Trajectory Secures $45M Series A to Democratize Continual AI Model Training

Trajectory, a year‑old startup founded by three Stanford alumni with backgrounds at Google DeepMind and Apple, announced a $45 million Series A round led by Sequoia and joined by Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The deal values the company at $300 million.

General‑purpose models from OpenAI, Anthropic, or open‑source repositories do not learn from their own errors unless they are retrained. Enterprises typically bridge this gap by fine‑tuning, a process that demands costly research staff or forward‑deployed engineers.

Trajectory’s platform automates that loop. Customers connect an open‑source model to their applications, the system gathers performance feedback, and then retrains the model so it improves on specialized tasks — such as legal contract review or financial valuation — without requiring a dedicated ML team. Early adopters include Harvey, Rogo, Decagon, Mercor, and Nvidia.

Why it matters for GPU / AI infrastructure

Continual post‑training creates a steady stream of inference‑plus‑training workloads that run on GPU clusters. By lowering the expertise barrier, Trajectory expands the addressable market for GPU cloud providers and increases utilization of existing accelerator fleets, driving demand for scalable, on‑demand compute.

The startup faces rivals such as Thinking Machines Lab, Applied Compute, and Prime Intellect, but its founders’ deep research pedigree — DeepMind acqui‑hire, Apple generative modeling, and hires from OpenAI, Stripe, and Netflix — gives it a talent edge in turning continual learning into a production‑grade service.