My View on This

inference is a margin-structure bet, not a technology bet. Winners will be decided by utilization economics, how efficiently platforms sweat heterogeneous GPU fleets against bursty demand, rather than raw benchmark speed. Baseten's two-tranche structure also hints at a market still negotiating what this category is worth; diligence should focus on gross-margin durability once hyperscalers compress serving prices.

Sources