The video’s core call is that AI’s next bottleneck is electricity, not GPUs: power demand is outrunning grid buildouts, so money rotates from memory into on-site generation, turbines, utilities, and nuclear. Elon Musk’s roughly $1B APR Energy purchase is framed as the quickest fix for about 1 gigawatt of behind-the-meter data center power, while longer-dated power names like GEV, Bloom Energy, utilities, IPPs, and modular nuclear are positioned as the durable trade. Memory is treated as the first 2026 trade, but the market is said to be shifting as DRAM and HBM pricing stay tight, LTAs lock up supply, and electricity becomes the scarce input for AI infrastructure.
Elon Musk bought APR Energy for about $1B to secure about 1 gigawatt of power for data center campuses, enough for roughly 600,000 H100-class GPUs or about 750,000 homes.
APR Energy’s trailer-mounted gas turbines can power GPUs in days, bypassing grid integration, permitting, and transformer delays that are said to take 5 to 7 years.
U.S. data center power demand is described as doubling over 24 months from 31 GW to 66 GW, while total U.S. electricity demand rises from 1% to 3%.
By mid-2026, only 5 of the 12 GW planned for the year is said to be on track, even as demand for data center energy is said to have risen from 23 GW in 2023 to 46.5 GW by 2026.
Memory prices are said to be up an average 3% to 500% over the last 9 months, while memory stocks dropped about 20% from highs after running up about 20% a day for months.
DRAM is said to be up almost 20% in July even as memory stocks are down about 20%, with demand described as unchanged.
SK Hynix is said to have 13 to 15 customers locking in about 40% of projected profit next year, with supply sold out for all of 2027 and 40% of that 40% already committed.
MU is described as down around 24% over the last month despite 7x forward earnings, 350% revenue growth, and 85% gross margins.
South Korean investors are said to be overleveraged by about $1B, and the market reportedly lost 1.5T in value over the last week.
The power stack is split into quick fixes and longer-dated buildouts: solar, nuclear, and on-site natural-gas turbines are named as the three modular routes.
Bloom Energy’s fuel-oxide-cell units are described as trailer-portable, more efficient natural-gas conversion that can last 4 to 7 years and avoid 5- to 7-year transformer waits.
GEV is positioned at the center of the trade, up 300% in 3 years, with orders through 2031 and 2025 orders doubling year over year to $7.1B.
GEV is said to already power data centers for MSFT and OpenAI, with the company compared to the 'TSMC of power' because it has a proven method and recurring demand.
Utilities and IPPs are framed as the next layer: big buyers are locking 10 to 20 years of electricity at fixed terms, and MSFT is said to have a 20-year Three Mile Island deal for 100% of output.
The bullish end-state is that electricity is now a currency: the trade is said to stay bullish because AI demand is agnostic to model origin, with Chinese open-source models like Chimera 3 and more than 56% of OpenRouter tokens going to Chinese open-source models.
Valor is called early because permitting is still unsolved, but modular nuclear is still framed as a speed-up versus typical nuclear timelines of 2031 to 2035.
The last-line view is that the 2026 first trade was memory, but the durable trade is power because more energy means more productivity, innovation, and prosperity.
Limitless says it has gone independent and is looking for a partner.
APR Energy, GEV, Bloom Energy, MU, SK Hynix, Microsoft, OpenAI
APR Energy's trailer gas turbines are used to power data centers under Clean Air Act constraints and are described as legal when permitted on-site, behind the meter.
GEV is said to be taking orders through 2031 and to be a predictable, steady-growth power supplier rather than an exponential ETF-style trade.