The end of the GPU era
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Unlock all features
FREE: Get instant access to 10 AI summaries, chats, or transcripts per day.
Related videos
This is bad...
Theo - t3․gg
104.3k views
I’m done.
Theo - t3․gg
185.8k views
This model is kind of a disaster.
Theo - t3․gg
166.2k views
Claude Mythos and the end of software
Theo - t3․gg
287.6k views
I’m serious.
Theo - t3․gg
95.7k views
OpenAI is lying
Theo - t3․gg
81.9k views
The drama never ends...
Theo - t3․gg
91.9k views
The Future of TypeScript
Theo - t3․gg
152.4k views
The end of the Clawdbot saga
Theo - t3․gg
107.7k views
Fumble of the decade
Theo - t3․gg
80.5k views
Top Comments (10)
TSMC depends on EUV lithography machines from a company called ASML. I reckon ASML that is the more critical company
Reminds me of the evolution of Bitcoin mining, CPU -> GPU -> ASIC
The AI bubble is basically Nvidia and OpenAI
The reason RAM prices went up 4x is that a massive amount of not-yet-manufactured memory was bought with money that doesn't really exist to be put into GPUs that haven't been made yet, to be installed in data centers that haven't been built, powered by infrastructure that may never exist, to satisfy demand that isn't actually there, in order to generate profits that are mathematically impossible.
Okay then the question is what makes TSMC possible? ASML - a real monopoly on lithography solutions that makes TSMC possible. I am sure it would be hard, but with the right experts a secondary TSMC Fab can be built. Fun fact: in case of an invasion, TSMC has remote kill switches for their plant equipment. Not sure if they are literal explosives. However the CEO said that this would render the entire production plant useless.
Yup, inference rather than training is becoming increasingly important. This is where NVIDIA struggles.
Remember when Apple was the first Trillion dollar company, like, one year ago... lol
GLM 4.7 flash is running on a single GPU at home now and it's agentic and capable.
imagine having a flood of 80GB gpus for cheap after whole industry figures out a beter way of compute for datacenters xD
0:38 They are working with Meta to make TPUs compatible with PyTorch so that the community can train models using PyTorch framework as good as JAX & Tensorflow works with TPUs. Thanks Theo for covering about Nvidia ❤
Unlock the Data Inside
Turn Videos into Knowledge
- Get FREE 10/day: transcripts, summaries, chats
- Chat with videos, export text & PDF
- $1 free API credit for RAG, chatbots & research
Free forever plan • All features unlocked
Top Comments (10)
TSMC depends on EUV lithography machines from a company called ASML. I reckon ASML that is the more critical company
Reminds me of the evolution of Bitcoin mining, CPU -> GPU -> ASIC
The AI bubble is basically Nvidia and OpenAI
The reason RAM prices went up 4x is that a massive amount of not-yet-manufactured memory was bought with money that doesn't really exist to be put into GPUs that haven't been made yet, to be installed in data centers that haven't been built, powered by infrastructure that may never exist, to satisfy demand that isn't actually there, in order to generate profits that are mathematically impossible.
Okay then the question is what makes TSMC possible? ASML - a real monopoly on lithography solutions that makes TSMC possible. I am sure it would be hard, but with the right experts a secondary TSMC Fab can be built. Fun fact: in case of an invasion, TSMC has remote kill switches for their plant equipment. Not sure if they are literal explosives. However the CEO said that this would render the entire production plant useless.
Yup, inference rather than training is becoming increasingly important. This is where NVIDIA struggles.
Remember when Apple was the first Trillion dollar company, like, one year ago... lol
GLM 4.7 flash is running on a single GPU at home now and it's agentic and capable.
imagine having a flood of 80GB gpus for cheap after whole industry figures out a beter way of compute for datacenters xD
0:38 They are working with Meta to make TPUs compatible with PyTorch so that the community can train models using PyTorch framework as good as JAX & Tensorflow works with TPUs. Thanks Theo for covering about Nvidia ❤