Breaking! The AI company that Nvidia re-invested in suddenly wiped all Web3 traces—are retail investors still waiting for token issuance and airdrops?

You know what’s funny? An AI infrastructure company that was founded just two years ago: right after it got money from three hardware giants—NVIDIA, Intel, and Dell—those Web3 traces on its own body, like “issuing tokens, mining, and air drops,” were cleaned up completely.

The company is called Prime Intellect. On July 8, 2026, it announced that it raised $130 million in a Series A round at a $1 billion valuation—led by Radical Ventures, with participation from NVIDIA Ventures, Intel Capital, and Dell Technologies Capital. These three “dads” are themselves in GPUs, CPUs, servers, and data centers. When they team up to back you, how “hard” is that?

The official statement also added: annual recurring revenue (ARR) has already exceeded $100 million, and more than 6,000 enterprises and startups are using their platform. Not just boasting? Let’s break it down.

First, the people. The CEO, Vincent Weisser, used to work on decentralized science (DeSci). He is a co-founder of Bio Protocol, VitaDAO, and CryoDAO, and served as ecosystem and AI lead at Molecule. The CTO, Johannes Hagemann, worked as a research engineer at Germany’s AI company Aleph Alpha, focusing on distributed AI and brain-computer interfaces. In October 2025, they also hired a risk investor named Ash Arora, who specializes in productization and revenue. The company now has only 40 full-time employees—so the productivity per employee is extremely impressive.

Where’s the money? In the seed round, it raised $5.5 million, led by Distributed Global and CoinFund, with the CEO of Hugging Face also investing. In March 2025, it raised another $15 million, led by Peter Thiel’s Founders Fund, with Andrej Karpathy (an early member of OpenAI), Tri Dao (Chief Scientist at Together.AI), and Emad Mostaque (co-founder of Stability AI) among the participants. This $130 million Series A brings the total to $150 million+.

But what’s truly interesting isn’t how much they raised—it’s how they turned the money into technology, and then into revenue.

In November 2024, Prime Intellect released a 10 billion-parameter model, INTELLECT-1, trained across five countries and three continents, with globally distributed collaborative training. Compute utilization across continents reached 83%, and within the U.S. it reached 96%. In 2025, they released INTELLECT-2, with 32 billion parameters, and built an asynchronous reinforcement learning framework called PRIME-RL. Then in November 2025, based on Zhipu GLM-4.5-Air, they built a 106 billion-parameter MoE model, INTELLECT-3. It was trained on 64 nodes with 512 H200s for two months, and the whole thing is fully open-source. What does this prove? It proves they can turn “distributed training” from a paper idea into production-grade tooling.

In February this year, Prime Intellect launched a full-stack AI training platform called Lab, helping people train models themselves without buying a GPU cluster. It opened fully in May. In June, prime-rl 0.6.0 supported trillion-parameter-level MoE. With 28 H200s, it can handle a 131k token sequence, and single-step training takes less than 5 minutes. Optimization methods include FP8 low precision, DeepEP, DeepGEMM, KV Cache tiered offloading, Router Replay, and more. Sounds complex, but the essence is: maximize GPU utilization and minimize customer costs.

In July this year, prime-rl added a unified algorithms layer, with six training methods built in, including GRPO, MaxRL, and self-distillation. You can use algorithm A for math tasks and algorithm B for terminal operation tasks with the same agent—without changing the underlying layer. Prime Intellect has shifted from “doing training for you” to becoming an “RL operating system.”

Big hardware giants aren’t investing for nothing. NVIDIA’s Blackwell, Blackwell Ultra, and the NVL72 systems are provided directly for Prime Intellect, and the Dynamo inference framework is also integrated. Prime Intellect is also testing NVIDIA Vera CPUs for RL sandboxing. They claim that each Vera slot can stably run 176 virtual machines, with multithread throughput about 30% higher than AMD Zen 5. Of course, this data comes from joint testing and can’t be treated as a universal conclusion.

Commercial cases have also come out. The fintech company Ramp used Prime Intellect Lab to train a retrieval intelligence agent called FastAsk, based on the Qwen3.5-35B-A3B model. The results: an accuracy rate of 66.25%, higher than Claude Opus 4.6’s 61.88%, with an average time cost 27% lower. This shows that small models can be trained into domain-specific experts—large models aren’t the only solution.

Now back to that $100 million ARR. Note that the original wording is “exceeding $100 million in annualized revenue,” not “actual revenue in the past twelve months.” Annualized revenue extrapolates using the speed from a recent month or quarter. If the business growth is fast, this number can be far higher than actual revenue. Also, the GPU compute market is pay-as-you-go by usage, not SaaS subscription. There are no automatic-renewal annual contracts. Prime Intellect’s revenue comes from four parts: GPU instances billed by the hour; Lab hosted training billed by tokens; inference and evaluation billed by usage; and sandboxes billed by resource time. The GPU cluster itself has a high ticket price, and customers spend not only on “renting GPUs,” but across the whole chain: building environments, running inference, doing evaluation, reinforcement learning, and deploying to production. Reinforcement learning naturally consumes lots of compute. With 6,000+ customers and cases like Ramp, it at least suggests the business isn’t just spinning. But as a private company, it hasn’t had audited financial statements and hasn’t disclosed monthly revenue, customer pay-through rates, revenue breakdown, or customer concentration. It also hasn’t said whether compute-market revenue is recognized as the customer’s total spending or as the platform’s net revenue. Plus, the platform currently doesn’t provide a formal SLA (service level agreement). So that $100 million figure—look at the direction, don’t treat it as a fixed number.

Finally, the most detail-worthy warning for retail investors: Prime Intellect’s official documentation previously stated “contract deployment on the Base Sepolia testnet,” “future migration to a self-developed chain,” and “allocating token rewards to the compute pool via the RewardsDistributor contract based on active time.” All of that has been removed. The timing was right after the Founders Fund-led round in March 2025. By then, the underlying logic of the project changed: issuing tokens, pulling in retail compute, and air-drop incentives—those narratives are a minefield for traditional venture capital compliance. To attract mainstream capital, they had to wash “Crypto-first” into “AI-first.” But the core kernel of distributed training didn’t change—only the decentralization focus shifted from retail-focused token speculation to an invisible pipeline for B2B businesses: low-cost scheduling of globally idle compute. Today, Prime Intellect looks like a pure AI SaaS company, and the endgame is likely an IPO or acquisition by a hardware giant.

Retail investors, are you still expecting it to issue tokens and do air drops? Go wash up and sleep.

—

Follow me: get more real-time analysis and insights from the crypto market!

#LAB两日腰斩53% $BTC $ETH $SOL # World Cup champion prediction #Iran announces closure of the Strait of Hormuz