On March 21, in Austin, Elon Musk held a press conference in an abandoned power plant, announcing something no one in the AI industry dared to do: make their own chips.
Tesla, SpaceX, and xAI have jointly built Terafab—a $25 billion 2-nanometer chip manufacturing plant. Chip design, lithography, manufacturing, packaging, and testing will all be completed under one roof. The first product, the Tesla AI5, boasts performance close to the NVIDIA H100, with inference costs reportedly 10 times cheaper. Samples are planned for the end of 2026, with mass production starting in 2027. The target capacity is 1 terawatt of AI computing power per year—50 times the current total global AI computing power.
80% of these chips will be sent into space, mounted on satellites to run AI, and then distributed back to Earth via Starlink. Only 20% will remain on Earth. This is because the power required for 1 terawatt of computing power is simply too much for Earth's power grid—solar radiation in space is five times that on Earth, and the vacuum of space provides much more efficient heat dissipation, eliminating the power grid bottleneck.
The news sent shockwaves through the industry.
He's going to challenge another giant that no one dares to mess with.
In the AI chip market, NVIDIA basically holds an absolute monopoly.
Global capital expenditure on AI infrastructure this year is projected to reach $400-450 billion, with $250-300 billion allocated to chip procurement. The vast majority of this money flows to one company—NVIDIA. NVIDIA has a backlog of $1 trillion in orders, a market capitalization exceeding $3 trillion, and a gross profit margin of over 55%. Its H100 chip sells for $30,000 per chip, and demand far exceeds supply.
Everyone knows that the entire AI industry is being held hostage by one company. Google has TPUs but only uses them internally, AMD is catching up but still lags far behind, and Intel's attempt at contract manufacturing has been a complete disaster. No one has truly stepped forward to say: "Let's fight head-on."
Then Musk stepped in. He did the design himself, the manufacturing himself, and even built his own factory.
What is he thinking?
The apparent reason is easy to understand—there aren't enough chips. Tesla's Autopilot, Robotaxi, Optimus robots, and xAI's Grok are all computing power-consuming monsters. Queuing up to buy NVIDIA's H100? Even with large quantities, there's no guarantee of getting one. Using TSMC for manufacturing? Apple is ahead of them in the queue.
But spending $25 billion just to solve the supply chain issue is too much. He sees something much bigger than "buying chips".
Let's lay out the cards in his hand: Tesla has millions of cars and robots on the ground. xAI has the Grok model. SpaceX has rockets to send things into space. Starlink has a global satellite network to transmit data from space back to Earth. Now Terafab is adding the last piece—chip manufacturing.
From chip manufacturing to running AI models, to launching into space, to global distribution, the entire chain is controlled by one person.
The last person to do this was Rockefeller—he controlled the entire oil supply chain, from extraction and refining to transportation and retail, becoming the most powerful person of his time. Only this time, the resource has changed from oil to computing power.
Whether it can be done is a matter of much debate.
This matter is highly controversial in the industry. Those who are optimistic say Musk has done too many "impossible" things—SpaceX, Starlink, Tesla—and each time he was questioned, he ultimately succeeded. Wedbush directly gave Tesla a target price of $600.
Those who are skeptical have strong arguments. Jensen Huang responded with a subtle statement: "Building your own chip factory is an extremely difficult challenge." TSMC's decades of accumulated process technology and know-how in over 2,000 processes cannot be bought with money alone. Bernstein calculated that achieving the 1 terawatt target could ultimately cost $5 trillion. Moreover, in the past decade, every advanced process chip factory—TSMC Arizona, Samsung Taylor, and Intel Ohio—has exceeded its budget and been delayed. There have been no exceptions.
Whether it can be done, nobody knows right now. But the more worthwhile question to consider is—
If it succeeds: a disruptive change
First, NVIDIA's monopoly will be breached.
Everyone in the AI industry has experienced the feeling of being held hostage by NVIDIA. If a near-performing alternative with inference costs ten times cheaper appears on the market, even if it's only supplied to Musk's own company, it means NVIDIA has lost a major customer. Competitive pressure will force NVIDIA to lower prices or accelerate innovation. The overall price of AI computing power may therefore decrease.
Second, AI computing power is moving from the ground to space.
The plan to send 80% of chips into space, if realized, would fundamentally change the physical basis of AI computing. Previously, everyone assumed computing power resided in data centers, limited by power grids, cooling systems, and land availability. Space deployment breaks this ceiling. Thousands of AI satellites operating in orbit, selling computing power worldwide through Starlink—the potential of this business model is immense.
Third, the power structure of the AI industry needs to be reshuffled.
Currently, the power distribution in AI is as follows: NVIDIA makes chips, TSMC handles manufacturing, Meta/Google/OpenAI provides models, and AWS/Azure offers cloud services. Each layer has different players. But if Musk integrates chip manufacturing, AI models, space deployment, and global distribution, he will be operating across four layers. This vertical integration will make all the other players uncomfortable.
Finally, there's the geopolitical aspect. Currently, over 90% of the world's advanced chips are manufactured by TSMC, whose factories are located in Taiwan. If a conflict were to occur in the Taiwan Strait, global AI production would grind to a halt. Terafab, built in the US, represents a $25 billion investment in domestic advanced chip production capacity—a seemingly worthwhile deal for Washington.
My opinion
Will it be completed on time? Most likely, it will be delayed. How much will it ultimately cost? Most likely, well over $25 billion.
But the direction is right.
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 of the puzzle represents a trillion-dollar business, and no one else has the resources to replicate it.
The power map of the 20th century was drawn with oil. This century is being redrawn by computing power.
Musk has marked out a large area on this new map. Whether anything can be grown there remains to be seen in the next few years. But the location he chose is certainly good.
