Groq Inc. is raising approximately $600 million in new funding at a valuation of about $6 billion, according to people familiar with the discussions. The round is being led by venture capital firm Disruptive, which has committed more than $300 million. The fundraising, still in the final stages, brings Groq’s total capital raised to more than $2 billion.
The latest round more than doubles the company’s valuation from last year, when it raised $640 million in a deal led by BlackRock funds. The increased investor interest underscores Groq’s growing relevance in the AI infrastructure market, particularly as companies look beyond Nvidia for scalable and cost-efficient inference solutions.
Founded by Jonathan Ross, a former engineering leader at Google, Groq is focused entirely on inference, the process of executing trained AI models rather than training them. Its proprietary chips, called Language Processing Units (LPUs), are designed for ultra-low latency and high-throughput inference, avoiding many of the hardware bottlenecks that affect GPU-based systems.
“Every time we launched a new AI product at Google, the compute required to run it in production was typically 10 to 20 times more than what was needed to train it,” Ross said in a recent podcast interview. “Inference at scale is a completely different challenge.”
While Nvidia remains dominant in AI training, its graphics processing units rely on high-bandwidth memory and advanced packaging technologies like CoWoS, both of which are in limited supply. Ross said Groq intentionally avoids those dependencies, enabling faster ramp-up times for new capacity and less exposure to long-term hardware shortages.
“Training will always be GPU-dominated and high margin,” Ross said. “But inference needs scale and predictability. Our hardware avoids bottlenecks by skipping exotic components. That lets us ramp up capacity within months, not years.”
Groq has already started executing on that model. Just four weeks after deciding to build a new data center in Helsinki, the company began installing racks. It expects to begin serving European customers imminently. The move follows Groq’s announcement of its first European data center deal and adds to a growing global footprint aimed at bringing low-latency inference closer to users.
“Even from Finland, users across Europe will get better performance than with a GPU in their own region,” Ross said. “Our latency is that low.”
The company’s infrastructure strategy includes partnerships with Bell Canada on its Bell AI Fabric project and with Meta to power the official Llama API. These collaborations have positioned Groq as one of the few players outside Nvidia with deep integrations into enterprise-scale AI deployments.
Ross has repeatedly emphasized that inference is not simply a scaled-down version of training. It demands a different architecture, one that is optimized for throughput, energy efficiency, and deployment agility. Groq’s LPUs deliver approximately three times the compute per watt compared to traditional GPUs, according to Ross, which is especially important in power-constrained markets like Europe.
While many AI hardware startups have emerged in the inference space, Ross expressed skepticism about their ability to compete. Developing a new chip architecture can cost between $250 million and $500 million, and most competitors, he said, are simply replicating GPU designs rather than creating purpose-built inference systems.
“If you’re building inference chips that still use external memory, you’re basically copying GPUs and you won’t win by copying Nvidia’s playbook,” Ross said. “They’ve already mastered it.”
Groq’s focus on inference also shapes its talent strategy. Unlike model development firms that compete for the same pool of AI researchers, Groq hires engineers with a background in systems and semiconductors. While companies like Meta and OpenAI have offered compensation packages approaching $100 million to lure talent, Ross said Groq has had an easier time recruiting due to its differentiated focus and attractive equity offerings.
Still, scaling inference capacity remains a challenge, particularly when the compute demand is increasing faster than suppliers can respond. Ross pointed to Nvidia CEO Jensen Huang’s public comments that customers must order GPUs two years in advance. Groq, by contrast, can deploy new capacity in roughly six months thanks to its streamlined supply chain and modular data center strategy.
Internally, the company is also recalibrating expectations. As first reported by The Information, Groq recently reduced its 2025 revenue forecast by more than $1 billion. According to a person familiar with the matter, that revenue is now expected to materialize in 2026. The company declined to comment on sales projections but confirmed it continues to expand data center operations in North America and Europe.
The brand has also drawn occasional confusion due to its phonetic similarity to “Grok,” the AI chatbot developed by Elon Musk’s xAI. When asked about the potential for brand conflict, Ross responded with a touch of humor.
“Everyone in kindergarten learns the concept of dibs,” he said. “And we’ve called dibs.”
Groq’s latest funding comes at a time when the AI hardware market is rapidly fragmenting, with entrants like Cerebras, Positron, and SambaNova attempting to challenge Nvidia’s dominance by offering alternatives tailored specifically for inference. While many of these players tout raw performance or energy efficiency, Groq’s edge lies in its ability to deploy capacity quickly, scale globally, and deliver consistent low-latency performance without relying on constrained components. “Compute is becoming the electricity of AI,” said CEO Jonathan Ross. “But if someone can produce it more efficiently and reliably, that’s not a commodity, that’s a competitive advantage.”
