Prediction Open AI assessment confidence: 95% certainty 4/5 debate potential 3/5

Sridhar: Inference will account for 90% of AI power load within four years

K.R. Sridhar · Frontier Forum: The new power map for AI infrastructure · Oct 21, 2025 · at 14:50

Bloom Energy CEO K.R. Sridhar discusses the shifting power requirements of AI computing with host Stephen Lacey.

0:00 / 0:11exact quote · 11.3s

720p mp4 · rendered on demand · StarZero watermark

“Three to four years from now, I predict that 90% of the load will be inference and 10% will be training, even though the 10% in training then will be significantly larger than the 90% training today.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Within three to four years, 90% of AI power load will be for inference and 10% will be for training, even as absolute training capacity continues to expand.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

On the October 21, 2025 podcast, K.R. Sridhar predicted that within three to four years (setting a target window between October 2028 and October 2029), 90% of AI power load would be dedicated to inference versus 10% for training, even as absolute training capacity continues to grow. As of the evaluation date of September 2, 2026, the stated prediction horizon remains more than two years away. Current mid-2026 industry research confirms that absolute training capacity is expanding significantly (via multi-hundred-megawatt and gigawatt cluster deployments), while projections for inference share by 2028–2030 vary widely—ranging from 65% in Gartner and IaaS reports up to 80–90% in lifecycle operational load and reasoning model studies. Because the target timeframe has not arrived, the claim cannot yet be resolved.

The claim specifies a 3-to-4-year timeline from October 2025, setting the resolution window in late 2028 to late 2029 (horizon: 2029-10-21), which is in the future as of September 2, 2026.
2025–2026 reports from EPRI, Epoch AI, and LBNL confirm that absolute training capacity is expanding rapidly with gigawatt-scale clusters.
Forecasts for inference share of AI compute/power by 2029–2030 remain mixed across authoritative sources, with estimates spanning 65% (Gartner, Introl compute capacity) to 80–90% (Iron Mountain, AIMultiple enterprise draw, full lifecycle operational metrics).

Checked against

[1] introl.com [2] introl.com [3] ironmountain.com [4] tensormesh.ai [5] aimultiple.com [6] digitaledgedc.com [7] civo.com [8] mckinsey.com

inference training power split datacenter AI EPRI 2024 2025 2026 AI inference share of datacenter power load training vs inference AI inference vs training power compute load percentage Gartner IDC EPRI SemiAnalysis AI inference 90 percent training 10 percent power consumption forecast

assessed as of 2026-09-02

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

⏳ prediction date: 2029-10-21; checked again automatically on the next resolution wave

Mentioned

AI load K.R. Sridhar

⚑ Spot an error? Report it

More from K.R. Sridhar

Prediction Not checkable as stated

Sridhar: AI revolution cannot be powered without natural gas

“And without natural gas, we are not going to be able to power the AI revolution. That's a fact.”

K.R. Sridhar Oct 21, 2025 ▶ 19:59 Frontier Forum: The new power map for AI infrastructure

Prediction Not checkable as stated

Sridhar: Politics will block taxpayer-funded grid expansions for hyperscalers

“Now, if the grid were to build excess capacity purely to meet this huge demand and expects the taxpayer to pay for it so the hyperscaler can benefit, That politics is never going to fly.”

K.R. Sridhar Oct 21, 2025 ▶ 5:55 Frontier Forum: The new power map for AI infrastructure

Prediction Not checkable as stated

Sridhar: Grid upgrades cannot match AI power demand within 5–7 years

“Realistically, over a five, six, seven year period, it cannot catch up with the demand base that exists out there.”

K.R. Sridhar Oct 21, 2025 ▶ 7:58 Frontier Forum: The new power map for AI infrastructure

Prediction Not checkable as stated

Sridhar: High-power AI data centers must adopt captive on-site power

“We are finally building high electricity intensity AI factories. These are, these data centers are AI factories. They have been the exception and not the rule. They are now going to have to follow the rule.”

K.R. Sridhar Oct 21, 2025 ▶ 10:47 Frontier Forum: The new power map for AI infrastructure

Prediction Held up

Sridhar: Data centers will shift to 800-volt DC architecture

“The second pain point that they are working on as we speak is shifting the mindset to an 800 volt DC bus going into Iraq. Let me explain this to you. When you are a large power plant that produces 50 megawatts, 500 megawatts worth of power, because so much pow…”

K.R. Sridhar Oct 21, 2025 ▶ 17:39 Frontier Forum: The new power map for AI infrastructure

Opinion

Sridhar: Sustainability has dropped to bottom-three priority for data center operators

“I would say for pretty much all the data center operators today, the sustainability goal has shifted from being in the top three to being in the bottom three.”

K.R. Sridhar Oct 21, 2025 ▶ 19:12 Frontier Forum: The new power map for AI infrastructure

Sridhar: Inference will account for 90% of AI power load within four years

More from K.R. Sridhar

Turn any episode into a week of clips.