Intel Stay of Execution Has a Name Now
Tesla signed a contract for a chip that does not exist yet, in a fab that is mostly a spreadsheet, at a cost Bernstein estimates could reach \$13 trillion.
Building enough chip capacity to back what Elon Musk is describing costs between $5 trillion and $13 trillion, according to Bernstein estimates. Tesla just committed to being the first outside customer Intel Foundry would need to write that check.
Tesla has signed to be that customer, Reuters confirmed on April 22. The contract marks Intel's first major outside customer for its 14A process node, the technology that determines how small a transistor can get. Intel shares rose 3.6%. Intel CEO Lip Bu Tan had said publicly that Intel would exit chip manufacturing entirely if it failed to land an external customer for next-generation production. Tesla answered that call.
The check size is not incidental. Bernstein's range is large because nobody has built what Musk is describing. The analysts derived it from current AI accelerator power consumption and wafer cost assumptions applied to the stated output target. The gap between a $3 billion research fab at GigaTexas, capable of a few thousand wafers per month, and the full-scale ambition of one terawatt of compute per year, is where most ambitious chip projects die. Intel joined Terafab on April 7. The silicon does not.
The 14A process node does not tape out until risk production in 2028 and high-volume manufacturing in 2029, according to Intel's own roadmap as presented by CEO Lip-Bu Tan at the Cisco AI Summit in February. Intel announced that timeline before Tesla appeared on the order book. Adding a high-profile anchor customer does not compress it. It raises the stakes when the date arrives.
What Tesla is actually buying is a slot. A place in the queue when 14A is ready. For Intel Foundry, that slot is existential. For Tesla, it is an option on leading-edge manufacturing that does not yet exist, from a foundry that has never delivered chip volumes to an outside customer. SpaceX, co-located in Terafab, is separately planning to make its own GPUs for AI model training, according to Reuters.
"The question is whether they can actually deliver the volume," said Seaport Research Partners analyst Jay Goldberg. "It is not equivalent to Apple or Nvidia in terms of chip volumes. But it is a real customer. It can be real volumes."
Real volumes are the thing Intel has not yet demonstrated it can deliver to anyone outside its own product divisions. The shares went up. The question now is whether 2028 looks more like 2024 or more like 2018.
