Aishwarya Mahesh

Worthwhile to read this SemiAnalysis piece on whether AI datacenters are actually driving up electric bills. In PJM, household bills are up ~15% but the main driver is a capacity auction where prices spiked 9.3x based on a forecast that keeps getting revised down. PJM overestimated datacenter load by 800MW in 2024 and another 1.1GW in 2025. Meanwhile Texas has higher forecast AI buildout with no price shock because ERCOT lets real-time market signals do the work instead of a central planner's simulation.

who is hyped for this and which vc fund is going to host watch party when it comes out :)

The memory required by Nvidia AI chips is surging massively, with the new Rubin needing 288GB. Their upcoming Rubin chip needs 288GB of RAM just to operate. To put that into perspective, that is 800% more memory than you will find in a top-tier desktop computer. It is also 2300% more than what is inside a high-end smartphone. Large-scale purchases by OpenAI and Alphabet have depleted supply, driving 16GB DDR4 prices up 2352% to $76.90 and causing a massive global memory shortage. The global memory chip shortage is completely out of control. Hardware manufacturers simply cannot make memory chips fast enough to keep up with the demands of these new AI processors. Chart from Bloomberg.

Worthwhile to read this SemiAnalysis piece on whether AI datacenters are actually driving up electric bills. In PJM, household bills are up ~15% but the main driver is a capacity auction where prices spiked 9.3x based on a forecast that keeps getting revised down. PJM overestimated datacenter load by 800MW in 2024 and another 1.1GW in 2025. Meanwhile Texas has higher forecast AI buildout with no price shock because ERCOT lets real-time market signals do the work instead of a central planner's simulation.

Worthwhile to read this SemiAnalysis piece on whether AI datacenters are actually driving up electric bills. In PJM, household bills are up ~15% but the main driver is a capacity auction where prices spiked 9.3x based on a forecast that keeps getting revised down. PJM overestimated datacenter load by 800MW in 2024 and another 1.1GW in 2025. Meanwhile Texas has higher forecast AI buildout with no price shock because ERCOT lets real-time market signals do the work instead of a central planner's simulation.

Founder of Chicago-based prop trading firm DRW says compute will be world's top commodity in 10 years. People will spend more on GPUs than an oil, which means that there should be a futures financial market for GPUs.  Interesting implications for startups and cloud providers: