Speaking with Ron Baron at the Baron Investment Conference on 6 November, Elon Musk put global output at “around 20 gigawatts of AI chips, maybe 25” this year and “maybe around 35 to 45 gigawatts of AI chips next year”, according to our transcript of the broadcast.
Treating those 35 to 45 gigawatts as IT load and applying a power usage effectiveness (PUE) of 1.2, an assumption analyst Shanu Mathew used in an X post on 8 November, gives 42 to 54 gigawatts of data-center facility capacity. We checked the multiplication; the 1.2 PUE is an assumption, not a measured figure.
Business Pill · Grid Cost Allocation
An evergreen Business Engineer explainer on who pays for new grid lines. It teaches the general idea only and does not discuss Musk, Terafab or PJM.
The key insight:Musk estimated 20 to 25 gigawatts of AI chips this year and 35 to 45 gigawatts next year, speaking at the Baron Investment Conference on 6 November. At the 1.2 PUE assumed by Shanu Mathew, the 2027 range is 42 to 54 gigawatts of data-center facility capacity, which by our arithmetic is 2.6 to 3.3 times the 16,436 megawatts of forecast data-center load growth in PJM’s 2028/2029 auction.
Musk was speaking by video link to Ron Baron on Friday, 6 November, at the 33rd Baron Investment Conference in New York, Farzad’s write-up says. Musk said a terawatt of yearly chip production is the goal behind Terafab and means “30 times what current global production will be roughly next year”, per our transcript.
A transcription posted by @pequityresearch renders the this-year estimate as “We’ll make this year around 20 GW of AI chips, maybe 25”. In that exchange Baron compared the plant’s output with the GDP of the 20th largest country, and Musk replied, “Yeah, it might even make the top 10”.
Elon Musk’s estimates of AI-chip output for 2026 and 2027, from the Baron Investment Conference on 6 November 2026, and the 2027 range as data-center facility power at a 1.2 PUE, the assumption used by Shanu Mathew on X. Facility figures and the PJM multiple are our arithmetic.
What Was Posted on X
Per this talk, @elonmusk estimates 35-45 GW of AI chip production in 2027. Treating that as IT load implies ~42–54 GW of facility capacity at 1.2 PUE.
If the U.S. absorbs 20–25 GW IT and represents 50-67% of global demand, the rest of the world would need:
>~10–25 GW IT >~12–30 GW of facility capacity at the same PUE
The U.S. allocation itself would require ~24–30 GW of facility capacity. Tall task even for next year before we start weighing 2028+.
Shanu Mathew posted on 8 November that, per the Baron conference talk, Musk estimates 35 to 45 GW of AI chip production in 2027, and that treating that as IT load implies about 42 to 54 GW of facility capacity at a 1.2 PUE. The piece uses this arithmetic; the 1.2 PUE is the post’s assumption.
Elon Musk believes Terafab’s production value can be the equivalent to the GDP of a top 10 country:
“We’ll make this year around 20 GW of AI chips, maybe 25. Then next year, depending on who whose projections you believe, it’s maybe around 35 to 45 GW of AI chips next year. But if you want to do a terawatt, you need 30 times what current global production will be roughly next year for just us.”
Ron Baron:
“I mean we could sell it to other people too, but basically we’re building 30 times what the annual production is in one plant, which you think is going to do as big production as the GDP for the 20th largest country. One one plant.”
— P Equity Research 📰 (@pequityresearch) Nov 08 05:44
A transcription posted by @pequityresearch on 8 November of Musk and Ron Baron at the 33rd Baron Investment Conference, including Musk’s remark on 20 to 25 GW of AI chips this year and 35 to 45 GW next year.
The Power Arithmetic
Mathew’s post treats the chip figure as IT load: 35 gigawatts times 1.2 is 42 gigawatts, and 45 gigawatts times 1.2 is 54 gigawatts. The post adds that if the U.S. absorbs 20 to 25 gigawatts of IT load, the U.S. allocation “would require ~24-30 GW of facility capacity”.
For scale only, PJM’s independent market monitor counted forecast data-center load growth of 16,436 megawatts in the 2028/2029 capacity auction. By our arithmetic, 42 to 54 gigawatts is 2.6 to 3.3 times that figure, though one is a global chip estimate and the other a forecast for a single U.S. grid region.
Between Musk’s two ranges, the midpoint rises from 22.5 gigawatts this year to 40 gigawatts next year, an increase of 78%, also our arithmetic.
As we read the numbers, Musk’s figures are stated in chip gigawatts, while grids plan in facility megawatts; the PUE multiplier is the bridge between the two, and the 1.2 used here is an assumption from Mathew’s post.
As we read the comparison, the PJM figure covers one U.S. grid region and a single auction year, while Musk’s estimate is global, so the multiple shows scale, not a like-for-like gap.
As we read the transcripts, Musk gave ranges, not a method, so the estimates cannot be checked against reported production data from the documents we read.
Elon Musk, Baron Investment Conference, 6 November 2026 (our transcript)
“maybe around 35 to 45 gigawatts of AI chips next year”
Three Implications
FOR GRID PLANNERS PJM’s monitor counted 16,436 megawatts of forecast data-center load growth in one auction; the facility figure from Musk’s estimate is 42 to 54 gigawatts at the assumed PUE, by our arithmetic.
FOR THE PUE ASSUMPTION Mathew’s post uses 1.2; a higher PUE raises the facility figure and a lower one reduces it.
FOR THE U.S. SHARE Mathew’s post says a U.S. allocation of 20 to 25 gigawatts of IT load would need about 24 to 30 gigawatts of facility capacity.
The analysis says: “Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”
As we read it, Musk’s estimate puts that power question in gigawatts: chips rated at 35 to 45 gigawatts need facilities of 42 to 54 gigawatts at a 1.2 PUE, by our arithmetic. The analysis does not discuss Musk’s Baron remarks.
What Is Not Established
Musk’s figures are estimates, not measured production totals, and the transcripts we read do not give his method. The 1.2 PUE is an assumption from Mathew’s post; a different PUE changes the facility figure. Chip power ratings are not the same as the electricity a data center draws in operation.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Musk estimated 35 to 45 gigawatts of AI chips next year at the Baron Investment Conference on 6 November. At a 1.2 PUE, as Shanu Mathew assumed, that is 42 to 54 gigawatts of data-center facility capacity, which is 2.6 to 3.3 times the 16,436 megawatts of forecast data-center load growth in PJM’s 2028/2029 auction, by our arithmetic.
A note on sourcing. Musk’s figures are from our transcript of the Baron Investment Conference broadcast on 6 November 2026 and a transcription posted on X by @pequityresearch. The 1.2 PUE is the assumption in Shanu Mathew’s post on X. The PJM figure is from PJM’s independent market monitor. We did not contact Musk or his companies. The facility conversion, the PJM multiple and the midpoint change are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
