Written by: KarenZ, Foresight News
An AI infrastructure company founded just over two years ago, on one hand announces support from investment arms of Nvidia, Intel, and Dell, and on the other claims its annualized revenue has already exceeded $100 million—these two figures combined are enough to make Prime Intellect one of the most noteworthy AI projects to re-examine recently.
July 8, 2026, the decentralized AI infrastructure network Prime Intellect announced the completion of a $130 million Series A funding round at a $1 billion valuation, led by AI-focused venture capital firm Radical Ventures, with rare joint participation from investment arms of Nvidia, Intel, and Dell, bringing its total funding raised to over $150 million.
While disclosing the massive funding, Prime Intellect officially announced that in less than a year, its annualized revenue (ARR) has rapidly jumped to over $100 million, and the platform serves more than 6,000 enterprise and startup clients.
What's the Background?
As mentioned in "OpenAI Founding Member Steps In! Quick Read on Decentralized AI Dark Horse Project Prime Intellect" in March 2025, Prime Intellect was founded in January 2024 by co-founders Vincent Weisser and Johannes Hagemann.
- CEO Vincent Weisser was previously long involved in the intersection of decentralized science (DeSci) and AI, having co-initiated projects like Bio Protocol, VitaDAO, and CryoDAO, and served as the Ecosystem and AI Lead at the DeSci platform Molecule.
- CTO Johannes Hagemann focused on distributed AI and semi-automated engineering, brain-computer interfaces, and previously worked as an AI Research Engineer at the German AI company Aleph Alpha.
Additionally, in October 2025, venture capitalist Ash Arora joined Prime Intellect as Head of Applied Go-to-Market (Applied GTM), responsible for product strategy, commercialization, revenue, and applied AI products in post-training processing and reinforcement learning. Ash Arora recently pointed out that Prime Intellect's full-time team size has now reached 40 people.
In terms of funding, Prime Intellect has raised over $150 million cumulatively. A $5.5 million seed round in April 2024 was co-led by Distributed Global and CoinFund, with angel investors including Hugging Face CEO Clem Delangue.
Less than a year later, in March 2025, Prime Intellect completed another $15 million funding round led by Peter Thiel's Founders Fund, with investors including OpenAI founding member and former Tesla AI Director Andrej Karpathy, Together.AI Chief Scientist Tri Dao, Stability AI co-founder Emad Mostaque, and other heavyweight figures in the AI field.
The latest round is different in nature. In the $130 million Series A round, NVIDIA Ventures, Intel Capital, and Dell Technologies Capital are not just financial investors; their parent companies hold key positions in GPU, CPU, server, and data center infrastructure respectively.
Intel Capital's explanation of this investment also indicates: The reason hardware giants are buying in is that Prime Intellect is attempting to bring underlying computation, training environments, evaluation, reinforcement learning post-training, and upstream inference together on a unified control plane.
What Are the Substantive Developments?
An early notable achievement of Prime Intellect was proving that long-distance, heterogeneous GPUs could also collaborate on training. Following its technical iterations over the past two years, one can see how the platform gradually transformed research experiments into commercial product lines.
In late November 2024, Prime Intellect released the 10-billion parameter model INTELLECT-1, with training nodes spanning five countries and three continents. The company claimed it achieved an overall compute utilization of 83% across continents at that time, and when training using only nodes distributed across the United States, compute utilization reached 96%.
Less than half a year later, Prime Intellect released INTELLECT-2, advancing the goal to globally distributed reinforcement learning with 32 billion parameters. To achieve this, the team developed the asynchronous reinforcement learning framework PRIME-RL, SHARDCAST for propagating model weights, and TOPLOC to verify if inference nodes are "working honestly."
A more critical change occurred with INTELLECT-3. In November 2025, Prime Intellect released a 106-billion parameter MoE model based on Zhipu GLM-4.5-Air, fine-tuned with supervision and reinforcement learning. The model was trained for about two months on 64 nodes with 512 NVIDIA H200 GPUs; model weights, training framework, data, RL environments, and evaluation methods were all open-sourced. The significance here is not just releasing another model, but the company validated an entire production system with its own research project: PRIME-RL handles asynchronous training, Verifiers and Environments Hub provide unified tools and a community ecosystem to build and host RL environments and evaluations, Prime Sandboxes isolate execution of agent-generated code, and the compute orchestration layer manages clusters, storage, and monitoring.
In February of this year, Prime Intellect launched a full-stack AI training platform called Prime Intellect Lab, specifically designed to help individuals, engineers, and AI companies train and optimize their own models (especially agentic models) without needing to build expensive GPU clusters themselves. On May 7th, Lab ended its beta and officially opened fully.
In June, Prime Intellect released prime-rl version 0.6.0, claiming to push the engineering limit to trillion-parameter scale MoE (Mixture of Experts) models. Prime Intellect disclosed that on GLM-5 series software engineering tasks, it could process sequences up to 131,000 tokens using 28 H200 nodes, with single-step training time under 5 minutes.
The key behind this is not a single algorithm, but the joint optimization of training and inference systems: the inference side uses FP8 low-precision computation and components like DeepEP and DeepGEMM to increase throughput; pre-filling and decoding are separated to avoid long tool outputs slowing down generation; KV Cache hierarchical offloading improves concurrency. The training side also adopts block-scaled FP8 and reduces routing discrepancies between MoE model training and inference via Router Replay, combined with FSDP, expert parallelism, and context parallelism. These optimizations ultimately impact GPU utilization, training time, and customer costs.
In July this year, prime-rl added a unified algorithm layer, built-in with six types of training methods: GRPO, MaxRL, On-Policy Distillation, self-distillation, SFT Distillation, and ECHO, and allows selecting different algorithms for different environments within the same training run. Simply put, the same agent can use one learning method for math tasks and another for terminal operation tasks without rewriting the underlying trainer. This moves Prime Intellect from "running training for clients" closer to a scalable RL operating system.
Hardware-Software Synergy: Nvidia is More Than Just an Investor
Looking at the Series A investor lineup, the binding between hardware giants and Prime Intellect goes beyond capital, extending deep into hardware-software architecture co-construction.
The collaboration between Prime Intellect and Nvidia spans both hardware and software layers. On the hardware side, its training and serving workloads already use NVIDIA Blackwell, Blackwell Ultra, and NVL72 rack-scale systems, which the company claims are more efficient than previous Hopper clusters.
On the software side, NVIDIA Dynamo is used for global inference orchestration, auto-scaling, request routing, and KV Cache offloading, and is integrated with Prime Intellect's large-scale LoRA (Low-Rank Adaptation, a fine-tuning technique for large language models) deployments.
Nvidia's own technical blog also confirms that Prime Intellect has deployed the NVIDIA Dynamo inference framework in its production workflows and participated in co-designing and integrating LoRA Adapter support.
Prime Intellect stated in March this year that it would test RL sandbox workloads around the NVIDIA Vera CPU and plans to migrate some sandboxes and provide GPU sandboxes on Vera Rubin systems once Vera is publicly available. The company's self-tests claim each Vera CPU socket can stably run 176 VMs in parallel; in its defined RL sandbox workloads, with multi-threading enabled, throughput is on average about 30% higher than the baseline of AMD Zen 5 with only physical cores enabled on AWS.
These numbers show potential cost advantages, but they currently come from collaborative testing between the parties, and the comparison environments are not identical, so they cannot be taken as independent general performance conclusions. References to Vera Rubin and GPU sandboxes should be stated as "planned adoption," not already large-scale commercial deployment.
Along with product maturity, real commercial monetization is occurring. According to Prime Intellect's disclosure, fintech company Ramp uses Prime Intellect Lab to train the retrieval sub-agent FastAsk for Ramp Labs: Ramp turned its AI spreadsheet editor Ramp Sheets into a trainable RL environment, then performed reinforcement learning training based on the Qwen3.5-35B-A3B foundational model.
Results published by Prime Intellect show FastAsk's accuracy at 66.25%, higher than Claude Opus 4.6's 61.88%, with average response time about 27% lower.
Since the test set and evaluation were defined by the collaborating parties, this does not mean this 35B model outperforms Opus in general capabilities, but it proves a narrower yet more commercially valuable proposition: enterprises can train smaller models to become experts in specific workflows.
Is the $100 Million 'ARR' Real?
It must be clarified that Prime Intellect's official statement uses the phrase "over $100 million in annualized revenue," not "has earned $100 million in revenue in the past year."
Annualized revenue is typically extrapolated from recent monthly or quarterly revenue speed to a full year; if the business is growing rapidly, it may be significantly higher than the actual revenue over the past twelve months. For GPU, training, and inference businesses charging based on usage, this metric also does not represent clients signing automatically renewable annual contracts of equivalent value.
From Prime Intellect's announcements and launched paid products, the company's commercialization mainly covers four categories: first, the compute marketplace, including GPU instances billed per usage hour, multi-node clusters, and reserved clusters; second, Lab hosted training, charging based on model input, output, and training tokens; third, inference and hosted evaluation, also related to model call volume; fourth, Sandboxes, charging based on CPU, memory, disk, and runtime.
The growth drivers of this revenue structure are not hard to understand. First, GPU clusters themselves are high-price-per-client, continuously consumed resources billed hourly, allowing revenue scale to climb faster than pure software subscriptions. Second, Prime Intellect is extending the customer consumption path from "renting GPUs" to "building environments—running inference—conducting evaluations—reinforcement learning training—deployment," allowing the same client to generate usage across multiple stages. Third, agent reinforcement learning inherently requires extensive parallel rollouts, long-context inference, and isolated sandboxes, naturally consuming more compute power than ordinary API Q&A.
Prime Intellect's disclosed over 6,000 clients and the Ramp case at least indicate the platform is no longer just a research demo. However, when scrutinizing the $100 million figure, several boundaries remain. Prime Intellect is a private company; currently, there are no publicly audited financial reports, the monthly or quarterly revenue basis for calculating annualized revenue, customer payment rates, revenue breakdown, or customer concentration. Whether compute marketplace revenue is recognized based on total client expenditure or platform net revenue has also not been clarified by the company.
Furthermore, Prime Intellect's compute marketplace currently does not offer formal Service Level Agreements (SLAs), with the company stating the reason is the underlying infrastructure comes from multiple suppliers. The official suggestion is for users with higher stability requirements to choose Secure Cloud; if supplier-side failures occur, refunds or platform credits may be provided.
Compared to a single financial number, more easily verifiable progress is that Prime Intellect has turned originally scattered distributed collaborative training into a true full-stack infrastructure "with proprietary models, an open-source ecosystem, backing from hardware giants, and actual enterprise billing for implementation."
Token Cues Erased from Documentation
One detail that cannot be ignored is that as Prime Intellect now steps into the $1 billion valuation club and loudly announces $100 million ARR, the author discovered: The once highly Web3-colored statements in the official documentation—"contracts deployed on Base Sepolia testnet," "future migration to a proprietary chain," and "distributing token rewards to compute pools based on active time via the RewardsDistributor contract"—have been completely erased.
This deletion at the documentation level was foreshadowed as early as March 2025 in that initial official tweet.
At that time, Prime Intellect announced the completion of a $15 million funding round led by Silicon Valley powerhouse Founders Fund, with a core investor roster even featuring top figures like Andrej Karpathy (OpenAI co-founder), Clem Delangue (Hugging Face CEO), and Balaji Srinivasan. It was from this moment that the project's underlying logic was deconstructed.
The previously grassroots-flavored narrative of "issuing tokens, pooling retail computing power, airdrop incentives" immediately became a red-line compliance risk zone for traditional venture capital. To receive ammunition from mainstream capital markets, Prime Intellect had to superficially complete a thorough cleansing from "Crypto-first" to "AI-first."
However, its distributed model training still retains the P2P network topology kernel, but decentralization is no longer a token narrative aimed at retail speculation; instead, it has become an invisible pipeline for B2B enterprises to "schedule idle global compute at low cost."
Now, Prime Intellect more closely resembles a pure AI SaaS company, with its endgame likely being an IPO or a high-premium acquisition by traditional hardware giants.
