COHUCohu Inc.

Investment Thesis

Cohu supplies the equipment that gates whether a fabricated chip is shippable: semiconductor test handlers, automated test equipment, and test interface products — contactors and power probe cards — plus inspection/metrology and AI-driven software analytics. The company reports one aggregated segment, 'Semiconductor Test & Inspection,' and its active installed base exceeds 25,000 systems across 280+ high-volume manufacturing facilities, serving 108 customers in 31 countries.

The thesis is that rising AI compute intensity turns test into a bottleneck with pricing power. As accelerator power climbs into the 700W–1,000W+ range, a handler must actively sink that heat while holding die temperature in spec during test — otherwise the device is throttled (under-tested, slower throughput) or the junction overheats (false yield loss). Cohu's Eclipse test handler with T-Core active thermal control is the named product for exactly this problem, and management attributes the AI-compute momentum to its adoption 'for high-power processors used in data centers.'

The financial inflection is real and filed: Q2 FY2026 revenue was $149.0M, +38% YoY, gross margin 45.4% (up from 43.7% a year ago), and non-GAAP net income of $14.1M versus $0.7M a year earlier — roughly a 20x improvement. Management raised FY2026 high-performance-computing revenue to $100–110M and the annual AI-driven compute pipeline to roughly $850M, and guided Q3 to $170M ± $7M. With a $2.39B market cap, Cohu is a small-cap with direct leverage to the AI test-and-inspection buildout. Deep-research verdict: STRENGTHENED — scores ts 7 / ec 7 / ttm 8. Research-only at POOL stage.

Physical AI / Value-Chain Relevance

Layer 1 — AI Factory & Cloud Training Infrastructure. Cohu sits at the back-end gate of the AI hardware stack: every AI accelerator, HBM stack, and wide-bandgap power device is revenue only if it passes test, and rising test intensity — longer test times, more parallelism, tighter thermal control — monetizes directly into Cohu's product line. The 10-K names the growth programs explicitly: high-bandwidth memory (HBM) inspection, high-performance processor test in AI applications (CPUs, embedded NPUs, discrete and integrated GPUs, ASICs and xPUs), SiC and GaN wide-bandgap test, and AI-driven software analytics — plus 'physical AI' itself as a growth category.

The moat is test-cell qualification friction plus consumable lock-in, not a one-time product sale. Test contactors and power probe cards are specific to individual device designs, need frequent replacement, and scale in size and complexity with the number of devices tested in parallel — a recurring-revenue stream the 10-K calls a key contributor. Once a thermal-capable handler is qualified into a datacenter accelerator's test cell, the interface consumable stream follows for the life of that device design. An AI-driven software analytics layer (digital twin, scripting) adds switching cost on top of the installed base.

Catalysts

• Q2 FY2026 results (reported): +38% YoY revenue, 45.4% gross margin, non-GAAP income of $14.1M; FY26 HPC revenue guidance raised to $100–110M; AI pipeline raised to roughly $850M.
• Q3 FY2026 guide: $170M ± $7M — roughly +14% sequential, signaling the ramp is not a one-quarter event.
• Test-cell utilization at approximately 80% and rising ('low 80s' at end of June): historically, sustained high-80s% utilization is when handler pricing power and contactor pull-through accelerate.
• Sell-side validation: Stifel raised its price target to $70 on AI compute test becoming a top-3 segment in CY2026; AEHR's AI test beat (July 14) lifted the whole test-equipment sector.
• $850M AI-driven compute pipeline: at $100–110M current-year HPC revenue, the pipeline is roughly 8x this year's HPC revenue — the bull case is conversion velocity, not demand.

Positioning / What the Market May Be Missing

The market's default frame for Cohu is a small, cyclical test-equipment name that posted a GAAP operating loss of -$69.8M in FY2025 and is still digesting a restructuring. What is under-appreciated:

First, the AI-compute mix is already re-rating the business: HPC revenue of $100–110M on a run-rate tracking roughly $620M is about 16–18% of revenue, and it is growing while auto and consumer stay weak — the mix shift is visible in the 45.4% gross margin and the 161bp YoY expansion. Second, the consumable architecture means revenue durability: contactor/probe-card attach follows handler installs with a lag, so the systems sold for AI compute today produce a recurring interface stream for the life of the device design. Third, the roughly 80% test-cell utilization is a capacity-tightness signal that typically precedes pricing power.

One divergence to flag: the consensus target in our database ($47.82) sits below the latest price ($54.37) — either the sell-side has not caught up with the AI-test re-rating, or the stock has run ahead of fundamentals; the vault note cites Stifel's $70 target, so the DB consensus figure appears stale. The crowding flag is YELLOW, and the short ratio (6.52 days to cover) is elevated — positioning that can squeeze or burn depending on execution.

Risks and What Invalidates the Thesis

• Customer concentration: the ten largest customers are 60% of 2025 revenue; a single major program loss or a capex pause at a top hyperscaler or memory customer moves the whole company.
• Competitive intensity: the 10-K warns of 'several participants resulting in intense competitive pricing pressures' — Advantest and Teradyne can price against Cohu's handlers, and the thermal moat is a 12–18-month head start, not a permanent monopoly.
• Cyclicality: auto and consumer were still weak in Q2; the AI-compute ramp could be masked if legacy end-markets keep dragging.
• Profitability history: FY2025 was a GAAP loss year; the return to non-GAAP profitability is recent and untested across a downturn.
• Valuation sensitivity: GAAP EPS is negative (the DB shows -$0.79) and the stock carries a rich multiple on recovering earnings; a guide-down on HPC conversion would be punished hard.

The cleanest disconfirmer: Q3 revenue missing the $170M guide with HPC conversion below the $100–110M run-rate, or an Advantest/Teradyne active-thermal product win at a named hyperscaler customer.

What to Watch Next

1. Q3 FY2026 results versus the $170M guide — the first clean test of the ramp's durability.
2. HPC revenue conversion: whether $100–110M becomes $150M+ as the roughly $850M pipeline converts.
3. Contactor pull-through: does interface revenue lag handler installs by 2–4 quarters, as the model implies?
4. Test-cell utilization into the high-80s% — the pricing-power trigger.
5. Competitive response: any Advantest/Teradyne thermal-handler announcements at major test conferences.
6. Auto/consumer recovery or continued drag — the offsetting variable in the mix.