THT Insights
86 posts
THT Insights
@THT_Insights
ThoughtAI and THT news, commentary, and community-focused insights. Not endorsed by ThoughtAI. Not financial advice. DYOR
Judging AGI by how well it can mimic us is a category error, because mimicry isn't intelligence and isn't general. We should judge AGI by how well it learns to do things we didn't teach it (including things we don't know how to do ourselves).
Automation executes. Autonomy reasons. 🧠 Using Agentic AI, operators are now deploying intelligent agents that can self-optimize, self-heal, and self-configure. ➡️ nvda.ws/4vEuuKS
I was doing some research today into Energy for Ai and found some interesting stuff that nobody is really aware of. Within an Ai data centre energy must be double converted, it’s delivered in AC from the grid then converted into DC so it can be efficiently stored and uninterrupted. It’s then converted back into AC to send it through the building and towards the server racks. And then once again converted back into DC because silicon GPU’s can only run on low voltage DC. Around 30-35% of energy according to Jensen Huang is lost on overheads such as Energy Conversion and Cooling the server racks before it even reaches the GPU’s
The most dangerous financial opinion you can hold is the one you inherited without questioning it.
the binding constraint on AI is not chips. it is electrons. one AI inference task burns up to 1,000 times more power than a single google search. multiply that by every chatbot query and every image generation and every coding assistant call happening in parallel right now. IEA forecasts global data center electricity demand at ~945 terawatt hours by 2030. that is more than double the 2024 baseline of around 415 TWh. anthropic estimates the US alone needs roughly 50 gigawatts of new capacity dedicated to AI by 2028. that is twice the peak demand of new york city. just for AI. the gartner number is the one that bond desks should be staring at. by 2027, 40 percent of AI data centers will face power constraints that limit their growth. last july 60 data centers in northern virginia tripped offline simultaneously when a voltage event caused a 1,500 megawatt swing on the grid. the average large transformer in the US is over 40 years old. the lead time on a new one is measured in years. ireland already pulled the brake. data centers consumed 20 percent of total irish electricity in 2024 and dublin stopped issuing new permits. this is the part where the macro story gets unexpectedly old fashioned. utilities have the largest order backlog in their history. nuclear restarts are getting signed. nextera and google are reopening an iowa plant together. the trade in 2025 was the GPU. the trade in 2026 might be the substation. #AI #Energy