Nuclear power (including small modular reactors) will not be a meaningful power source for AI data centers for the next three to four years from January 2026 (i.e., through roughly 2029-2030), even though it is the right long-term direction.
“I think it's both the obviously the right thing to do and probably not the source of data centers for the next three or four years.”
View on YouTube
Explanation
This prediction concerns nuclear power's role in data centers through 2029-2030, a timeframe not yet elapsed.
Elon Musk's concept of placing data centers/chips in space will take roughly 8 to 10 years to become reality, not 3 to 5 years, due to unresolved technical challenges like inter-satellite communication and cooling.
“I I think it's in the 8 to 10 year category, not in the 3 to five.”
View on YouTube
Explanation
This prediction concerns space-based data centers on an 8-10 year timeline that has not elapsed.
The current AI-driven memory (DRAM/HBM) shortage will take about 18 months from January 2026 to resolve, with memory prices remaining elevated throughout that period (into roughly mid-2027).
“And what's happened is the signal into the the makers has exploded and it will take us about 18 months to digest. The prices will stay high.”
View on YouTube
Explanation
This prediction concerns the memory shortage resolving by around mid-2027, a timeframe not yet reached.
Cerebras's inference performance will improve by more than 2x over the next 18 months, beating the traditional Moore's Law pace.
“We broke doubling every 18 months... and we've carved out a whole new trajectory. And uh my view is in the next 18 months we'll be way over 2x.”
View on YouTube