Investing.com -- Macquarie analysts said in a note Monday that they estimate the development cost of R1 was $2.6 billion, which is 467 times higher than DeepSeek reported.
“We estimate DeepSeek's R1 development cost was US$2.6b, said the firm. “This is based on prior work and implies a development cost 467x higher than reported.”
The report highlights that "new markets grow by volume, not price" and that as computing costs decline, adoption will accelerate.
The analysts argue that training compute is a commodity with a defined cost curve, where improvements in hardware efficiency increase the supply of compute power, while software advancements reduce the demand for it.
"Efficiency gains in hardware increase the supply of 'units' of compute per MW. Efficiency gains in software lower the demand for 'units' of compute," the note states.
Macquarie also points out the structural advantage of open-source AI models, which benefit from free development, MIT license use, and widespread adoption.
"This will continue to lower barriers to entry for foundational model builders," the analysts wrote.
Despite concerns about cost, demand for computing power continues to rise. The report identifies Jevons' Paradox in AI, where increased efficiency drives higher overall consumption.
"Inferencing cost reductions drive Jevons' Paradox. Efficiency gains drive total compute consumption growth," Macquarie noted.
From an investment perspective, Australian stocks NextDC and Megaport are seen as medium-term beneficiaries of lower inferencing costs.
Macquarie asserts that "Capex intentions are still the key driver for data centre operators", emphasizing that AI-related spending is becoming a "material driver of current revenues."
The analysts conclude that "a bet against AI is a bet against the largest balance sheets on the planet," suggesting that hyperscaler investments in AI infrastructure will continue to expand, despite speculative risks.