The AI trade in one sentence: we're turning energy into intelligence. The exchange rate — accuracy per joule — improved 18x in 16 months. About 6x from better hardware, 3x from better models. Every company pushing this ratio forward is set up to win.
Here's the investible stack:
$NVDA: Each GPU generation multiplies intelligence per watt. Hopper to Blackwell was the big jump. Efficiency is the business model and why customers keep upgrading for training.
Memory ($MU, SK Hynix, Samsung): HBM keeps the chip fed so no joule is wasted waiting. Every HBM generation moves more data per watt. Memory demand scales with every GPU sold, and we need more memory per GPU too.
$IFNNY: World's largest maker of power semiconductors — the chips that convert and control electricity itself. An AI rack pulls over 100 kilowatts, and every power conversion between grid and chip loses energy as heat. Nvidia's moving datacenters to 800-volt power later this year, which means fewer conversions and more Infineon content per rack.
$AVGO: Attacks the energy lost between chips. Training isn't one chip thinking — it's tens of thousands of GPUs talking constantly, and that conversation is pure energy cost. Broadcom's switch chips move roughly double the data per watt each generation, and co-packaged optics convert signals to light right at the chip, cutting power cost dramatically. Faster fabric also means less GPU idle time, so every chip in the rack gets more productive per joule.
Custom silicon ($GOOGL TPUs, $AMZN Trainium): A GPU is flexible but you pay for that in wasted joules. Custom chips strip out everything except the exact math AI needs, so nearly every joule goes into useful work. Design work flows back to $AVGO and $MRVL.
This is the AI efficiency trade. Own the stack that turns electricity into intelligence.
Here's the investible stack:
$NVDA: Each GPU generation multiplies intelligence per watt. Hopper to Blackwell was the big jump. Efficiency is the business model and why customers keep upgrading for training.
Memory ($MU, SK Hynix, Samsung): HBM keeps the chip fed so no joule is wasted waiting. Every HBM generation moves more data per watt. Memory demand scales with every GPU sold, and we need more memory per GPU too.
$IFNNY: World's largest maker of power semiconductors — the chips that convert and control electricity itself. An AI rack pulls over 100 kilowatts, and every power conversion between grid and chip loses energy as heat. Nvidia's moving datacenters to 800-volt power later this year, which means fewer conversions and more Infineon content per rack.
$AVGO: Attacks the energy lost between chips. Training isn't one chip thinking — it's tens of thousands of GPUs talking constantly, and that conversation is pure energy cost. Broadcom's switch chips move roughly double the data per watt each generation, and co-packaged optics convert signals to light right at the chip, cutting power cost dramatically. Faster fabric also means less GPU idle time, so every chip in the rack gets more productive per joule.
Custom silicon ($GOOGL TPUs, $AMZN Trainium): A GPU is flexible but you pay for that in wasted joules. Custom chips strip out everything except the exact math AI needs, so nearly every joule goes into useful work. Design work flows back to $AVGO and $MRVL.
This is the AI efficiency trade. Own the stack that turns electricity into intelligence.