Price Wars Don't Build Better AI — They Just Make It Cheaper

This week, OpenAI and Anthropic announced price cuts of up to 80% on their foundation models. The reason? Chinese competitors like DeepSeek and Moonshot are undercutting them so aggressively that corporate customers are jumping ship to manage their AI bills. On the surface, this looks like great news for enterprises struggling with runaway compute costs. Dig deeper, and it reveals something more troubling about where AI development is headed.

The price war isn't really about better technology. It's about market share in an increasingly commoditized space. When Google simultaneously launches Gemini 3.7 Flash at half the price of its predecessor while OpenAI introduces Ultrafast mode running 14 times faster, we're not witnessing innovation races—we're watching optimization races. These are important, but they're fundamentally different animals.

Here's what gets lost in the celebration of cheaper tokens: the economic model that funded the last generation of AI breakthroughs is collapsing. The massive capital investments that produced GPT-4, Claude, and Gemini were justified by premium pricing and the promise of market dominance. If foundation models become commodity infrastructure sold on razor-thin margins, where does the funding for genuinely novel research come from?

The articles announcing new model releases this week read like spec sheets for slightly better processors. Faster inference. Lower costs per token. Better benchmarks on coding tasks. These are engineering wins, not scientific ones. Meanwhile, the research that could lead to actual capability jumps—the kind that defined the leap from GPT-3 to GPT-4—requires patient capital and tolerance for uncertainty. Those aren't characteristics of companies fighting price wars.

Look at where serious money is flowing instead. Defense tech companies like Cambridge Aerospace and Hadrian are raising billions at astronomical valuations. Why? Because they're solving hard, specific problems with clear customers willing to pay premium prices. The defense sector isn't racing to the bottom on cost—it's racing to deliver capabilities that didn't exist before. That's a very different competitive dynamic.

The AI price war also raises uncomfortable questions about sustainability. Chinese competitors can undercut US companies partly because of different cost structures and government support. But they're also benefiting from years of open research and model releases from Western labs. If the response to that competition is to commoditize everything faster, we're potentially accelerating a cycle where nobody can afford to do the foundational research that makes future breakthroughs possible.

This doesn't mean cheaper AI is bad. Broader access to powerful models will unlock countless applications and democratize capabilities that were previously limited to well-funded organizations. But there's a difference between making AI accessible and making it so cheap that developing the next generation becomes economically irrational.

The real test will come in 18 to 24 months. Will we see genuinely new capabilities emerging from labs engaged in this price competition? Or will we see incremental improvements to existing architectures while the actual innovation moves to better-funded domains like robotics, defense, and specialized industrial applications?

Right now, the message from the market is clear: AI models are becoming infrastructure, priced like utilities. That's a maturation of the industry, but it's also a warning. Infrastructure businesses don't typically fund moonshots. They optimize what exists and compete on efficiency. If we want the next wave of AI breakthroughs, we might need to look beyond the companies currently racing to offer the cheapest tokens.