NVIDIA's future hinges on $400b revenue and AI software dominance

Nvidia will hit $400b annual revenue in 5 years time. ~90% of revenue will come from data centre customers. This equates to $90b / quarter, or equivalent to 30,000 Blackwell racks (at ~$3m per rack).

At 150kW per Blackwell rack, data centres will need to expand at 4.5gW/quarter to keep up. Global data centres are expected to increase wattage at 15% per year, which in 5 years time will be close to the 18gW annual increase required. AI (GPU) data centres have higher yields than other data centres, so some amount of use conversion is also expected.

Main challenges for achieving these targets:

• $400b annual revenue assumes Nvidia continues to be dominant in GPU design and AI software stack - successful competition from AMD, Intel, or a Chinese firm could undermine this.
• Uptake of an open-source/cheaper/better platform than Nvidia's CUDA would heavily undermine Nvidia's moat and enable any sizeable firm to directly engage semiconductor manufacturers, such as TSMC, to produce their own chips, stealing away Nvidia's high margin products (similar to what Apple did with its M-Series chips)
• $400b revenue assumes Nvidia and its partners (particularly TSMC and ASML) can continue innovating at sufficient speed to require data centre customers to continue buying the latest and greatest products. In 5 years time Nvidia will be selling a Blackwell successor which needs to be sufficiently better than Blackwell that data centres consider replacing the still perfectly functioning Blackwell racks for the successor racks.
• $400b revenue also assumes data centre customers (mainly AI model producers/operators) continue to find sufficient revenue and financing to continue investing in better models at pace. Nvidia risks being like a $2,000 smart phone - many customers just find that to be too expensive relative to their incomes, especially when their current phone is almost as good.
• AI, Semiconductors and electricity are fairly politicised and are seen as "strategic" for defence purposes. This invites a lot of political uncertainty as regulations could be imposed that restrict AI model growth, or the size of the AI model customer base; that increase the cost of producing chips or reduce the customer base (e.g China banned from buying Blackwell); regulations could also place barriers on increasing data centre power consumption, particularly as government priorities surrounding energy swing heavily with each administration. Each time risking heavy-handed regulation, but also each time causes disruption which slows the total expansion of energy production. E.g. anti-fossil fuel regulations caused power prices to spike as there was a lack of alternatives. Eventually markets adjusted and renewables, such as solar, became much more prominent. Now the US government (and others globally) are rolling back anti-fossil fuel regulations whilst also making it harder to build further green energy projects, causing less efficient and higher cost energy to be encouraged. As regular voters experience higher energy prices, governments are more encouraged to restrict electricity availability to high-electricity users, such as data centres. Many data centres are "switching" to Nuclear, in an attempt to generate lots of "clean" energy with a very small physical footprint - likely hoping to avoid that anticipated backlash - which will definitely have a higher cost than using renewables, but may not work out as the data centres hope.