Nvidia & Nebius – The GW Game.

Has Nebius cracked the compute business model?

• Nvidia’s investment of $2bn into Nebius gives it crucial fuel to expand into the large market that awaits it and Nebius appears to have found a way to reduce the cost of compute which will improve the economics of every gigawatt of capacity that it builds.
• Nvidia will invest $2bn in Nebius, which I presume will be in new equity, a good proportion of which will go straight back to Nvidia to purchase its AI systems.
• This is very similar to the deal that Nvidia struck with OpenAI and Nokia, although given its much smaller scale, I suspect that this one will not be cancelled.
• Nvidia and Nebius will also collaborate to ensure that Nebius can optimally deploy 5GW of Nvidia-powered capacity by 2030.
• This is where the question marks arise as this will cost around $160bn to deploy at today’s prices, and Nvidia’s investment will barely make a dent in the financing that will be required to meet this total.
• It is crucial to note that the 5GW is an intention and does not represent a firm commitment by Nebius, meaning that if market conditions change it will be able to change this target with no penalties or problems.
• For 2026, Nebius intends to spend $16bn – $20bn (of which 60% is financed) and end the year with 800MW to 1GW of connected power which is an intriguing forecast.
• This is because 1GW of Nvidia capacity costs around $40bn to build and yet Nebius is implying that it can build around 730MW for $18bn.
• This would imply a price of around $25bn per GW, which in my opinion, changes the financial dynamics materially.
• Using RFM Research’s estimates, this would increase the 5-year EBIT margins from 13% (at $40.8bn) to 45% with revenues of $10bn per GW, which is roughly what the industry currently earns.
• Ignoring the depreciation schedule (where I expect extensive fiddling by companies trying to flatter their earnings), the return rises from 1% over 5 years to 17% which is pretty healthy.
• This leads me to conclude that either Nebius has found a way to greatly reduce the cost or the next generation of Nvidia equipment is going to improve the economics in a way that its previous generations have failed to do.
• Looking at Nebius’ guidance for 2026, it expects to have an $8bn annual revenue run rate (ARR) with around 900MW of active capacity putting it pretty much bang on the industry average of $10bn per GW.
• Hence, this is not about revenue, it is about cost, and I suspect that Nebius is banking on the improvements that Vera Rubin is expected to deliver.
• Vera Rubin should deliver 3.6 ExaFLOPS per rack (or cabinet) which is 3.3x more than Blackwell Ultra does and Rubin consumes 125kW per rack compared to 100kW for Blackwell Ultra.
• This means that Vera Rubin should deliver around 2.6x more tokens per unit of energy consumed.
• Even with a price rise from Nvidia, Nebius should still be getting a lot more tokens per GW, and as long as the price per token remains stable, its economics should improve materially.
• This is the problem that has hampered the economics of previous generations as despite huge increases in token production per GW, revenue per GW has remained stable implying significant price declines.
• This does not explain how Nebius has managed to reduce the cost per GW of capacity as with a price rise, the total cost per GW should be rising and revenue per GW with it.
• This is the problem with the GW measure, as it does not take into account how much compute is actually produced which would be a much better measure.
• The net result is that if Nebius has indeed managed to reduce the cost to $25bn per GW, then its economics are going to improve materially even if the price stays flat at $10bn per GW.
• In my opinion, Nebius’ management are pretty hardy and inventive, having endured the fall-out from Russia’s invasion of Ukraine and the break-up of Yandex leading to the formation of Nebius.
• Hence, if I was to back any of the neo-clouds, it would be this one.