Before a silicon wafer is cut into chips, every chip on it is tested. A probe card is the interface: a custom board with thousands of microscopic needles that land on each chip's contact pads while a tester checks it. Every new chip design needs its own card, and the cards wear out, so demand tracks the number of new designs and wafers. FormFactor is the largest US maker of probe cards, and it also sells the probe stations engineers use in the lab.
Source: FormFactor Q2 2026 results (8-K Ex 99.1) and Form 10-Q; Investor Day (11 May 2026); consensus from MarketBeat and StockAnalysis (1–2 Oct 2026); peer data from company filings.
Why I chose it
1. A consumable on every AI wafer
Probe cards are custom to each chip design and wear out with use, so demand rises with every new design and every new wafer start, not only with factory expansions. FormFactor sells into foundry and logic (CPUs, GPUs, custom accelerators), into DRAM, and through a Systems business of engineering probe stations and cryogenic test equipment that its probe-card rivals do not match.
In plain termsEvery new AI chip needs its own set of test needles, and they have to be replaced as they wear.
2. A lead in HBM, and three new fronts
FormFactor's SmartMatrix contactors test hundreds of finished high-bandwidth memory stacks at once at HBM4 speeds, the high-value step where it describes its share as "quite strong". Three new fronts are opening: advanced probe cards for GPUs, now qualified and shipping production units from Q3; custom accelerators, mainly from 2027; and co-packaged optics test, now expected to significantly exceed $20m in 2026.
In plain termsIt already leads in testing AI memory, and is now moving into testing AI processors and optical chips.
3. The 2030 plan is arriving early
At the May Investor Day management set a 2030 model of $1.6bn of revenue, a 55% gross margin and earnings per share above $5. Q2 already annualises above $1bn of revenue at a margin above 50%. Guidance has not been cut in six quarters, the last four results beat consensus earnings by 27% to 34%, and consensus earnings for this year have risen about 2.8 times since February.
In plain termsThe company is already most of the way to the targets it set for four years from now.
4. The credit lens: the cleanest structure in the group
FormFactor carries $334m of net cash, an undrawn $150m credit line running to 2030, and one $11.7m building loan fixed at 2.75%. There is no convertible and no share sale in prospect. It is paying for a new plant in Farmers Branch, Texas, reaching initial capacity in early 2028, out of operating cash flow and a $24.2m state grant. That structure gives it options in a downturn that a levered rival would not have. What it cannot do is protect the multiple.
In plain termsThere is nothing to worry about in the balance sheet. The risk is in margins and in the price, not in the debt.
What would break it
The balance sheet is not the risk. The risks sit in the quality of this year's margin, the price the market is paying, and a two-front contest with a larger rival. Each is observable in public data.
| Signal | What I would see | Where it shows first |
|---|---|---|
| The margin gain proves partly temporary | The Q3 guide of a 54% gross margin includes about 300bp management calls non-recurring: $7m to $9m of tariff refunds and precious metal reclaimed from a closed plant. The underlying rate is near 51%, and Texas start-up costs run into 2027. | Q3 results, ~28 Oct to 4 Nov 2026 |
| The valuation leaves no room | At 49x this year's and 39x next year's consensus earnings, the shares trade above the analyst mean target. On growth-adjusted terms (PEG 1.56) FormFactor is the most expensive probe-card name: Technoprobe 0.91, Micronics Japan 0.75, MPI 0.38. | Forward P/E versus peers |
| Memory concentration | SK hynix is 24 to 27% of revenue. HBM, about two-thirds of DRAM revenue in Q2, shifts toward lower-value conventional DDR in Q3, and the memory cycle can turn. | SK hynix Q3 results; DRAM mix in the Q3 print |
| Technoprobe wins the two-front war | FormFactor is attacking Technoprobe's GPU franchise while Technoprobe enters HBM with its first qualified customer. Technoprobe is already larger in probe cards, earns a higher gross margin (57.2% in Q2) and is adding capacity faster. | Technoprobe and Micronics Japan Q3, Nov 2026 |
| Capacity arrives late | Until the Texas plant reaches initial capacity in early 2028, capacity rather than demand limits share gains in CPUs and custom accelerators. | Farmers Branch first production, late Q4 2026 |
| AI capex sentiment | The shares fell 16% in a day on 2 July 2026 on fears about hyperscaler returns, and move 1.4 to 1.8 times the semiconductor index. | Hyperscaler capex commentary, late Oct |
Known noise, not a thesis breaker: the CEO sells shares monthly under a pre-set trading plan, and insiders have been sellers with no purchases; this is routine at this size. Short interest is about 4% of the float, so this is not a squeeze story. The stock is at the top of its six-month $82 to $160 range for the third time, and short-term indicators say it is stretched.
How I value it
The operating plan is delivered (2028 revenue near $1.45bn at a 54% gross margin, earnings per share near $4.75), but the multiple compresses to about 31 times, in line with the large test-equipment makers Teradyne and Advantest.
The same operating plan, valued at about 36 times 2028 earnings, the median forward multiple of the wider die-test peer group, as the Texas plant ramps and the 2030 model arrives about two years early.
2028 revenue near $1.65bn at a 55.5% gross margin and earnings per share near $5.80, at 35 times. Upside stays open if GPU and custom-accelerator share gains and co-packaged optics test arrive faster.
Where I sit against the street, and what the range leaves out. All three of my cases sit above the analyst mean target of about $133.50 (median about $157.50; eight buys, four holds, no sells), and my bull matches the street high of $200. The three cases vary the multiple more than the earnings: each assumes the margin structure holds. So I also test a stress case outside the weighting, at about $75: this year's margin gain proves temporary and the memory cycle turns, leaving 2028 revenue near $1.15bn at a 49% gross margin and earnings per share near $3.00 at 25 times. It is not a target; it is the outcome a reader should weigh against a 49x starting multiple.
Entry discipline and dated checkpoints
The stock is testing the top of its six-month range for the third time, after a 56% rise in a month. The levels below frame where the risk and reward improve, what would confirm a breakout, and what would invalidate the idea.
Glossary
- Probe card
- A custom board with thousands of fine needles that contacts every chip on a wafer during testing.
- MEMS probe card
- A probe card with micro-needles made by lithography, used for fine-pitch, high-pin-count GPUs, custom chips and HBM.
- Wafer sort
- The electrical test of each chip on a finished wafer, before the wafer is cut.
- SmartMatrix
- FormFactor's full-wafer contactor for testing many HBM stacks in parallel at speed.
- HBM
- High-bandwidth memory: stacked DRAM placed beside an AI processor.
- DDR
- Conventional DRAM used in servers and PCs; lower value to test than HBM.
- Known-good die
- A chip proven good before it is assembled into a multi-chip package.
- Probe station
- Lab or production equipment that positions the wafer under the probe card.
- Co-packaged optics
- Optical engines built on silicon and placed next to switch or processor chips, which need their own test step.
- Non-GAAP gross margin
- The company-adjusted share of revenue left after direct costs, excluding items such as stock compensation and amortisation.
- Tariff refunds
- Repayment of 2025 tariffs later ruled unlawful; a one-off boost to this year's margin.
- Net cash
- Cash and securities minus total debt.
- P/E and PEG
- Price over earnings per share; PEG divides the P/E by the expected earnings growth rate, to compare companies growing at different speeds.
- Stress case
- A deliberately severe outcome used to test risk; not a forecast and not part of the weighted value.