Valued at $1 Billion, Nvidia Makes a Big Bet! Is Prime Intellect Shedding Its Web3 Label?

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

Written by: KarenZ, Foresight News

A company that has been established for just over two years in the AI infrastructure space has announced support from investment arms of Nvidia, Intel, and Dell while claiming its annual revenue has exceeded $100 million—these two figures together make Prime Intellect one of the most noteworthy AI projects to revisit recently.

On July 8, 2026, the decentralized AI infrastructure network Prime Intellect announced the completion of a $130 million Series A funding round at a valuation of $1 billion, led by Radical Ventures, a venture capital firm focused on AI, with rare participation from the investment arms of Nvidia, Intel, and Dell, bringing total funding to over $150 million.

While disclosing this massive funding, Prime Intellect officially announced that in less than a year, its annual recurring revenue (ARR) has rapidly surged to over $100 million, with more than 6,000 enterprise and startup clients served by the platform.

What’s the background?

I previously mentioned in March 2025 in "OpenAI Founding Members Step In! A Quick Read on the Decentralized AI Dark Horse Project Prime Intellect" that Prime Intellect was co-founded in January 2024 by Vincent Weisser and Johannes Hagemann.

* **CEO Vincent Weisser** has long been involved in the intersection of decentralized science (DeSci) and AI, having co-founded projects such as 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 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 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 noted that Prime Intellect's full-time employee count has reached 40.

In terms of funding, Prime Intellect has raised over $150 million in total, including a $5.5 million seed round in April 2024 led by Distributed Global and CoinFund, with angel investors including Clem Delangue, CEO of Hugging Face.

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 Andrej Karpathy, one of the founding members of OpenAI and former AI director at Tesla, as well as Tri Dao, chief scientist at Together.AI, and Emad Mostaque, co-founder of Stability AI, among other heavyweight figures in the AI field.

The nature of the latest round is somewhat different. In the $130 million Series A funding, NVIDIA Ventures, Intel Capital, and Dell Technologies Capital are not just financial investors; their parent companies are positioned at critical points in GPU, CPU, server, and data center infrastructure.

Intel Capital's explanation for this round of investment also indicates that the hardware giants are willing to invest because Prime Intellect is attempting to consolidate the underlying computing, training environment, evaluation, and post-training reinforcement learning with the upper-level inference all within the same unified control plane.

What substantial progress has been made?

One of Prime Intellect's early notable achievements is proving that long-distance, heterogeneous GPUs can also collaborate in training. Following its technological iterations over the past two years, it is evident how the platform has gradually transformed research 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 an overall computation utilization rate of 83% across continents, while training using only nodes distributed across the United States achieved a utilization rate of 96%.

Less than six months later, Prime Intellect released INTELLECT-2, pushing the target to 32 billion parameters for global distributed reinforcement learning. To achieve this, the team developed the asynchronous reinforcement learning framework PRIME-RL, SHARDCAST for propagating model weights, and TOPLOC to verify whether inference nodes are "doing their job" accurately.

A more critical change occurred with INTELLECT-3. In November 2025, Prime Intellect released a 106 billion parameter MoE model based on the Zhiyu GLM-4.5-Air, which was trained for about two months on 64 nodes with 512 NVIDIA H200 GPUs; the model weights, training framework, data, RL environment, and evaluation methods were all open-sourced. The significance here is not just the release of another model, but the company validated a complete production system with its research project: PRIME-RL is responsible for asynchronous training, Verifiers and Environments Hub provide unified tools and community ecosystems to build and host RL environments and evaluations, and Prime Sandboxes isolate the execution of code generated by intelligent agents, while the computation orchestration layer is responsible for clustering, storage, and monitoring.

In February this year, Prime Intellect launched a full-stack AI training platform, Prime Intellect Lab, specifically designed to help individuals, engineers, and AI companies train and optimize their models (especially agentic/intelligent models) without needing to build expensive GPU clusters themselves. On May 7, the Lab concluded testing and officially opened to the public.

In June, Prime Intellect released version prime-rl 0.6.0, claiming to push engineering limits to trillion-parameter MoE (Mixture of Experts) models. Prime Intellect disclosed that it could handle sequences of up to 131,000 tokens on 28 H200 nodes for software engineering tasks in the GLM-5 series, with single-step training times of less than five minutes.

The key behind this is not a specific algorithm, but the joint optimization of training and inference systems: the inference side uses FP8 low-precision computing and components like DeepEP and DeepGEMM to improve throughput, pre-filling and decoding separation to avoid long tool outputs slowing down generation, and KV Cache layered offloading to enhance concurrency; the training side also adopts block-scaling FP8 and reduces routing differences between the MoE model's training and inference sides through Router Replay, further adding FSDP, expert parallelism, and context parallelism. These optimizations ultimately affect GPU utilization, training time, and customer usage costs.

In July this year, prime-rl added a unified algorithm layer, incorporating 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. In simple terms, the same intelligent agent can use one learning method for mathematical tasks and another for terminal operation tasks without needing to rewrite the underlying trainer. This brings Prime Intellect closer to a scalable RL operating system rather than just "running training for customers."

Hardware-Software Collaboration: Nvidia is More Than Just an Investor

From the lineup of investors in the Series A round, the binding of hardware giants with Prime Intellect goes beyond mere capital involvement and delves into collaborative construction of hardware and software architecture.

Prime Intellect's collaboration with 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 are 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 offloading, 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 also confirms that Prime Intellect has deployed the inference framework NVIDIA Dynamo in production workflows and participated in the co-design and integration of LoRA Adapter support.

Prime Intellect previously stated in March this year that it would test RL sandbox workloads around the NVIDIA Vera CPU and plans to migrate some sandboxes once Vera becomes publicly available, providing GPU sandboxes on the Vera Rubin system. The company claims that each Vera CPU socket can stably run 176 virtual machines in parallel; under its set RL sandbox workload, the throughput is about 30% higher than the AMD Zen 5 baseline on AWS with only physical cores enabled.

These numbers demonstrate potential cost advantages, but they currently come from collaborative testing between both parties, 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 to adopt" rather than already being in large-scale commercial use.

As the product matures, 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, using Qwen3.5-35B-A3B as the base model for reinforcement learning training.

The results disclosed by Prime Intellect show that FastAsk has an accuracy rate of 66.25%, higher than Claude Opus 4.6's 61.88%, with an average time reduction of about 27%.

Since the test set and evaluation are defined by both parties, this does not mean that the 35B model surpasses Opus in general capabilities, but it proves a narrower and 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 the original wording used by Prime Intellect is "over $100 million in annualized revenue," not "$100 million in revenue achieved in the past year."

Annualized revenue typically extrapolates the revenue rate from a recent month or quarter to a 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 that charge based on usage, this metric does not represent that customers have signed equivalent amounts of automatically renewable annual contracts.

From Prime Intellect's announcements and the products that have gone live, the company's commercialization mainly covers four types of products: the first is the computing market, including GPU instances billed by usage time, multi-node clusters, and reserved clusters; the second is Lab-hosted training, charging based on model inputs, outputs, and training tokens; the third is inference and hosted evaluation, also related to model call volume; the fourth is Sandboxes, billed based on CPU, memory, disk, and runtime.

The growth drivers of this revenue structure are not difficult to understand. First, GPU clusters are high-ticket, continuously consumed resources billed by the hour, allowing revenue scales to rise faster than pure software subscriptions. Second, Prime Intellect is extending the customer consumption path from "renting GPUs" to "building environments --- running inferences --- conducting evaluations --- reinforcement learning training --- deploying online," allowing the same customer to generate usage across multiple stages. Third, agent reinforcement learning requires a large number of parallel rollouts, long-context inference, and isolated sandboxes, which naturally consumes more computing power than ordinary API Q&A.

The disclosure of over 6,000 customers by Prime Intellect, along with the Ramp case, at least indicates that the platform is no longer just a research demonstration. However, when reviewing the $100 million figure, several boundaries must be retained. Prime Intellect is a private company and currently does not publicly audit financial reports, calculate the monthly or quarterly revenues on which the annualized revenue is based, customer payment rates, revenue splits, or customer concentration. It is also unclear whether the computing market revenue is recognized based on total customer spending or platform net income.

Additionally, Prime Intellect's computing market currently does not provide formal service level agreements (SLAs), with the official reason being that the underlying infrastructure comes from multiple suppliers. The official recommendation for users with high stability requirements is to choose Secure Cloud; if a supplier-side failure occurs, refunds or platform credits may be offered.

Compared to a single financial figure, a more easily verifiable progress is that Prime Intellect has transformed its originally loose distributed collaborative training into a full-stack infrastructure that "has self-developed models, an open-source ecosystem, endorsements from major hardware players, and real enterprise landing invoices."

The erased clues of token issuance

One detail that cannot be ignored is that as Prime Intellect enters the $1 billion valuation club and publicly announces a $100 million ARR, I found that the official documents previously contained highly Web3-colored statements: "Contracts deployed on the Base Sepolia testnet," "future migration to self-developed chains," and "distributing token rewards to the computing power pool based on active time through the RewardsDistributor contract" --- have been completely erased.

This document-level deletion was foreshadowed in a tweet released by Prime Intellect in early March 2025.

At that time, Prime Intellect announced the completion of a $15 million financing led by Silicon Valley's top Founders Fund, with core investors including Andrej Karpathy (co-founder of OpenAI), Clem Delangue (CEO of Hugging Face), and Balaji Srinivasan, among other top figures. It was from this moment that the underlying logic of the project underwent deconstruction.

The previously grassroots narrative of "issuing tokens, pulling retail computing power, and airdrop incentives" immediately turned into a minefield that touches the compliance red lines of traditional venture capital. To accommodate the ammunition of 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 the P2P network, but decentralization is no longer a token narrative aimed at retail speculation; instead, it has become 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 the future trajectory is likely to lead to an IPO or a high-premium acquisition by traditional hardware giants.