While many tech giants continue to focus on training ever larger models, there’s a new silent yet equally strategic race: deploying these models quickly, efficiently, and sustainably. In this arena, the American startup Groq is gaining ground. This week, they launched their first data center in Europe, located in Helsinki, Finland, in partnership with Equinix. Their clear mission: to dominate the AI inference market, where the real contact with users happens.
In the AI lifecycle, there are two main stages: training—where the model learns to predict words, images, or tokens—and inference—when it responds to a user request, such as a prompt on ChatGPT or an instruction in Copilot. Until now, NVIDIA has reigned with its powerful GPUs, essential for training. But inference demands a different kind of power: speed, stability, low latency, and energy efficiency.
That’s where Groq enters the scene with its LPUs (Language Processing Units), chips designed from scratch to respond in record time without the need for massive hardware deployments. Their goal isn’t competing in labs but rather making an impact in the daily lives of millions of users. Europe is their next big stage.
Why Finland? Because it offers clean energy, a favorable climate for cooling, and, most importantly, a strategic position to serve companies and administrations across Europe that need speed and regulatory compliance. According to the company, the Helsinki data center will operate with ultra-low latency, energy efficiency, and compliance with digital sovereignty—all based on existing infrastructure, without building a new campus from scratch.
Groq states that their deployment in Helsinki was completed within weeks and can scale quickly. “We’ve built a global network already processing over 20 million tokens per second, and this expansion allows us to reach European clients without compromising speed,” they said in an official statement.
This move from Groq isn’t isolated. McKinsey estimates that the inference hardware market will double in size compared to training hardware in data centers in the coming years. Barclays goes further, predicting that within just two years, major players will spend more on inference chips than on training. The consequence? NVIDIA could see up to $200 billion in market share slipping away.
Groq, for its part, isn’t aiming to compete by building larger models. Their strategy is that as AI integrates into all devices and processes, what will matter most is instant operation at an acceptable energy cost. Their LPU chip—based on deterministic processing and without the need for HBM—has a distinct advantage here.
In a moment when companies are not only asking “what can AI do?” but also “what does it cost to use at scale?”, Groq positions itself as a realistic alternative to the increasingly energy-hungry and computationally demanding models.
Moreover, by not relying on traditional GPUs, they can bypass supply chain bottlenecks, cut costs, and deploy in regions with strict regulations like the European Union. This makes their arrival in Europe more than just expansion—it’s a statement of intent.
Perhaps Groq isn’t yet as recognized as NVIDIA, but it’s already shaping up as a disruptive player in practical AI. Their strategy is clear: not trying to replace everything, but specializing in what’s coming next. And if everyday AI usage grows as expected, it’s not far-fetched to imagine that in a few years, when someone requests an instant AI response, that response might not come from an NVIDIA GPU…but from a Groq LPU.