Prime Intellect

Overview

Prime Intellect sells the pieces a company needs to post-train its own models with reinforcement learning. The stack runs from GPU capacity aggregated across dozens of providers, through a public hub of RL environments and an open-source asynchronous RL trainer, to hosted training, evaluation and deployment of the resulting models. The company began in 2024 as a decentralized training project that pooled GPUs across continents to train open models, and it still releases open models and open code. Its revenue, though, comes from a more ordinary place. Companies such as Ramp, Zapier and Mixedbread use it to turn an open-weight model into a specialist for one task. In July 2026 it raised $130 million at a $1 billion valuation and reported more than $100 million in annualized revenue and more than 6,000 customers.

Founding date per Wikipedia and TechCrunch. Sacra and some crypto press give 2023. Sacra’s last headcount estimate was 23 full-time staff, which predates the Series A.

Thesis

For most of the last three years a company that wanted better AI performance had one practical move, which was to call a better frontier model and improve the prompt. Training a model of its own meant pre-training, and pre-training was a capital contest that only a handful of labs could enter. Reinforcement learning changed the arithmetic. A capable open-weight base model is now free to download, and RL against a well-specified task can push it past a frontier generalist on that task. Ramp’s spreadsheet agent, discussed below, is the company’s favorite example, a 35 billion parameter open model that beat Claude Opus 4.6 on Ramp’s own evaluation after about 26 hours of RL training.

The obstacle is that RL is an infrastructure problem. A run needs a trainer that keeps GPUs busy while slow agent rollouts complete, an inference fleet, sandboxes in which agents can act safely, environments that define and score the task, and evaluations. The frontier labs built all of this in-house. Epoch AI reported in January 2026 that labs were paying $200 to $2,000 per RL task, about $20,000 per replicated website and around $300,000 for a complex product clone, with some vendor contracts running to seven figures a quarter, and TechCrunch reported that Anthropic had discussed spending more than $1 billion on environments over a year. Outside the labs, almost no company can assemble that stack.

Prime Intellect’s bet is that the stack can be sold. Its investors frame the pitch as ownership. Dell Technologies Capital’s investment note says intelligence “should be owned, not rented,” and Radical Ventures describes the product as owning the learning loop. The commercial version is narrower. A company that sends millions of calls a month to a frontier API for one well-defined job can train a smaller open model that does that job better, faster and cheaper, and keep improving it on its own production traces.

The thesis rests on two conditions. The first is that post-training remains worth doing, which OpenAI publicly doubted in May 2026 when it began winding down its fine-tuning platform on the grounds that newer models make most fine-tuning unnecessary. The second is that the value accrues to a software layer rather than to whoever sells the GPUs underneath it.

Founding Story

Vincent Weisser and Johannes Hagemann founded Prime Intellect in January 2024, according to Wikipedia. Weisser came from decentralized science. He was a founding steward of VitaDAO, a collective that funds longevity research, and led AI work at the biotech funding platform Molecule. Hagemann was a machine learning engineer at Aleph Alpha, the German foundation model company, where he worked on large-scale training, and had been a strategy advisor to VitaDAO, where the two first worked together.

The original idea was that if training could tolerate slow, unreliable links between machines, idle and fragmented GPU capacity could be pooled to train open models no single participant could afford. The company raised a $5.5 million seed in April 2024, co-led by the crypto funds Distributed Global and CoinFund, with Protocol Labs founder Juan Benet and Hugging Face chief executive Clem Delangue among the backers. Its first product was a compute marketplace aggregating GPUs from many providers.

The research came next and gave the company its reputation. In November 2024 it released INTELLECT-1, a 10 billion parameter model pre-trained on 1 trillion tokens by 30 compute contributors across five countries and three continents, using a scaled version of DeepMind’s DiLoCo method with a reported 400-fold cut in communication bandwidth and 83% compute utilization across continents. In February 2025 it raised $15 million led by Founders Fund, with Menlo Ventures and angels including Andrej Karpathy, Tri Dao and Dylan Patel. In May 2025 it released INTELLECT-2, a 32 billion parameter reasoning model post-trained with RL on QwQ-32B, in which rollouts were generated by permissionless inference workers around the world and verified with the company’s TOPLOC hashing scheme.

INTELLECT-2 was the turn. To run RL across untrusted, distant machines, the team had built prime-rl, a fully asynchronous RL trainer, along with tools to broadcast weights and verify rollouts. Research lead Will Brown, previously an ML researcher at Morgan Stanley, had written verifiers, an open library for defining RL environments. Those tools were more generally useful than the decentralized setting that produced them. In August 2025 the company launched the Environments Hub, arguing that environments were becoming “fragmented, closed, and hard to share” as startups built them exclusively for closed labs. In November 2025 it released INTELLECT-3, trained on a conventional colocated cluster of 512 H200 GPUs. In February 2026 it launched Lab, the hosted product that turned the tools into a service.

The company still describes itself as building “the open superintelligence stack.” What changed is which part of that stack customers pay for.

Product

Prime Intellect sells modules that can be bought separately or combined. The open-source components are free and are the main route by which developers find the company. The hosted components are what it charges for.

Compute

The compute business aggregates GPU capacity from other providers. The website lists on-demand prices from $0.47 to $4.99 per GPU-hour across hardware generations, H100 spot capacity at $0.94 per hour, and reserved clusters of 512 B300 GPUs at about $5 per GPU-hour, sourced through a product it calls Liquid Reserved Clusters from more than 50 providers. Buyers can run single nodes or clusters of 256 or more GPUs with Slurm or Kubernetes. This is the original 2024 marketplace, and the component most exposed to competition.

Environments Hub and verifiers

An RL environment packages a dataset of tasks, the harness an agent uses to attempt them, such as tools, a sandbox and context management, and a rubric that scores the outcome. The verifiers library defines a standard format for all three, and the Hub lets anyone publish, version and download environments in that format. The company seeded the Hub with a bounty program paying $100 to $5,000 or more per environment, and reported 400 environments crowdsourced in the first two months. By the Lab launch in February 2026 it reported more than 1,000 environments from more than 250 creators and more than 100,000 downloads, and its homepage now lists more than 2,500.

In July 2026 the company also published about 365,000 agentic tasks, roughly 198,000 in software engineering, 28,600 in terminal use and 137,600 in search, standardized from 23 existing open datasets. That is useful repackaging rather than original work, and anyone can replicate it.

prime-rl

prime-rl, the asynchronous trainer built for INTELLECT-2, separates the trainer, the inference fleet that generates rollouts, and an orchestrator between them, so training does not wait on the slowest rollout. In June 2026 it reported RL on the 1 trillion parameter GLM-5.1 across 28 H200 nodes with rollouts up to 131,000 tokens and step times under five minutes, and in August it described moving GLM-5.2 weights from trainer to inference in about four seconds. The results are company-reported, but the code is public.

Sandboxes

Agentic RL needs somewhere for thousands of agents to run code or browse at once. The company sells sandboxes for this purpose and in September 2026 described a move to microVM isolation for them, a month after publishing research on an offline sandbox escape it had found.

Lab

Lab is the hosted product that ties the rest together. A customer picks an environment from the Hub or writes one with verifiers, picks an open base model, and runs RL training with LoRA adapters on Prime Intellect’s infrastructure, then evaluates and deploys the result on multi-tenant inference built on NVIDIA’s Dynamo stack. The company reported that the private beta completed more than 3,000 RL runs. Pricing is per token, with rates not published in the launch post. The roadmap lists supervised fine-tuning, full fine-tuning, online distillation and continual learning from production traces, and the homepage already leads with the last of these, “turn production traces into the next training run.”

Open models and research

The INTELLECT series remains the company’s public proof of competence. INTELLECT-3 is a 106 billion parameter mixture-of-experts model with 12 billion active parameters, built by supervised fine-tuning and RL on the GLM-4.5-Air base, and released with its weights, trainer, environments and training recipe. In August 2026 the company open-sourced Prime Agent, a self-improving coding and research harness.

INTELLECT-3 shows both what the product does and where it stops. Against the same base, post-training added six points on AIME 2025 and eight on LiveCodeBench, the kind of gain on verifiable tasks that customers are paying for. It barely moved knowledge-heavy benchmarks such as GPQA and Humanity’s Last Exam, and a larger open model, DeepSeek v3.2, still led on most of them. RL sharpens what a base model can already do. It does not supply knowledge the base lacks.

Customer

The company reports more than 6,000 customers, a figure that covers everyone from individual researchers renting a GPU to enterprises running training jobs. Radical Ventures calls them “6,000+ teams,” and the company describes them as AI startups, labs and enterprises. The count is company-reported, and no split between compute-only and training customers has been published.

The named customers fall into three groups.

The first is software companies with one high-volume, well-defined agent task. Ramp, the corporate card company, trained a retrieval subagent for spreadsheets. Its main agent was wasting 17.8% of its tool calls on retrieval overhead, so Ramp trained FastAsk on Qwen3.5-35B-A3B in a custom environment of 14 finance task types and got a model that scored 66.25% on its held-out set against 61.88% for Claude Opus 4.6, at close to Haiku 4.5 latency. Zapier used Lab and verifiers to turn its AutomationBench evaluation, 600 tasks drawn from a 9,000-app ecosystem, into a training environment, so that the same environment measured and improved its agents. Mixedbread, a search company, reported that its Toast 1 model beat Claude Fable 5 by about ten points on the OfficeQA Pro V2 retrieval benchmark at 27% of the cost and roughly twelve times the speed.

These results come from the customers and the vendor, on evaluations the customers built. They are also exactly the case the product is designed for, a narrow task with a scorable answer and enough volume that cost and latency matter.

The second group is AI research organizations. Goodfire, an interpretability company, used the Hub, verifiers and compute to study reward hacking across open models, reporting that 50% to 96% of rollouts showed it in capable models. Character.AI, Arcee, Browserbase, Standard Intelligence and Inception appear as customers or partners on the company’s site and in Radical’s note, and Dell Technologies Capital names Perplexity.

The third group is the open-source community that builds on the free libraries and publishes to the Hub. It pays little, but it is the top of the funnel for the other two.

Enera describes the enterprise sweet spot as a company spending more than $100,000 a year on frontier APIs with more than 10,000 monthly calls for a task with a clear success metric. That is an analyst heuristic rather than a company figure, but it matches the named cases.

Market Size

There is no clean market definition for what Prime Intellect sells, and the available figures cover different slices.

The narrowest published figure is Research and Markets’ estimate for custom LLM training platforms, $2.82 billion in 2025 and $3.52 billion in 2026, forecast to reach $8.63 billion by 2030 at 25.1% a year. On that basis Prime Intellect’s $100 million run rate would be close to 3% of the 2026 market, which is plausible for a leader in a young category and also suggests the forecast undercounts the category, since several rivals below report run rates of similar or larger size. The broader figure is the training share of cloud AI infrastructure, which Gartner put at $19.0 billion of a $42.3 billion total in 2026. Much of that is frontier-lab pre-training that Prime Intellect will never serve, but its compute business competes for the rest.

The investor view is the most bullish. Dell Technologies Capital partner Radhika Malik wrote that the market “was effectively zero” two years ago, is now “into the several billion dollars a year range,” and should grow tenfold in three to five years, an unsourced estimate from an investor in the round. Frontier-lab environment budgets are larger still, but those contracts are mostly exclusive and closed, and Prime Intellect’s open Hub does not compete for them.

Competition

Prime Intellect competes with at least four kinds of company, and each competes with a different part of its stack. The broad platforms compete for training and inference workloads, the custom-model firms compete for the enterprise buyer, general compute providers compete for GPU spend, and decentralized AI projects compete for the open-source research community.

Post-training platforms

Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, launched Tinker in October 2025, an API that lets researchers write their own training loops while Thinking Machines runs the distributed infrastructure, with LoRA fine-tuning and support for RL. It prices per million tokens across prefill, sampling and training, and training Qwen3-8B costs about $0.40 to $0.44 per million tokens. The company raised $2 billion at a $12 billion post-money valuation in July 2025 and was reported in September 2026 to be in talks at $40 billion or more, with at least a few hundred million dollars of annualized revenue. Tinker is the closest product to Lab, backed by more than ten times the capital, though its trainer is not open source.

Fireworks AI raised $1.5 billion at a $17.5 billion valuation in July 2026 and reported more than $1 billion in annualized revenue. Its core business is inference, but it sells reinforcement fine-tuning billed per GPU-hour alongside hosting, so a customer can train and serve in one place. Together AI, which Forrester lists among post-training tooling providers, sells similar fine-tuning on its own cloud.

Custom-model builders

Applied Compute was founded in 2025 by former OpenAI researchers Yash Patil, Rhythm Garg and Linden Li, who worked on Codex and o1. It trains custom models for each enterprise on its own data, closer to a service than a platform. It raised $80 million at $1.3 billion led by Kleiner Perkins in April 2026, and in August The Information reported it was in talks near $3 billion with about $50 million in annualized revenue. It runs its RL sandboxes on Modal. It sells the result where Prime Intellect sells the tools, which suits enterprises without ML teams.

OpenPipe, which built the open-source Agent Reinforcement Trainer, was acquired by CoreWeave in September 2025, after CoreWeave’s purchase of Weights and Biases. That put an RL toolkit inside a GPU cloud, Prime Intellect’s combination arrived at from the other direction.

Compute and serverless infrastructure

Modal raised $355 million at $4.65 billion in May 2026 and reported more than $300 million in ARR, up fivefold since September 2025. It is general-purpose serverless compute, but it markets sandboxes and elastic inference for RL and counts Applied Compute and Cognition as RL customers. Neoclouds such as CoreWeave and marketplaces such as SF Compute and Vast.ai sell the same GPUs that Prime Intellect resells. The hyperscalers are moving up the stack as well. Amazon Bedrock added reinforcement fine-tuning for open-weight models such as Qwen3-32B and gpt-oss-20b in February 2026, behind an OpenAI-compatible API.

Decentralized training projects

Nous Research is the closest analogue to Prime Intellect’s origins. It releases the open Hermes models and runs Psyche, a decentralized training network on Solana using its DisTrO communication method. It raised $50 million led by Paradigm in April 2025 at a $1 billion token valuation and a $75 million Series B at $1.5 billion in July 2026, per Blockchain Council. It competes for researchers and contributors more than for enterprise customers. Prime Intellect has launched no token, which Crypto Briefing notes despite its crypto-native early investors, and its commercial product now runs mainly on conventional clusters.

Frontier labs

The labs compete in two ways. They sell the models that Prime Intellect’s customers are trying to replace, and each generation narrows the gap a specialist model has to close. And they decide whether to sell customization at all. OpenAI’s reinforcement fine-tuning supported only o4-mini, and in May 2026 the company closed its fine-tuning platform to new users, with new jobs ending for all customers by January 2027. Its stated reason was that newer models such as GPT-5.5 follow instructions well enough that prompting is now “cheaper and faster” for most use cases. That removes a competitor, and it also tells enterprises that the largest lab thinks most of them do not need what Prime Intellect sells.

Business Model

Prime Intellect makes money in three ways, and it has disclosed the split for none of them.

The first is compute resale, billed per GPU-hour on demand or under reserved cluster contracts. The company buys capacity from more than 50 providers and resells it. Gross margin on GPU resale is typically thin, because the buyer can compare prices directly and the underlying product is identical. Reserved clusters can carry better economics, and the website models the profit a customer could make reselling idle reserved capacity, but the company has not published its own margins.

The second is hosted training and inference through Lab, billed per token. This is where the software premium lives. A customer paying per token for RL training is paying for the trainer, the environments, the sandboxes and the orchestration as well as the GPUs underneath, and a customer serving its trained LoRA adapter on Prime Intellect’s multi-tenant inference keeps paying for as long as the model is in production.

The third is enterprise work with larger customers, which the case studies suggest involves hands-on support. The Series A funds are going, per the company, to long-horizon agents, autonomous research and continual learning, which are the features that would keep an enterprise training on the platform after its first model ships.

The illustration below shows why the mix matters, using assumed margins rather than any disclosed figure.

Illustrative. Assumes 15% gross margin on compute resale and 60% on hosted training and inference. Neither figure is company-reported.

The same $100 million run rate supports very different businesses depending on the first column. The open-source libraries do not appear in the table, but they are what makes the software share possible. Developers learn verifiers and prime-rl for free, publish environments to the Hub, and later pay to run those environments at scale on the hosted service.

Traction

All figures are company-reported.

The revenue figure is the headline. The company’s Series A post and Menlo Times say demand grew to more than $100 million annualized “in less than a year,” and a newsletter analysis puts it at about nine months. Neither says when the clock started. Lab launched publicly in February 2026, five months before the round, and the compute marketplace has sold GPUs since 2024, so the growth likely includes compute revenue that predates the hosted training product. A run rate is also a snapshot, typically the latest month multiplied by twelve, and in a business with usage billing and large reserved-cluster contracts, one month can overstate the year.

The customer count needs the same care. Six thousand customers producing $100 million implies an average of under $17,000 a year, which means the base is mostly small accounts with a few large ones carrying the revenue. That is typical of developer platforms and says little about concentration, which is undisclosed.

The technical record is the more verifiable part. The INTELLECT models, prime-rl, verifiers and the Hub are public, and verifiers had about 4,200 GitHub stars according to RL List. The case studies from Ramp, Zapier, Mixedbread and Goodfire describe production workloads, and Ramp published its result on its own research site with training details.

Valuation

The Series A priced the company at $1 billion, about ten times its reported run rate. Total funding is a little over $150 million across three rounds, the $5.5 million seed in April 2024, the $15 million round in February 2025 and the $130 million Series A in July 2026. The earlier valuations were not disclosed.

The comparables sit well above it. Fireworks AI trades at about 17.5 times its reported $1 billion run rate, and Modal at about 15.5 times its reported $300 million ARR. Applied Compute’s reported talks near $3 billion on about $50 million of annualized revenue imply a multiple around 60. Thinking Machines was reported to be seeking $40 billion or more on a few hundred million dollars of revenue, a multiple in the high double digits at the least.

Against that set, ten times revenue for a company reportedly growing from near zero to $100 million in about a year is a low multiple. The likeliest explanation is the revenue mix. Investors pricing Fireworks and Modal are pricing inference and serverless compute with established margins. If most of Prime Intellect’s revenue is resold GPU time, a multiple near ten is closer to what neoclouds receive than to what software platforms receive, and the round would be pricing the company as a compute reseller with a promising software attachment. If the software share is large, the round looks inexpensive. The lead investor, Radical Ventures, and the corporate participants NVIDIA, Intel and Dell would have seen the split, and the public has not.

Key Opportunities

Continual learning from production traces

The homepage’s lead promise is turning production traces into the next training run. A customer that deploys a trained model on Prime Intellect’s inference, logs its traces there, and retrains on them every few weeks is running a loop that is hard to move. That would convert one-off training jobs into recurring revenue and give the company the retention that per-job fine-tuning lacks.

The open-source standard

If verifiers becomes the default format for RL environments and prime-rl the default open trainer, Prime Intellect’s hosted service becomes the easiest place to run them, in the way that open-source projects often anchor a managed service. The Hub’s growth to 2,500 listings and adoption by research groups such as Goodfire suggest that the format is spreading. A standard environment format is also the natural interface for a marketplace in which environment builders are paid when their work is used in training runs.

The open-model gap left by OpenAI

OpenAI’s exit from fine-tuning removes the easiest option for an enterprise that wanted a customized model without leaving a frontier vendor. Customers that still need customization now have to go to open models, which is the only kind Prime Intellect trains. Anthropic and Google still offer managed fine-tuning through Bedrock and Vertex, but on narrower terms than a platform built for RL on any open model.

Backers along the hardware chain

NVIDIA, Intel and Dell all invested in the Series A. For each of them, a large population of companies training their own models is more demand for GPUs, servers and on-premise clusters than a world in which a few labs sell API access. Dell in particular sells on-premise infrastructure to enterprises that want to keep training data in-house, a channel for a software stack that can run anywhere.

Key Risks

The revenue may be mostly compute

This is the strongest bear case. Prime Intellect began as a GPU marketplace, still sells GPUs by the hour, and has not said how much of its $100 million run rate is resale. Resold compute is a thin-margin business in which the company competes with CoreWeave, Lambda, SF Compute, Vast.ai, the hyperscalers and every other broker for the same chips, and in which customers leave for a lower price. If most of the revenue is compute, the $1 billion valuation rests on a software business much smaller than the headline suggests, and the growth rate says more about the GPU market in 2026 than about demand for RL tooling.

Frontier models may make specialization unnecessary

OpenAI’s stated reason for leaving fine-tuning was that newer models do most tasks well enough with prompting. Each frontier release shrinks the set of tasks on which a specialist open model can win, and it raises the bar a customer’s evaluation has to clear. Ramp’s FastAsk beat Opus 4.6 by about four points on a 128-example synthetic set. A later frontier model could erase a margin that size, leaving only cost and latency, which a cheaper frontier tier can also offer.

Better-capitalized platforms with the same product

Thinking Machines has more than ten times Prime Intellect’s capital and a directly comparable training API. Fireworks and Together already host the inference that trained models will run on and can bundle training into it. Amazon has put RL fine-tuning on Bedrock. Applied Compute, growing quickly, sells finished models to buyers who would rather not operate a stack at all. Open libraries can be run by any of them.

Open source cuts both ways

Because verifiers, prime-rl and the environments are open, a customer that learns the stack on Prime Intellect can move the same workload to another GPU provider without rewriting it. The 365,000-task agentic collection is repackaged public data. The company’s moat therefore rests on operating the hosted service better than others can operate the same code, which is a genuine but narrow advantage.

Unverified metrics

Revenue, the customer count, the Hub listing count and the case-study results are all company- or customer-reported. The run rate has no disclosed start date, definition or concentration, and the case-study evaluations were designed by the customers who ran them. None of the central numbers has been tested from outside.

Research ambitions against a platform business

The Series A priorities are long-horizon agents, recursive language models, autonomous AI research and continual learning. The revenue comes from narrow agents for well-defined tasks. Pursuing both risks underinvesting in the enterprise work that pays for the research.

Security at scale

Running thousands of agent rollouts that execute untrusted code is a hard isolation problem, and the company’s own sandbox escape research shows the problem is live. A breach in a multi-tenant environment holding enterprise data and weights would be serious for a company selling ownership and control.

Summary

Prime Intellect sells the infrastructure for training models with reinforcement learning, from GPUs and sandboxes to a public environment hub, an open asynchronous trainer and a hosted service that trains, evaluates and deploys open models. It started in 2024 as a decentralized training project, proved that approach with INTELLECT-1 and INTELLECT-2, and turned the tools that research produced into a commercial platform in 2025 and 2026.

As of September 2026 it had raised a little over $150 million, including a $130 million Series A at $1 billion led by Radical Ventures with NVIDIA, Intel, Dell and ICONIQ, and it reported more than $100 million in annualized revenue and more than 6,000 customers. Its strongest evidence is public. Ramp published a trained open model that beat Claude Opus 4.6 on its own task at close to Haiku latency, the INTELLECT models and training code are open, and the environment hub reports more than 2,500 listings.

The question is what kind of company the revenue describes. Prime Intellect still resells GPUs, has not disclosed how much of its run rate that accounts for, and is priced at about ten times revenue, well below inference and compute peers. It also faces rivals with far more capital selling similar training APIs, and a frontier lab that has publicly concluded most customers no longer need fine-tuning. Its advantages are an open stack that researchers already use and a head start in packaging RL for companies outside the labs.

If continual training on production traces becomes normal enterprise practice, a platform that owns the environment format, the trainer and the deployment has a strong position. If frontier models keep absorbing specialized tasks, Prime Intellect is a compute broker with an excellent open-source research arm. The published numbers do not yet settle which.