Musk Must Buy GlobalFoundries, Cambricon, and Cerebras

Elon Musk has stated repeatedly and publicly that he needs between 100 billion and 200 billion chips per year for Tesla (NASDAQ: TSLA), Optimus, Starlink, and xAI. “If they cannot provide 100 billion to 200 billion AI chips within the time we need,” he’s proclaimed, “then I will do it myself.” My bet is he will do exactly that.

Musk’s businesses together demand more silicon than any private organization in history. As Optimus scales, Starlink proliferates, and Tesla transitions to full autonomy, no existing foundry, including Taiwan Semiconductor Manufacturing Co. (NYSE: TSM) and Samsung, can afford to allocate the needed volume to a single customer. This reality forces a strategic conclusion, as yet unmentioned on Wall Street. Musk must acquire or partner closely with three very specific companies. He needs GlobalFoundries to manufacture his mid- and mature-node chips. He needs China-based Cambricon, or an equivalent inference-architecture designer, to shrink chip footprint and reduce wafer requirements. And he needs Cerebras to accelerate AI training for xAI and Optimus to break his dependence on Nvidia (NASDAQ: NVDA).

Recent reporting indicates that Tesla discontinued the original Dojo supercomputer project and folded its development efforts into the AI5/AI6 accelerator program, which now serves as the unified architecture for both training and inference across Tesla and xAI. Many readers believe Dojo is still active because some of its compiler work, interconnect research, and ASIC concepts were absorbed into AI6, creating the appearance of continuity even though the standalone Dojo initiative has ended. Musk’s chip needs are “sui generis.” Traditional autos use several hundred chips. Teslas use more than a thousand. A humanoid robot needs nearly the same number. A Starlink terminal uses several hundred RF (radio frequency) and control devices. And an AI lab like xAI consumes thousands of high-performance accelerators for training large models.

Across these businesses, Musk’s consumption is already between four and five billion chips per year and will surge to more than 25 billion by 2030. His public comments about needing one hundred to two hundred billion chips annually refer not to AI GPUs, but to the billions of small inference processors, motor controllers, sensor ASICs, RFICs, and PMICs required to embed intelligence in vehicles, robots, and satellites at scale.

Toting up the bill

To understand why Musk must go shopping, we need to count up what chips Musk’s businesses will need over the next five years across node sizes, die sizes, and volumes. The best way to count is by application rather than by business segment.

Table 1 shows that Tesla dominates current demand, with each vehicle containing between 1,500  and 2,000 chips. Optimus robots will require between 800 and 1,500 chips, many of which involve distributed inference and motor-control silicon. Starlink terminals require 300 to 500 chips apiece, heavily weighted toward RF and mixed-signal parts. SpaceX launch systems depend on radiation-tolerant control ASICs (application specific integrated circuits). xAI requires the highest-performance chips of all: thousands of wafer-scale or advanced-node accelerators for training.

The table shows that most of Musk’s semiconductor needs lie in the 12- to 40-nanometer nodes, not the three-nanometer bleeding edge. Thus, GlobalFoundries is strategically essential precisely because of its limitations. Tesla, Optimus, and Starlink depend on vast quantities of mature-node silicon; no leading-edge manufacturer will dedicate this capacity to Musk, sacrificing other higher-margin customers.

Building a Musk bundle

No single semiconductor company can satisfy Musk’s vast and varied needs. He needs a combination of acquisitions and strategic partnerships, especially in three categories: advanced AI training, small-footprint inference engines, and high-volume mature-node manufacturing.

Consider Table 2. Cerebras provides the high-performance AI training required for xAI and Optimus. Its wafer-scale processors are manufactured at TSMC and represent the fastest path to training LLMs and humanoid-robot control systems. Cambricon and Horizon Robotics, in contrast, offer tiny inference processors designed for mature- and mid-nodes, allowing Tesla and Optimus to collapse hundreds of discrete chips into a handful of tightly integrated NPUs. GlobalFoundries and United Microelectronics Corporation (UMC) provide the mature-node volume needed to produce billions of chips annually. Intel (NASDAQ: INTC) contributes advanced packaging necessary for future wafer-scale systems.

The required combination becomes obvious. Cerebras for training. Cambricon for distributed inference. GlobalFoundries for massive, mid-node wafer supply. Together, these companies satisfy all categories of Musk’s semiconductor needs, except for the small number of advanced-node chips already supplied by TSMC and Samsung.

But they’re too small

Today, however, these companies can’t produce the volume Musk needs. GlobalFoundries and UMC cannot allocate this volume to Musk without displacing established customers, and TSMC and Samsung prioritize Apple (NASDAQ: AAPL), Qualcomm (NASDAQ: QCOM), Advanced Micro Devices (NASDAQ: AMD), and Nvidia.

According to Table 3, TSMC produces more than 15 million wafers per year across all process nodes, while Samsung produces around 10 million. GlobalFoundries and UMC each produce around two million wafers annually. Musk will need between five and 10 million mature-node wafers annually by 2030, and up to 20 million by the mid-2030s if Optimus reaches scale.

Clearly, external suppliers cannot currently provide Musk with enough wafers. Even if GlobalFoundries dedicated 100% of its output to Musk, the volume would be insufficient. That is why Musk announced a Texas fab with an initial capacity of 100 thousand wafers per month, scaling to one million wafers per month.

These numbers make the strategic conclusion self-evident. To meet his long-term chip demand, Musk must control a high-volume mature-node fab, increase its capacity, and acquire or partner with an inference-chip company capable of minimizing die sizes to get more chips per wafer. Without Cambricon-level IP, the die count per Tesla and Optimus remains too high. Without GlobalFoundries-level manufacturing and then some, the wafer supply remains insufficient. And without Cerebras, the training compute layer will depend on Nvidia forever.

Timeline to silicon independence

Crucial to understand is the timeline Musk faces. Training compute must scale immediately. Inference architectures must be internalized within two years. Wafer production must expand over the next five years. Packaging must become increasingly domestic. These transitions must occur concurrently.

As Table 4 shows, Musk must secure Cerebras-level training compute almost immediately. He must integrate Cambricon-class inference architecture beginning in 2026 to shrink chip count. He must expand wafer supply through GlobalFoundries and his own Texas fab beginning in 2027. By 2030, he must produce several hundred thousand wafers per month internally.

With this timeline, acquisitions or partnerships with GlobalFoundries, Cambricon, and Cerebras are no longer optional.

The investment implication is straightforward. Musk’s ecosystem is transitioning from a demand-taker to a silicon-sovereign producer. The companies capable of enabling that transition represent the new foundational layer of the next wave of robotics, autonomy, and AI infrastructure. They are no longer semiconductor suppliers. They are strategic assets.