DeepSeek, a Chinese artificial intelligence lab, released R1, an open-weight reasoning model that matched or approached the performance of leading closed models on several benchmarks, according to Reuters. The company reported training costs a small fraction of those of comparable Western models, a claim that reshaped the conversation about how expensive frontier AI is to build.

FOUNDER. has not independently verified the training-cost figures, which the company reported and which are difficult to compare across labs with different accounting for compute.

Why it mattered

The release challenged the assumption, dominant for two years, that frontier model quality requires frontier capital. If a capable reasoning model can be trained at a materially lower cost, the moat around the best-funded labs narrows.

The open question

Whether the cost figures generalise is contested. Several Western researchers argued the comparison understated the full cost of development, including failed runs and prior research. The honest position is that the cost of training a frontier-class model is falling, that it is falling faster than most investors expected, and that the implication for the economics of the largest labs is the most important open question in the sector.