· Venture Capital Tracker · investment-strategies · 3 min read
Trajectory’s $40M Series A at $300M: Continual Learning Infra for Production Agents
Sequoia led Trajectory’s $40M Series A at a $300M post with NVIDIA and Bessemer — infrastructure so product agents improve from real usage, not frozen offline models.
Trajectory raised $40 million Series A at a $300 million post-money around August 17, 2026, led by Sequoia, with NVIDIA and Bessemer Venture Partners participating — per FinSMEs and Dealroom (The Information).
Unexpected truth: the company is not selling another chat model. It sells the learning loop so production agents stop being frozen software.
Key facts
| Field | Detail |
|---|---|
| Company | Trajectory (San Francisco; founded May 2026) |
| Founders | Ronak Malde, Michael Elabd (ex-DeepMind); Arjun Karanam (ex-Apple) — Dealroom |
| Round | $40M Series A @ $300M post |
| Lead | Sequoia |
| Participants | NVIDIA, Bessemer |
| Prior | ~$15M seed @ ~$115M post (Dealroom); Bessemer also on seed |
| Product thesis | Continual learning from traces/corrections; SDPO and related training stack |
| Named customers (site/press) | Clay, Decagon, Harvey |
Who uses the product — and for what job
Users: AI-native product companies that ship agents into real workflows (sales, support, legal) and need models/harnesses to improve from production mistakes.
Job: observe usage → capture privileged corrections → retrain/deploy — continuously — so quality rises without a giant offline research cycle each time.
Harvey’s public-facing partnership narrative (legal workflows on efficient models) is the archetype: regulated domains want cheaper, specialist performance that keeps learning.
Why now
- Tool-using agents fail in ways offline benchmarks miss; production feedback is the new training data.
- Closed-model API costs push teams toward customized open models and better harnesses — Trajectory sits on both jobs per Dealroom framing.
- Sequoia-led infra checks cluster where many vertical AI apps share the same learning bottleneck.
Why Sequoia / Bessemer / NVIDIA — portfolio fit
| Investor | Fit |
|---|---|
| Sequoia | Lead on AI infra platforms that become default plumbing for other startups |
| Bessemer | Cloud/dev-tools franchise; returning from seed |
| NVIDIA | Strategic alignment on post-training / NeMo-class workflows |
Likely founder rationale: raise from the infra franchise that other AI founders already trust — plus NVIDIA adjacency for training stack credibility — rather than a vertical-only healthcare or legal fund.
Competitive map
| Player | Difference |
|---|---|
| Classic MLOps / eval vendors | Observability and eval; less end-to-end continual learning productization |
| Frontier labs’ fine-tune APIs | Powerful but not always productized learning loops for every app team |
| Inference platforms (Fireworks) | Serve/run models; Trajectory emphasizes continual learning from production |
| In-house research teams | Only feasible for the largest AI apps |
When not to over-read
- Primary sourcing is FinSMEs / Dealroom / The Information — treat as reported until a company post mirrors figures.
- Customer names are design-partner style evidence, not disclosed ARR.
- Seed→Series A step-up in ~2 months is capital-cycle heat; diligence retention of learning gains.
Practical takeaway
- Founders (AI apps): If your agent quality plateaus after launch, continual-learning infra is becoming a buy-vs-build decision.
- Investors: Sequoia + Bessemer + NVIDIA is a high-signal infra syndicate — underwrite data rights and customer learning deltas.
- Operators: Compare Trajectory against building an internal post-training team before your third vertical agent ships.
Sources
Last updated:
Editorial note: AI tools assisted with research, structure, or drafting. Venture Capital Tracker retains human editorial responsibility for factual accuracy, relevance, and source quality before publication.