Musk first poured cold water on the project, and then even his supposed tech partner Intel started running to TSMC for capacity. Musk’s own Tesla is also apparently ready to walk away.
Why can’t these strong partners produce enough AI chips?
Friends, two very strange signals have appeared recently.
The first came last week: in a SpaceX filing for a possible public listing, Musk openly admitted to investors that he may not be able to secure chip supply.
He basically said:
We hoped TerraFab would solve our chip supply problem, but the project could still fail. We cannot guarantee TerraFab will meet its goals on time—or even that it will be completed.
Even worse, another filing added that Intel and Tesla may both withdraw from the project.
If those two core partners pull out, TerraFab would lose both a key customer and crucial process support.
Then what is left? Just an empty shell.
But that was not the end.
A few days later, on June 1 at Computex 2026, Intel CEO Lip-Bu Tan said something that shocked analysts:
Intel’s transformation will take 5 to 10 years, and we would be very happy to be TSMC’s top-9 customer.
That’s right.
The same Intel that once built its identity around designing, manufacturing, and defining standards; the same Intel that had been dueling TSMC with 18A; is now openly admitting it will hand part of its key production to TSMC and even talking nostalgically about its relationship with TSMC.
Why? Is he crazy—or finally realistic?
In the last two episodes, we discussed Jensen Huang’s $150 billion and Lisa Su’s $100 billion move.
They were all doing the same thing: fighting for TSMC capacity.
Now Lip-Bu Tan has joined the line, and Musk—who once wanted to bypass TSMC—is also bowing to reality.
Today we’ll place this crucial piece into the puzzle and map out the full picture of this war.
Why might TerraFab fail? Why did Lip-Bu Tan concede? And why are these AI-era giants all rushing to TSMC?
More importantly, as investors, what should we actually see in this silent transfer of power?
This is not an isolated event—it is the AI war forcing the market to revalue everything.
Upstream materials, advanced process, advanced packaging, optical communications, memory, and power are all on this line.
In fact, many companies we discussed before are part of this rally’s rotating cast.
Why Did Intel Suddenly Stop Insisting on Making Itself?
Let’s start with Intel.
Why did it suddenly stop insisting on making everything itself and instead run to TSMC for capacity?
Many people find this odd.
In fact, time has pushed Intel into a corner.
Lip-Bu Tan knows manufacturing matters. In fact, he understands better than most that 18A and 14A are the lifeline of Intel’s manufacturing comeback.
But advanced process is not something that instantly works just because you say “I’m back.”
Intel 18A is showing yield improvement, but it is still in the ramp-up and customer-validation stage, while ECA is the next major investment step.
In other words, Intel’s manufacturing revival is underway, but not fully on the platform yet.
The problem is: time waits for no one.
On the other side, the CPU comeback is already here, and the AIPC wave is at the doorstep.
AI is moving from training to inference, from chatbot hype to agentic AI, and from cloud models to enterprise systems, edge devices, and AIPC.
Once AI agents really start working, they do much more than answer questions: they break tasks down, call tools, read files, run code, access databases, manage permissions, handle network traffic, and coordinate multiple models and apps at once.
GPUs can’t do all that alone.
GPUs are great at massive matrix compute—raw horsepower.
But CPUs are better at system scheduling, complex logic, I/O and memory management, network communication, and enterprise software ecosystems.
So in the AI training era, GPUs were the absolute stars. But when AI agents scale, CPUs move back to the front stage.
And that is Intel’s territory.
This is not Intel fan service; the data already shows it.
In Intel’s Q1 2026 report, CCG (client computing group, where PC chips sit) posted $7.7 billion in revenue; DCAI (data center and AI) posted $5.1 billion, up 22% year over year.
Together, these two businesses are Intel’s core product revenue engines.
Intel’s total Q1 2026 revenue was $13.6 billion, and CCG plus DCAI alone contributed $12.8 billion.
So Intel’s real money and real market presence still come from CPUs and CPU-adjacent product lines.
That is good news for Intel.
But the good news arrived too fast.
Intel needs time to prove itself and regain advanced manufacturing rhythm, while customers want products now—cloud providers want CPUs now, PC makers want solutions now.
And AMD, Intel’s old rival, will not stand still and wait for Intel’s yield to recover.
AMD’s CPU share has already surpassed Intel’s, so it cannot be underestimated.
AMD has also announced more than $10 billion of investment into the Taiwan ecosystem, expanding strategic cooperation around advanced manufacturing, advanced packaging, and AI infrastructure, and tying its EPYC and Helios platforms to Taiwan’s supply chain.
What does that mean? It means Lisa Su is not only chasing Nvidia in GPUs—she wants to keep hitting Intel in CPUs, AI servers, advanced packaging, and system platforms.
Meanwhile, Jensen Huang is watching closely too.
As we discussed yesterday, he plans to bring RTX Spark—his superchip for the next wave of AI agents—into the PC battlefield.
And the painful part is that RTX Spark is based on ARM.
Huang is holding Microsoft with one hand and bypassing x86 with the other, moving straight into Intel’s most traditional and core territory.
So Lip-Bu Tan is facing not one enemy, but multiple battlefields at once.
If Intel insists on waiting until its own process node is fully mature before moving, it may lose the chance window for PCs and AI servers to others.
So he has to be realistic.
That is why he wants TSMC capacity.
Because what TSMC can give him is not just wafers—it is time.
And his turn is clearly a public setback for TerraFab.
It may not be the heaviest blow, but it exposes the most awkward truth: even Intel—the company supposed to provide the imagination for advanced manufacturing—still needs TSMC to underwrite it.
So how can TerraFab replicate TSMC in such a short time?
Why TerraFab’s Dream Team Still Can’t Escape Failure
We’ve talked about TerraFab before.
Even though I personally don’t like it, I have to admit the enthusiasm around it is not groundless.
On paper, TerraFab is one of the boldest attempts in the last decade to build a TSMC-like company from scratch.
Think about who is behind it:
-
Demand side: Musk’s own AI empire. Tesla needs FSD, Robotaxi, and Optimus; xAI needs to train and deploy GROK; SpaceX wants Starlink, mobility services, AI compute, satellite networks, and even future space data centers.
-
In other words, TerraFab was not built first and then looking for customers; it was born with customers, tied directly to some of Musk’s most chip-hungry, compute-hungry, growth-story-heavy companies.
On the tech side, it also had Intel.
The original plan was not to start from zero, but to rely on Intel’s 14A process route.
Intel may have been under pressure from TSMC for years, but its foundation is still there, and since the Gelsinger era it has been pushing an aggressive “five nodes in four years” plan to catch up.
It already has years of engineering know-how and process accumulation.
And it also carried the policy halo of U.S. domestic advanced manufacturing.
TerraFab was not just a random team building chips; it tried to bind itself to one of the few players still willing to push the most advanced nodes.
That also solved the equipment side in theory: through Intel’s supply chain, TerraFab could secure some equipment allocation.
And on talent? Musk could just throw money at engineers in Hsinchu.
For Musk, TerraFab’s target is 2027 mass production of 2nm chips.
He does not have ten years to grow a team slowly; he needs to gather people who already know how to do the job.
And TSMC veterans are exactly that group.
He even put Tesla and SpaceX equity behind it.
Money? Even less of a problem.
If TerraFab really launches, it could be backed by SpaceX IPO financing, Tesla cash flow, xAI fundraising, and the U.S. reshoring narrative.
Most companies fear not having money; Musk’s biggest strength is convincing capital markets he can change the world.
So in theory, the chessboard looked luxurious: customers, tech, equipment, talent, and money.
But why is it still not working?
Less than three months after the March 21 launch, why have both sides turned pessimistic?
What is missing?
In the end, it still comes down to Musk underestimating the difficulty of wafer manufacturing.
Musk likes speed, disruption, and moving fast.
But wafer manufacturing is one of the slowest, heaviest, and most discipline-driven industries in the world.
It is not a matter of buying the best parts and assembling them into a TSMC.
First, the partnership itself was never truly locked in.
TerraFab looks on the surface like a triad of SpaceX, Tesla, and Intel, but the filing still wasn’t a final binding joint-venture agreement.
If Intel decides not to cooperate one day, TerraFab would lose Intel 14A, the engineering team, and even ASML lithography allocation.
So the two major advantages we just mentioned—tech and equipment—were never as solid as they looked.
If they disappear, the so-called super fab instantly turns from an “AI chip rescue plan” into an empty lot without a foundation.
And even if Intel stays, the equipment problem remains.
Getting some allocation through Intel’s supply chain does not mean you can receive enough tools on your own schedule, and it definitely does not mean equipment equals mass production.
EUV capacity is scarce, and TSMC, Samsung, and Intel are already in line.
Can Musk spend more money to push forward? Maybe a little.
But the supply chain is already scheduled by quarter, by year, and by customer qualification. It is not something ASML can air-drop to Texas just because you say “I want to change the world” today.
And advanced process is not just lithography.
Lithography is only the brightest jewel. There is still deposition, etching, ion implantation, CMP, cleaning, inspection, metrology, defect analysis, automation, clean systems, and chemical supply.
If any link breaks, the whole line waits.
So the equipment problem is not about whether you can afford it; it is about whether you can get it on time, in a complete set, steadily, and with priority.
That depends not on cash, but on industry trust and supply-chain position.
Talent is similar—you cannot just poach people and think it will work.
Musk can offer high pay, but a leading fab needs thousands of engineers, equipment techs, process integration staff, yield teams, packaging teams, and supply-chain teams working together for years.
Taking one engineer may bring one person’s experience, but not the TSMC platform—the organizational ability to bind thousands of people into one system and turn decades of know-how into standard operating procedures.
Even more dangerous is yield and localization.
Intel 14A sounds advanced, but advanced does not mean mature, and mature does not mean stable mass production.
The harsh truth of chipmaking is this: making one chip and mass-producing a batch of chips are two totally different skills.
-
Making a sample in the lab is a technical breakthrough.
-
Shipping hundreds of thousands of wafers every month with stable yield, controlled cost, and reliable lead time is what real commercial capability looks like.
And that is not easy.
One dust particle, one temperature fluctuation, one cut too deep, one film too thick, one metal interconnect defect, or one packaging stress change can ruin a whole batch.
Moving yield from 20% to 50%, from 50% to 70%, and from 70% to 85% takes countless engineers, equipment vendors, materials suppliers, and customer design teams grinding together.
What makes TSMC scary is not only its technology, but its almost obsessive manufacturing culture: customer requirements never slip, yield is inspected daily, problems are traced to the root, and engineers and suppliers grind together.
Musk is great at rapid iteration, but wafer manufacturing cannot be learned like rockets—blowing up along the way is not an option.
Chip customers do not want to hear “we learned a lot.” They want to hear “you delivered on time.”
More importantly, TSMC sits atop years of industrial collaboration and an entire Taiwan semiconductor ecosystem.
An AI chip is not designed and delivered by TSMC alone:
-
EDA and IP tools define the design rules
-
Equipment, materials, photoresist, wafers, gases, and chemicals come in the middle
-
Advanced packaging, testing, boards, and HBM integration come next
-
Then Quanta, Wistron, Inventec, and Foxconn turn chips into servers, racks, and AI factories
That is the hardest thing for Musk to copy.
He can poach a few engineers, but he cannot poach the decades of supply-chain density in Hsinchu Science Park.
He can buy equipment, but he cannot buy away the long-term coordination among TSMC, ASML, Applied Materials, Lam Research, and Tokyo Electron.
He can work with Intel, but he cannot quickly duplicate the process database TSMC built through generations of products from Apple, Nvidia, AMD, Qualcomm, and MediaTek.
He can build a Texas factory, but he cannot build the kind of industrial network Taiwan has—where packaging, testing, board assembly, server ODM, cooling, and power are all just one phone call away.
That is also why Jensen Huang, Lisa Su, and Lip-Bu Tan are all rushing to Taiwan.
Because they all understand one thing: in the AI era, the real scarce asset is not a building—it is the network behind the building.
TSMC is only the brightest node in that network. The real power is that Taiwan’s semiconductor supply chain has become the most important infrastructure of the AI era.
Jensen Huang says Taiwan is the center of the AI revolution not because he likes slogans, but because Nvidia’s throne needs a full chain:
-
Working with MediaTek on RTX Spark
-
Working with TSMC to make chips
-
Advanced packaging connecting GPU, CPU, and high-speed memory
-
Quanta, Wistron, and Foxconn turning them into servers and racks
-
And finally optical communications, power, and cooling to actually run the AI factory
Lisa Su’s $10 billion is not random spending either.
AMD named ASE, SPIL, and PTI because it knows that if it stays stuck at the back of TSMC’s queue, it will struggle to catch Nvidia.
So she wants to support more advanced packaging partners so MI450, EPYC Venice, and Helios systems can ship on time.
Lip-Bu Tan is more pragmatic: Intel wants a manufacturing comeback, but the AIPC window is already open.
He cannot wait for 18A and 14A to fully mature before fighting, so he needs TSMC capacity and Taiwan’s ecosystem to buy time.
AI Investing Enters the Second Half: From GPUs to Running AI Factories
So the real winners in this round are not one isolated concept, but the hard infrastructure links of AI: advanced process, advanced packaging, AI server ODMs, high-speed interconnect, power, and cooling.
-
TSMC earns certainty of delivery
-
ASE, SPIL, and PTI benefit from packaging spillover
-
Quanta, Wistron, and Foxconn benefit from AI servers and rack-level delivery
-
Corning, Marvell, Ciena, and Lumentum benefit from optical communication and high-speed interconnect
That is what it means for AI investing to move from the first half to the second half.
In the first half, the market only asked: who has the strongest GPU?
In the second half, it will ask: who can turn GPUs into a fully running AI factory?
And the answer is not only on Nvidia’s stage—it is in the wafer capacity tables, advanced packaging lines, server factories, and optical-comms and cooling order books.
