Frontier AI is easy to use: building it is still hard

Frontier AI systems are a lot easier to get your hands on now. You can reach them through APIs, low-code tools, and open-weight downloads. Training a frontier model from scratch, though, is still out of reach for almost everyone.

ChatGPT Download

The original Transformer model cost about $670 in 2017. GPT-4 is estimated at $78 million to $100 million, and Gemini Ultra at about $191 million. And that’s before you factor in the other requirements: elite researchers, massive datasets, chips, and repeated retraining. Those numbers go a long way toward explaining why the United States produced 61 notable models in 2023, compared with 21 in the European Union and 15 in China.

So if you want to put AI into a product or a workflow, today’s tools are worth using. But the limits are still there. Open-weight releases don’t change who controls the base model, real tuning takes compute, clean domain data, and specialized expertise, and most startups, hospitals, and mid-sized manufacturers are still integrating what already exists while incumbents keep benefiting from scarcity.

That concentration is getting tighter across the whole stack. Amazon Web Services held 30% of the cloud market in Q1 2026, Microsoft Azure 20%, and Google Cloud 13%. Nearly half of global 2025 funding went to those players. Qualcomm bought Modular. d-Matrix bought Wallaroo.ai. Research stayed concentrated in the Global North, women and racial minorities were still underrepresented, and underserved regions could end up with weaker systems while a small group keeps deciding what comes next, even if access to models now runs through the cloud, APIs, and downloads.