I can think of two possible explanations, one malevolent and one benevolent, but both of them require accepting intent to deceive, so I am very much hoping someone has a third explanation of why A) Tesla was so unprepared and B) Tesla was unable or unwilling to take delivery of H100s as planned and put them to use via the same or similar means that the other company was able to use (many months earlier than planned).
The most powerful cognitive bias is narrative bias. If someone tells you a story, you believe it, and then you view all future information on the basis of that narrative.
Even if the narrative is false. Even if the other information contradicts the narrative.
The current narrative is unrealistically simplistic. For these reasons:
1. There are multiple data center projects going on at Tesla and XAI simultaneously for the past several years, including DOJO and NVIDIA infrastructure in NY, California and Texas at various stages, plus a half dozen or a dozen third party data centers with rental agreements and buildouts at various stages of negotiation and completion.
2. There is significant lead time for all of these efforts of at least a year. Rental agreements can take a year, construction in a year is fast and then deploying 10,000 H100s or whatever also takes months if not a year from beginning to end.
3. Given multiple years of lead time for chips, hosting/power and other infrastructure projects, you have to predict 1, 3, 5 years in advance how much you will need of construction, location, power, capital equipment. So everything is an estimate based on estimates.
YOU DO NOT REALLY KNOW HOW MUCH YOU ACTUALLY NEED that far in advance.
4. There are multiple efforts that significantly affect the amount of compute needed at any given point in time. The cost of training is going down every year with improvements in software. The amount of training needed goes down every year or becomes more efficient with improvements in the software. The extent of retraining goes down. The future size of the model needed changes as all of the above becomes more efficient.
5. Dojo and the FSD software have had significant advancements.
In the normal course of business, any one of the items in any of the 5 previous points alone is sufficient to shift the needs by an amount of 12,000 H100s for 6 months.
If you think there's any evidence of deception here you are reading way too much into it, or you are locked not a narrative, or you just don't understand business and how it works when you are innovating.
Almost all Tesla analysis these days comes from a very simplistic viewpoint of people who have clearly never worked for an innovative technology company... and most journalists don't know the first thing about business.
Elon can't give a 14 paragraph explanation for the switch and if he did people would not read it and it would reveal operational details that he should not reveal. So Elon is forced to give a simplistic answer.