With financing from Nvidia, Intel, and Dell, quietly erasing the traces of token issuance, how did Prime Intellect report an ARR of 100 million dollars?

Written by: KarenZ, Foresight News

An AI infrastructure company established just over two years ago announced support from investment institutions under Nvidia, Intel, and Dell, while claiming an annual revenue exceeding 100 million dollars—these two figures combined are enough to make Prime Intellect one of the most worthy AI projects to re-examine recently.

On July 8, 2026, the decentralized AI infrastructure network Prime Intellect announced the completion of a Series A financing of 130 million dollars at a valuation of 1 billion dollars, led by the AI-focused venture capital firm Radical Ventures, along with rare joint participation from investment institutions under Nvidia, Intel, and Dell, bringing total financing to over 150 million dollars.

While disclosing this massive financing, Prime Intellect officially announced that within less than a year, its annual recurring revenue (ARR) has quickly soared to over 100 million dollars, and the number of enterprise and startup customers served by the platform has exceeded 6,000.

What's the background?

The author previously mentioned in March 2025 in the article titled OpenAI Founding Members Take Action! Read Fast the Decentralized AI Dark Horse Project Prime Intellect that Prime Intellect was co-founded in January 2024 by co-founders Vincent Weisser and Johannes Hagemann.

• CEO Vincent Weisser has long been involved in the intersection of decentralized science (DeSci) and AI, previously co-founding projects like Bio Protocol, VitaDAO, and CryoDAO, and served as the head of ecology and AI at the DeSci platform Molecule.
• CTO Johannes Hagemann focuses on distributed AI, semi-automated engineering, and brain-machine interfaces, previously working as an AI research engineer at the German AI company Aleph Alpha.

Additionally, in October 2025, venture capitalist Ash Arora joined Prime Intellect as the head of Applied GTM, responsible for product strategy, commercialization, revenue, and applications of AI products in post-training processing and reinforcement learning. Ash Arora recently pointed out that current Prime Intellect's full-time employee count has reached 40.

In terms of financing, Prime Intellect has raised over 150 million dollars, including a 5.5 million dollar seed round financing in April 2024, co-led by Distributed Global and CoinFund, with angel investors including Clem Delangue, CEO of the machine learning tool Hugging Face.

Less than a year later, in March 2025, Prime Intellect completed another financing round of 15 million dollars, led by Peter Thiel's Founders Fund. Investors also included one of the founding members of OpenAI, former AI director at Tesla Andrej Karpathy, Together.AI's chief scientist Tri Dao, and Emad Mostaque, co-founder of Stability AI, among other prominent figures in the AI field.

The latest round has had a different nature. In the 130 million dollar Series A financing, NVIDIA Ventures, Intel Capital, and Dell Technologies Capital are not just financial investors; their parent companies are respectively positioned in key areas of GPU, CPU, server, and data center infrastructure.

Intel Capital's explanation for this round of investment also indicates: the reason why hardware giants are investing is that Prime Intellect is trying to unify underlying computing, training environments, evaluation, and post-training reinforcement learning with upper-level inference all under the same control plane.

What are the substantial advancements?

One noteworthy early achievement of Prime Intellect is proving that long-distance, heterogeneous GPUs can collaborate in training. Looking back at its technological iterations over the past two years, one can see how the platform is systematically translating scientific experiments into commercial product lines.

At the end of November 2024, Prime Intellect released the 10-billion parameter model INTELLECT-1, with training nodes spanning five countries and three continents. The official claim at that time was that it achieved an overall computational utilization rate of 83% across continents, while utilizing nodes distributed solely in the United States resulted in a computational utilization rate of 96%.

Less than half a year later, Prime Intellect released INTELLECT-2, pushing the goal to 32 billion parameters for global distributed reinforcement learning. To this end, the team developed an asynchronous reinforcement learning framework called PRIME-RL, a SHARDCAST responsible for disseminating model weights, and a TOPLOC used to verify whether inference nodes "are working truthfully".

A more crucial change occurred with INTELLECT-3. In November 2025, Prime Intellect released a 106-billion parameter MoE model, supervised-finetuned and reinforced learning based on the Zhipu GLM-4.5-Air. The model was trained on 64 nodes using 512 NVIDIA H200 GPUs for about two months; model weights, training frameworks, data, RL environments, and evaluation methods were all open-sourced. The significance here is not merely about releasing another model but that the company used its research project to validate a complete set of production systems: PRIME-RL responsible for asynchronous training, Verifiers and Environments Hub providing unified tools and community ecology to build and host RL environments and evaluations, and Prime Sandboxes isolating the execution of code generated by agents while the computational orchestration layer took care of 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/intelligent agent models), without needing to set up expensive GPU clusters themselves. On May 7, Lab concluded testing and officially opened to the public.

In June, Prime Intellect released the prime-rl 0.6.0 version, officially stating that it pushed the engineering limit to trillion-parameter MoE (Mixture of Experts) models. Prime Intellect disclosed that for GLM-5 series software engineering tasks, it could use 28 H200 nodes to process sequences of up to 131,000 tokens, with single-step training times under 5 minutes.

The key behind this is not a specific algorithm, but rather a combined optimization of the 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 separation to avoid long tool outputs slowing generation, and KV Cache layering to enhance concurrency; the training side similarly adopts block-scaling FP8, and reduces routing discrepancies between the training and inference sides of MoE models through Router Replay, enhanced further by FSDP, expert parallelism, and context parallelism. These optimizations ultimately affect GPU utilization, training time, and customer costs.

In July of this year, prime-rl added a unified algorithm layer, with built-in six types of training methods: GRPO, MaxRL, On-Policy Distillation, self-distillation, SFT Distillation, and ECHO, allowing different algorithms to be selected for different environments in the same training session. Simply put, the same agent can use one learning method for mathematical tasks and another for terminal operation tasks without rewriting the underlying trainer. This makes Prime Intellect move from "running training for customers" closer to a scalable RL operating system.

Synergy of Software and Hardware: Nvidia is more than just an investor

From the lineup of participants in the Series A round, the binding of hardware giants and Prime Intellect is not only at the capital level but also deeply embedded in the co-construction of hardware and software architecture.

The collaboration between Prime Intellect and Nvidia covers both hardware and software layers. On the hardware side, its training and service workloads are already utilizing NVIDIA Blackwell, Blackwell Ultra, and NVL72 rack-level systems, which the company claims to be more efficient than the previous Hopper clusters.

On the software side, NVIDIA Dynamo is used for global inference orchestration, automatic scaling, request routing, and KV Cache unloading, and is integrated with Prime Intellect's large-scale LoRA (Low-Rank Adaptation, a technique for fine-tuning large language models) deployment.

Nvidia's own technical blog has also confirmed that Prime Intellect has deployed the inference framework NVIDIA Dynamo in production workflows, and is involved in the co-design and integration of LoRA Adapter support.

Earlier this year, in March, Prime Intellect stated it would test RL sandbox workloads around NVIDIA Vera CPU and plans to migrate some sandboxes after Vera becomes publicly available, providing GPU sandboxes on the Vera Rubin system. The company self-tested and claimed each Vera CPU slot can stably run 176 virtual machines in parallel; for its set RL sandbox workload, with multi-threading enabled, throughput was approximately 30% higher than the AMD Zen 5 baseline on AWS with only physical cores enabled.

These numbers demonstrate potential cost advantages, but they come from collaborative testing and the comparative environments are not entirely the same, so they cannot be taken as independent general performance conclusions. Vera Rubin and GPU sandboxes should also be described as "planned for adoption," rather than already in large-scale commercial use.

With the maturity of the products, real commercial monetization is occurring. According to Prime Intellect, the fintech company Ramp uses Prime Intellect Lab to train the retrieval sub-agent FastAsk for Ramp Labs: Ramp has turned its AI spreadsheet editor Ramp Sheets into a trainable RL environment, base model for reinforcement learning training being Qwen3.5-35B-A3B.

The results disclosed by Prime Intellect show that FastAsk achieved an accuracy of 66.25%, higher than Claude Opus 4.6's 61.88%, with an average time consumption around 27% lower.

Since the test set and evaluation were defined by both collaborating parties, this does not mean that the 35B model surpasses Opus in general capabilities, but it validates a narrower and more commercially valuable proposition: enterprises can train smaller models to become specialists in specific workflows.

Is the 100 million dollar "ARR" real?

It must be clarified that the original words used by Prime Intellect officially is "over 100 million dollars in annual recurring revenue," rather than "having achieved 100 million dollars in revenue over the past year."

Annual recurring revenue usually extrapolates the revenue rate from a recent month or quarter over a year; if the business is rapidly growing, it can be significantly higher than the actual revenue over the past twelve months. For usage-based charging in GPU, training, and inference businesses, this metric does not imply that clients have signed equivalent amounts of annual contracts that can be automatically renewed.

From the Prime Intellect announcements and the already launched charging products, the company's commercialization mainly covers four types of products: first is the computing market, including GPU instances charged based on usage duration, multi-node clusters, and reserved clusters; second is Lab-hosted training, charged based on model input, output, and training tokens; third is inference and hosted evaluation, also related to model invocation volume; fourth is Sandboxes, charged based on CPU, memory, disk, and runtime.

The growth drivers of this revenue structure are not difficult to understand. Firstly, GPU clusters themselves are high-value, continuously consumed resources which can lead to faster revenue growth compared to pure software subscriptions. Secondly, Prime Intellect is extending the customer consumption path from "renting GPUs" to "building environments—running inference—doing evaluations—reinforcement learning training—deployment," enabling the same customer to generate usage across multiple links. Thirdly, agent reinforcement learning requires extensive parallel rollouts, long-context inference, and isolated sandboxes, naturally consuming more computing power than regular API Q&A.

The disclosed figure of over 6,000 customers and the Ramp case at least indicate the platform is no longer just a research demonstration. However, several boundaries must be maintained when reviewing the 100 million dollar figure. Prime Intellect is a private company and currently has no public audited financial reports, the monthly or quarterly revenues that the annual recurring revenue is based on, customer payment rates, revenue segmentation, or customer concentration. There is no official explanation regarding whether computing market revenues are recognized based on total customer spending or platform net income.

Moreover, Prime Intellect's computing market currently does not provide formal Service Level Agreements (SLA); the official reason stated is that the underlying infrastructure comes from multiple suppliers. The official recommendation for users with high stability requirements is to choose Secure Cloud; if there is a supplier-side failure, refunds or platform credits may be offered.

Compared to a single financial figure, it is easier to verify advancements: Prime Intellect has truly built a "full-stack infrastructure with self-developed models, an open-source ecosystem, giant hardware endorsement, and real enterprise landing with actual invoices" from originally loose distributed collaborative training.

Deleted clues of token issuance from the documents

One detail that cannot be ignored is that as Prime Intellect now steps into the 1 billion dollar valuation club and boldly announces an ARR of 100 million dollars, I noticed: the official documents previously containing strong Web3 elements such as: "Contracts deployed on the Base Sepolia test network," "future migration to self-developed chains," and "allocation of token rewards to the computing pool based on active time through the RewardsDistributor contract"—have been completely erased.

This kind of deletion at the document level was foreshadowed in the official tweet published by Prime Intellect in early March 2025.

At that time, Prime Intellect announced the completion of a 15 million dollar financing led by the top-tier Silicon Valley Founders Fund, with key investors such as Andrej Karpathy (co-founder of OpenAI), Clem Delangue (CEO of Hugging Face), and Balaji Srinivasan among others. It was at this moment that the foundational logic of the project underwent deconstruction.

The previously grassroots narrative of "token issuance, pulling retail computing power, airdrop incentives" immediately turned into one of the most striking boundaries of traditional venture capital compliance. In order to engage with mainstream capital markets, Prime Intellect had to superficially complete a comprehensive cleansing from "Crypto-first" to "AI-first."

However, its distributed model training still retains the topological core of P2P networks, but decentralization is no longer a narrative aimed at retail speculation; instead, it has turned into an invisible pipeline for B-end enterprises to "schedule global idle computing power at low cost."

Now, Prime Intellect resembles a pure AI SaaS company, and its future ending is likely to trend toward an IPO or high premium acquisition by traditional hardware giants.