Goldman Research: Power gen shortage can be mitigated by utilizing existing capacity. Key points:
(1) AI computing is different from clouding computing in terms of uptime. LLM training workload can be paused and resumed. This would allow a shift of AI workload away from a peak load.
(2) AI inferencing workload, e.g., agentic AI query, could be re-routed geographically to wherever power capacity is available. Latency may not be a critical issue since inferencing compute time itself is much longer than latency. Re-routing the workload globally? That could be between day and night.
(3) This concept could result in more distributed AI data centers domestically and globally.
(4) It's good for traditional utility companies because their reserved capacity will be running more, instead of sitting idle for most of time.
https://t.co/os5it75Cmo
@RevShark It can be resolved by fed cutting the rate. Increasing rate on the other hand will cause more inflation in housing cuz new house and apt rents going up with borrowing cost.
@Ische2@biancoresearch@elonmusk Exactly! Comparing Tesla with other car makers is like comparing Apple to Orange. Tesla would want to be compared to $AAPL etc.
$AMD, $NVDA
AMD vs. NVDA
It's interesting to see how NVDA and AMD view their competitors.
NVDA views large cloud companies, such as Amazon, Google, as competitors in custom HPC/AI chips, while AMD views them as customers or partners.
@bpoppenheimer Same business philosophy played out at AMD. Lisa Su led a remarkable turn-around with a focus on very selective product lines and it worked.
AMD vs. NVDA
(at the end FY '22, per 10-K)
- Headcount
25,000 vs. 26,000
(Xilinx acquisition +5k)
- Fixed Asset Property/equip
1.5B vs. 3.8B
- R&D Expenditure
5B vs. 7.3B
- Revenue
23.6B vs. 27B
China wind power auction doubled in 2019 from 2018. Total award ~68GW (52GM on shore and 16 off shore). Majority (~50GW) went to 8 domestic developers. Via https://t.co/nqfAiJ6yV9.
https://t.co/Px4kfIOc75