Cohu’s beat is not the turn the headline suggests; it is an AI-systems option wrapped in a still-subscale recovery

The print should not be bought as a clean semi-cap recovery, because the quarter says demand is improving in selected sockets rather than across the model. Street numbers were set for loss minimization, not acceleration: EPS at -$0.06 versus -$0.19 was a +68.4% surprise, and revenue at $126.2 million versus $122.1 million was a +3.4% surprise. What was priced in was a modest revenue beat off a depressed base and some expense help; what actually surprised was that the beat came with management already pointing to Q4 revenue of roughly $122 million plus or minus $7 million. That split matters for portfolio construction. The variant perception is that Cohu is not yet a broad handler-and-contactor upcycle, but it is building an underappreciated call option in HBM inspection and edge AI or data center-related systems that can matter if 2026 demand converts faster than the current guide implies.

That distinction is visible in the financial trajectory, where the revenue line has stopped falling but has not regained the scale needed to absorb fixed costs cleanly. Revenue has rebounded from the trough to $126.2 million, and the latest quarter grew +17.2% sequentially and +32.4% year over year. That is the right direction, but it is still below the earlier-cycle revenue base of $179.4 million. Gross margin on the street-comparison history was 35.7%, which is better than the prior-quarter 34.4% but still far from the 48.1% level at the start of the displayed cycle. The market may be extrapolating the revenue slope; the margin series argues that mix, volume, and utilization have not yet caught up.

The capacity story explains the margin guide, because the company is not committing to a fourth-quarter system ramp despite the third-quarter beat. Jeffrey Jones put a hard stake in the ground when he said, “Overall, we expect Q4 revenue to be about $4 million or 3.5% lower than Q3, driven by systems revenue.” That wording is important because it attributes the sequential decline to the part of the business that should carry upside if AI-related capex were already broadening through Cohu’s order book. It also separates this event from the higher-beta peers showing larger year-over-year growth rates. ATEYY posted revenue YoY of +43.8% and 6871.T posted +48.3%, while Cohu’s latest reported revenue YoY was +32.4%. Cohu is participating in the recovery, but the comparison says it is not yet the cleanest expression of the test-and-assembly rebound.

The more interesting part of the print is that the narrow AI-related pockets are becoming large enough to monitor, even if they are not yet large enough to drive the consolidated guide. The HBM inspection line is the cleanest evidence. Jones said Cohu had “repeat orders for Neon HBM inspection tools,” and that those orders raised the year’s revenue forecast for the systems to “between $10 million and $11 million.” That is not a company-wide inflection on a $126.2 million quarterly revenue base, but it is specific, repeat-order demand in the part of the market where customers are still spending. David Duley’s framing adds a second piece: 2025 system revenue associated with edge AI or data center-related AI was “approximately $40 million,” with growth expected into 2026. The variant view is that investors should value this as a visible product wedge rather than dismiss it as AI language layered onto a cyclical trough.

Still, the wedge has to fight the mix arithmetic, and that is why the stock reaction to the beat should be disciplined. Cohu’s own call basis showed Q3 gross margin at 44.1%, while the quarterly history basis shows 35.7%; those are different reporting bases and should not be blended. On the company’s own account, operating expenses were $48 million, or $2 million lower than guidance, which helped the quarter look better than revenue alone would suggest. But management guided operating expenses to about $50 million in Q4, including around $2 million for variable R&D product development prototype materials. That tells investors the company is spending to seed products before the revenue guide proves the payback, a reasonable strategy but not one that supports a near-term operating leverage thesis unless orders accelerate.

The balance sheet actions reinforce that management is buying time for the product cycle rather than declaring the trough definitively past. Cash and investments decreased by $11.2 million during Q3, and Jones tied that to cash used in operations to support a 17% growth in sales and a $33 million increase in accounts receivable. That working-capital absorption is not a red flag if Q4 collections normalize, but it reduces the quality of the third-quarter beat because revenue growth consumed cash. In early Q4, the company raised gross proceeds of $287.5 million through an upsized offering at a 1.5% interest rate and with a 5-year term. The capped call language also matters for equity holders: Cohu purchased a 100% capped call intended to limit dilution until the stock price exceeds $41 per share. The financing is therefore less a distress signal than a strategic extension of runway, but it makes the next two quarters a test of whether capital was raised into a product inflection or simply ahead of another uneven demand period.

That unevenness also shows in customer concentration, which is the main read-through for the OSATs. During the third quarter, 3 customers each represented more than 10% of sales, with 1 in mobile and 2 in automotive. For ASE Group and Amkor, both identified customers for Cohu test handlers and contactors, the implication is not that OSAT capex has broadly reopened; it is that selective handler, contactor, and inspection demand is being placed where end markets justify it. The $2.3 million precision analog test contactor order at a U.S. customer is a small but concrete positive for contactor demand, while the first Diamondx order from a long-standing handler customer came with roughly $1.7 million in new business. Those amounts are not large enough to shift ASE Group or Amkor capex models by themselves, but they do show that Cohu is getting paid for specific upgrades inside existing customer relationships rather than relying only on a cyclical replacement wave.

The recurring-revenue mix is the other reason not to overstate the systems rebound. Jones said recurring revenue was 55% in Q3 and is expected to be about 60% of total revenue in Q4. A higher recurring mix can stabilize revenue, but in this case the step-up coincides with the sequential decline in systems revenue. That is defensive quality, not necessarily growth quality. The company’s longer-term margin case depends on systems recovering enough to lift factory utilization while recurring revenue cushions volatility. If recurring becomes 60% because systems are soft, the model may become less cyclical but also less explosive. If recurring stays near that level while AI-related systems rise from the approximately $40 million 2025 base, the market is likely underpricing the earnings torque.

The call delivery supports that nuanced read: management sounded more directional on guidance, but not more relaxed about the environment. The tone history shows guidance_tone improved to 0.49 in Q1 FY2026 from 0.33 in Q4 FY2025, while uncertainty rose to 82.1 from 76.6. That combination is unusual. It says management’s prepared forward language became more constructive even as the transcript carried more uncertainty. The call-over-call delta makes the conflict sharper: guidance_tone rose +0.17, but qa_sentiment fell -0.14. PMs should read that as a management team willing to frame the next leg, while analysts on the call still pressed on what has to convert for the numbers to work.

The tone data also helps explain why the market can misprice this print in both directions. A simple beat-and-raise screen misses the fact that the Q4 revenue guide is lower than Q3, while a simple sequential-decline screen misses the fact that Cohu has named repeat HBM inspection orders and approximately $40 million of AI-related system revenue for 2025. Prepared_sentiment in the tone history was 0.65 in Q1 FY2026, but tone_confidence was only 0.17. That is not a contradiction to ignore; it is the setup. The company is more constructive in what it can point to, but the confidence score says the proof remains order conversion and margin response rather than management conviction alone.

The peer comparison frames the opportunity cost. Cohu’s latest gross margin in the quarterly history was 35.7%, while 7729.T was 42.4% and 6315.T was 36.2%. Against those peers, Cohu is not obviously mispriced because it has already caught the group’s margin structure; it is potentially mispriced because a small amount of AI-system revenue could have a disproportionate effect if it lifts utilization from a depressed base. That is why the $10 million to $11 million Neon HBM inspection forecast matters more than its absolute size suggests. It gives investors a measurable product line to track, and it connects the Cohu story to HBM capacity additions without requiring a heroic assumption that all semiconductor test spending has turned.

The bear case is straightforward and should not be dismissed. Q4 revenue is guided at approximately $122 million plus or minus $7 million, operating expenses are expected to be about $50 million, and recurring revenue is expected to be about 60% of total revenue. If systems revenue declines as guided and recurring mix rises mainly because system demand is soft, Cohu remains a cost-control and product-option story rather than an earnings recovery story. The tax provision also flattered Q3 by coming in about $3.5 million lower than forecast at $11.7 million, due to the reversal of tax reserves after a jurisdictional audit. That is legitimate accounting, but it is not evidence that end demand improved. Investors should therefore separate the EPS surprise from the order-quality debate.

The bull case, however, is more actionable than the headline guide implies. Revenue exceeded the street by +3.4% despite a still-weak base, and management identified repeat Neon HBM inspection orders plus approximately $40 million of AI-related system revenue for 2025. If those system categories grow into 2026 while quarterly operating expenses move toward approximately $49 million when revenue is around $130 million per quarter, the earnings bridge becomes visible. That operating-expense target is the most important medium-term number in the call because it defines the revenue level at which the restructuring begins to matter. Cohu does not need a return to $179.4 million revenue to change the debate; it needs proof that revenue around $130 million can carry better mix and lower expense intensity.

What to watch next is therefore concrete. For Q4, the first test is whether revenue lands above the $122 million guide midpoint and closer to the high end of the plus or minus $7 million range; a result below the midpoint would confirm that Q3 systems strength was lumpy. The second test is mix: recurring revenue should not merely rise to about 60% because systems fall, while Neon HBM inspection should stay on the “between $10 million and $11 million” annual track with evidence of repeat demand into 2026. The third test is cost: operating expenses need to stay near the about $50 million Q4 guide and then move toward the approximately $49 million level at the beginning of 2026. Break any two of those, and the stock is just a cyclical stabilization trade. Hit them, and the market will have to revalue Cohu as a small-cap test name with a measurable AI-systems wedge and a recovering cost base.