March 21, Austin. Elon Musk held a launch event in an abandoned power plant, announcing something no one in the AI industry had dared to do: building their own chips.
Tesla, SpaceX, and xAI are jointly building Terafab—a $25 billion 2-nanometer chip manufacturing facility—integrating chip design, lithography, fabrication, packaging, and testing under one roof. The first product, the Tesla AI5, delivers single-chip performance approaching the NVIDIA H100, with claimed inference costs 10 times lower. Sampling is planned for late 2026, with mass production targeted for 2027. The target capacity is 1 terawatt of AI computing power annually—50 times the current global total AI computing capacity.
Eighty percent of the chips will be sent to space to run AI on satellites, with data distributed back to Earth via Starlink. Only 20% will remain on Earth, as the power grid cannot support the electricity required for one terawatt of computing power—space receives five times more solar radiation than Earth, offers superior vacuum-based heat dissipation, and has no grid constraints.
As soon as the news broke, the industry was shaken.
He’s taking on another giant that no one dares to touch.
In the AI chip market, NVIDIA is essentially in absolute monopoly.
Global AI infrastructure capital expenditures this year are projected at $400–450 billion, with $250–300 billion allocated for chip purchases. The vast majority of this spending is flowing to one company—NVIDIA. It has over $1 trillion in backlog orders, a market capitalization exceeding $3 trillion, and a gross margin above 55%. Each H100 chip sells for $30,000 and is in high demand.
The entire AI industry is being strangled by one company—everyone knows it. Google has TPUs but uses them exclusively for itself; AMD is catching up but still lags far behind; Intel wants to be a foundry but has messed it up badly. No one has truly stepped forward to say: “I’ll take you head-on.”
Then Musk stepped in—designing it himself, manufacturing it himself, and even building the factory himself.
What is he thinking?
The surface reason is easy to understand—there’s not enough chip supply. Tesla’s autonomous driving, robotaxi, Optimus robot, and xAI’s Grok are all power-hungry monsters. Trying to buy H100s from NVIDIA? Even in large quantities, you might not get through the queue. Going to TSMC for manufacturing? Apple is ahead in line.
But $25 billion is too much just to address the supply chain. What he saw was much bigger than "buying chips."
Lay down his cards: Tesla has millions of vehicles and robots on the ground. xAI has the Grok large model. SpaceX has rockets capable of sending payloads into space. Starlink has a global satellite network that transmits data from orbit back to Earth. Now Terafab completes the final piece—chip manufacturing.
From designing chips, to running AI models, to launching into space, to global distribution—all in one person’s hands.
The last person to do this was Rockefeller—control over the entire chain from oil extraction to refining, transportation, and retail made him the most powerful man of his era. This time, the resource has shifted from oil to computing power.
Can it be done? Opinions vary.
This issue is highly controversial in the industry. Supporters point out that Musk has repeatedly accomplished what many deemed impossible—SpaceX, Starlink, Tesla—each time proving critics wrong. Wedbush has set a $600 price target for Tesla.
Skeptics have strong arguments. Jensen Huang responded with a subtle remark: "Building your own chip factory is an extremely difficult challenge." TSMC’s manufacturing expertise has been accumulated over decades, encompassing thousands of process steps—know-how that cannot be bought with money alone. Bernstein ran the numbers: achieving the 1-terawatt goal could ultimately cost $5 trillion. Moreover, over the past decade, every advanced chip fabrication plant—TSMC’s Arizona facility, Samsung’s Taylor plant, and Intel’s Ohio site—has exceeded its budget and been delayed. Not a single exception.
Can it be done? No one knows right now. But a more worthwhile question to consider is—
If achieved: a disruptive landscape
First, NVIDIA's monopoly will be cracked open.
The entire AI industry has felt the pinch of being held back by NVIDIA—everyone in the field has experienced it. If a substitute emerges on the market with performance close to NVIDIA’s but at one-tenth the inference cost—even if only supplied to Musk’s own companies—it would mean NVIDIA loses a major customer. This competitive pressure would force NVIDIA to lower prices or accelerate innovation, potentially driving down the overall cost of AI computing power.
Second, AI computing power is moving from Earth to space.
If the plan to launch 80% of chips into space is realized, the physical foundation of AI computing will change. Previously, everyone assumed computing power was confined to data centers, limited by power grids, cooling, and land availability. Space deployment breaks through this ceiling. Thousands of AI satellites operating in orbit, selling computing power globally via Starlink—this business model offers enormous potential.
Third, the power structure in the AI industry must be reshuffled.
Currently, AI power is distributed as follows: NVIDIA makes the chips, TSMC handles manufacturing, Meta/Google/OpenAI develop the models, and AWS/Azure provide cloud services. Each layer has different players. But if Musk connects chip manufacturing, AI models, space deployment, and global distribution all in one, he alone spans all four layers. This kind of vertical integration will make every other player uncomfortable.
Finally, geopolitics. Over 90% of the world’s advanced chips are currently produced by TSMC, whose factories are located in Taiwan. If a crisis were to occur in the Taiwan Strait, global AI operations would grind to a halt. Terafab, built on U.S. soil, offers Washington a strategic advantage: for $25 billion, it gains domestic production capacity for advanced chips—a deal that makes sense no matter how you calculate it.
My opinion
On schedule? Likely to be delayed. Final cost? Likely to far exceed $25 billion.
But the direction is correct.
Musk is currently the only person in the world who simultaneously controls chip manufacturing, AI models, rocket launches, global satellite communications, and millions of hardware terminals. Each of these five pieces represents a business worth trillions, and no one else has the conditions to replicate them.
The power map of the 20th century was drawn by oil. This century is being redrawn by computational power.
Musk has marked out a large area on this new map. Whether anything can be grown there will depend on the next few years. But the location he chose is indeed excellent.