The fact that Brad Gerstner and Bill Gurley devoted an hour of their podcast to nuclear energy is a clear signal that Silicon Valley is entering a new era of AI infrastructure.
AI is the next big thing in Silicon Valley. NVIDIA is selling chips faster than they can make them. If GPUs are the brain of AI, electricity is the food that keeps it going and growing. But there may simply not be enough power for the next generations of LLMs that will cost billions to train.
Tech companies are running into real world problems that can’t be solved with a software upgrade: getting access to a lot of stable, reliable power. These are infrastructure problems. Policy problems. Regulatory problems. So what will happen when one of the biggest commercial opportunities meets the slow-moving world of infrastructure? Here are my 6 predictions:
- Big tech will ramp up lobbying efforts to governments around the world. They will push regulators to get more nuclear built, to let them buy existing power plants, to make it easier to bring new fossil fuel generators online and to build out the grid.
- Data centers will compete with consumers for base-load power. Once that narrative makes it into main stream media, consumers will push back and kick off a broader discussion around access to affordable electricity and who should pay for the infrastructure investments needed.
- Since latency is not an issue for training LLM models, a lot of it will move abroad where restrictions on buying baseload from a grid operator are lower and it’s easier to get access to power.
- In the US, datacenters will go where power is available or can be brought online quickly. Or in other words ”time to power” is more important than “cost of power”. This is pushing datacenters to places like Texas where it’s easier to get interconnection approvals and build new gas power plants.
- The cloud providers are abandoning their decarbonization goals as fast as they can buy GPUs. Stable clean energy just isn’t available at large scale and datacenters need reliable power. They might buy renewable energy credits (RECs) but they are a bit of a cheat since they don’t address long-term intermittency challenges of renewables (i.e. buying solar RECs from solar that was generated between 9am-6pm can’t power your data centers from 7-9pm when the load on the grid is the highest).
- Tech companies will invest in emerging clean technologies like long-duration storage, geothermal or SMRs (small modular reactors) but only a small percentage of the total incremental power will be supplied by these new technologies in the next 5 years as they are still early in their maturity cycle.
Big tech companies are forced to become AI infrastructure companies, but infrastructure is notoriously hard and slow to scale. The race to build a “digital god” is so existential to big tech that they are willing to spend ungodly sums on not losing the battle.