There are claims in the media that Microsoft is going to build a $100 billion computing facility for AI that will use 5 GW of power. Since the power is needed 24/7 they want to use nuclear power rather than renewables:
https://www.datacenterdynamics.com/en/news/microsoft-openai-consider-100bn-5gw-stargate-ai-data-center-report/
https://www.reuters.com/technology/microsoft-openai-planning-100-billion-data-center-project-information-reports-2024-03-29/
That is of course, over 100 times the cost of any current supercomputer, and over 100 times the power budget, enough power for several million homes.
Naturally, if this is true it may be the case that other large tech companies in the US and China build similar facilities, each dwarfing the computing resources of the government laboratories unless they follow suit.
9 comments:
That is bigger than LLNL and LANL computers! Wow
Microsoft could easily foot the bill for this, they have 80 billion cash on the books, their business throws off about 70 billion cash a year, they are the world's most valuable company with a perfect AAA credit rating, meaning they could borrow many times that amount if they had a need to.
How will Microsoft afford this, by selling our data to the highest bidder?
What version of ethernet has less latency than infiniband? Time equals speed.
They're going to get a nuclear reactor OK'd and built in that time frame? - truly warp speed, Mr. Scot.
I'm sure they mean Terabit ethernet:
https://en.wikipedia.org/wiki/Terabit_Ethernet
There would be latency anyways in the links, for a computer center spread out over 1 square mile it would be 8 microseconds due to the speed of light in the optical fibre.
It looks like this new facility, if built, would be regulated under Biden's executive order via the defense production act, since it would exceed certain thresholds:
https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/
https://www.theregister.com/2023/11/05/biden_ai_reporting_thresholds/
That of course, would set Microsoft up with some monopoly power, and could establish tight government control over the results they produce. These regulations would benefit Microsoft of course, as they would provide a barrier to entry through regulatory capture:
https://www.datasciencecentral.com/regulatory-capture-why-ai-regulation-favours-the-incumbents/
This could bring about vast profits, of course, if the machine provides valuable breakthroughs such as AI systems with superhuman intelligence, and perhaps also breakthrough results in biotechnology, materials science, energy, or if the machine can be used to design successor machines with far greater capacity -- creating an unsurpassable advantage that will permanently establish Microsoft's (and thereby the United States') control of all cutting-edge science and technology.
Another thought on building such a facility of course, is that one might anticipate that due to the progress in computing technology, it would require constant purchasing of equipment and upgrades. This, along with the cost of maintaining and using the facility, could easily run to say another $30 or $40 billion a year. (The very high costs mean that Microsoft is anticipating even more return on its investment, of course, something which would transform their company). Nuclear plants on the other hand have a high upfront cost, but might be rated for up to 60 year service lifetimes, and the power costs little to produce once the reactor is built:
It does seem like the United States lacks key parts of the nuclear supply chain needed to build and operate nuclear power plants, including those it already possesses.
Most nuclear plant designs require the production of large steel forgings such as reactor vessels. Over half of the world's capacity to produce these is in Russia and China, it seems, there is no production capacity in North America, the free world produces these mainly in Japan.
Russia is believed to have approximately half of the world's uranium enrichment capacity, and much of our nuclear fuel comes directly from Russia or is bought through third countries like Kazakhstan with enrichment supplied by Russia. We are still importing enriched uranium directly from Russia two years into the Ukraine war, and currently have only 13% of the world's enrichment capacity while producing 33% of the world's nuclear power.
Over half of the world's uranium is mined by Muslim states in central Asia, including Kazakhstan, there are large exploitable deposits in Canada and Australia but these each contribute less than 10% to world supply. Kazakhstan is part of China's belt and road initiative, and also subject to Russian influence.
There are easy ways to address some of these issues, for example CANDU reactors do not require enriched uranium, due to the properties of the heavy water moderator, and also can be fueled with thorium blended fuel I believe. Most of Canada's reactors use this technology which is fully mature and it has a good safety and operational record.
This is a nice article too that goes into some of the practical reasons it is difficult to pursue nuclear fission energy projects in the United States:
https://thebreakthrough.org/blog/advanced-nuclear-energy-is-in-trouble
Particular regulations may impede licensing of new designs, or require expensive construction materials and techniques that in some cases do not contribute to safety. High interest rates penalize the up-front capital investments needed for nuclear construction, while raising the cost of materials needed. Small modular reactors only make economic sense when mass-produced but no firm orders whatsoever are placed, or token orders for a few reactors, which then fall through. Startup companies have issues of course, obtaining enriched fuel since there are shortages, or using low-cost suppliers in China.
Despite the US having the largest fleet of nuclear plants, with the best operational history, most nuclear plants being constructed or proposed worldwide are of Chinese or Russian design it seems. China produces about a third of the world's electricity, with a recent growth rate of about 7% a year, creating a strong need for new power sources.
https://www.world-nuclear.org/information-library/current-and-future-generation/plans-for-new-reactors-worldwide.aspx
One thing that is interesting to consider, is that it seems Amazon's cloud is already the same approximate scale as the Microsoft proposal. They have invested about $150 billion dollars in it, and it is a business with over $100 billion annual revenue. Also, of course, Amazon has a 20 GW portfolio of renewables to cancel out its carbon footprint, and assuming reasonable capacity factors for wind or solar this might equate to something like 5 GW on 24/7. A web search turns up other numbers like 25 million square feet in data centres which is 1 square mile and so on.
Microsoft's existing Azure cloud is the second largest after Amazon -- it also exceeds $100 billion in annual revenue. It is harder to track down the exact details on power usage or square feet, but it must be comparable.
What is novel about the Microsoft plan, is to locate that much computing equipment in a single location, rather than many data centres located around the planet, as it seems they are already making cloud investments roughly on that scale.
Post a Comment