For two years, the story of the semiconductor shortage was a story about cars waiting for parts. Then it became a story about data centers. This week the two stories met again.
Several chipmakers warned that competition from artificial intelligence data centers for mature-node capacity and memory is lengthening lead times for automotive-grade components and printed circuit boards. Some tier-one suppliers say the allocation priority for older-generation memory has slipped behind data-center customers, and that production planning for 2027 model-year vehicles is now being adjusted.
The mechanism is simple and hard to fix. Automotive chips are made on mature manufacturing processes, the same ones used for power management, sensors and microcontrollers across consumer devices, industrial equipment and cars. Fabs add capacity slowly and prefer long-term contracts with large buyers. Data centers buy in volumes that dwarf a single automaker’s order and commit for years.
Memory is the tighter constraint. The factories that make the older memory parts used in vehicles have been converting capacity to the high-bandwidth memory that AI accelerators require, because the price per wafer is many times higher there. When capacity moves, the parts left behind get scarce first.
Circuit boards are a quieter version of the same problem. Server boards use more layers, more copper and more sophisticated materials, and the laminates that go into them come from a supply chain that has not grown much in five years. Automotive boards compete for the same material, and they lose.
The result is familiar to anyone who worked in procurement during the pandemic. A car is a few thousand parts organized around a handful of controllers, and a shortage of one controller can idle an assembly line. Automakers learned in 2021 to hold more inventory and to sign direct agreements with chipmakers instead of leaving sourcing to suppliers. That lesson is being tested again.
The difference this time is the direction of the market. In 2021, demand for cars rebounded faster than anyone expected and the chip industry was caught flat. Now the shock is on the other side: AI is pulling capacity away from everything else, and the car industry is the customer with the weakest bargaining position, because it buys the least.
Qualification rules make it worse. Automotive components have to be validated for temperature ranges, vibration and long service life, and the validation takes months. A supplier cannot simply switch a vehicle to a different memory part when one runs short. The replacement has to be tested, documented and approved, and that schedule is measured in quarters, not weeks.
For the 2027 model year, the calendar is unforgiving. Vehicle programs lock their specifications years ahead, order tooling, and build pilot fleets before volume production. A part that slips by a quarter in that sequence does not just delay a shipment. It can change what gets built: features dropped, trim levels altered, options bundled differently to reduce the number of distinct parts on the line.
Suppliers have been here before and are responding the way they did last time. Some are pre-buying components and holding them, which protects production and ties up working capital. Others are redesigning boards around parts they can actually get, which is a slow way to solve a fast problem.
What they cannot do is outbid the data centers. A wafer of mature-node logic sold to a cloud customer earns more than the same wafer sold to a tier-one automotive supplier, and the cloud customer will sign a multiyear agreement that makes planning easier. Nothing in the pricing system makes an automaker’s order more attractive than that.
Governments have noticed. The American and European subsidy programs written after 2021 were aimed at leading-edge logic, on the argument that the most advanced chips were the strategic ones. The current squeeze suggests the vulnerability was also in the ordinary parts: the controllers, the power semiconductors, the memory that goes into a car assembled in the Midwest or the South.
The auto industry has some insulation it did not have then. Inventories are higher, contracts are more direct, and procurement teams have spent four years building visibility into suppliers two and three levels down. That visibility is also how they learned this week that the queue got longer.
There is a version of this that resolves quietly. Memory makers finish their conversions, mature-node fabs bring on the capacity they announced, and the automotive allocation returns by the middle of 2027, which is late but not fatal to the model year.
There is another version in which the AI build-out keeps absorbing capacity for another two years, and the car becomes the product that waits. That is not a technology failure. It is a price signal doing what price signals do, in a market where the customer with the smallest order gets served last.
The shortage of 2021 arrived with a fire at a plant in Japan and a pandemic that scrambled everything. This one has a simpler cause. Someone else is willing to pay more for the same wafer.
