Research Publication – Reality Bytes – AI – The Economics of Compute

14th April 2026: Radio Free Mobile extends its AI coverage with an examination of the economics of AI compute. RFM subscribers will receive their copy by email.

AI compute is currently uneconomic. Spending $41bn to build 1GW of compute that lasts for 5 years and generates only $10bn per year offers a very poor return. Nvidia has promised that Vera Rubin should lift that by 5x to $50bn, but there is likely to be substantial price erosion from competition, as the example of history dictates. RFM concludes that revenue per GW must rise to at least $15bn to prevent a substantial correction in AI build-out and valuations. RFM proposes a base case of $25bn/GW and is closely tracking revenue / GW in the industry as an indicator of a correction. 

• Compute business model is broken: At today’s pricing ($10bn/GW), no one is making money on selling AI compute as the AI data centre only has 5 years to recoup its investment of $40.8bn, which leaves returns at 1%. Unless returns can be improved to 8% or better, a large correction is likely.
• $bn/GW: is the key metric of AI data centre profitability. This is because the cost to build a data centre is pretty much fixed and so any incremental revenue per GW that can be generated will significantly enhance the profitability and viability of the AI data centre.
• Vera Rubin: is supposed to be the answer to the $bn/GW problem, as a 50% increase in the cost to roll it out, will deliver a 3x increase in the number of tokens that can be produced per GW of energy. Assuming no change in token pricing, a 3x increase in revenue per GW for a 50% increase in costs should be very attractive indeed.
• Poor history: However, previous generational shifts from A100 to H100 and H100 to Blackwell have failed to deliver any increase in revenue per GW, which, according to OpenAI and CoreWeave, has been stubbornly stuck at $10bn/GW for at least 3 years. It would appear that production increases per GW from successive chip generations have been eaten up in token price declines, raising the possibility that history is about to repeat itself.
• Scenarios: There is great uncertainty over how token pricing will evolve and so RFM has looked at 5 possible scenarios, including a base case representing a reasonably optimistic outcome where the AI data centre becomes profitable enough to avoid a messy correction.
• Nvidia GTC: Nvidia’s claim that Vera Rubin would increase revenue / GW 5x was the highlight of GTC 2026. Should this prove accurate, the return available from investing in AI compute would become so attractive that aggressive competition would almost certainly follow. This would lead to a fall in token pricing, resulting in lower margins and returns.  A 5x increase in revenue looks to be an overoptimistic assumption.
• Outlook: Revenue per GW must improve to at least 50% to $15bn/GW to avoid a large financial correction, but something closer to $25bn per GW will be required for returns to be healthy enough to continue attractive investment. This is the metric to watch.