#16
What limits AI scaling (chips, power, transformers)?
Interviews
10
Answers
10
X posts
4
Span
2023 to 2026
First asked Dec 21, 2023, most recently Oct 20, 2026. Excerpts below are from machine transcripts with filler words (um, uh, you know) and stutters removed; … marks a cut. They are not checked word for word, so use the link to hear the exact words, or the transcript link for the raw text. Timestamps: 4 cross-checked against YouTube captions, 4 from the archive transcript only, 1 checked against our transcript of the X replay, 1 unverified (archive time, not checked against the X video). X has no link-to-time option, so X links open at the start.
How the answer evolved · oldest first
Dec 21, 2023 · X Spaces: Cathie Wood & Elon Musk · asked by ARK Panelist
Q: Does diffusion help Elon, in the things that transformers are not necessarily that good at yet, like planning and remembering rather than sort of a current state?
I think transformers are important for all aspects of AI. Yeah, I mean computers are very good at remembering things. Your phone remembers the video that you took down to the last pixel, whereas most humans cannot remember who they met last week. So memory is the easy. I mean, we've already outsourced memory to computers in that, like you say, how much of human knowledge is in digital form versus in contained in …
Apr 8, 2024 · In Good Company · asked by Nicolai Tangen
Q: Is it the supply of chips? Just what is it?
Yeah, last year it was chip constrained and the hardware deployment to break it down into the three areas of people, data and hardware. Starting with hardware. Last year it was about a shift supply. People could not get enough Nvidia chips. Particularly this year it's starting to transition to a voltage transformer supply that actually getting enough voltage transformers put in place. So my sort of very niche joke …
Jun 22, 2024 · First Look Inside SpaceX's Starfactory with Elon Musk · asked by Tim Dodd
Q: So you guys are trying to move down to three grid fins, is that right?
It's not a high priority. It's a mass optimization. You definitely only need three, and I think technically you could do it with two if you're willing to oscillate.
Jan 22, 2026 · Davos · asked by Larry Fink
Q: And once again, with all that expenditures, again, how can we make sure that it's broadened?
I just think the natural thing is it's going to be very broad because AI companies will seek as many customers as they possibly can and the cost of AI will get is already very low and it's plummeting every year. I mean, almost. The cost of AI is almost meaningfully changing on a month to month basis.
Feb 5, 2026 · Dwarkesh Podcast · asked by Dwarkesh Patel
Q: Is there a way to make the GPUs even more power dense than what Nvidia and TPUS and et cetera are planning on doing that would be especially privileged in the space based world?
Well, I mean the basic math is if you can do about a kilowatt per reticle and then you'd need 100 million full radical chips to do 100 gigawatts. Yeah. So yeah, depending on what your yield assumptions are, you know that tells you how many trips you need to make. But cool. You need if you want if you're going to have 100 gigawatts of power. You need 100 million chips running that are running a kilowatt sustained. …
Feb 10, 2026 · xAI All-Hands · asked by xAI Team
Q: So one of the most incredible things about what we're doing here at Macroheart is how fast we do it, right?
All right, thank you. All right, so that was live from the front lines in Memphis. So fundamental to any AI company success is the compute advantage. And what we've demonstrated over and over again is that XAI can actually deploy more AI compute faster than anyone else. And actually, as Jensen Huang of CEO of Nvidia has said many times in interviews, there is no one faster at getting AI compute online than xai. So …
Aug 4, 2026 · SpaceX Q2 2026 Earnings Call · asked by Analyst
Q: How confident is your line of sight on permitting chips and turbines for that?
So we're actually aiming to far exceed that gigawatt number in terms of power online, power, cooling and electrical equipment. So our tentative target is to actually have 20 gigawatts at the power and cooling level online by the end of next year. Now, I don't think we're going to achieve 20 gigawatts, but we want to have a series of projects that cumulatively come to 20 gigawatts by the end of next year. Some of …
Watch on X at 0:31:23 ↗Transcript ↗Unverified timestamp: from archive transcript, not checked against the X videoSep 1, 2026 · G20 Innovation Ministerial Summit: Remarks in North Carolina · asked by Moderator
Q: And as government leaders here think about how to prepare their economies and build the right power and data infrastructure, how should they be thinking about where sort of construction, compute and power needs are going to be in the next …
Well, there actually is quite a crisis of power. So this is something if you just sort of follow the AI topic on the X platform, which by the way is where almost all of the AI discourse takes place, I think you get a very good sense for where things are headed. So that's how I get my news and it's incredibly good. Everyone who's anyone in AI posts on X, so that's why I'd recommend just go on the AI topic on X and …
Sep 29, 2026 · Elon Musk, Jensen Huang and Tom Brown on the AI Revolution · asked by Moderator
Q: Helping make the turbines. So can you talk about that?
Yeah, it's, I mean just to frame this, the average power consumption in the United states is about 500 gigawatts. So every 5 gigawatts of incremental steady state production is a 1% increase in the power used and I think actually will materialize to be approximately 1% increase in GDP because the intelligence for what keeps increasing. So if you just plot on a curve, how is the useful intelligence per watt …
Oct 20, 2026 · SpaceX Q3 2026 Earnings Call · asked by Opening remarks
Q: SpaceX superintelligence compute update (prepared remarks)
At this point I would say I'm fairly confident of being able to bring an incremental 10 gigawatts of compute online in 2027. That would be in addition to the roughly 3 gigawatts … that we'll have online by the end of this year. So I would be surprised if we're brought on less than 13 gigawatts of compute by the end of next year. And as I mentioned earlier, power, cooling, electrical equipment infrastructure is somewhat of a limiting factor. So to make sure that that is not a constraint for our build out, we're actually going to aim for an incremental 20 gigawatts of power cooling infrastructure next year.
Watch on X at 0:18:37 ↗Timestamp checked against our transcript of the X replay
In his own posts on X
@elonmusk · Apr 2, 2025
As I said a few years ago, the AI scaling constraint will move from chips to voltage transformers to electricity generation. That is worrying for US leadership in AI long-term. https://t.co/eSv2yfAsJT
View on X ↗@elonmusk · Aug 29, 2026
Consensus estimate is that ~15GW of AI compute produced in 2027 cannot be turned on in 2027. This is harder than just finding power, as you also need to build out all the transformers, wiring, liquid-cooling, (massive) chillers & complex networking. https://t.co/cRXJ1Wrkri
View on X ↗@elonmusk · Sep 1, 2026
To bring large AI data centers online requires building massive power plants, transformers, liquid cooling loops & chillers, as well as incredibly complex networking (especially for training clusters) https://t.co/pRBdDsOY1Q
View on X ↗@elonmusk · Oct 13, 2026
It’s not that supercomputer data centers are not in high demand, as they definitely are. However, getting money to build complex hardware systems with massive power, cooling & networking is not the same as doing it. SpaceX has a super cracked team that is outperforming the rest of the world by an order of magnitude in mass to orbit and even we find it very difficult!
View on X ↗
Legend: Timestamp cross-checked; Timestamp from archive, not cross-checked; Unverified: caption-only timestamp; Timestamp checked against our transcript of the X replay; Unverified timestamp: from archive transcript, not checked against the X video. Report an error →