The once and future AI data center

Welcome to Runtime! Today: How the hyperscalers are adapting their data-center design strategies as demand for AI workloads, electricity, and water takes off, Google's quantum-computing "breakthrough," and the latest funding rounds in enterprise tech.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Power to the processors

Cloud hyperscalers have understood for several years that the demands of AI workloads would require new ways of thinking about data-center design, and have made substantial changes to how they architect and supply the buildings that house their massive server farms. But those changes haven't been enough to address worries about the substantial increase in electricity and water consumption that will be needed to accommodate plans for a massive AI data center buildout.

This week Microsoft and Google tried to address some of those concerns by unveiling new strategies for reducing water consumption and increasing electricity supply, respectively, as they bring new AI capacity online. Both initiatives will take several years to roll out as both companies plan to invest tens of billions in new data-center construction over that period of time.

Microsoft announced Monday that it began rolling out zero-evaporation liquid-cooling technology in its latest AI data centers as of last August.

Liquid-cooling technology is considered essential for AI servers by just about everybody at this point, as traditional air cooling or evaporative-cooling approaches can't handle the heat coming off AI servers working with Nvidia's latest chips.
Microsoft's approach, however, introduces liquid cooling at the chip level in a closed-loop system, which means "we can deliver precise temperature control without water evaporation," it said in a blog post.
That means Microsoft won't have to replace water used for cooling systems as it evaporates, which the company estimated could save 125 million liters (33 million gallons) of water a year based on its current consumption.
However, this approach will result in "a nominal increase in our annual energy usage compared to our evaporative datacenter designs across the global fleet," requiring additional research to eliminate that increase, Microsoft said.

Meanwhile, Google announced a partnership with Intersect Power on Tuesday to build new data-center complexes in the U.S. next to sources of clean energy.

Intersect Power builds large arrays of solar panels and operates four facilities in California and Texas, with plans to build more than a dozen sites over the next few years.
Under the new partnership, Google Cloud and Intersect "will develop industrial parks with gigawatts of data center capacity in the U.S." by 2026.
"When Intersect Power builds new clean energy assets in regions and projects of interest, Google will be able to provide power offtake as an anchor tenant in the co-located industrial park that would support data center development," Google said in a blog post.
Given the difficulty of building new power transmission lines in the U.S., the idea of co-locating data centers with power suppliers is gaining steam among the cloud providers, as seen in the separate deals for nuclear-powered data centers in Pennsylvania struck by AWS and Microsoft.

The next wave of data-center construction is not going to be nearly as easy as the first wave that introduced cloud computing as we know it, despite the ambitious plans that the hyperscalers have to accommodate what they believe will be a huge increase in demand for AI workloads over the rest of the decade. For one, there simply isn't as much power — even dirty power — available as they would like, and rural towns and cities with cheap land are increasingly wary of the effects that data centers can have on their communities.

That means AWS, Microsoft, and Google are going to need to reduce the impact of their AI footprint sooner rather than later, especially after acknowledging that the AI buildout of the last two years shredded their climate goals.
However, we'll find out over the next few years if companies actually need and want that much computing power as they start to put their generative AI applications into production.
If that demand doesn't materialize as expected there will be a lot of excess AI data center capacity on hand, and while that outcome would be devastating for the hyperscalers it could make their climate goals achievable again.

As the year winds down, now is a great time to sponsor Runtime and get your message in front of the enterprise tech industry leaders and decision makers that are looking for new solutions in 2025. Extend your 2024 budget into next year: Book a weekly sponsorship of Runtime before 12/31 and get a second week free! See our rate card and formats here.

Qubita be kidding me

For all the hype around generative AI, ChatGPT is a real thing you can go to right now (read the rest of this newsletter first) and use in interesting ways. Backers of quantum computing, on the other hand, have been insisting for decades that they are right around the corner from upending computing as we know it, only to acknowledge when pressed that the princess is actually in another castle.

It was Google's turn to drive such a news cycle Monday with the introduction of its Willow quantum chip, which the company said was able to perform a calculation in five minutes that would take today's most powerful supercomputers 10 septillion years to process. Turns out, however, that the calculation in question is just a random circuit sampler that "has no known real-world applications," according to IEEE Spectrum.

Google does appear to have made strides when it comes to error correction, one of the biggest issues preventing quantum computers from being anywhere close to useful for most applications. But it also said that Willow "lends credence to the notion that quantum computation occurs in many parallel universes, in line with the idea that we live in a multiverse, a prediction first made by David Deutsch," which … sigh.

Enterprise funding

Nscale raised $155 million in Series A funding to expand construction of its AI data centers, which will be used as both a public cloud service as well as private cloud infrastructure.

Astrix landed $45 million in Series B funding for its identity management service focused on "non-human identities," used to verify computers and other devices on a network.

Stainless scored $25 million in Series A funding to help companies generate software-development kits for users of their own products.

Gentrace raised $8 million in Series A funding for its testing software, which allows customers to fine-tune performance and detect errors in their AI agents and other generative AI applications.

The Runtime roundup

Oracle missed Wall Street expectations for revenue and profit, and its stock fell nearly 7% Tuesday.

MongoDB beat expectations for its third quarter, but its stock also fell Tuesday upon the news that its chief financial officer would be leaving the company.

Snowflake will no longer allow customers to use single-factor authentication methods to access its service after November 2025, closing a loophole that allowed attackers to make off with boatloads of sensitive data from improperly secured accounts earlier this year.

Chicago is considering an increase to a tax on cloud computing services used by companies based in the city, which is hard to believe exists in the first place.

Thanks for reading — see you Thursday!