Geography is Back in Artificial Intelligence: What Does the NVIDIA–Groq Move Mean?

Prof. Dr. Mustafa Ergen

Before the year ended, NVIDIA effectively hollowed out Groq for $20 billion. It took all the intellectual property rights and the human organization into its fold. It left the company as an empty shell with just its customers and business agreements. NVIDIA’s move with Groq says this: The future of artificial intelligence cannot be won solely through “training.” Why did the owner of the GPU, which has swept the world, feel the need to acquire another chip company? This is actually a continuation of the historical process:

• CPU → FPU
• FPU → DSP
• DSP → GPU

Now, a new link is being added: GPU → IPU (Inference Processing Unit). This is not the end of GPUs. But it has become an indicator that GPUs alone are not sufficient. Because if a GPU is a Swiss army knife, an inference chip is like a factory conveyor belt.First, let’s briefly explain how artificial intelligence works! Artificial intelligence first learns, then thinks! The easiest way to understand AI is to think of it like a human. A human also goes through two stages: learning and using what they’ve learned. In AI, these two stages are called: Training and Inference. Training is rare. Inference is continuous. For example, a language model like ChatGPT or Gemini:

• It is trained once a month
• But it performs inference billions of times a day

It’s like a doctor. Finishing medical school = training, diagnosing a patient = inference. A doctor only needs to graduate from one school in any country once. But they treat thousands of patients, and they have to be present when treating a patient. The same applies to AI.Groq designed a completely custom architecture from scratch for AI inference. The team that developed this architecture consists of names who previously worked on developing TensorFlow at Google.The inference architecture is deterministic (produces no surprises), compiled once, and then runs at the same speed millions of times. Control logic is minimal. That’s why Groq chips offer lower latency, higher throughput, and lower cost. In short, Groq adopted an approach of “not doing everything” but “doing one job perfectly.”Because in the long term, the energy bills of data centers, the profitability of AI services, and the sustainability of national AI infrastructures all depend entirely on the inference architecture.That is, the real volume game is in inference. The next front is edge data centers and robotics. Because the real explosion will be:

• Autonomous vehicles
• Warehouse robots
• Smart factories
• Medical robots
• Defense systems
• Wearable devices shifting to the edge.

Here, latency is not a matter of milliseconds but potentially life and death. For systems that cannot go to the cloud and need to make instant decisions, inference architecture is an existential issue.Therefore, training can be done remotely, but inference must be nearby.Training:

• In large data centers
• In a few locations
• Can be done globally

Inference, however:

• Must be close to you
• Has no tolerance for latency
• Runs continuously

That’s why:

• On your phone
• In your car
• In the factory
• In the hospital
• In defense systems, inference is local.

This is also a geopolitical issue. Because inference produces decisions, and decisions are power. If a country uses AI but its inference is dependent on another country, that country slows down, becomes vulnerable, and cannot make independent decisions. That’s why the real issue in AI is not “who trained the model?” but “where is the decision produced?”For the US, inference means centralized control in military systems, intelligence, and financial infrastructures. Therefore, training can be left global, but inference architecture and standards want to be kept under control.The EU regulates not the model but the behavior. But behavior emerges not in training but in inference. That’s why deterministic, auditable, and certifiable inference architectures are ideal for Europe. The real impact of the AI Act will shape inference more than training.Geography has sort of returned in artificial intelligence. For a long time, it was said that “geography is unimportant in the digital world.” Now, the geopolitics of AI will be shaped not by “who trained the model?” but by “who produces the decision, and where?” In this regard, the first to wake up in our region was Saudi Arabia. They established the HUMAIN company. To turn their old oil power into the new oil — data — with a strategy based on local data, local models, and local AI assistants, they were Groq’s first customer.With the newly established AI institution in the public sector, Turkey’s position could be unique here! Low latency to Europe, proximity to the Middle East, access to Central Asia, and the intersection of energy lines could stand out. This geography could turn Turkey into a regional inference hub.