Macquarie analysts said in a note on Monday that they estimate R1's development costs at $2.6 billion, 467 times the cost announced by DeepSeek. The company said: We estimate the development cost of DeepSeekR1 to be $2.6 billion. This is based on previous work and means development costs were 467 times higher than reported.
The report highlights that emerging markets grow on volume, not price, and that adoption will accelerate as computing costs fall.
Analysts believe that training computing is a commodity with a clear cost curve, with improvements in hardware efficiency increasing the supply of computing power and advances in software reducing demand for computing power.
The report states: Improvements in hardware efficiency increase the supply of computing ‘units’ per megawatt. Increased software efficiency reduces the need for computational ‘units’.
Macquarie also pointed to the structural advantages of open source AI models, benefiting from free development, use of the MIT license and widespread adoption.
Analysts wrote: This will continue to lower the barrier to entry for basic model builders.
Despite cost concerns, demand for computing power continues to grow. The report points out Jevons’ paradox in the field of artificial intelligence, where increased efficiency leads to increased overall consumption.
Macquarie points out that the reduction in reasoning costs drives Jevons' paradox. Efficiency improvements drive growth in total computing consumption.
Macquarie claimed that capital expenditure intentions remain the main driver for data center operators and highlighted that spending related to artificial intelligence is becoming an important driver of current revenue.
The analysts concluded that a bet on AI is a bet on the largest balance sheets on the planet, suggesting that investments in AI infrastructure by hyperscale companies will continue to expand despite speculative risks.
Related articles: