A genuinely elite ex-DeepMind team raised a Sequoia-led Series A at a $300M valuation to sell “continual learning.” The pedigree is real and the timing is perfect — but the product runs on other companies’ open-weight models and other companies’ rented compute, and its own pitch, “own your value function,” hands the data moat to the customer. Even the headline raise is unsettled.
The number isn’t stable. TechCrunch reports $45M, The Information $40M, and Trajectory’s own site “$60M Series A and Seed.” The cleanest read is a $45M Series A on top of a $15M seed — ~$60M total — at a $300M post-money. ProofStory does not treat any single raise figure as settled fact.
By a cofounder’s own admission, customer models retrain roughly weekly and are frozen in between — batch post-training, not the always-on learning the name implies. The “beats frontier labs on narrow tasks” claim ships with no public benchmarks.
Trajectory owns no base model and, by design, no cross-customer data. Its pitch — “own your value function” — deliberately gives the data advantage to the customer. The team is world-class; the defensibility layer underneath it is thin and borrowed.
Key Finding: Trajectory pairs an unusually strong team (DeepMind, Apple, Windsurf acqui-hire; Jeff Dean and Fei-Fei Li as seed angels) with a top-of-cycle theme. But it sits on two borrowed foundations — open weights it doesn’t control and compute it rents — while explicitly disclaiming the one asset, a cross-customer data network effect, that would normally create a moat. The valuation is running ahead of a waitlist-stage product with no disclosed revenue.
Trajectory captures the signal that agents normally throw away — edits, retries, corrections — and folds it back into a customer’s model through reinforcement learning. The pipeline is elegant. The question is what Trajectory actually owns at each step.
Real product-usage traces — edits, retries, corrections — from the customer’s own app. Owned by the customer.
The customer defines “what good looks like.” Trajectory’s layer, but the value function belongs to the client by design.
Nvidia Nemotron or Zhipu’s GLM — third-party open weights. Not Trajectory’s.
Runs on rented sandbox/compute (Runloop, Daytona). Retrain cadence is ~weekly, not continuous.
Customer approves updates before production. The improved model — and its data edge — stays with the customer.
Trajectory’s genuine IP is the orchestration — trace capture, reward specification, and an SDPO-style update algorithm. That is real and hard. But it is a thin layer sandwiched between open weights it doesn’t control and GPU infrastructure it rents, serving customers who keep the resulting data advantage.
The brand promises models that never stop evolving. The mechanism, per cofounder Michael Elabd, retrains customer models as often as every week and leaves them static in between — batch post-training on a cadence. WIRED notes critics could argue the company “hasn’t built true continual learning in the traditional sense.” Keeping a human in the approval loop is a real safety choice, but it is also an admission that this is periodic retraining, not live adaptation. And the headline performance claim — post-trained models “beat frontier labs on narrow tasks” — comes with no public benchmarks.
Strong logos line Trajectory’s site — Harvey, Rogo, Decagon, Clay, Mistral, Nvidia. Two things complicate them: the lists don’t match across sources, and the anchor name wears four hats at once.
Nvidia is investor, compute supplier, base-model supplier, and listed customer. Nvidia funds Trajectory; Trajectory post-trains Nvidia’s Nemotron; Nvidia’s CEO appears as a testimonial; and Nvidia is counted among the customers. Traction from such a backer is not arm’s-length, and inflates how independent the demand looks.
TechCrunch lists Harvey, Rogo, Decagon, Mercor, Nvidia; the website shows Clay, Mistral, Harvey, Decagon, Rogo; WIRED names only Decagon, Clay, Harvey. Mercor and Nvidia-as-customer appear only in TechCrunch. Customer lists are being tailored to the audience — a sign to discount the roster until revenue is disclosed.
For regulated buyers (Harvey in law, Rogo in finance), the unexamined exposure is privacy. Trajectory claims it samples data distributions rather than training on raw customer inputs, but there is no public subprocessor list and the privacy-policy text wasn’t retrievable. The “we don’t train on your raw data” method is material, legally untested, and unverified.
Six structural risks the $300M valuation does not resolve.
The entire product starts from third-party open weights — Nvidia Nemotron, Zhipu’s GLM. If open-weight releases slow, license terms tighten, or policy restricts Chinese-lab models, the foundation under Trajectory can shift without warning.
By design, the customer owns the value function and keeps the improved model. Trajectory accrues no compounding proprietary dataset across accounts — the one asset that normally creates defensibility is deliberately given away.
Frontier labs shipping native reinforcement fine-tuning would commoditize this layer overnight. So could the customers themselves — Harvey, Clay, Decagon are capable AI teams who could build post-training in-house.
Regulated customers in law and finance, no public subprocessor list, and an unproven “don’t train on raw data” claim. A material, untested liability for exactly the marquee accounts.
Investor + compute + base model + “customer” in one entity inflates apparent traction and makes demand look more independent than it is.
A ~3x markup ($115M→$300M) in months, no disclosed revenue, 11–18 employees, a waitlist-stage product, and “beats frontier” claims with no benchmarks.
Trajectory is a brilliant team renting its foundations. The ex-DeepMind pedigree and Sequoia syndicate are the real assets; the technology — trace capture and RL orchestration — is genuinely hard. But the company builds on open weights it doesn’t own, compute it doesn’t own, and a value function it deliberately gives to the customer, while a weekly retrain is marketed as “continual learning.” At $300M with no revenue disclosed and no benchmarks, the price is paying for the people — not yet for a moat.
Based entirely on publicly available information, including the TechCrunch announcement of November 2, 2026. The raise amount is contested across sources and no revenue figure is public; figures are labeled accordingly.
We build and maintain websites, help businesses of every size put AI to work, and run growth marketing that’s held to the same standard as our research.
