Cognex: Vision Quest - [Business Breakdowns, EP.207]

Summary

• Cognex is the #2 machine-vision player behind Japan’s Keyence, selling ruggedized factory cameras where “the software is the real value add” — guide, gauge, inspect, and ID applications with usually quick payback. Its serviceable market has grown from $2.9B in 2017 to $6.5B, likely reaching $8–9B at the next investor day; the industry has compounded at roughly 10%, with Cognex above that. Brett Larson’s caveat on the model: “Cognex doesn’t adjust out their stock-based comp, which is great for a company that’s essentially a software company, but you get stuck with the cyclicality of an industrial company as a trade-off” — the software is tied to the hardware and the sale is recognized upfront.
• The company’s DNA is “stacking S-curves essentially over 40 to 50 years”: OCR on semiconductor wafers for IBM in 1981, bottle-cap inspection for J&J, smart cameras onto factory floors around 2000, then barcode-reading ID products from 2010. The logistics line they hoped might one day hit $75M in sales — built hand-in-hand with Amazon — peaked at roughly $300M and 30% of the business in 2021.
• The next S-curve is AI: moving “from rules-based programming of vision systems to teaching by example,” via the ViDi and SUALAB IP acquisitions/acqui-hires. Deep learning addresses tasks difficult to program — 30 million people worldwide still do visual inspection, “something that humans actually aren’t great at” — while edge-learning products train on as few as 5–10 images and deploy in hours, opening less-sophisticated customer tiers that Keyence targets.
• Cognex is explicitly “taking Keyence’s playbook”: the Emerging Customer Initiative’s first cohort of young salespeople did 80,000 customer visits, added 3,000 new customers to a base of 30,000, and exited the year at roughly $1M/week in sales with accretive gross margins. Roughly 60–70% of those sales went to customers who “have never had a camera in their factory” — the ambition is to grow the base to hundreds of thousands.
• The business sits at the end of a long down cycle: operating margins were 13% last year versus a peak above 30%, including roughly 200bps of ECI headwind, against long-term targets of 15% constant-currency top-line growth and 40% incrementals. Valuation is “within reach of its 10-year low at five and a half times next-12-month sales” versus a 6–10x normal range — the 16x ZIRP print being, “like most things, just hilarious in hindsight.” The host noted that, unusually for a cyclical, the multiple has expanded when sales grew and compressed when they declined — a peak-on-peak pattern.
• Key risks: cyclicality (“early is the same thing as wrong — basically every incremental investor in Cognex the last year or two is probably feeling very early”), China at 18% of 2024 sales with the domestic-manufacturer chunk probably an uphill battle to grow over five to 10 years, and the AI transition itself opening “the window for disruption.”
• The takeaway lesson is culture: two CEOs in four and a half decades, “ministers of culture” in every office paid separately for the role, and founder Dr. Bob Shillman staying on as chief culture officer until 2021 through a long CEO transition — with voluntary attrition at half of industry peers.

Deep dive

1. Cameras are the delivery vehicle; software and application engineering are the product

• Brett’s setup: Cognex (“cognition experts,” founded 1981) sells ruggedized cameras with embedded processing — “like two cell phones stacked together with the Cognex yellow” — but the software analyzing images at high speed is the real value add. Applications split into guide, gauge, inspect, and ID: the DataMan family reads barcodes and printed characters (picture an Amazon sorting facility scanning tens or hundreds of thousands of packages daily), while In-Sight covers robot guidance, dimensioning, and quality inspection such as circuit-board solder checks.
• The value proposition — improving quality and throughput, reducing costs and waste, and increasingly addressing labor constraints — carries a usually quick payback. After upfront programming, “the human’s kind of out of the loop”; the camera feeds decisions to a PLC like Rockwell’s. SAM: $6.5B, up from $2.9B in 2017, probably $8–9B with new categories, while the industry has grown at roughly a 10% CAGR through cycles.

2. The Keyence enigma: 2% R&D, mid-80s gross margins, and a relentless scripted salesforce

• Cognex is #2 to Keyence, and the two historically split the pyramid: Cognex at the top with trained application engineers doing technical, spec-level sales — “they’re not winning on price” — differentiating on read rates 100–300bps better than peers, which at 100,000 packages a day means thousands fewer human interventions. The yellow cameras themselves signal something to a new COO walking a factory.
• Keyence, “very much an enigma,” spends roughly 2% of sales, or low single digits, on R&D with mid-80s gross margins versus Cognex’s mid-teens R&D and roughly 70% gross margins — and perhaps only 20% of its sales are comparable. Its edge is a process-oriented, KPI-tracked sales machine of college graduates so relentless “it’s become a meme in the community”: download a spec sheet with “your buddy’s email and phone number, not your own.”
• Rounding out the field: Hikrobot, a division of Hikvision, is #3 at roughly half Cognex’s sales, driven by domestic Chinese manufacturers; legacy players include Teledyne’s DALSA/Point Grey, SICK, Basler, Datalogic, and Zebra’s Matrox. Go-to-market is 70% direct (factory floor or machine builders/OEMs), roughly 15% systems integrators (mostly logistics), and roughly 15% distribution.

3. Forty years of stacking S-curves, from IBM wafers to Amazon tunnels

• The origin story is the pattern: the world’s first industrial OCR system reading serial numbers on semiconductor wafers for IBM, then J&J asking, “can you do some of these other novel applications like verifying the caps were on the bottles.” Semiconductors and electronics capital equipment were as much as 80% of sales going into the dot-com bubble, fell to 54% following the hardware bust, and reached 15% by the end of 2010 as smart cameras took vision into automotive, consumer electronics, food and beverage, and packaging.
• The ID business is the best specimen: launched around 2010 to displace laser scanners, with management saying they’d be happy if it eventually reached $75M in sales. Developed closely with Amazon, it hit roughly $300M and 30% of Cognex at the 2021 peak — “just an enormous new S-curve.”

4. Deep learning opens new applications; edge learning reaches lower tiers

• The technology shift began with the 2017 ViDi and 2019 SUALAB acquisitions — “essentially IP and acqui-hires” — moving “from rules-based programming to teaching by example.” Rules-based systems are ill-suited to subtle, high-variation tasks: inspecting a phone case for scratches and paint blemishes that a human can judge quickly, or deboning chickens where “every chicken’s different” and it is difficult to quantify where a robot should grab. With 30 million people still doing visual inspection — work humans can do quickly but at which they fatigue and miss things — the potential new-application opportunity is large.
• Edge learning is “more financially tangible for the investor community right now”: pre-programmed products trained on 5–10 images, running in hours, and sellable by a much less technical salesforce. Hence the Emerging Customer Initiative — the first cohort of sales Cognoids did 80,000 visits, added 3,000 customers to a 30,000 base with accretive gross margins, and exited at roughly $1M/week; cohort two entered in 2025. They sometimes run into Keyence in its bread-and-butter segment, but roughly 60–70% of sales are to customers who’ve never had machine vision at all.

5. Five end markets, one rhythm: upfront capex sales waiting on the next wave

• The stickiness/lumpiness trade-off: customers hope cameras last 10–20 years, so there is no regular replacement cadence, and installed bases are sticky because customers standardize on one vendor’s software. But revenue is recognized upfront and tied to large buildouts. Matt’s framing, which Brett endorsed: if Apple keeps producing the same phones, “you’re not going to see a big step up in revenue” — you need a new form factor or feature. Consumer electronics was 17% of sales in 2024; Apple was once 20% of Cognex, now mid- to high-single digits. The market is at maintenance levels awaiting a possible cycle tied to LLMs, AR/VR, humanoid robots, or another manufacturing shift. The long-term growth target for the end market is mid-teens.
• At the low end, new edge-learning systems may cost $1,000–$2,000 each and produce orders around $10,000; complex implementations can run into the hundreds of thousands. Smaller purchase orders are book-and-ship, while strategic factory builds have more lead time.
• Logistics is largest at 23%, having returned to growth and grown 20% in 2024; its long-term target is 20% growth. Non-Amazon vision tunnels are growing quickly, with add-ons such as damage inspection and dimensioning. Automotive is 22% and disappointed — EV-battery capex did not come through as expected, and sales fell mid-teens in 2024; its long-term target is 10% growth. Semiconductors are 10–15%, boosted by the late-2023 Moritex bolt-on, which sells optics and lighting; the outlook is positive. PMI is the catch-all for the remaining 20–25%.
• Interesting cycle mechanic: given short-cycle orders, “that business will inflect before we even know what it is they’re spending on” — Cognex can act as a leading indicator on capex, especially in consumer electronics.

6. Trough margins, trough multiple, and a culture with two CEOs in 45 years

• The financial frame: over the pre-downturn decade, Cognex grew about 13% excluding M&A and at mid-teens rates in constant currency. Long-term targets are 15% constant-currency top-line growth and 40% incrementals (Brett thinks “low double digit topline longer term” is the right bogey). Operating margins of 13% versus 30%+ peak reflect the depth and duration of the downturn and investment through it — the 2019 downturn only took margins from 27% to 20%. Last quarter’s double-digit organic growth already showed leverage falling through.
• Valuation two ways: the implied-FCF analysis says Cognex needs to compound free cash flow at a low-double-digit rate against an internal hurdle rate of at least 10% — below the long-term model, from a cyclical low with hopefully margin recovery ahead (“that’s the bull case”). EV/NTM sales is 5.5x, within reach of the 10-year low, versus a 6–10x normal range. The balance sheet has net cash, with cash and investments equal to roughly 10% of market cap; over time, the company has converted roughly 100% of net income to free cash flow and returned over 100% of free cash flow to shareholders, one-third through dividends and two-thirds through buybacks.
• The host noted that Cognex’s multiple has historically risen with sales growth and fallen with sales declines, unlike many cyclicals; Brett called the peak-on-peak phenomenon painful on the way down.
• Risks as Brett ranks them: cycle timing (“early is the same thing as wrong”), China at 18% of sales — two-thirds Western multinationals such as Apple and Foxconn, for which it seems unlikely they would install Hikrobot “for obvious reasons,” while the domestic remainder is probably an uphill battle to grow over five to 10 years — and the technology transition cutting both ways as a disruption window.
• The closing lesson is culture, deliberately engineered to survive its founder: Dr. Bob — who jokes that he doesn’t believe in exercise because basically your body is a bunch of mechanical joints with finite use — built the “work hard, play hard, move fast” Cognoid culture, then stayed as chief culture officer until 2021 while Danaher alum Rob Willett was CEO from 2011. Ministers of culture in every office are paid separately for the incremental role; leap-year skydives, armored-vehicle bonus deliveries, themed annual reports, and attrition at half of peers reinforce the point.