Responsible AI/ML Governance & Risk Management | EU AI Act, GDPR, EO 14100 | NIST AI RMF 1.0 & 2.0 | Senior Non-Resident Fellow, AI and Global Governance | Opinions are my own
Everyone is focused on hoarding or building their own #AIchips (hello Google, Meta, Microsoft, Amazon, Musk, Andreessen!) and frothing for the latest foundational model release but few understand the power generation macro forces at play required for scaling the technology. This poses questions about #sovereignAI for future US national security and seasonal climate extremes placing demand on an already taxed grid (Northern Virginia, Texas, Georgia, Oregon, etc.). #GoldmanSachs July 2024 Research Newsletter Issue 129 is worth a read especially Brian Janous and Carly Davenport interviews (pg 15 to 18) about challenges to unlock new power grid capacity, which Sachs cites as one of the growth speed spike strips to scaling AI (in the US). 👵 The US power grid is aged and suffers from legacy equipment. ☁ Migration of on prem data to the cloud resulted in significant increase in computation with almost no rise in electricity usage. 📈 Utilities are fielding hundreds of requests for huge amounts of power as everyone chases the AI wave, but only a fraction of demand will be realized. AEP, one of the largest US electricity companies has received 80 to 90 GW of load requests; only 15GW is likely real because many AI projects that companies envisioning never see the light of day. 15GW is massive give AEP currently owns/operates 23GW of generating capacity in the US. 🚄 Utilities have not experienced a period of load growth in almost 2 decades and are not prepared for- or even capable of matching- the speed at which AI is developing. 🕯 Regulatory lags, interconnection and supply chain constraints are also impediments to meeting rising power demand. Total capacity of power projects waiting to connect to the grid grew nearly 40% last year, with wait times ranging from 40 to 80 months and lead times for critical electrical components such as transformers and switchers have substantially increased. Until issues can be resolved and the grid can catch up, a significant power crunch will force utilities and states to pick and choose who receives power. 🗓 The power constraint issue cannot be resolved without a significant buildout of electric grid infrastructure. Power project developers start the process 5 to 7 years in advance. ✖ Data centers will likely more than double their electricity grid use by 2030; power demand growth will rise to levels not seen since the turn of the century. ⏱ Data centers that power AI models must run 24/7 given the nature of AI workloads so that requires a constant energy source like natural gas that can be dispatched on demand, rather than renewables which is intermittent. The US historically hasn’t demonstrated the best track record of building new capacity in the areas of nuclear energy.
John Bartol
5d
Thank you for calling this out and providing such a focused commentary. This conversation continues to be one of the more interesting, pressing, and concerning when I am talking with fellow AI and GRC practitioners/observers, and it just adds to the irony of early statements from previous years that AI would help us better understand and solve climate and energy problems. Not sure we are getting any closer to either.
Founder @ The Cyber Boardroom, Chief Scientist @ Glasswall, vCISO and GenAI expert
5dI think the cleaver ones in this play are Groq who are creating micro-processors for inference and not model generation. More and more the models will become a commodity and the action will be happening at the prompt level (which is where we need processors that are energy efficient at running the modes, not in creating the models)