Musk's Terafab: The Last Vertical Mile from Sand to Token

01 Trigger Event

Bloomberg reports that Elon Musk has confirmed TSMC and his companies are in discussions about a chip manufacturing collaboration in Texas, which could make TSMC a chip supplier for Tesla, SpaceX, and xAI. The project code name "Terafab" hints at enormous scale, but public details remain scarce—no capacity numbers, investment amounts, timeline, and no indication whether TSMC would build a new fab, form a joint venture, or repurpose existing capacity. Musk's characteristically vague treatment of details is hardly surprising.

02 What This Really Means

On the surface: "TSMC may have landed a major client." What's actually happening is Musk pushing AI compute stack verticalization down yet another layer.

Take stock of the assets already under Musk's control: xAI handles the model layer (Grok), X provides data plus traffic distribution, Tesla has Dojo in-house silicon plus an edge inference fleet, SpaceX provides capital plus prospective orbital data centers. Now Terafab may give him direct control of the fab.

If TSMC truly enters via joint venture or long-term underwriting, Musk's "sand to inference API" pathway becomes nearly fully self-contained. What remains uncontrolled: EUV lithography (ASML monopoly), HBM memory (Micron/SK Hynix), advanced packaging. This isn't a capacity story—it's an inflection signal in industry organizational structure.

What actually gets priced: when a buyer grows large enough to "invite" the world's largest foundry to build on its turf, the "foundry-neutral, diversified-customer" supply assumption begins to fail. The foundry-neutral model TSMC has built since the Morris Chang era fundamentally assumes customers are sufficiently diversified, with no single-customer dependence exceeding some threshold. A client like Musk who wants "Tesla's autonomous driving chips, SpaceX's radiation-hardened chips, and xAI's training GPUs all stuffed into one fab" is a stress test of this model itself.

I should admit I may be misjudging here—I can't verify whether Terafab's process roadmap can truly be unified.

04 Historical Analogies

Two reference frames.

Reference Frame A: Apple Silicon.

Apple has gradually replaced Intel/Qualcomm with self-developed silicon since 2014. The key turning points weren't M1 (2020), but 2013's in-house GPU IP development and the 2017 cancellation of Imagination. Apple controls the full stack from application to chip design, but the fab has always stayed with TSMC. Apple never attempted to build its own fab, because the capital intensity of an advanced fab (roughly $20 billion per facility) and the pace of process iteration are beyond what an applications company can chase.

Musk's path differs from Apple's. Apple aimed to substitute the design layer; Musk aims to lock down the capacity layer. Apple's solution was "long-term underwriting of TSMC advanced capacity + self-developed design." Musk's solution may be "joint venture + tie-up."

Reference Frame B: AWS Graviton.

Amazon began developing in-house Arm server chips in 2018. AWS's success wasn't from controlling the fab, but from its buyer position in the cloud market guaranteeing return on chip investment. Musk goes further than Amazon: he's not just a large customer, he's building his own "satellite layer" (Starlink), "automotive layer" (Tesla), and "model layer" (xAI). Demand diversification itself is why TSMC is willing to negotiate.

An earlier structural comparison: TSMC in the 1990s was itself the product of vertical reverse-integration—Philips supplied the technology, Morris Chang built the specialized foundry model, and from then on the fabless-plus-foundry division of labor became the default. Thirty-five years later, Musk wants to walk that path in reverse—either because this foundry model has hit its limits, or because customer concentration has reached the critical threshold for reverse integration.

04 What This Means for AI Builders

Short term (6-12 months): no direct impact. Capacity won't come online for at least 2-3 years, and TSMC is unlikely to put all of Musk's eggs in one basket. OpenAI/Anthropic/Google's existing supply won't be immediately affected.

But several meta-assumptions need recalibration.

First, the AI infra "capacity is a scarce resource" assumption strengthens. If TSMC truly expands for Musk in Texas, the geographic concentration of global advanced capacity may disperse from Taiwan—which is itself good for geopolitical risk. But short-term dispersion may buy Musk priority access during mid-to-long-term capacity tightness. AI startups' optimistic stance that "I can scale to 100,000 cards anytime" needs a haircut for the next 12-24 months.

Second, "vertical integration of compute" among hyperscale buyers is a trend, not an anomaly. Google has TPU plus its own cloud; Amazon has Trainium plus AWS; Microsoft has Project Athena; Meta has MTIA. Now Musk wants to bring in the fab itself. Competition among large buyers is no longer just about model capability, but about "total stack cost from wafer to token."

Third, for mid-scale players (Anthropic, Mistral, model companies other than xAI): divergence will intensify. Either huddle with a major backer (Anthropic-Google tie-up, Mistral relying on self-developed IP plus multi-foundry diversification), or get caught in the middle with neither self-developed chips nor major customer lock-in.

For token gateway layers like opcx.ai: upstream supply diversification is actually good news—multiple competitors means more stable capacity access. But if Musk consumes more tokens internally (X platform plus Tesla ecosystem plus SpaceX internal use), overall supply to the external API market may be neutral to tight.

05 Counterarguments / Risks

Several places where I may be misjudging.

First, Terafab may be just another of Musk's "announce-delay-downsize" projects. Cybertruck, Robotaxi, Starship have all followed similar curves. Between an announced fab and a built fab, there may be 5 years and $30 billion. I'm only 50% confident on this myself.

Second, TSMC negotiating with Musk doesn't equal TSMC committing. TSMC's core strategy is customer diversification—founding for Apple, NVIDIA, AMD, Qualcomm. Locking in capacity long-term for a politically charged client with a rich history of Twitter-driven governance is a hard sell to the TSMC board.

Third, I may be overestimating the consistency of Musk's "vertical integration" agenda. Musk simultaneously manages 6-7 companies, each with fundamentally different fab needs (Tesla needs automotive plus inference; SpaceX needs radiation-hardened; xAI needs training GPUs). Stuffing these into one Terafab, the process roadmaps may be entirely incompatible and may not actually combine. Intel's failed effort over the past decade to unify PC, server, and AI chips under IDM is a cautionary tale.

Fourth, the geopolitical logic of "fab localization" may outweigh commercial logic. TSMC's U.S. fab construction is mainly a response to the CHIPS Act and export control countermeasures. If so, Terafab's economics may be far inferior to building in Taiwan—I'm not sure on this, but worth tracking. If Musk ultimately gets quoted uncompetitive pricing, Terafab may devolve into a political showpiece.

The final wildcard: whether Musk can truly compress compute costs to "lower than everyone" as he claims. If yes, the unit economics of the entire AI application layer must be rewritten. If not, this is another high-profile PR play. Historically, Musk has lost a few times on the second front (Model 3 production hell, multiple Starship explosions), but when he wins, he wins big (Starlink, Model Y). I can't predict this one—only track production timelines.