Generative AI doesn’t burn calories, it burns kW
Compute-hungry generative AI workloads will consume 80% of data centre power over the next 15 years driving the infrastructure market to dwarf today’s $300bn global public cloud market, according to DigitalBridge CEO Marc Ganzi.
Speaking at the company’s Q2 2023 earnings results Ganzi said “[Public cloud] is about a $300 billion annual business with large cloud service providers representing about two-thirds of the market. Based on the feedback we’re getting from our CEOs and their conversations with customers around ramping capacity requirements, we believe this opportunity is going to be at least as big as the public cloud market was over a decade ago.”
“Public cloud, which really has been building and leasing space to the data centre marketplace over the last 10 years, is at about 13 gigawatts,” he said. Ultimately, to drive AI and to get networks to where…we believe they can go, we believe the opportunity set is close to 38 gigawatts.
We are toast
Using numbers that would get most ESG professionals to review their companies’ carbon neutral plans, Ganzi said new specialised AI chips, GPUs that NVIDIA, AMD and Intel are producing, consume 2-3 times the power of prior generations. “To put that into context, the latest chips literally consume as much power as a toaster,” he said.
The second driver was the size of AI models. Large language models like ChatGPT have billions of parameters with the latest models, GPT4 reaching almost 1 trillion. “Data centre power consumption is set to rise dramatically and increasingly be dominated by these AI workloads,” he said. “It’s estimated that 80% of the data centre power will…be consumed by AI over the next 15 years.”
Ganzi gave an example of how this power-hungry AI world will impact infrastructure. “It starts with AI model training, which will happen principally in large public and private clouds,” he said adding it would mostly happen at data centres like DigitalBridge’s Vantage, Switch or Scala DCs building 100-, 200-megawatt 400-megawatt campuses.
“Instead of drawing 50 megawatts today, it will be 200 megawatts in the future, but with the same footprint,” he said. “There’s going to be higher power density on a per-rack basis with each rack filled with power-hungry GPUs, drawing 40 or more kWs compared to traditional data centre racks, which draw 10kWs or less.”
Edge gets a boost too
Beyond supercharged DCs, Ganzi expects an uplift right across the network to users. Once an AI model is cloud trained, it’s ready to be utilised by consumers, enterprises and machines, a process Ganzi called AI inference.
“Gen AI is edge delivered. This is an important concept for all investors to get their minds wrapped around. For the actual applications, we use our phones, we use our laptops and speed and latency matter. It’s not efficient to send data back and forth to Ashburn, Virginia from Boston or Miami,” he said. “Edge data centres, fibre, cell towers and small cells is going to become increasingly relevant over the next three to five years.”
“We’re seeing it selectively already, to be honest, with inbound interest from things like metro fibre capacity that are in the order of magnitude larger from cloud service providers, which is why you’ve seen some of our fibre revenues pick up in the growth in fibre and enterprise fibre pick up,” he added.
Results at a glance
The company reported Q2 2023 total revenues of $425m, GAAP net loss attributable to common stockholders of $22m and distributable earnings of $10m.
Recurring revenues across the data centre portfolio were up 22% compared to Q2 2022. “Towers also driving substantial growth, up 21% year-over-year as carriers proliferate their 5G networks around the world,” he said. “Fibre has seen a dramatic increase, up 15%.” He added that small cells quite didn’t deliver the double-digit growth the company was hoping for.
Year-to-date, we’ve deployed $4bn supporting the growth of our portfolio companies in a challenging macro environment. I believe this enforces the consistent need for new digital infrastructure by our customers despite some of those headwinds,” he added.
Ganzi’s vivid picture of an AI world eclipsing the public cloud was presented in a backdrop of the digital infrastructure giant announcing it was on track to meet its next fund-raising targets. He revealed that this quarter DigitalBridge’s fundraising hit $2.7bn – far exceeding Q1’s $700m total according to Infrastructure Investor. Part of that, the Digital Bridge Partners III, brought in an initial $1.2bn, which is 15% of the $8bn target.