A procurement manager at an AI server maker reportedly placed an order this month for multilayer ceramic capacitors. Delivery: 52 weeks. Not a specialty part, not a premium accelerator — the MLCC, the component the industry calls the rice of the electronics industry, the dam that stores charge and releases it in measured amounts. A full year's wait for rice.
This week the scarcity is below the top of the stack, and the news is that the shortage has inverted: the binding constraints on AI capacity are now commodity passives, IC substrates, and the unglamorous parts inside the machines that make chips.
Samsung Electro-Mechanics' MLCC lead time stretched to 52 weeks this month, per The Herald Business. Customers ordering today wait a year. Only two suppliers — Samsung Electro-Mechanics and Murata — can meet AI-server demand, and AI servers take more than ten times as many MLCCs as conventional ones. Utilization at Samsung's component division hit 91 percent in the first half, effectively full. The Philippines expansion won't produce at mass scale until 2027 or later.
Only two suppliers can meet AI-server MLCC demand, and AI servers take more than ten times as many as conventional ones.
It isn't only MLCCs. TrendForce puts ABF substrate lead times at 48 to 56 weeks — up to 4.7 times the 12-week equilibrium, meaning an order placed today can take 13 months. MLCCs sit at 30 weeks per TrendForce, with a separate TrendForce reading as of the 19th putting them at 35 weeks. HDDs at 50 weeks. DRAM at 20. Five of six tracked categories are in shortage; GPUs are the lone exception.
The customer response is the real story. Since May, Samsung Electro-Mechanics has disclosed seven long-term supply contracts totaling about 4.3 trillion won in cumulative orders. MLCC-related contracts disclosed this year alone reached roughly 2.4 trillion won, or $1.81 billion. And the contract structure is spreading into substrates: a 360.5 billion won package-substrate deal with a global memory company — the first substrate long-term agreement Samsung Electro-Mechanics has disclosed this year, and a signal that the seller's market has reached even commodity BT-based BGA substrates.
The practice is mutating further. Nvidia is reportedly investing around 2 trillion won to expand substrate production at Samsung Electro-Mechanics' Sejong and Busan sites, with a multi-year supply agreement attached. Tesla reportedly put up roughly 1 trillion won for a dedicated Sejong line, contracting substrates for self-driving AI chips and the Optimus robot from 2029 through 2033. Customer prepayments are reportedly becoming a key funding mechanism. LG Innotek is reportedly in final-stage expansion talks with multiple Big Tech customers, expected to fund much of the investment with customer money.
Here is where it gets structurally interesting. Morgan Stanley tallies $14.3 billion in substrate capacity investment across major makers, with Samsung Electro-Mechanics taking about 35 percent. But most new capacity arrives after the second half of 2028 — Sejong targets September 2028 mass production, Vietnam June 2030. Unimicron chairman S.C. Chien says equipment lead times for more than 80 percent of tools have lengthened, with over 20 percent delayed by more than double, some out to 30 months. Low-CTE glass fiber cloth, copper-clad laminate, and ABF face a projected 20 percent material shortfall in 2027.
Now go one layer deeper, inside the tools themselves. TrendForce reported in September that deposition equipment parts went from a 4-month lead time to as long as 10 months; packaging equipment parts from 4 months to 6–8. One laser processing maker faces up to 40 months for a key Japanese-made part and can't pull it forward with a premium. Imported RF power equipment has stretched to 24 months.
One equipment executive's summary: "We're willing to pay a premium for faster delivery, but we've been told the parts supplier's production capacity is limited and lead times can't be shortened."
The causes are boring and therefore durable. Parts makers didn't expand before the boom and hesitate now because they don't know how long the supercycle lasts — Japanese suppliers especially. Custom treated parts lead times as the business of the second- and third-tier supply chain. Government support focused on fab construction and equipment localization; supplier diversification, parts interoperability, inventory transparency, and financing for small parts makers got relatively little. Korea's equipment localization rate sits around 20 percent and has barely moved since 2011.
This is why Intel's CTO answered the question the way he did. Asked on the Dwarkesh Podcast whether the world has enough tools for Intel to make 10 times more leading-edge wafers, Pushkar Ranade said: "Not 10X, but probably 2X, yes." Tenfold is impossible, he explained, because the tooling ecosystem — equipment suppliers, their suppliers, and their sub-suppliers — cannot scale. Tools ordered today take about 18 months to arrive, taking buyers into 2028; building a fab on top means the end of the decade. He said Intel would have spent more on equipment had it known a year ago what it knows now.
Read that against the demand curve. AI compute has grown about threefold a year for a decade. Ranade's ceiling on feasible expansion is 2x. The arithmetic doesn't fit, and it doesn't fit because of valves, ceramics, and RF power supplies.
The inversion is now complete. Commodity passives and second-tier components dictate system-level schedules. A GPU may be in balance at 20–30 weeks, but an AI server needs every component at once, so a 56-week substrate or a 40-month laser part sets the delivery date. Leverage has migrated to suppliers the industry treated as second- and third-tier business: the specialty valve maker, the ceramics shop, the Japanese laser component house reportedly facing a backlog stretching into 2029.
Some of this pressure has release valves. Chien projects CoWoS capacity growing 6.8-fold by 2028 from 2024, against a 6.3-fold rise in hyperscaler capex — but IC substrate capacity merely doubling and DRAM 1.8-fold. Nvidia's three-year plan aims to raise substrate capacity 2.5x and double test capacity. None of it lands before 2028.
Parts are now the binding constraint — on fab timelines and data-center fit-out alike. Power and liquid cooling still gate site energization, and transformers, gas turbines, and grid interconnection remain slow. But a wafer fab cannot start without its deposition tools, and a deposition tool cannot ship without a valve that used to arrive in four months.
The strategic consequence is that the industry's spending is aimed at the visible bottlenecks while the invisible ones set the ceiling. Jensen Huang is reportedly right that chips are no longer the hardest part. The correction is that the parts inside the machines that make the chips may be the hardest part of all — and the suppliers of those parts, obscure a year ago, now hold leverage that used to belong to fabs.