Discovery › screen
Screen semi test ate and tech sweep synthesis
Date: 2026-07-29 · Command: /screen (round 2 — tech & emerging tech)
Framework: analysis_notes.md §0–§4
Context: Round 1 (screen-unswept-fields) covered healthcare/energy/industrials/staples and was rejected by the user as uncompelling business models. This round pivots to software, hardware, tech and emerging technology, on the hypothesis that "the recent massive downslide might help with surfacing value."
Companion files from this sweep (written separately): - screen-robotics-automation-machine-vision-physical-ai-2026-07-29.md - screen-gaming-adtech-consumer-internet-2026-07-29.md - screen-japan-tech-semis-2026-07-29.md
PART A — Cross-Sweep Synthesis
A1. How the fields were chosen
The library had already run 44 screens, ~15 of them on enterprise software, plus semi-cap equipment, memory, analog/auto semis, photomasks, OSAT and quantum. Three screens from today alone already covered the drawdown (screen-tech-crash-hidden-value, screen-quality-smallmid-value, screen-global-vs-us-value). Re-flying those was pointless.
Four genuinely unswept tech angles were selected:
| Angle | Why unswept | Thesis tested |
|---|---|---|
| Semi test / ATE | Semi-cap, memory, analog, photomasks, OSAT all covered — never the test layer | Test intensity tracks device complexity, not fab capacity |
| Japan tech/semi | Our own 7/29 global screen flagged Japan and never converted it to names | Cheapness migrated from multiple into shareholder returns |
| Robotics / physical AI | Datacenter AI covered exhaustively; the physical layer never | Automation still in a capex trough while industrials re-rated |
| Gaming + adtech | ~15 software screens, every one enterprise/B2B | Unharvested "AI will disrupt this" discount |
Two guardrails were hard-coded into every brief, both learned from the same day's AMAT analysis: 1. "Down a lot" ≠ cheap. Measure every candidate against its own historical multiple band, not its 52-week high. 2. Story vs revenue. Reject anything whose valuation rests on a TAM slide.
A2. ⭐ The meta-finding
The drawdown did not surface value in tech, because the run-up was larger than the drawdown.
Three of four founding hypotheses were killed by the evidence. In each field the pattern was the same: names down 15–68% from their highs that are still expensive against their own history, or whose declines were fully earned.
| Field | Hypothesis | Outcome |
|---|---|---|
| Semi test | Priced as capacity-levered when it's complexity-levered | ❌ Thesis TRUE, trade FALSE. All 11 names trade above their own historical bands |
| Japan | −8.4% drawdown = opportunity | ❌ The drawdown is 100% currency. TOPIX hit a record high July 6 |
| Robotics | Still in a capex trough | ❌ Cycle already turned. Machine tool orders +27.8% YoY Q1'26 |
| Gaming | Unharvested AI-disruption discount | ❌ False positive. The big drawdowns had non-AI causes and were earned |
| Adtech | Same | ✅ Partly true — discount real and partly unearned, but moats are genuinely thin |
Only three names survived §0 across ~250 screened. That low hit rate is itself the answer to the user's hypothesis.
A3. The survivors
| # | Ticker | Field | Price | Fwd P/E | vs own history | FCF yield | Conv | Status |
|---|---|---|---|---|---|---|---|---|
| 1 | ZBRA | Machine vision / warehouse robotics | $288.94 | ~15.6x adj | vs own 3yr avg ~40x | 6.0% | 7.0 | ✅ −18% off high |
| 2 | NTES | Gaming (China) | — | 12.5x | vs ~26x peers | — | 7.0 | ✅ great + cheap |
| 3 | SHECY | Silicon wafers (Japan) | $18.41 | — | 6.41% total shareholder yield | — | 6.0 | ⚠️ entry $15.00–16.75 |
ZBRA — the cleanest mispricing found. Zebra is a machine-vision and warehouse-robotics platform still priced as a fading barcode-scanner company: ~15.6x adjusted forward against its own three-year average near 40x, 6.0% FCF yield, −18% off high. The mispricing is in the market's label, not its arithmetic. It is also a genuine diversifier — 3PL logistics and industrial end-markets, not datacenter capex. The named flaw: the "device refresh vs. secular automation" split is unresolved, and Zebra does not disclose the size of its machine-vision/robotics business — which is the entire optionality.
NTES — the only name in the "great + cheap" quadrant. 41% operating margin, 22% ROE, 25% FCF CAGR, net cash, a growing dividend and a shrinking share count, at 12.5x forward vs ~26x for peers. The key distinction: it is cheap because it is Chinese, not because of AI — a risk that can be underwritten and sized, and one entirely decoupled from the thesis this screen was built on. ⚠️ Concentration: this would be a third Chinese position alongside BABA and TCEHY in the Structural-Risk sleeve. Size the sleeve, not the name.
SHECY — quality, but not in zone. The only Japanese name passing both §0 gates: 6.41% total shareholder yield (buyback +158% YoY), 11% ROE, 26% operating margin, net cash, −25.6% off high, and the one semi-linked name that is early-cycle rather than peak (three of five wafer oligopolists synchronously revised 2026 shipments up). Entry $15.00–16.75 vs $18.41 today. ⚠️ Adds to existing AI/semi concentration. SONY is the only Japanese name of the three that genuinely diversifies — and it is the only one with frictionless access (NYSE, 6.5M shares/day, $0.00 spread, options).
A4. Watchlist candidates with hard triggers
| Ticker | Field | Now | Trigger | Why worth waiting for |
|---|---|---|---|---|
| FORM | Probe cards | $103.19 | <$70, ideally <$60 | Purest consumable annuity in semi test — 81% probe cards, #1 share, near-zero debt, no dilution |
| CGNX | Machine vision | ~$62 | <$48 | Best switching-cost moat in automation; 68% GM, net cash, 7 straight quarters of margin expansion |
| DV | Adtech | — | — | 10.2% FCF yield, 82% GM, only small-cap adtech buying back more than it issues |
| SONY | Diversified tech | — | $19.50–21.50 | 19.1x fwd, 3.0% growing yield, the only frictionless Japan access |
A5. The most valuable rejections
| Name | Headline | Why rejected |
|---|---|---|
| Nidec | Cheapest multiple in Japan (10.7x fwd) | 1,000+ instances of improper accounting, ¥160.7B restatement, TSE "special alert." Cleanest §0 illustration of the sweep. Permanently disqualified. |
| IPGP | 46% off high | Still 36x forward. Gross margin collapsed 55%→38%, FCF negative, revenue −11%/yr. Chinese fiber-laser makers permanently destroyed the price umbrella. Structurally impaired, not cyclically depressed. |
| RBLX | −68% | Not AI. Age-verification cost 12M DAUs (144M→132M); FY26 bookings guide cut 22–26% → 8–12%; FCF guided down; 5%/yr dilution. The drawdown is earned. |
| ONTO | Re-checked from 7/13 | The crash made it worse — P/E 95 → 114.7, because earnings fell 48%. Rejection stands. |
| AEHR | Burn-in / KGD story | P/S 48x, EV/Rev 40.5x on a loss-making ~$60M-revenue company, +365% off its low, beta 3.18. Uninvestable at any story. |
| EA | Screens cheap | Not investable — a $210/sh take-private trading at a 0.24% merger spread, stuck at CFIUS. Pure arbitrage. |
| Humanoid robotics | Entire category | <3,000–4,000 units in productive commercial deployment globally, ~90% Chinese, 23% buyer satisfaction. No investable US-listed pure-play with revenue. |
| TER / MKS | Best robotics assets (UR/MiR) | Disqualified by the diversification constraint — 104% revenue growth is semi test riding the exact AI wave already heavily owned. Would have doubled an existing bet while feeling like diversification. |
A6. Two findings that correct our own prior research
1. Japan's drawdown does not exist. The 7/29 global screen recorded EWJ at −8.4% off its high and concluded Japan was the one foreign market worth pursuing. In fact: - TOPIX set a record closing high of 4,101.96 on July 6, 2026 - USD/JPY is ~163.5 — the weakest yen in four decades, −8.41% over 12 months - EWJ's −8.4% and the yen's −8.41% are the same number
There is no index-level Japan drawdown to buy. That table row is an FX chart. If a Japan allocation is wanted rather than a name, the honest instrument is DXJ (yen-hedged), not EWJ. Also worth noting: the global screen's own Toyota pick remains a better expression of the Japan thesis than anything in Japanese tech.
2. The Japan governance thesis is name-specific, not country-specific. Computing buyback+dividend yield from cash-flow statements, the range inside one field is 0.03% to 4.48% — Shin-Etsu returns 5x TDK and 25x Advantest. Buying "Japan" gets you Advantest's 0.25% as easily as Shin-Etsu's 6.4%. Japan is now a stock-picker's market, not an allocation. The re-rating leg is over: P/B 1.7x is a post-2008 high with the index at a record.
A7. An irony worth recording
The clearest way AI is currently damaging a gaming company is Nintendo's bill of materials, not its content pipeline. NTDOY is −53%, and the cause is memory-chip inflation — the AI datacenter buildout bidding up DRAM — plus the Switch 2 cycle, with FY27 net profit guided −27%. This directly corroborates the 7/29 memory screen's finding that DRAM contract prices are still +13–18% QoQ and that memory consumers are being squeezed, not helped.
PART B — The Semiconductor Test (ATE) Screen
The only field in this sweep without its own file. Never previously screened by this agency.
B1. The thesis — confirmed at the industry level
| Evidence | Source |
|---|---|
| ATE spend rose from ~7% of WFE (2023) → ~15% (2025) — test intensity more than doubled as a share of capex | SEMI, 2026 |
| 2026 forecast: test equipment +12.0% vs WFE +9.0% — test outgrows capacity spend | SEMI, 2026 |
| Probe card TAM $3.5–5.5B by 2026, 6.5–9.5% CAGR, driven by KGD protocols forced "leftward" by chiplets/2.5D/3D | HDIN Research |
| Teradyne Q2'26: revenue +104% YoY, AI >60% of sales, record memory test $212M | Teradyne Q2 2026 |
Test genuinely decouples from wafer capacity. The mechanism is real and better-evidenced than the analog "wrong-cycle" trade it superficially resembles.
B2. The market got there first
| Ticker | Price | Off 52w high | Off 52w low | P/E ttm | P/E fwd | P/B now | P/B own 2yr band |
|---|---|---|---|---|---|---|---|
| TER | $360.85 | −26.0% | +252% | 49.6 | 34.9 | 17.97 | 3.67 – 11.45 |
| FORM | $103.19 | −35.6% | +296% | 118.6 | 31.8 | 7.74 | 1.69 – 5.18 |
| COHU | $44.99 | −39.7% | +153% | n/a (loss) | 30.5 | 2.76 | 0.75 – 2.00 |
| CAMT | $136.53 | −36.8% | +80% | 142.2 | 30.2 | 9.21 | — |
| NVMI | $389.56 | −36.8% | +67% | 48.9 | 30.1 | 8.92 | — |
| ONTO | $245.38 | −36.5% | +174% | 114.7 | 24.7 | 5.72 | — |
| ATEYY | $189.14 | −15.0% | +191% | 60.2 | — | 28.14 | — |
| AEHR | $74.14 | −41.4% | +365% | n/a (loss) | 53.6 | 10.97 | P/S 48x |
| ICHR | $73.54 | −35.3% | +461% | n/a (loss) | 27.3 | 3.83 | capacity-levered |
| UCTT | $81.20 | −43.7% | +282% | n/a (loss) | 20.6 | 5.79 | capacity-levered |
| KEYS | $311.39 | −17.0% | +104% | 50.2 | 26.1 | 8.40 | — |
The three names with hard historical bands are the proof: - TER: FY2025 traded $65.77–$205. Today $360.85 — 76% above its FY2025 high. P/B 17.97 vs an FY2025 low of 3.67. - FORM: FY2025 traded $22.58–$61.10. Today $103.19 — 69% above its FY2025 high. - COHU: FY2025 traded $12.57–$28.03. Today $44.99 — 60% above its FY2025 high, while still loss-making.
Even the most beaten-up name in the field sits at the richest valuation of its recent life. "Down 40%" here means "down 40% from a bubble."
B3. The peak-earnings problem (§0)
The two flagship names had zero growth for three years, then exploded in 2026:
| Rev 2022 | Rev 2025 | 3yr rev CAGR | 3yr NI CAGR | 3yr FCF CAGR | FY25 FCF yield | |
|---|---|---|---|---|---|---|
| TER | $3.16B | $3.19B | +0.4% | −8.2% | +2.8% | 0.80% |
| FORM | $748M | $785M | +1.6% | +2.3% | −43.9% | 0.15% |
| COHU | $813M | $453M | −17.7% | n/m | −52.2% | 0.50% |
| CAMT | $321M | $496M | +15.6% | −14.1% | +36.9% | 2.00% |
TER is capitalized at 35x forward on a business that compounded revenue at 0.4% over the prior three years — and its Q3'26 guide is non-GAAP EPS $1.85–2.15 against Q2 actual of $2.47, a ~19% sequential decline on ~6% lower revenue. Margin compression at the top, one quarter after "record everything."
B4. Top 3 — as WATCH candidates, not buys
None clears §0 today. Ranked by business quality × structural fit × distance-to-entry.
1. FORM — FormFactor · Probe cards · $8.04B · $103.19 · −35.6% off high · Conviction 5/10
The purest expression of the thesis. 81% of revenue is probe cards — the highest-value consumable in wafer sort. Cards dull, bend and break; the harder you test HBM to protect yield, the faster you burn them. A razor-blade annuity levered to units × test-seconds, not fab capex. ~22% global share (#1 vs Technoprobe ~16%); top 5 control 62%. Q2'26 revenue $258.2M (+14% sequential), record DRAM probe cards on HBM4 Smart Matrix at-speed stack test. Metrics: P/E fwd 31.8; P/B 7.74 vs own 2yr band 1.69–5.18; FY25 FCF $11.7M = 0.15% FCF yield; FCF CAGR −43.9% (capex tripled to $103.7M); 3yr rev CAGR +1.6%; Debt/Assets 2.6% (pristine); shares flat at 78.3M (no dilution); ROE 7%. Moat — switching costs (real) + intangibles. Probe cards are designed per-device and qualified into a customer's test program; requalification is expensive and risks yield during a ramp. Note Teradyne bought 10% of Technoprobe for $516M in Apr 2025 — the incumbents believe value accrues at the interface layer. Bear case. Three years of ~zero growth preceded this. The entire re-rating rests on HBM4/HBM5 content, and our own memory screen flagged 2027 supply risk — probe-card burn is a derivative of memory wafer starts, so a digestion year hits volume and price together. GAAP earnings are a fraction of non-GAAP. FCF is ~zero while capex funds a boom that may not persist. At P/B 7.7 there is no asset floor. If HBM normalizes this is a $700M–1.0B revenue business earning ~$2/share GAAP — worth $50–60, not $103. Fair value $60–90 · entry <$70 (want <$60 for real margin of safety).
2. COHU — Cohu · Test handlers, contactors, interface consumables · $2.12B · $44.99 · Conviction 3/10
The only name not on peak earnings. Revenue collapsed $813M (2022) → $453M (2025) as auto/industrial test went into a two-year depression; earnings are ahead of it, not behind. Best recurring mix in the field: 60% of Q1'26 revenue from consumables, interfaces, software, spares and services. Test utilization recovered to 78–79%, near the capacity-buy threshold; FY2026 guide +20–25% revenue. Metrics: P/E fwd 30.5 (ttm n/m, EPS −$1.19); P/B 2.76 vs own 2yr band 0.75–2.00; GM 43%; 3yr rev CAGR −17.7%; Debt/Assets jumped 1.9% → 26.4% ($328M raised in 2025); shares −1.4%/yr. Bear case. A cyclical recovery story dressed as a value story, and the recovery is already paid for — P/B 2.76 is 38% above the top of its own two-year band while still posting losses. Management levered the balance sheet at the bottom. If auto/industrial recovery stalls there is no earnings floor, and now there is debt. Fair value $35–55 · entry <$30.
3. CAMT — Camtek · Advanced-packaging inspection · $6.38B · $136.53 · Conviction 4/10
The only name with real cash generation — FY25 FCF $127M on $14.8M capex (asset-light), FCF CAGR +36.9%, 2.0% FCF yield (best in field). Genuinely complexity-levered: inspection content scales with dies-per-package. 51% GM, 22% op margin, +15.6% 3yr revenue CAGR. ⚠️ Data gap: FY25 net income fell $118.5M → $50.7M while operating income rose $108M → $128M — the drop is entirely below the operating line, likely tied to the $500M convertible issuance. The ttm P/E of 142 is misleading; confirm in the 20-F before acting. Bear case. Revenue growth decelerated to +3% YoY — does not support 30x forward. Debt/Assets 42.1% (was ~29%). Free float is thin (28.5M of 46.7M shares), amplifying both directions. Fair value $75–110 · entry <$85.
B5. Field verdict
The complexity-vs-capacity thesis is real and was confirmed with evidence. But it is fully discounted, and then some. This is not a mispricing; it is a correctly-priced-then-overpriced secular story that has taken a ~25% haircut off a ~130% twelve-month melt-up. The premise in the brief — "priced as if levered to capacity" — is false as of today's prices. These stocks are priced as complexity winners.
The trap here is worse than in analog semis because it compounds two problems: peak earnings (TER and FORM each grew revenue ~0–2%/yr for three years, so there is no organic base to fall back to) and HBM single-point-of-failure (the same 2027 memory supply risk our memory screen flagged is the sole driver of the incremental earnings being capitalized at 30–35x).
No gems. Do not buy this field today. The correct output is a watchlist with hard entry triggers, because the secular claim is one of the better ones in semis — it just needs a real de-rating, most likely arriving with a 2027 memory digestion year.
Most worth /analyze: FORM — so an entry plan exists before the de-rating rather than after. Focus on GAAP-vs-non-GAAP earnings quality, the capex cycle suppressing FCF, and HBM revenue concentration by customer.
Data gaps and hazards flagged across the sweep
fin.pyreturns broken EV for foreign ADRs — confirmed on every Japanese ADR (negative EV), and on NTES and NTDOY. All EV/EBITDA and P/S figures for ADRs are unusable; finalists were hand-computed.- yfinance EV also unreliable for TER and FORM — implies net cash of $6.7B and $1.8B against balance sheets showing $294M and $103M. P/B, P/E and P/S were used instead, which reconcile.
- Ticker resolution errors:
SCRYYresolves to SCOR SE (French reinsurer), not Screen Holdings;CPCAYresolves to Cathay Pacific, not Capcom (correct: CCOEY);CAJreturns nothing. Capcom, Square Enix, Konami, Bandai Namco, CD Projekt and Krafton went unscreened. - Access is a first-order constraint in Japan. Roughly half the best assets (Disco, Lasertec, Screen, SUMCO, Keyence) have no tradeable US line. Ibiden — a genuine AI substrate bottleneck at −39.5% off high — trades 314 shares/day at a 5.3% spread. Untradeable. KYOCY's quote is dislocated 7% from last trade.
- Sony's trailing figures are contaminated by the Financial Group spin-off.
- Framework shortfall:
fin.pycarries only ~4 annual periods, so all CAGRs in this sweep are 3-year, not the §1-preferred 5–8yr. roic.ai is on the free plan (2 years) and could not close the gap. This is the main reason no conviction here exceeds 7.0. - GMED share-count figure in the robotics screen is almost certainly a source error — see that file's §5.
- claude.ai connectors Alpha Vantage, FactSet, S&P Global unauthorized this session; no cross-validation available.
Scouts: Semi Test/ATE · Japan Tech & Semis · Robotics & Physical AI · Gaming & Adtech. ~250 names screened. Quantitative data from .mcp/fin.py (Yahoo Finance), roic.ai (FY2024–25) and Public.com live quotes, retrieved 2026-07-29.