By Marc J. Elzenbeck
Earlier in 2024, Microsoft’s current leadership proposed a plan to reduce power consumption from multiple groups within the company in order to re-route its capacity to its US-based Graphical Processing Unit (GPU) servers. GPUs power AI. The plans feature shifting the “compute” for core services, for products like MS Office, to other countries. Not a known recipe for service quality success.
Why the frenzy? The idea is to put AI on top of enterprise subscription software, or Software as a Service (SaaS). In sum, it provides a way to search and organize institutional memory, then make compelling stories with it. The more data you generate and collect, the more data trails you have to store, the more permutations you’ve got to mine and re-arrange into Business Intelligence. Stare into the void and it stares back at you. Exponentially.
Not feeling enthralled yet? To understand Big Tech’s desperate commercial affair with AI, all you have to do is look at its flat or falling SaaS revenues. AI is the first new-new thing Salesforce and like providers have had to sell in 20 years. It’s a justifiable upgrade to squeeze out another $30-$50 license per seat a month.
Here is Sales Scenario One: someone may have looked at your ad for floormats on Facebook back in 2016 … and they just leased a new Mazda! AI spots it and sends them a 20% discount coupon. Due to improved ad-targeting efficiencies, the climate crisis is averted.
Management has decided to scrap its energy free-up plan. Not because it was seen as dumb and generated internal resistance, but because it wasn’t aggressive enough. Telling executives to turn off their lights and charge laptops at home is good practice, but a single Nvidia H100 GPU blows through about the same daily kilowatts as a household of five people. That’s just raw consumption of one processor. It does not account for the construction, building, cooling, maintenance, and finance costs spread across the typical 10,000 GPUs in one hyperscaled data center.
A creative person at Microsoft hatched a cunning plan: re-open the Three Mile Island nuclear power plant, aka America’s Chernobyl, closed since 1979 after 80% or so of the uranium in its Unit 2 reactor burned off into the air above the Susquehanna River. Its owner, Centennial Energy, was somewhat reluctant to reopen the disaster site until Microsoft committed to pay them more than double (according to industry analysts) the going electricity rate for 20 years. Three Mile Island is scheduled to start back up in 2028.
At a mere .85 Gigawatts, however, Three Mile Island’s output isn’t up to powering the New Industrial Revolution. Something more along the lines of an AI Manhattan Project is afoot. Led by OpenAI’s co-founder, top execs from Anthropic, Microsoft, Google, and Nvidia visited the White House last week and informed officials they’re going to require a round number of 5-Gigawatt Mega-Centers to remain competitive in the global AI race against space aliens, or possibly China. 5 Gigawatts? That’s what it takes to power the city of Miami, Florida.
US Energy Secretary Jennifer Granholm is doing a PR tour to explain that fast-tracking permits for small nuclear reactors is the path forward, per a CNBC interview: “They’re committed – telling us – to bringing that power with them, which is why the need for small nuclear reactors, collocating data centers with small reactors, or partnerships.”
One suspects this level of investment isn’t simply so customers can get chatbots to write haikus about how bad their company is (an actual case).
It might be wise to look for a higher-level motivation. Siri was originally a large Defense Advanced Research Projects Agency (DARPA) weapons control project, one that heavily employed AI. The basis for Facebook was once a well-funded DARPA project based on digital pioneer Gordon Bell’s research. Google? DARPA. Boiled down to purely technical steps, targeting weapons with AI is almost indistinguishable from targeting advertising.
Targeting sorts through available data to form a probabilistic picture of what someone is likely to do when presented with opportunity within its capabilities. Here is Sales Scenario Two: Cellular and financial data shows a subject participates in a known terrorist cell, but when home, visits children at bedtime and then has a 98.4% chance of stepping onto the outdoor balcony at 9:15 p.m. local time. A pre-programmed explosive drone is sent to intercept the threat. The Western Way of Life is saved.
According to the technology news site Freethink, in order to secure adequate energy sources, Microsoft has been tuning its own AI to automatically draft and submit the documentation for government proposals required to gain regulatory approvals. These approvals will be in turn used to build nuclear plants to power its future generative AI data centers.
