Elon Musk has never been shy about setting absurdly ambitious targets, but his latest vision for Tesla’s artificial intelligence hardware goes far beyond simply keeping up with NVIDIA and AMD. In his telling, Tesla must ultimately ship more AI chips than every other specialist combined if it wants to dominate what he calls “real-world AI” – the software that drives cars, robots, and autonomous services in messy, unpredictable environments.
On X, Musk has been reminding followers that Tesla has quietly been designing its own AI accelerators for several years. The company is now on a yearly release rhythm, rolling out fresh platforms roughly every 12 months, a cadence more reminiscent of smartphone chips than the slow, conservative timelines of traditional automakers. The next generations, often referred to as AI5 and AI6, are meant to feed everything from Tesla’s Full Self-Driving (FSD) software to future robotaxi services like the Cybercab and the humanoid Optimus robot.
Behind the showmanship, however, sits a staggering number: Musk has floated internal targets of up to 200 billion AI chips per year. That figure is so outsized that even many Tesla fans blinked. Forum threads filled up with people accusing him of FOMO on the AI boom, joking that he sounds like he is on a never-ending stimulant bender, and pointing out that he already struggles to deliver on past promises like the hyperloop, flawless autopilot, or unbreakable glass. If Tesla can barely get door handles, glass roofs and basic build quality right, critics ask, why should anyone believe it will suddenly out-ship NVIDIA on cutting-edge silicon?
The capacity problem is real. Today, most high-end AI chips come from a very small advanced foundry ecosystem, centered on Taiwan’s TSMC, with Samsung as a smaller but important player. These fabs are already booked solid by customers like Apple, NVIDIA and AMD who pay billions up front for priority access. As one skeptic put it, those who spend the most lock in capacity first – and Tesla, despite its market hype, is not at the top of that queue. Building new leading-edge fabs in the United States is painfully difficult, heavily subsidized, and protected by enormous walls of patents and hard-won expertise. High-end chips are not commoditized EVs; they are among the most complex products on the planet.
Musk’s response is to go even bigger. He has been talking about a “TeraFab” concept: a mega-scale, highly automated fabrication complex capable of churning out an unheard-of volume of AI accelerators tailored to Tesla’s workloads. In parallel, Tesla is spreading its bets across multiple foundry partners, working with TSMC and Samsung today while openly courting Intel’s fledgling foundry business for future generations. In Musk’s view, no combination of external suppliers will ever fully satisfy Tesla’s hunger for compute, so building significant in-house capacity is not a vanity project but a survival strategy.
Yet many AI researchers and engineers argue that Musk is chasing the wrong metric. You do not win the AI race purely by drowning rivals in silicon, they say; you win by building better models, smarter architectures and more reliable systems. Simply throwing more chips and energy at the problem may create flashy demos, but it also risks exhausting capital and attention that could have gone into safety, software quality and new ideas. Tesla’s own track record fuels this skepticism: FSD remains controversial, with users sharing everything from glowing praise to frustrated stories of phantom braking and inconsistent behavior. Musk’s separate AI venture, the Grok chatbot, has likewise been criticized as more edgy marketing than breakthrough technology.
Commenters also question whether Tesla, a company still battling very physical design quirks like awkward door handles, limited emergency releases and sun-blasting glass roofs, should really be spending its political and financial capital on competing with semiconductor giants. Some call Musk a brilliant salesman who constantly moves the goalposts, stacking grand promises – from hyperloop tunnels to robotaxis – on top of each other faster than they can be delivered. Others defend him, arguing that every major disruption looks impossible right up until it becomes obvious in hindsight, and that scaling compute is the only way to train AI systems on the full chaos of real-world driving.
There is also a geopolitical and supply-chain dimension. Relocating a meaningful slice of advanced chip production to U.S. soil would be a win for industrial policy, but history shows that politicians and corporate leaders alike have spent decades offshoring critical manufacturing. Reversing that trend will require more than tweets and grandiose slides. It will demand long-term partnerships, brutal transparency about costs and delays, and likely collaboration with some of the very competitors Musk loves to mock.
Still, it would be a mistake to dismiss the entire vision as pure delusion. Musk has repeatedly bet that vertical integration can unlock products competitors struggle to match, from batteries to software updates beamed over the air. If Tesla can gradually assemble the missing pieces of a chip supply chain – leveraging SpaceX experience, multiple foundries and a tidal wave of real-world driving data – it could carve out a unique position in “real-world AI,” even if it never literally outproduces every chip maker on Earth.
In the end, the real question is not whether Tesla will build exactly 200 billion chips a year. It is whether the company can balance Musk’s appetite for scale with the discipline required to build safe, trustworthy AI systems and sensible cars that drivers actually enjoy using. Outproducing NVIDIA and AMD might make headlines, but in the long run, winning the real-world AI race will depend just as much on reliability, safety, model quality and user trust as on the number of wafers rolling off a futuristic TeraFab line.
1 comment
you can’t win this race just by outspending everyone, at some point you need better models and safer ideas, not just bigger data centers burning gigawatts