Two Companies Build an Estimated 90% of the World's Chip-Testing Machines. An AI Accelerator Now Spends 40 Times Longer on One Than a Phone's Chip Does.
Every bottleneck this newsletter has covered so far sits somewhere you could point a camera at — a fab, a port, a mine. This one hides in the last room a chip visits before it ships, inside a machine built by one of exactly two companies on Earth — and it spends roughly forty times longer checking an AI chip than it spends on the one in your phone.
Same machine, forty times the wait: a tester throughput race
Run one identical tester for a chosen stretch of simulated time — once testing phone chips (~30s each), once testing AI accelerators (~20min each; both figures are simplified, rounded industry estimates). Both bars share one scale, set by the phone lane's total, so the gap is drawn to scale. Pick a run length, then press Run.
The plain version
Before an AI chip or a phone chip leaves the factory, it has to prove it actually works. It gets plugged into a machine called a tester, which pokes at every circuit with tiny electrical probes and checks that the whole thing passes. Here's the part almost nobody outside chip manufacturing knows: essentially two companies build nearly all of these machines worldwide — a Japanese firm, Advantest, and an American one, Teradyne. Between them, they're estimated to supply somewhere around 85 to 90 percent of the global market. That's already the kind of chokepoint this newsletter tracks every day — a mineral refiner, a lithography maker, a design-software vendor. This one is just quieter, because nobody argues about testers the way they argue about GPUs.
The twist is what's happening inside that machine. A chip for your phone is a relatively modest electrical puzzle and clears the tester in about half a minute. A modern AI accelerator — dense with transistors and stacked memory — is a much bigger puzzle to check, and it can sit in that same machine for 20 minutes or more. That's roughly 40 times longer, per chip. A tester doesn't care what it's testing; it can only handle one chip at a time. So if an AI chip ties up the machine 40 times longer than a phone chip would, that one machine's effective output for AI chips collapses to a sliver of what it could do for ordinary ones — and you can't just buy more testers overnight. The specialized probe cards they need, the needle-like connectors that actually touch the chip, now take months to build.
Everyone's been watching the chip fabs for the AI boom's next shortage. The quieter one might be the room where the chips get checked on their way out the door.
The expert version
Automated test equipment (ATE) qualifies every die before packaging and shipment: functional and parametric test against the pin/pad interface, speed-binning, and — for high-power parts — a burn-in step at elevated voltage and temperature to screen early-life failures. The ATE market is a structural duopoly. Industry trackers and sell-side research put Advantest's and Teradyne's combined share of global tester revenue at roughly 85–90%, with the exact figure varying by methodology (some trackers include probe-card and handler revenue, others don't). Within that, Advantest's share of the AI/SoC-tester sub-segment specifically is reported to have risen from about 56% to 66% between calendar 2024 and 2025, as its V93000 platform became a default qualification tool for Nvidia- and Broadcom-class accelerators.
The AI-specific constraint is dwell time, not unit count. A mainstream mobile SoC clears a tester in roughly 30–60 seconds. A reticle-limited AI accelerator paired with multiple HBM3E stacks — more I/O to exercise, more logic to pattern, and memory that itself needs roughly twice the per-stack test time of HBM2E — can occupy the identical machine for 20 minutes or longer: an estimated 20–40x multiplier over a mobile part (figures are industry estimates, not vendor-published benchmarks, and vary with die size and test-pattern depth). Because effective tester throughput is fleet size divided by per-unit dwell time, a 20–40x increase in dwell time compresses capacity by roughly the same factor, unless the installed base grows at a matching rate or multi-site/parallel test schemes claw some of it back.
Capacity isn't elastic on a useful timescale. Teradyne's Semiconductor Test segment generated $1.111 billion in Q1 2026 — about 87% of the company's record $1.282 billion quarterly revenue, per its SEC filing — crossing $1 billion in a single quarter for the first time, with management attributing roughly 70% of total revenue to AI-related demand. One level further downstream, Advantest's probe-card subsidiary has said a single advanced probe card for a large AI die now takes five to six months to build. Front-end wafer and HBM capacity has been 2026's headline AI supply constraint; the tester-and-probe-card layer sitting directly behind it is the less-visible one.
Why it matters for tech + supply chain: Everyone's watching wafer fabs and HBM lines for the AI boom's next shortage. The quieter one is the room at the very end of the line, built almost entirely by two companies, where an AI chip can tie up the machine roughly forty times longer than the one in your phone.
Why it matters for tech + supply chain: Front-end capacity (CoWoS, HBM) gets the headlines, but ATE dwell time is a multiplicative constraint one step further downstream — and because it scales with die complexity rather than wafer starts, it doesn't ease just because fabs add capacity.