> ## Content Index
> Fetch the complete content index at: https://www.theinvestorsociety.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Trajectory Raises US$40M Series A at US$300M Valuation to Keep Enterprise AI Agents Learning After Deployment
- URL: https://www.theinvestorsociety.com/trajectory-raises-us-40m-series-a-at-us-300m-valuation-to-keep-enterprise-ai-agents-learning-after-deployment/
- Published: 2026-08-18T00:00:37.000Z
- Updated: 2026-08-18T00:00:37.000Z
- Author: The Editor

Trajectory, a San Francisco company building what it calls a continual learning platform for AI agents, has raised US$40 million in Series A funding at a US$300 million post-money valuation. Sequoia Capital led the round, with participation from NVIDIA and Bessemer Venture Partners. The deal was first reported by The Information on August 11, 2026, and has been carried by trade outlets since. The company was founded in May 2026 by chief executive Ronak Malde and chief technology officer Michael Elabd, both former Google DeepMind researchers, together with Arjun Karanam, previously a researcher at Apple. Dealroom reports the company has 12 employees; the broader team is drawn from DeepMind, OpenAI, Apple, Meta Superintelligence Labs, Amazon AGI, Scale AI, Stripe and Figma. Trajectory says it will use the capital to scale its automated model-tuning infrastructure, expand research and engineering headcount, and build enterprise tooling that optimizes the programmatic harnesses through which AI agents call external tools.

The round comes roughly two months after a US$15 million seed at a US$115 million post-money valuation, led by Conviction, with Bessemer, Radical Ventures and BoxGroup participating alongside angel investors including Jeff Dean and Fei-Fei Li and founders from Notion, Dropbox, Braintrust and Hugging Face. That sequence puts total funding at about US$55 million and marks a valuation increase of roughly 2.6 times in two months. The product addresses what the founders describe as an experience gap: models keep getting more capable in the abstract but do not accumulate knowledge from how a specific company's users actually work. Trajectory's platform continuously post-trains open-weight models on real product usage, using a self-distillation approach that ingests human-in-the-loop corrections and feeds them back into the agent's future decision-making. Named customers are AI-native software companies rather than traditional enterprises: Clay, Decagon and Harvey per the funding announcement, with Mercor and Rogo added in a June partnership release. Neither revenue nor customer count has been disclosed.

## Market Context

The round sits on a genuine spending shift. Gartner forecasts worldwide AI-optimized infrastructure-as-a-service spending will grow 96% in 2026 to US$42 billion, with inference spending of US$23.3 billion overtaking training spending of US$19 billion for the first time — a signal that money is moving from building models toward running and improving them in production. As access to closed frontier models has grown more expensive, enterprises have increasingly customized open-weight models on their own data, which is the demand Trajectory and its peers are selling into.

The category is crowded and stratified. Industry maps place Trajectory in an RL-as-a-service tier alongside Applied Compute, cgft, Metis and Osmosis, sitting above a platform-infrastructure layer that includes Tinker, OpenPipe and Prime Intellect, and alongside a separate cluster of environment-building startups. The clearest benchmark is Applied Compute, founded a year earlier by three former OpenAI researchers. It priced at a US$1.3 billion post-money valuation in an April 2026 round led by Kleiner Perkins, has raised roughly US$160 million from Kleiner Perkins, Benchmark, Sequoia Capital and Lux Capital, generates around US$50 million in annualized revenue after roughly quadrupling in nine months, and was reported in August to be in talks at about US$3 billion with Elad Gil in discussions to lead. Its customers include DoorDash, Cognition and Mercor. Trajectory is competing for the same buyers at a fraction of the scale and, so far, without a disclosed revenue figure of its own.

## Signal

> "it's their first day on the job"

> **Arjun Karanam**, co-founder, Trajectory — describing how today's models behave in conversation despite steadily rising benchmark scores

Karanam's framing is the clearest statement of what the company is actually selling, and it deliberately sidesteps the capability race. His argument is that model intelligence and accumulated job experience are separate axes, and that the industry has been optimizing only the first. Whether that gap is a durable business or a temporary artifact of how models are currently deployed is the question the US$300 million valuation is betting on.

## Regional Relevance

**For the United States.** The deal is a compact illustration of how American AI capital is being allocated right now: a three-month-old, twelve-person company priced at US$300 million on the strength of research pedigree and a thesis about where the next bottleneck sits. Two features make it strategically legible beyond the dollar figure. First, NVIDIA's participation continues a pattern of the chipmaker taking positions in companies whose products increase demand for its hardware — continual post-training is, by design, a workload that consumes GPUs indefinitely rather than once. Second, the bet is explicitly on open-weight models as the enterprise default. If customization of open weights becomes the standard path for U.S. companies deploying AI, the economics of the closed frontier labs shift, since customization becomes a competitive layer someone else owns rather than an add-on they can bundle.

**For San Francisco and the wider category.** Every party to this transaction is within a few miles of one another: Trajectory in San Francisco, Sequoia and Bessemer on Sand Hill Road and in the city, and a founding team assembled from DeepMind, Apple, OpenAI and Meta's superintelligence unit. That concentration is the point — the round was priced on relationships and reputational signal, not on operating history, which is possible in very few places. Globally, the same cost pressure is pushing companies toward open-weight customization regardless of geography, which is why the tooling layer for post-training has drawn capital from Silicon Valley, Toronto and elsewhere at once. For markets outside the United States, the practical stake is whether this layer stays a service sold by a handful of Bay Area companies or becomes commoditized software that local teams can run on their own infrastructure.

## The Other Side

**What is being priced at US$300 million, exactly?** Trajectory was founded in May 2026 and has 12 employees. It has disclosed no revenue, no customer count beyond five named logos, no retention data and no benchmark results. Its valuation rose from US$115 million to US$300 million in roughly two months, on a base of two months of operating history. The instructive comparison is the category's revenue leader: Applied Compute reached a US$1.3 billion valuation with around US$50 million in annualized revenue and a year of operations, and its reported US$3 billion discussions follow a roughly fourfold revenue increase. Trajectory's price is not anchored to anything comparable — it is anchored to the founders' research credentials and to Sequoia's willingness to preempt. That is a legitimate strategy in a competitive market for AI talent, and preemptive pricing of research teams has produced real outcomes before. But it means the round conveys information about investor conviction and about competition among funds, not about product-market fit, and it should not be read as the latter.

**How much of the technology is Trajectory's own?** The company's own public statements complicate the moat question. Elabd, its CTO, appears in Thinking Machines Lab's marketing for Tinker describing that product as "core infrastructure for Trajectory's continual learning platform." A June partnership release states that Trajectory runs its continual learning workloads on Runloop's execution environments. And the self-distillation fine-tuning technique that underpins this approach was introduced in published academic work by Shenfeld, Damani, Hübotter and colleagues, supported by a Tinker research grant — not developed in-house. Trajectory's contribution, on the evidence available, is the automation, orchestration and productization layer that sits on top. That is a real and often valuable place to be. It is also structurally exposed: the platform-infrastructure providers underneath it, Tinker among them, have every incentive to move up the stack, and a company whose differentiation is workflow automation over publicly documented methods has a narrower defensive position than one that owns the underlying research. Trajectory has not published its own benchmarks or a technical paper establishing that its implementation outperforms alternatives.

**Whose money is really funding this, and how correlated is the customer base?** Two structural features deserve scrutiny rather than accusation. NVIDIA is an investor in a company whose entire value proposition is that models should be retrained continuously — a workload that consumes NVIDIA silicon on an ongoing basis. This is a legitimate and disclosed strategic investment, and NVIDIA has made many, but it is the kind of arrangement where the investor's return does not depend solely on the startup's equity performing. Separately, Sequoia both led this round and publishes founder interviews promoting the continual-learning thesis, which is normal venture practice but means some of the most prominent third-party validation of the idea comes from a party holding the position. On the demand side, the named customers — Clay, Decagon, Harvey, Mercor, Rogo — are all venture-funded AI-native startups. That is a fast-moving, well-capitalized buyer set with unusually short procurement cycles, which is why early revenue in this category can look excellent. It is also highly correlated: these companies' own budgets depend on continued venture funding for AI applications, so a slowdown in that market would hit Trajectory's customer base simultaneously rather than gradually. Traditional enterprise customers, the buyers who would prove durability, are not yet on the list.

## Sources & Transparency

- [Trajectory Raises $40M in Series A Funding at $300M Post-Money Valuation](https://www.finsmes.com/2026/08/trajectory-raises-40m-in-series-a-funding-at-300m-post-money-valuation.html?ref=theinvestorsociety.com)
- [Trajectory, Founded by Ex-Google and Apple Researchers, Raises Funding From Sequoia in Back-to-Back Round](https://www.theinformation.com/newsletters/ai-agenda/trajectory-founded-ex-google-apple-researchers-raises-funding-sequoia-back-back-round?ref=theinvestorsociety.com)
- [Trajectory raises $40M Series A at $300M valuation](https://dealroom.co/news/144435-trajectory-raises-40m-series-a-at-300m-valuation/?ref=theinvestorsociety.com)
- [Trajectory raises $40M at $300M valuation in Sequoia-led round](https://cryptobriefing.com/trajectory-sequoia-funding-300m-valuation/?ref=theinvestorsociety.com)
- [Applied Compute in talks to double valuation to $3B on open-source demand](https://dealroom.co/news/144691-applied-compute-in-talks-to-double-valuation-to-3b-on-open-source-demand/?ref=theinvestorsociety.com)
- [Applied Compute eyes $3 billion valuation as open-model demand grows](https://www.cryptopolitan.com/applied-compute-eyes-3-billion-valuation-as-open-model-demand-grows/?ref=theinvestorsociety.com)
- [AI Infrastructure Roadmap: Five frontiers for 2026](https://nextbigteng.substack.com/p/ai-infrastructure-roadmap-five-frontiers-for-2026)