1. Introduction — Starting With a Turnaround Number
In August 2026, SpaceX’s AI division, SpaceXAI, released second-quarter results that stunned observers. A unit that had posted a $609 million loss just one quarter earlier flipped to an $1.10 billion profit within a single quarter. Revenue reached $2.56 billion, and the division signed $14.1 billion in new cloud computing contracts. CFO Brett Johnson said the business was “on a trajectory to a $100 billion annualized revenue run rate by year-end,” and Elon Musk went further, calling it “not a question mark — a number we hit even if we do nothing else.”
Trace the roots of this dramatic reversal back far enough, and you land on an unexpected decision Tesla made a year earlier. In August 2025, Musk abruptly announced the company was shutting down Dojo, its in-house AI training supercomputer. At the time, it looked like a question mark hanging over Tesla’s AI hardware strategy. In hindsight, it wasn’t a retreat — it was a realignment toward a much bigger picture: the in-house AI5 and AI6 chips, and Terafab, a semiconductor megacomplex valued at $16.8-119 billion.
2. Why This Story Matters
Full Self-Driving, robotaxis, and the Optimus humanoid robot — Tesla’s future businesses ultimately hinge on how much AI compute the company can secure, and how cheaply. Every generational leap in chip performance widens the range of perception and decision-making that robots and self-driving cars can handle in real time, which means the timeline for Tesla’s entire business is tied directly to the production schedules of AI5, AI6, and AI7.
The ripple effects don’t stop at Tesla’s own walls. This is an automaker attempting to vertically integrate the entire semiconductor value chain, from chip design through manufacturing to memory. For Samsung Foundry, landing a customer of this size is a clear win — but if Tesla eventually pushes into producing its own memory, Samsung could become both a partner and a competitor over the long run. Intel’s decision to join Terafab as a partner fits into the same broader trend of rebuilding a leading-edge foundry ecosystem inside the United States, and it’s likely to accelerate realignments across the semiconductor industry more broadly.
3. Dojo’s Death, AI5’s Revival
The logic behind killing Dojo turned out to be straightforward: rather than designing a dedicated supercomputer chip completely from scratch, it made more sense to build around the AI5 and AI6 chips Tesla was already planning to produce anyway. Then, in January 2026, came an interesting twist. As the AI5 design matured faster than expected, the project was revived under a new name — “Dojo 3” — this time built entirely around Tesla’s own silicon.
The approach works like this: instead of building a purpose-made supercomputer chip, Tesla mounts multiple AI5 or AI6 chips on a single board, cutting network cabling cost and complexity by several times over. In other words, Dojo’s “death” wasn’t really an ending — it was a transfer onto a more general-purpose engine called AI5.
4. AI5 and AI6 — Design and Dual-Source Manufacturing
AI5, Tesla’s next-generation chip for FSD and inference workloads, completed tapeout — design finalization — in April 2026. Compared with AI4, it delivers roughly 5x effective compute performance, 8x raw compute performance, 9x memory capacity (up to 192GB of LPDDR5X), and 5x memory bandwidth.
Production isn’t concentrated in one place. AI5 is being manufactured simultaneously at Samsung Electronics’ Taylor, Texas fab (on a 2nm process) and at TSMC’s Arizona fab, with mass production targeted for late 2026 to early 2027. For the follow-on AI6 and AI6.5 chips, Tesla and Samsung Electronics signed a $16.5 billion supply agreement running through 2033, with the Taylor fab serving as the core production base.
5. Terafab — Owning the Entire Semiconductor Value Chain
In March 2026, Tesla and SpaceX (including its SpaceXAI subsidiary) announced plans to jointly build Terafab, a massive semiconductor complex. Intel subsequently joined as a partner. Phase 1 investment alone totals $16.8 billion, with the possibility of scaling up to $119 billion across multiple expansion phases. The manufacturing footprint would span more than 100 million square feet — roughly ten times the size of the existing Gigafactory Texas. The plan calls for integrating chip design, fabrication, packaging, and testing all on a single site, with at least 3,000 jobs expected.
And it doesn’t stop there. Terafab has recently begun hiring specialists to develop advanced DRAM processes — a clear signal of intent to push into memory, long the domain of Samsung Electronics and SK hynix, as part of a broader supply-security strategy. Given that total investment could exceed $100 billion, execution speed and financing remain the key variables to watch.
6. SpaceXAI — The Turnaround, By the Numbers
SpaceXAI, the other pillar of the Terafab story, has separately been rapidly scaling up its Colossus data center in Memphis. Securing a third building in January 2026 brought total capacity to 2 gigawatts, roughly 550,000 GPUs, and around $18 billion in cumulative investment, with an additional $659 million expansion permit filed in March 2026. The stated end goal is 1 million GPUs.
SpaceX’s IPO (Form S-1) filing, disclosed in May 2026, revealed the scale of the business is even larger than it appeared. Rival AI lab Anthropic signed a deal to buy up an entire 300 megawatts of Colossus 1’s data-center compute capacity — about 15% of the full 2GW total. Under the agreement, Anthropic pays SpaceXAI $1.25 billion per month (with a discount for the first two months) through May 2029, working out to roughly $15 billion on an annualized basis and more than $40 billion over the life of the contract. With usage of Grok, its own chatbot, recently slowing, SpaceXAI effectively expanded into a “neocloud” model — monetizing spare compute capacity by leasing it to a competitor.
That dynamic blew past the company’s original revenue guidance entirely — which had called for $2 billion in 2026 revenue, roughly $500 million in annualized revenue run rate as of early 2026, and projections of $14 billion in revenue and $13.1 billion in EBITDA by 2029. The second-quarter 2026 results cited earlier — $2.56 billion in revenue and a $1.10 billion profit — are the clearest evidence of just how far those original numbers were exceeded. AI-related capital expenditure also more than doubled, from $7.7 billion to $15.8 billion, and in the first few weeks of the third quarter alone, an additional $6.7 billion in cloud service contracts were signed.
Separately, xAI was folded into SpaceXAI in February 2026 through an all-stock merger with SpaceX, a deal that valued the combined entity at roughly $1.25 trillion. (xAI had already merged with X, formerly Twitter, back in March 2025, meaning the company effectively doubled up twice — first as xAI+X, then as xAI+SpaceX.)
7. Conclusion
Put it all together and the picture is this: design (AI5, AI6) → contract manufacturing (Samsung, TSMC) → in-house fabs (Terafab) → memory — Tesla is attempting to gain control of nearly every link in the semiconductor value chain. Shutting down Dojo wasn’t a dead end; it was one waypoint on the road to this much larger plan. If it succeeds, it’s a bet that could reshape not just autonomous driving and robotics, but the semiconductor industry itself.
SpaceXAI’s dramatic earnings turnaround lends real weight to that possibility. But with total investment ranging from $16.8 billion to $119 billion, execution risk and how Tesla manages its relationship with existing partners like Samsung remain open questions worth watching closely.
References
- Tesla and SpaceX earnings disclosures
- SpaceX IPO (Form S-1) filing
- Related industry reporting
