Dylan Patel

“Datacenter power capacity growth will accelerate from a 12-15% CAGR to a 25% CAGR over the next few years. Global Datacenter Critical IT power demand will surge from 49 Gigawatts (GW) in 2023 to 96 GW by 2026, of which AI will consume ~40 GW.”

Deadline is end-2026. SemiAnalysis itself later claimed the forecast 'proved very accurate'; final resolution requires 2026 full-year global critical IT power capacity data.

“Google might have no profits in 2027: with roughly $180 billion (Google) and $200 billion (Amazon) of AI infrastructure capex, hyperscaler profits disappear by 2027.”

Alphabet has guided to $175–185B of 2026 capex against ~$165B of 2025 operating cash flow, so the test is whether reported 2027 profit/free cash flow collapses toward zero.

“2029 is the real ramp for co-packaged optics — the Street expects 2027, but true mass production only happens between late 2028 and 2029. It is a manufacturing thing: the volumes and yields are not there, so people are going to stay in copper all they can. Rubin is all copper, and Feynman on the GPU is still copper.”

Resolves on when CPO actually reaches volume production; note that later 2026 reporting suggests Nvidia may pull CPO into the Feynman generation (2028), which would cut against the copper-through-Feynman part of the call.

“The sleeping giant, Google has woken up, and they are iterating on a pace that will smash GPT-4 total pre-training FLOPS by 5x before the end of the year.”

Google did unveil Gemini Ultra on December 6, 2023, and it did exceed GPT-4's training compute — but Epoch AI estimates Gemini Ultra at ~5e25 FLOP versus ~2e25 FLOP for GPT-4, roughly 2.5x rather than 5x (and the model was not generally available until 2024).

“The path is clear to 20x [GPT-4's total pre-training FLOPS] by the end of next year given their current infrastructure buildout.”

20x GPT-4 (~2e25 FLOP) implies a ~4e26 FLOP training run by end-2024; Epoch AI's dataset records the first model above 1e26 FLOP as xAI's Grok-3 in February 2025, and no Google model publicly known or estimated to be near that scale was trained by the end of 2024.

“Production is increasing rapidly. Nvidia is going to ship over a million H100s plus A100s this year.”

Omdia estimated Nvidia shipped nearly half a million H100 and A100 GPUs in Q3 2023 alone and expected the figure to exceed half a million again in Q4; TechInsights put Nvidia's total 2023 data center GPU shipments at about 3.76 million units, comfortably above the one-million H100+A100 threshold.

“Okay, there's all these GPUs going into production. NVIDIA is going to sell well over 3 million total GPUs next year, over a million H100s this year alone. There's a lot of GPU capacity coming online.”

Nvidia's data-center GPU shipments were already ~3.76 million units in 2023 per TechInsights and rose further in 2024 (toward roughly 4 million) as data-center revenue more than doubled, with Omdia estimating Microsoft alone bought ~485,000 Hopper chips in 2024 and Meta, ByteDance and Tencent 224,000–230,000 each. Shipments comfortably exceeded the 'well over 3 million' bar.