This is absolutely brutal for the bear case on server depreciation.
The bear argument was that GPUs are worthless after 3 years and that hyperscalers are hiding it with accounting. Meanwhile the operators are saying the opposite: Amazon from 3 years to 5-6, Microsoft from 4 to 6, Google from 3-4 to 6, Meta from 3 to 5.5, and CoreWeave sits at 6.
These are the companies that run the fleets and see the utilization every single day, and they keep stretching the life because the old chips keep earning. A GPU that leaves the training cluster goes on to serve inference for years, and with compute sold out everywhere there is always a buyer for that capacity.
For $NVDA it means every chip sold carries more lifetime revenue for the customer, which is what justifies the next order.
For $NBIS it means the fleet they are building today keeps paying long after the first contract ends, and that is where the margin comes from.
Longer life, better returns, more capex.
I hope Michael is ok... (Not really)
Bullish ⚡️
The bear argument was that GPUs are worthless after 3 years and that hyperscalers are hiding it with accounting. Meanwhile the operators are saying the opposite: Amazon from 3 years to 5-6, Microsoft from 4 to 6, Google from 3-4 to 6, Meta from 3 to 5.5, and CoreWeave sits at 6.
These are the companies that run the fleets and see the utilization every single day, and they keep stretching the life because the old chips keep earning. A GPU that leaves the training cluster goes on to serve inference for years, and with compute sold out everywhere there is always a buyer for that capacity.
For $NVDA it means every chip sold carries more lifetime revenue for the customer, which is what justifies the next order.
For $NBIS it means the fleet they are building today keeps paying long after the first contract ends, and that is where the margin comes from.
Longer life, better returns, more capex.
I hope Michael is ok... (Not really)
Bullish ⚡️
