Elon Musk to Ron Baron: A 1 TW-a-Year Terafab Ambition

Ron Baron described himself as 83, with 56 years in investing. On Friday, November 6, 2026, he spoke with Elon Musk by video link at the 33rd Baron Investment Conference at the Metropolitan Opera House in New York. Farzad's video is a compilation that also inserts other interviews and earlier clips. Not every statement in it comes from that fireside chat.

Musk estimated that the rest of the world would produce AI chips with aggregate power ratings of around 20 to 25 GW in 2026 and 35 to 45 GW in 2027. These are his forecasts, not independently measured global production totals. Then he put the Terafab target on the table.

One proposed plant. The goal is to manufacture AI chips with a combined power rating of about one terawatt each year, not to generate a terawatt of electricity at the fab. Musk called that roughly 30 times his estimate of next year's global AI chip production. One terawatt divided by his 35 to 45 GW range is about 22 to 29 times. This is an ambition measured in hardware power capacity, not achieved output, chip count, wafer starts, computing operations or the whole semiconductor market. He said the chips would primarily be for internal use, though some could be sold.

Baron compared the proposed plant's output value with the GDP of the 20th-largest country. Musk said it might reach the top 10. They did not supply a calculation or a country ranking, and factory sales are not the same measure as GDP. Musk estimated a cost of "a trillion dollars probably more," phased over time, not a current funded construction budget. "If it fails it'll be crazy and if it succeeds it'll be audacious."

I covered why Terafab matters in Elon Just Made the Bet of the Century. The discussion expands on a project already described publicly by Tesla and SpaceX.

Why build a fab in America at all

Baron said nobody else is bringing silicon back at this size. Musk pushed back: TSMC and Samsung are building in the US. TSMC says its first Arizona fab entered high-volume production in late 2024. Musk also described a Samsung Texas fab as being built or already built for Tesla's AI5 and AI6. That does not establish operational volume production of those chips. Samsung's October 29 results still discuss Taylor operations commencing in 2027.

Relative to Asia, he said, US chip fabrication is still tiny. Even the optimistic case for every announced American fab is "perhaps on the order of 10%" of Asia. Taiwan does the heavy lifting. TSMC and Samsung are "amazing." Micron too on memory.

The concern, in Musk's account, is geopolitical. If that risk materializes before the US has a complete domestic supply chain, he warned, "we may not have any chips at all." He said Tesla and SpaceX work with TSMC and Samsung and want both to expand faster. He described being quoted around five years to bring a new fab to volume production, a timetable he considers too slow.

Even without a crisis, he argued the rate is simply too slow if the goal is a terawatt a year. That is why the project is called Terafab. He joked that people hear "terror" or "terra," but the root is tera, as in terawatt. "Arguably all three apply."

Musk described a research fab at Gigafactory Texas in Austin, followed by a production Terafab in Grimes County, near Texas A&M. Tesla's Q1 update independently confirms the research-fab plan at its Texas campus. The project website uses the spelling Terafab, lists a 100-million-square-foot ambition and labels output as 1 TW/year. Those are project targets, not operating results.

Musk said useful processing can occupy less than half a wafer's cycle time, sometimes around 20%, with much of the rest spent moving or waiting. Those are his estimates, not a measured industry-wide average established here. He argues that a research fab combining logic, memory, lithography, packaging and testing could shorten an iteration from a year to a month. That is his proposed improvement, not a demonstrated production cycle.

Baron contrasted roughly 550 GW in the US with three times that in China, and annual additions of about 300 GW versus 20 GW. Musk questioned the annual-addition figure and predicted China would move toward four times US power output. These remarks mix average output with installed capacity. EIA reports 4.43 trillion kWh of US utility-scale generation in 2025 and 1.28 TW of capacity. China's energy regulator published 10.58 trillion kWh and 3.89 TW respectively: roughly 2.4 times the generation and three times the capacity, using those series. Neither comparison validates the conference's annual-addition figures. Musk's claim that orbital AI is the way to compete is his judgment, not a conclusion established by these statistics.

Baron's stake, and the merger question

In the Baron clips included in the video, he said his firm's Musk-related investments had made about $30 billion and could make hundreds of billions more over the next decade. He described selling a quarter of a position for clients, followed by another doubling, and a roughly 40-fold gain from his basis. These are his accounts, not independently reconciled returns or realized gains. His further fivefold return is a forecast. He called Tesla his most remarkable investment in 56 years and described a contemplated 100-million-square-foot Terafab. The video includes separate television footage of Baron, so these lines should not all be assigned to the conference.

In another Baron clip, he said he would support Musk's decision about a possible SpaceX and Tesla combination. An interviewer discussed what independent committees might decide and at what price. That exchange does not establish a completed deal, approved terms or a committee outcome. I previously covered the possibility in Elon Musk Makes Shocking Predictions.

The Airstream next to the power plant

Baron praised Musk's use of an Airstream near the power plant. Musk said he likes the trailer and described the Memphis-area Colossus operation as producing over 1 GW, with a goal of 1.5 to 2 GW. He claimed a build time of about a year against a usual five. Those output and schedule comparisons are his account, not independently established plant performance. He said he walks the site, pulls heavy power cables, works on fiber and electrical switches, and uses plant-management software to understand the work. I covered the campus in Inside Colossus 2: The Race Is Time to Useful Work.

He also described what he called the Jose meme: one person digging the hole, eight people watching. Issuing more purchase orders does not create more work. Only more people who dig. SpaceX wants to be Jose. Actual physical work, not a PO machine. That is why he thinks half-finished data centers are sitting dark across the country while demand for compute is off the charts. Speculators think you hand a GPU purchase order to a contractor and rake in cash. "The reality is it's insanely hard."

The video also inserts a separate All-In discussion comparing compute payback periods. It is not part of the Baron fireside, and the clip does not establish audited, like-for-like returns. My earlier article, SpaceX CFO Just Put $100B ARR and Orbital Compute on a Clock, covers Johnsen's own short-payback argument separately.

SpaceX money: a $100B run-rate target, then the hard guess

Baron floated SpaceX ending 2026 around $100 billion of annualized revenue, then somewhere between $500 billion and a trillion next year. Elon tightened the language.

Musk said SpaceX was tracking toward an annualized revenue rate above $100 billion in December 2026. That is a target based on one month's revenue pace, not $100 billion of revenue already earned during 2026. He said he expects to meet it unless an extreme disruption intervenes.

For December 2027, his best guess is annual recurring revenue above $500 billion. A trillion is "not out of the question." He was careful: guess, not promise. Hyperscalers and dozens of smaller companies are "hounding" SpaceX for more compute. He said they could rent it several times over. When they tried to offload inference so they could train more, he compared the market to the Monty Python cheese shop. Every kind of cheese. No cheese.

The video also revisits the trillion-dollar discussion from the Q3 call. That older excerpt should not be mistaken for another new statement at the conference. Musk described aiming for deadlines he thinks have a 50% chance of being met. That explains his stated scheduling approach; it does not prove a measured on-time record.

Orbital AI is not a slide deck

Baron asked about data centers in space. Musk said SpaceX had launched satellites carrying Nvidia GPUs and a Google TPU, and that they were working to his knowledge. These are SpaceX-launched payloads, not necessarily SpaceX-owned satellites. Starcloud confirms an Nvidia H100 satellite launched in November 2025. Google reported successful TPU tests in orbit on October 30, 2026. For SpaceX's own planned Starmind computers, Musk guessed first launches around Q3 2027, followed by significant volume in 2028 and much larger volume in 2029. Nvidia's August 24 announcement confirms the collaboration and styles the name Starmind. It does not guarantee Musk's launch dates.

Musk sees a path to adding about 100 GW a year of solar-powered orbital AI. For scale, US utility-scale generation in 2025 averaged about 506 GW, while EIA lists about 1.28 TW of generating capacity. Average generation, electricity consumption and maximum capacity are different measures. His comparison of 200 GW of new continuous load with roughly 500 GW of average power yields about 40%, before accounting for other requirements. His prediction that space becomes the most economical place for large-scale AI in 30 to 36 months remains a forecast dependent on launch, power, cooling and networking, not established economics.

Musk illustrated the economics with 100 GW of added compute capacity each year. At an assumed annual revenue rate of $10 per watt, that would add $1 trillion of annualized revenue; at $50, $5 trillion. The arithmetic works under those assumptions. It is not a booked contract, profit estimate or a valuation of the hardware, and cumulative revenue would require the satellites to remain usable and rented. They have finite operating lives and costs. I wrote the broader thesis in SpaceX Bets Its Trillion-Dollar Future on Orbital AI, Not Rockets.

Starship, Starlink, and automated traffic

Full reusability is Musk's cost argument. A 2018 paper by NASA Ames researcher Harry Jones estimated average launch cost for systems first flown from 1970 to 2000 at about $18,500 per kilogram to low Earth orbit in 2018 dollars. That is not a universal 2010 price, and the early Falcon 9 cost decline predates routine booster reuse. Musk's goal of under $100 per kilogram with Starship is still a target. He described catching both stages with the tower arms to enable rapid reuse, and named Ship 42 as the next ship-catch candidate. In the clip he said "about a week from now." On November 5, reporting from UP.Summit put Flight 15's target at November 13. Neither reference guarantees a launch, a catch or a subsequent reflight.

Starlink V3, he said, has over 10 times the capability of V2, and they plan to launch far more of them. Bandwidth in a couple of years, in his framing, could be on the order of 100 times today's. Baron had been modeling Starlink as a consumer, government, and enterprise business. Elon's update: agentic demand changes the ceiling.

Musk cited Cloudflare as evidence for growing machine demand. Cloudflare's September 27 founders' letter says automated traffic passed human traffic in May 2026 and could reach 1,000 times human traffic in five years if current trends continue. Automated traffic includes more than AI agents, so this is not proof that agents alone already exceed humans. It is also a conditional forecast about traffic, not a measurement of satellite bandwidth demand. Musk thinks it may be conservative. Baron discussed a possible 100,000-satellite Starlink system; that is a longer-term ambition here, not confirmation of regulatory authorization.

Digital AI, robots, and what he wants on the tombstone

Excluding robotics, Musk guessed that digital AI could increase global economic output by 20% to 30%, described loosely as about $25 trillion, and could double white-collar productivity. Using the $130 trillion base Baron supplied, the exact arithmetic is $26 trillion to $39 trillion, not $25 trillion across the whole range. The base is part of their illustration, not a separately established GDP estimate here. The growth and productivity numbers are Musk's forecasts, not measured gains. Physical AI is a separate business he hopes his companies will lead.

Humanoid robots are the physical layer in Musk's forecast. He imagined machines working up to 168 hours a week and producing about five times a human's weekly output, with billions of robots over time and perhaps three or four industrial robots per human. A 168-hour week is 4.2 times a 40-hour week before assumptions about speed and breaks. Those are assumptions about future capabilities, not demonstrated Optimus productivity or uptime. His further predictions are universal high income and an abundance of goods and services that makes today's money system seem like seashells to a superintelligence.

Baron's last question was personal. When you are 83, how do you want people to think about you? Elon said perception and reality can diverge. What he wants in reality: "a significant contribution in the furtherance of civilization."

What has to go right

The plans depend on several things going right. Terafab would need to progress from research in Austin to volume production in Grimes County. SpaceX would need to meet Musk's December run-rate target and his much larger 2027 forecast. Starmind would need to launch and scale on the proposed schedule. Starship would need successful catches, reflight and much lower costs. The Memphis-area build would need to keep converting delivered equipment into working capacity.

Musk described his own forecasts as fallible and sometimes optimistic on timing. His confidence is part of the story, not independent evidence that the targets will be achieved. Baron's past returns, as he describes them, do not establish future performance.

I own SpaceX and Tesla stock. Nothing here is investment advice.