Musk's AI arms race accelerates
```
Author | Chai Xuchen
Editor | Zhou Zhiyu
In the tech industry, Moore's Law once served as a guiding beacon for the fast progress of the semiconductor sector. However, with the explosive increase in AI hardware demand, Musk is determined to bring down giants like NVIDIA and AMD with his sky-breaking ambition.
Recently, Musk has rolled out his own "power family bucket": from the AI5, whose design is complete, to the integrated training-inference AI6, then to the space-oriented AI7, and even the previously "written-off" Dojo supercomputing project will be revived.
While making big moves in the industry, Musk is simultaneously "directly recruiting" AI chip engineers, aiming to launch a new chip every 9 months, to overturn the market’s iteration speed. He’s also preparing to build his own wafer fab to control the lifeline of silicon from the source.
Because, behind "Iron Man" Musk's ecological blueprint, the demand for chip production capacity across various business lines has made AI chip supply chain delivery capacity and tech iteration speed the key factors limiting his expansion.
Piecing together these signals, a vast technological ecosystem covering autonomous driving, robotics, satellite communications, and brain-computer interfaces is surfacing. The man who once revolutionized the auto and rocket industries is now ready to rewrite the landscape of the AGI era.
Power Unification
On January 19, Musk made a major announcement, stating that the latest self-developed AI5 chip design was basically complete, prepared for both smart cars and robots. The next-generation "AI6" chip is also under development, integrating training and inference, usable for both robots and data centers. Musk also revealed Tesla will continue to launch AI7, AI8, AI9 chips, with the goal of completing each design cycle within 9 months.
"We expect the final chip output to surpass the total of all other AI chips combined," Musk flamboyantly stated, "I'm not joking."
To understand Musk's anxiety and ambition, one must first decode the technical paradigm shift represented by these three new cards in his hand: AI5, AI6, and AI7.
The currently (mostly) designed AI5, previously rumored as HW5.0, is the vanguard of this transformation. Tesla has predicted that AI5's performance would be 50 times greater than AI4. Musk claims it will be a very powerful chip, with single SoC performance roughly matching NVIDIA's Hopper, and dual-chip matching the Blackwell level, but at a much lower cost and significantly reduced power consumption.
In Musk's strategy, AI5's significance extends far beyond smart driving—he highlights that AI5 won’t just go in cars but also power Optimus robots, with future Tesla vehicles and robots sharing the same FSD algorithms and hardware.
In other words, AI5 is the key pivot of Tesla’s unified "car-machine brain" strategy.
As Tesla's humanoid robot Optimus rapidly evolves, Musk urgently needs a general computing core compatible with both the high-speed mobile scenarios in cars and the complex control scenes in robots.
The advent of AI5 means Tesla is eliminating the hardware barriers between cars and robots, attempting to use one "brain" to drive both wheels and legs, which will greatly dilute R&D costs and accelerate data reuse across diverse terminal forms.
If AI5 is still incremental within the traditional logic, AI6 is trying to upend the underlying industry architecture. Musk defines it as a "training-inference integrated" chip, waging war on existing AI infrastructure.
In current AI industry division, data center training chips (like NVIDIA H100) and end-device inference chips (like in-car FSD chips) are completely different species, with distinct accuracy, memory bandwidth design, and power requirements.
However, AI6 aims to break down this barrier, meaning the same silicon wafer can be installed in moving cars to process instant road conditions, or stacked in the thousands in data centers for day-and-night neural network training.
Once realized, Tesla would completely dismantle the computing barrier between edge and cloud, turning every Tesla car in a garage into a potential supercomputer node when idle—an astonishing vision for distributed computing.
The further-off AI7 nakedly displays Musk’s interplanetary ambitions. This chip is aimed at "space computing," not limited to the Earth's mild environment, but built to withstand cosmic radiation and vacuum cooling challenges.
AI7’s intended users are SpaceX's Starship and Starlink. In Musk’s ultimate vision, future intelligence shouldn’t be confined to fiber-connected data centers, but distributed worldwide—across Mars—via satellite networks. AI7 will be the neuron for this space-based internet, enabling distributed computing between Earth and sky, laying the computational foundation for humans to become an interplanetary species.
As for the Dojo project—previously rumored to be halted due to underperformance and critical resignations—its high-profile revival shows Musk realizes mere chip design isn’t enough; matched training cluster architecture is a necessity.
Dojo is seen as the cornerstone of Tesla’s AI ambitions, potentially delivering major boosts in autonomous driving video data processing and neural network optimization. Morgan Stanley has estimated full deployment of Dojo could add tens of billions to Tesla’s potential valuation.
Challenging Physical Limits
In the traditional auto industry, chip iteration cycles usually last three to five years—even consumer electronics giants like Apple stick to annual updates. Musk's proposition of "nine months per generation" sounds not only crazy, but counter to semiconductor engineering’s physical laws.
Behind this crazy acceleration, three irresistible driving forces lie hidden.
The primary reason: the algorithm’s demand for hardware has spiraled out of control. Tesla’s current FSD (Full Self Driving) tech has turned fully to an end-to-end neural network architecture—unlike the old rule-based codes, more a black box that feeds massive video data to "emerge" intelligence.
In this architecture, each order-of-magnitude jump in model parameters leads to a quantum leap in smart performance. The reality now is Tesla’s software team iterates algorithms far faster than hardware is allowed by Moore’s Law. Continuing on a three-year hardware cycle would mean Tesla's most advanced models remain restricted by old chip performance ceilings for up to two years.
Musk has bluntly stated that Tesla’s annual AI chip needs could be "1 billion to 200 billion pieces"—waiting for hardware to catch up to software is a strategically unacceptable delay.
Second, it’s the only way to seize the embodied intelligence time window. Musk has repeatedly asserted that the humanoid robot Optimus will be Tesla’s next trillion-dollar support, far surpassing the auto business in value.
Unlike cars that mostly operate in two dimensions, robots must manage extremely complex balance, grasping, and interaction in three-dimensional space, with much stricter requirements for real-time computing power, low latency, and energy efficiency. Musk predicts the next three to five years are the critical explosion and standards-setting window for humanoid robots, akin to the early smartphone free-for-all.
If Tesla cannot establish absolute technical gaps through rapid hardware iteration in this period, any catch-up by rivals will erase the first-mover advantage. The nine-month cycle is, at its core, about building a computational fortress before industry eruption.
Finally, it’s the anxiety of dependence on external computing power.
Though Tesla is a major NVIDIA customer, Musk knows that in the AI gold rush NVIDIA, the "shovel seller," commands absolute pricing and distribution power. As Tesla’s fleet size aims for tens of millions, and robot output for hundreds of millions, wholly externalized core computing would eat all the profits via high hardware costs.
More importantly, leaving the lifeblood of the company in Jensen Huang's hands conflicts with Musk’s "first principles" sense of security. Through nine-month cycles and custom ASICs, Tesla tries to outdo general GPUs in specific tasks to take back pricing power.
Ultimate Vertical Integration
Leaking a chip roadmap is just the beginning. This tech titan, holding all the wind vanes—AGI, autonomous driving, embodied intelligence, commercial aerospace, and brain-computer interface—has a new concept: build his own 2nm "TeraFab" (trillion-scale wafer fab).
In his view, TSMC and Samsung, though industry duopolies with money machine profits, are slow to expand capacity.
Globally, most tech giants follow the Fabless (no fab) model—they only design and outsource manufacturing to TSMC or Samsung.
But Musk is re-examining this labor division. The global "chip shortage" of the pandemic bruised the auto industry; those days of waiting with idle production left a mark on Musk’s memory.
Thus, the planned TeraFab—monthly output starting at 100,000 wafers, targeting up to 1 million per month—arises. It’s the challenge to global semiconductor capacity facing the collective eruption of xAI, Tesla, Optimus, SpaceX, and Neuralink around late 2025 or early 2026.
Industry insiders say that self-developed chips plus deeply attached manufacturing capacity—even exclusive lines—give Tesla supply chain sovereignty, no longer subject to contractors' scheduling and capacity allocations.
Deeper accounting lies in cost and energy efficiency squeezing to the limit. BYD's success shows that IDM (integrated design and manufacturing) mode, capital intensive as it is, delivers crushing cost advantages when scaled up.
When future Tesla needs chips for millions of cars, tens of millions of robots, and tens of thousands of satellites, it’s not just about procurement costs but energy optimization too.
Chip industry figures point out generic manufacturing processes compromise for all clients, while self-made chips let Tesla optimize from atomic-level transistor arrangement, cutting unnecessary circuitry and only keeping what’s essential for FSD and Optimus neural networks.
With battery energy density still not breaking through, efficiency gains from manufacturing directly determine robots’ endurance and cars’ driving range.
Bets and the Future
Through the dizzying technical terms and aggressive timelines, we see a tightly knitted AI ecosystem closed loop that Musk is constructing. Here, every link nourishes the next, mutually consequential.
At the front of the ecosystem are millions of Tesla cars on roads worldwide, serving as gigantic tentacles perpetually gathering real-world physical data—the most precious AI training fuel. Meanwhile, the soon-to-be-mass-produced Optimus robots will extend data collection into homes, factories, and more complex indoor scenarios—vastly enriching data diversity.
This huge data flow is constantly sent to the cloud, where the revived Dojo supercomputer and towers of AI6 chips stand ready. They devour data day and night, training ever stronger end-to-end neural network models. Those models are instantly OTA-deployed back onto cars and robots, making them smarter.
Above all this, the AI7-enhanced Starlink satellite network not only solves terrestrial base station coverage gaps, but more crucially, it’s building a space-based computing web. In the future, whether a Tesla drives through the desert or an Optimus works at a remote mine, both can call on space-based computing support in real time, no longer limited by local chip bottlenecks.
In this grand vision, chips are the blood flowing through the ecosystem, while the "nine-month generation" iteration speed is its heartbeat.
Musk knows full well—the core of AI competition is a race for computing power, but even more fundamentally, for the "speed of computing power evolution". Whoever can turn sand into computing fastest, convert power into intelligence most efficiently, defines the rules of the future.
Of course, Musk’s hyper-aggressive moves involve huge risks; building a wafer fab is a semiconductor "money-eater," and billions in investment may go years without returns. Abandoning NVIDIA’s general ecosystem for a closed Dojo hardware-software stack, if technical mistakes arise, Tesla faces massive sunk costs and lost time.
Yet looking back at Tesla’s history, from sticking to pure vision to even removing radar, Musk has always forged forward amidst controversy and bold gambles. He isn’t just making cars or robots; he’s striving to nurture a "silicon-based lifeform" with the capacity for self-evolution, through total mastery of physical computing foundations.
For global technology, Tesla’s current computing power sprint is both a wake-up call and a bugle charge. It announces the AI hardware war has upgraded from simple "spec battles" to a whole new dimension of "iteration speed" and "ecosystem closed loop". In this war, those failing to keep pace might not even get a seat at the table.
Musk is proving with near-paranoid commitment: On the road to artificial general intelligence (AGI), only those who claim sovereignty over computing power can grasp the key to the future.
Risk DisclaimerThe market carries risks; investment requires caution. This article does not constitute personal investment advice and does not consider the unique investment goals, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investments based on this article are at the user's own risk.
```