Next in Tech

The episode presents a coherent framing (AI data centers as energy technologies rather than liabilities) and specific metrics (50 gigawatts vs 25 gigawatts capacity, 100 gigawatts potential with flexibility), but relies heavily on repetition of core arguments rather than building dense new claims. The middle section discussing workload flexibility and orchestration contains novel thinking, but much of the latter half recycles the same value proposition without adding substantially new insights.

“the US wants to add, uh, 50 gigawatts or more of AI data centers just in the next three years between now and 2028...between now and 2028, just half of that capacity in AI data centers can actually get connected to, to the grid, 25 gigawatts or less”

“there's 100 gigawatts of capacity that we could connect right this second if we had flexible data centers”

The core reframe - treating AI data centers as flexible energy assets rather than fixed loads - is genuinely novel and contrarian to conventional data center planning. However, the execution largely applies existing concepts (workload flexibility, demand response) to a new domain. The comparison to Bitcoin miners and the energy technology framing are fresh, but the episode doesn't deeply challenge orthodoxies beyond this central thesis.

“think of AI not just as a liability on the grid, a big energy suck. Think of AI as the next great energy technology”

“think of them as an energy technologies, as virtual generators or virtual power plants and batteries”

Dr. Sivaram is exceptionally well-credentialed (Rhodes Scholar, physicist, former Chief Strategy and Technology Officer of Fortune 500 Ørsted) and is actively building a company (Emerald AI, less than two years old) that is executing at scale with real commercial deployments and partnerships with Nvidia, Oracle, and Digital Realty. This is a practicing operator with legitimate domain expertise, not a consultant or academic.

“Dr. Varun Sivaram, founder and CEO of Emerald AI, formerly the group Chief Strategy and Technology Officer of orsted, a Fortune 500 energy company, a physicist and Rhodes scholar”

“we've done now five demonstrations at, uh, real Commercial data centers all over the world”

The episode includes concrete metrics (50 GW, 25 GW, 100 GW, 4 trillion), named partnerships (Nvidia, Oracle, Silicon Valley Power, Digital Realty, National Grid UK, PJM), specific deployments (Santa Clara, Manassas Virginia, London), and technical demonstrations (30% power reduction in 30 seconds, halftime soccer game demand spike). However, it lacks granular data on cost savings, actual latency impacts on workloads, or detailed performance metrics from live deployments.

“First one's in Santa Clara with Nvidia and the utility Silicon Valley Power there...The Second one is 100 megawatts, a large scale multibillion dollar facility. A data center, Digital Realty and Nvidia are building in Manassas, Virginia”

“in London. When we did a demo, we proved that at halftime of a soccer game when all the British people turn on their tea kettles and there's a 1 gigawatt spike of energy on the grid, the data center can actually help to stabilize the grid by reducing its power use”

The host asks reasonable follow-up questions and pushes on practical challenges (e.g., battery limitations, workload coordination mechanisms), but misses opportunities to press on execution risks, customer adoption barriers, or the tradeoffs of workload slowdowns. The conversation is collegial rather than adversarial; the host rarely challenges claims or asks for evidence of the claimed 30% reduction or grid stability impact.

“But crypto mining is one of those workloads that you can start it, you can stop it and shut it down. Uh, typical data center workloads though, there's no coordination between when that workload runs and what power demands look like”

“I wanted to touch on one of the points you'd made about battery energy systems and energy storage, because this is one of those things that's been hung out there is all we need to do is have batteries as buffers”