[DigitalToday reporter Chi-gyu Hwang]"AI innovation is no longer about genius. It depends on how much data and computing power you can gather."

Martin Casado (마틴 카사도), a partner at Silicon Valley venture investment firm Andreessen Horowitz (a16z), highlighted in a recent interview with the Financial Times (FT) that the centre of gravity in AI is shifting from technology to capital.

In the past, whether a problem could be solved with AI depended on technical capability. Now it depends on whether enough money can be put in. He stressed that complex problems such as cancer treatment, climate change and new drug development were not the kind of things that could be solved simply by pouring money into them, but AI is lowering the barriers.

Capital power is also growing in building AI models. "Building AI models may not be as hard as people think," he said. "In the past, model development looked like something unique that only geniuses did because there were so few people in the field. Now I am seeing teams without PhDs and without Stanford backgrounds make good models because they have money."

He said people who deeply understand how AI actually works are still needed, including Mira Murati, Ilya Sutskever and Fei-Fei Li. Even so, Casado said what they have is access to knowledge and networks rather than genius. He added that over time such access will be open to everyone and made clear it will ultimately be a capital game.

On the stages of AI model development, Casado said the pretraining stage of pouring large amounts of money into training foundation models is effectively over. He sees reinforcement learning (RL) as driving the next stage of development.

"We are now at the stage where we are beginning to target specific problems where we can verify the results," he said. "The question is no longer whether AI can do a specific field. In most cases the answer is yes. The real question is whether it is cost-effective."

He said there is still uncertainty around the revenue structure of major AI model developers.

"Compared with the previous model training, it looks more economical, but if you look at the training underway now, it is terrible," he said. "A model is useful for only 3 to 6 months. Then you have to train again. No one can answer where an asset that loses value this fast and a company growing this fast will converge."

He also pointed out that the key is how long AI labs can raise cheap capital. He said that if Anthropic raises $30 billion, it would be difficult for the entire services sector based on models combined to attract funding on that scale.

"I do not know if it will be in 2 years or 6 years, but at some point the big model research companies will build a brand monopoly system, wield enormous influence in the enterprise market and become established incumbents," he said. "But from that point their growth will slow sharply, and value will move down a level and go to those targeting specific industries, applications and small models."