Trajectory's $45M Bet on Who Approves AI Changes

In November 2026, Trajectory raised $45 million for model post-training tools aimed at nontechnical customers. Adapting a pretrained model takes less data and compute than creating one, so more people can change its behavior—and each change gives someone another decision to make before deployment.

Enterprise teams can gain access to a generally capable model. Their harder problem is controlling whether it behaves reliably on specific tasks. Frontier training remains concentrated, but the downstream opportunity is a workflow in which domain owners specify acceptable behavior, test it on real tasks, approve releases and use production feedback to guide the next change.

Frontier compute concentrates training, not every adaptation

Epoch AI estimated that Google controlled about 25% of global AI compute, operating roughly 3.8 million TPUs and 1.3 million GPUs. An enterprise does not reproduce that infrastructure to make a model better at its own work. It starts from pretrained weights: fine-tuning generally needs less data and compute than pretraining, while post-training can optimize for outcomes that token prediction alone does not specify.

Builders and providers have widened that lower-resource path in different ways. In 2024, Mistral introduced a fine-tuning SDK for workstations, servers and small data-center nodes. OpenAI expanded its Custom Model program with assisted fine-tuning, customer data pipelines and evaluation systems. The two offerings differ in how much work the provider takes on, but both treat a base model as a starting point rather than the finished enterprise product. Snorkel AI’s move toward human-and-agent data development and evaluation belongs to the same broader post-training infrastructure shift.

Google’s compute position still matters. A customer that can tune a model does not thereby gain the ability to create or materially advance a frontier base model. Cheaper adaptation distributes some decisions about behavior without distributing the underlying training capacity.

Evaluation becomes a release gate when tasks diverge

A public benchmark helps a team compare broad capabilities. It cannot decide whether a revised model meets that team’s rules for a particular job. The distinction appeared early: in 2023, AWS previewed Model Evaluation on Bedrock with human testers involved before use. As model progress put pressure on older benchmarks, researchers introduced harder tests including FrontierMath, Humanity’s Last Exam and RE-Bench; OpenAI later introduced 13 evaluations for chain-of-thought monitorability. Those efforts answer different questions from a customer’s acceptance test.

A production team needs to define its own test cases, the order in which tests run, thresholds for promotion and the person authorized to accept exceptions. When a model changes, the team needs to run that process again. Human review is not merely a way to catch an error a benchmark missed. For consequential decisions, a recorded checkpoint identifies who accepted the behavior and on what evidence. That is deployment accountability, not another leaderboard column.

Trajectory is selling access to the decision, not a new base model

Trajectory makes the change in intended user unusually clear. In May 2026, it raised a $15 million seed round at a $115 million post-money valuation to build continual-learning models based on user interactions. Its November $45 million Series A, led by Sequoia at a $300 million valuation, accompanied a description of post-training tools for nontechnical customers. Trajectory secured both rounds, but neither report establishes broad customer adoption or documents an approval workflow already in use.

Patronus AI approached another part of the problem. It built simulated environments for evaluating agents that execute multi-step tasks and raised a $50 million Series B in 2026. A domain expert assessing a sequence of actions needs to see more than whether the final answer sounds right: the test must expose what the agent did along the way. Trajectory’s stated nontechnical audience and Patronus’s task simulations point to distinct products, not a single established vendor category.

Bringing subject-matter experts into the process also changes what software must represent. A business domain has its own concepts, exceptions and authorization rules. A tuning interface that accepts examples but cannot preserve those rules leaves the hard part with the customer—or with a bespoke services team.

Deployment makes local constraints part of model quality

In August 2026, SAP, Capgemini, Sopra Steria and OVHcloud reported stronger AI demand as enterprises moved from experiments toward deployment. Their positions differ—software, services and cloud infrastructure—but production gives each a reason to care about validation inside a customer’s environment, not just model selection on a public score.

Some customers also want control over where a system runs. Chinese suppliers sold on-premises “AI-in-a-box” products; Huawei estimated in 2024 that China’s market for such machines would reach about $2.3 billion that year. That was an estimate, not a realized sales figure. Separately, LatticeFlow’s checker found compliance issues when it assessed prominent models against elements of the EU AI Act. A checker finding is not a legal ruling, but it shows why a buyer cannot infer suitability for every deployment from general model capability.

The practical work falls to the people who know the domain and bear its risks. They must decide which failures are tolerable, which cases require review and which changes cannot proceed without authorization. Model performance remains necessary; the customer’s rules determine whether that performance is usable.

Arena and Weave show that evaluation spending is uneven

Demand for that work does not prove that every evaluation vendor has a large standalone business. In April 2025, The Information reported modest revenue at several evaluation startups. Weights & Biases had reached $50 million in annual recurring revenue by December 2024, but its evaluation product Weave accounted for only 2% of it.

In June 2026, Arena said it had reached $100 million in annualized run-rate revenue eight months after launching AI Evaluations, a performance-analytics offering. Arena already operated a widely used model leaderboard, giving it a distribution surface Weave did not have in the same form. Arena reported a company run rate, not audited annual revenue or a disclosed revenue figure for AI Evaluations alone. The comparison supports a narrower conclusion than a category-wide boom: evaluation can attract substantial spending when its seller already has a place in model-selection decisions, while other products have struggled to monetize the task.

Platforms can bundle the loop unless customers own the record

Infrastructure and model providers have reasons to keep the work close. CoreWeave acquired Weights & Biases, bringing developer workflow tools alongside its infrastructure business. AWS already offers model evaluation; OpenAI offers assisted fine-tuning and evaluation systems. Anthropic has described model-spec midtraining as an additional stage between pretraining and fine-tuning. Independent vendors are not competing against labs that have abandoned adaptation.

A vendor’s durable position therefore depends less on possessing a tuning technique than on holding a customer’s task cases, review decisions and history of approved changes in a form that remains useful across models. If customers can move that record with them, they can choose a base model without surrendering control of its release process. If the record stays inside a model or cloud platform, the platform has a straightforward way to bundle the workflow. The lower cost of changing models opened this field; who retains the authority to approve a change determines its boundaries.