Five Companies Make an Estimated 80% of the World's Advanced Silicon Wafers. One of Them Just Said a 10% Price Hike Won't Cover Its Costs.
Every chip this newsletter has covered — the AI accelerator, the HBM stack, the power chip, the capacitor riding next to it — starts life the same way: as a blank, mirror-polished disc of ultra-pure silicon. Five companies make nearly all of them. The last time demand spiked this hard, they overbuilt, got stuck with a glut, and lost money for years. This time, they're not blinking first.
The wafer squeeze: how much slack is left before 2029
Drag the slider to add new wafer capacity through 2029 and watch utilization and price respond. The three starred years (2026, 2028, 2029) use published UBS-cited estimates; 2027 is interpolated for a smoother line. The price curve is a simplified illustrative model built to show the mechanism UBS describes (a steep response once utilization nears 100%) — not an exact price forecast.
The plain version
Before a chip becomes anything — an AI accelerator, a phone processor, the sensor in your car — it starts as a blank disc of silicon, sliced thin and polished until it's flatter and smoother than almost anything else humans manufacture. That disc is a wafer: the canvas every chipmaker paints transistors onto. Nobody prints a single circuit until the canvas exists. Here's the part almost nobody outside the industry tracks: just five companies make nearly all of the advanced, 12-inch wafers modern chips are built on — two Japanese firms (Shin-Etsu and SUMCO), one Taiwanese (GlobalWafers), one German (Siltronic), and one South Korean (SK Siltron). Together, they're estimated to supply roughly 80% of the world's advanced wafer supply.
These five have been burned before. After the pandemic-era chip shortage, they built more capacity — then demand cratered from 2023 through 2025, prices fell, and some of that new capacity sat half-used, losing money. They learned a lesson: don't overbuild. Now AI chips and the memory stacked next to them are eating wafer capacity fast, right as demand for ordinary chips — phones, cars, appliances — is recovering too. Analysts at UBS think wafer factories could be running at 99% of capacity by 2028, and could be asked to produce more than they're physically able to by 2029, if nobody builds anything new.
So far, nobody's building. SUMCO has told customers that even a 10% price increase for next year's contracts "won't cover its costs" — supplier-speak for "we need a lot more than that before we'll even think about expanding." Shin-Etsu told analysts in late October that price increases have already started landing on some customers' invoices. The fabs get the headlines. The blank discs feeding them might be the quieter story.
The expert version
Every semiconductor device begins as a Czochralski-grown monocrystalline silicon ingot, sliced into wafers and polished to sub-nanometer flatness before a single lithography step occurs. Supply of the dominant 300mm (12-inch) format — used for essentially all leading-edge logic, DRAM, and NAND — is structurally concentrated: industry trackers put the combined share of the top five suppliers (Shin-Etsu, SUMCO, GlobalWafers, Siltronic, SK Siltron) at roughly 80% of 300mm volume, with the top ten reaching about 92%. Individual company shares vary by methodology and year, so treat any single-company percentage as approximate.
The current squeeze is a capacity-cycle story, not a feedstock story — polysilicon itself isn't scarce. After over-expanding into the 2021–2022 shortage, wafer suppliers absorbed a demand air pocket from 2023 through 2025 as PC, smartphone, and legacy-chip orders fell; utilization and pricing both slid, and new fab investment slowed sharply. AI accelerators and the HBM stacked alongside them now compete for the same 300mm substrate capacity as recovering legacy and automotive demand, and new ingot-pulling and polishing capacity typically takes two to three years to bring online — too slow to answer a demand spike within the same cycle. UBS estimated in a September 2026 note that 12-inch wafer utilization could climb from roughly 84% in 2026 to about 99% in 2028, and toward 103.7% by 2029 absent new capacity: demand nominally exceeding what installed capacity can physically supply. The bank argues suppliers would need price increases on the order of 40–50% to clear the cost-of-capital hurdle for new fabs; short of that, it expects wafer prices to rise over 20% annually in 2027 and 2028 even without major new supply.
Early price signals back the model: SUMCO said in mid-2026 that a 10% increase on 2027 contracts "won't cover its costs," and Shin-Etsu told analysts on an October 27, 2026 earnings call that price increases have begun for some customers, with broader resets tied to contract renewals in late 2027 or later. The main variable that could cap how far prices run is Chinese domestic wafer capacity, which is expanding faster than the incumbent five and gives chipmakers outside the leading edge some near-term leverage.
Why it matters for tech + supply chain: Everyone's watching GPUs and HBM for the next AI shortage. This one sits a layer further back — in the blank disc nothing else can exist without — and the five companies who make it are moving slower than demand, on purpose.
Why it matters for tech + supply chain: Wafer-fab capex decisions lock in roughly three years before they show up as shipped capacity, so the 2029 outcome is largely already determined by what the big five commit to in the next few quarters — which is exactly why their current reluctance to expand matters now, not later.