The Reliability Layer Behind AI Hardware.Original analysisNot investment advice

Semiconductor manufacturing has always had a hidden enemy: early failure. A device can pass basic electrical tests on the wafer prober and still fail after heat, voltage, current, vibration, or time exposes a defect. Burn-in exists to force those weak devices to fail before they reach the customer. In the past, this sounded like a niche test problem most readers could safely ignore. In 2026, it is becoming infrastructure, because chips are becoming too expensive to fail late.

The more expensive the package becomes, the more valuable it is to kill weak dies early.

This essay is about that shift. Aehr Test Systems is the case study; the real subject is bigger. As AI processors, silicon photonics, silicon carbide devices, and advanced packages grow more expensive and more integrated, reliability screening moves from a back-end testing detail to an economic necessity. The reliability layer becomes a tax on hardware quality. Whoever sits inside that layer gets paid every time someone wants to be sure expensive silicon will survive.

---

1. What burn-in actually means

Burn-in is the deliberate stressing of a semiconductor device under controlled conditions. The goal is to expose early-life failures before shipment. Devices are subjected to elevated temperature, voltage, current, or all three for a sustained period, on the theory that the weak ones will fail in the test chamber rather than in the field.

The reliability framework underneath this is the classic bathtub curve. New devices show a high initial failure rate, which falls quickly into a long stable useful-life period, and then rises again at end-of-life wear-out. Burn-in tries to consume the first part of the curve inside the factory. Pass a device through enough stress and the survivors are statistically more likely to behave themselves in the field.

Burn-in becomes especially valuable when field failure is expensive or dangerous. A silicon-carbide power device that fails inside an EV traction inverter is not just an RMA. A silicon-photonics laser that drifts after deployment is not a refund line item; it is a network event. An AI accelerator inside a USD 30,000-plus package that goes bad after integration is a multi-thousand-dollar problem that ripples through racks, schedules, and customer commitments. The cost of late failure is the part of this story that has changed in the AI era.

---

2. The 2021 Aehr thesis

The starting point for this essay is an uploaded 2021 SemiAnalysis piece on Aehr Test Systems. It is a useful historical anchor, not a final verdict, and the specific framing is worth restating in its own words before grading it.¹

The 2021 thesis was not only about test equipment. It was about moving failure discovery earlier in the manufacturing flow. Read carefully, the argument was a unit-economics argument: when the value of a packaged module is high relative to the value of a single die, screening before assembly is worth more than screening after assembly. No SemiAnalysis text, charts, screenshots, or images are reproduced here. The original piece is treated only as cited historical framing.

---

3. Why wafer-level burn-in matters

Packaged-part burn-in happens after the die has been encapsulated, lead-bonded, sometimes module-assembled, and otherwise made expensive. Wafer-level burn-in happens before any of that. If a weak die fails before packaging, the manufacturer avoids wasting packaging cost on a die that was never going to be usable. If a weak die survives wafer test only to fail after being placed into a module, an advanced package, or a multi-die assembly, the cost is dramatically higher.

The structural problem this solves becomes more acute as multi-die packages grow. A single bad die placed into a package with many good dies does not merely reduce yield by one die. It can degrade the whole assembly, depending on architecture and recovery options. In multi-die advanced packages, one bad die can ruin the economics of many good dies.

The cleanest way to think about this is that wafer-level burn-in is an option on packaging cost. The buyer of the option pays a per-die screening fee at the wafer stage in exchange for the right not to spend packaging cost on dies that would have failed anyway. The fee is fixed and visible. The avoided cost is variable and grows with the value of what would otherwise have been packaged. As the value of the package grows, the option becomes more valuable.

---

4. What changed by 2026

The 2021 thesis was built around SiC and silicon photonics. The 2026 thesis is broader. Aehr’s product line and customer base now point at multiple application categories: SiC power devices, silicon photonics for data-centre optical transceivers, AI ASICs at the package level, advanced packages, memory and flash reliability flows, and the general class of high-value chips where late failure has become uneconomical.

The simplest way to summarise the shift is that Aehr is not just a SiC testing company in 2026. It is trying to be a reliability-screening supplier for high-value semiconductors where late failure is unacceptable. Whether it succeeds is a different question, addressed below in the risks section.

---

5. The AI ASIC signal

The clearest sign that the opportunity has broadened past SiC is the AI ASIC order Aehr disclosed in early 2026. A record USD 41 million follow-on production order from its lead hyperscale customer for package-level burn-in of custom AI processor ASICs.³ The order included Sonoma high-power package-level test and burn-in systems, burn-in modules, and sockets. Aehr framed it as the largest order in company history and tied it to high-volume production burn-in for AI ASICs used in data-centre training and inference.

Read this carefully. It is a company announcement, not an independent measurement, and a single order from one hyperscaler is not the same as a durable market position. The reason it matters is the category. Hyperscaler-class custom AI ASICs are the chips where late failure is most expensive: large packages, long qualification cycles, scarce wafer and packaging slots, and customers whose cost per failed accelerator includes both the silicon and the cluster time it would have served. If burn-in is going to matter anywhere in AI infrastructure, this is one of the places it matters.

The same Aehr quarterly print added the texture. Q3 FY2026 revenue of USD 10.3 million, bookings of USD 37.2 million, a book-to-bill of more than 3.5x, and backlog of USD 38.7 million expanding to USD 50.9 million when later bookings are included. Aehr’s own remarks pointed at interest in wafer-level burn-in for devices going into CoWoS packages with HBM, AI processors, and photonic or electrical transceiver chipsets.² The orders are bigger than the revenue, which is the shape of a company trying to scale into a new category while still small.

---

6. Advanced packaging makes reliability more valuable

The macro reason the AI ASIC signal matters is the trajectory of advanced packaging. TSMC’s 2026 Technology Symposium framing is the cleanest single-source view. TSMC said it is in production with 5.5-reticle-size CoWoS and planning 14-reticle-size CoWoS by 2028, with an expectation that the 14-reticle version integrates approximately ten large compute dies and twenty HBM stacks. TSMC also said its co-packaged optics substrate solution starts production in 2026, with claimed power and latency improvements versus pluggable board-level optics.⁵

Treat the numbers as company claims, not independent benchmarks. The direction is what matters. The future AI package is not one big chip. It is a system of compute dies, memory stacks, photonic or electrical interconnects, substrates, power delivery, and thermal constraints. The more components inside the package, the more expensive late failure becomes.

A bad die is not just a bad die anymore. It can destroy the value of an entire package.

In that world, the value of catching infant mortality before packaging is no longer a back-end optimisation. It is a structural input to package economics. Wafer-level burn-in is one of the few tools that operates upstream of the cost stack, not downstream of it.

---

7. Silicon photonics is the second major pillar

Silicon photonics is the part of the 2021 thesis that aged most cleanly. AI data centres need faster, lower-power, lower-latency interconnects. Copper struggles as bandwidth and distance increase. Optical links become more important as AI clusters scale.

The 2021 article explained the reliability angle in plain terms: silicon-photonics lasers can shift during early life, and burn-in helps stabilise lasers, identify failures, and calibrate systems before final packaging. It connected the opportunity to 400G, 800G, and future co-packaged photonics.¹ The 2026 evidence supports the framing. In March 2026, Aehr announced an initial order from a major new silicon photonics customer in the data-centre optical transceiver market, developing transceivers for hyperscale AI and cloud data centres. The order included FOX-XP wafer-level burn-in, FOX-NP systems, WaferPak Auto Aligner, and WaferPak contactors. Aehr said the high-power FOX-XP configuration can test nine wafers in parallel and deliver up to 3,500 watts per wafer.⁴

The broader market context is consistent. Reuters reported that STMicroelectronics is launching a data-centre photonics chip developed with AWS, and cited a LightCounting market estimate that the relevant market was about USD 7 billion in 2024 with the potential to reach USD 24 billion by 2030.⁷ Treat those numbers as third-party research-firm estimates, not contracts. The shape is the part to keep. AI data centres are making optics strategic. Optical devices need reliability screening. That keeps the photonics pillar of the 2021 thesis alive on its own terms.

---

8. Silicon carbide, still important but more complicated

The 2021 piece anchored on silicon carbide. The 2026 update on SiC is more textured. SiC remains important for EVs, chargers, energy infrastructure, solar inverters, and power electronics. SiC devices face harsh operating environments and serious reliability requirements that map directly onto burn-in. None of that has changed. What changed is the cycle around SiC capacity and end demand.

EV demand softness made SiC cyclical and messy in ways the 2021 article did not anticipate. Wolfspeed, one of the most visible SiC names, filed for Chapter 11 restructuring in 2025 and aimed to reduce debt materially.⁸ Yole still sees long-term SiC device growth toward 2030 but has described a temporary slowdown from EV-market softness.⁹ Read together, the message is that the SiC growth story is intact in shape but bumpier than the 2021 view assumed.

This is precisely why the updated Aehr thesis should not rest only on SiC. SiC validates the burn-in problem and gave the company its commercial track record. AI ASICs and silicon photonics broaden the addressable opportunity. The company that depends on one cyclical end-market is more fragile than the company that sells reliability screening into three uncorrelated end-markets at the same time.

---

9. Wafer-level vs package-level burn-in

The technical distinction the spec asked for is worth making cleanly. Wafer-level burn-in screens devices before packaging. Package-level burn-in screens fully packaged devices or modules. They are not substitutes. They solve different problems at different cost levels in the manufacturing flow.

Aehr’s 2026 position covers both. FOX systems for wafer-level. Sonoma for high-power package-level. WaferPak as the consumable contactor layer underneath the wafer flow. The integration of Incal Technology, which brought the Sonoma capability, is what makes the AI ASIC story credible at all. Without package-level burn-in, the company would not have been the credible counter-party for the USD 41 million order.

---

10. ASML macro context

The macro reason any of this scales sits with ASML. ASML’s 2026 framing has three pieces that matter here. AI is driving stronger semiconductor capacity plans. AI demand for compute density and high-bandwidth memory is growing faster than the overall semiconductor market. AI pushes the industry toward more capacity, more integration, more memory, and more advanced systems.⁶

As hardware becomes more complex and more expensive, the unit value of reliability rises with it. ASML and TSMC describe an industry that is building bigger, more integrated, more expensive systems. Aehr sits in the less glamorous reliability layer that tries to prevent expensive failures inside those systems. The macro shift does not guarantee Aehr wins. It does make the category Aehr operates in matter more than it did five years ago.

In cheap chips, burn-in is a cost. In expensive AI hardware, burn-in becomes insurance.

---

11. The timeline

It is useful to mark the inflection points.

In cheap chips, burn-in is a cost. In expensive AI hardware, burn-in becomes insurance.

---

12. The risks

None of the above implies Aehr is destined to dominate reliability screening, or that the reliability layer is automatically a free lunch as packages get more expensive. A serious thesis carries a serious risk list.

The point is not that Aehr is guaranteed to dominate. The point is that the reliability problem is becoming more valuable as semiconductor packages become more expensive, and Aehr is one of the few public companies sitting inside that change.

---

13. Evidence ledger

Because this essay sits between an uploaded 2021 thesis and a 2026 reading, the load-bearing evidence is gathered here, paired with source and a one-line note on why it matters.

---

14. Final verdict

The 2021 thesis was right that wafer-level burn-in mattered for silicon carbide and silicon photonics. It correctly identified module-level economics as the reason early screening would become more valuable. What it could not yet say, because the world had not caught up to AI hardware yet, was that the same logic would extend to hyperscale AI ASICs and to a future of multi-die packages with ten compute dies and twenty HBM stacks per substrate.

The 2026 thesis is bigger and more honest. As AI hardware becomes more expensive, more optical, more power-dense, and more integrated, reliability screening moves from a back-end testing detail to an economic necessity. Aehr is a small company sitting inside that larger shift. The reliability layer is real. Whether Aehr is the company that consolidates it depends on execution across customer breadth, technology coverage, and competitive response. The category is the part that has been validated. The individual winner is still being decided.

In cheap chips, burn-in is a cost. In expensive AI hardware, burn-in becomes insurance.

---

Sources & footnotes

This essay treats the uploaded 2021 SemiAnalysis article as a cited historical anchor. Specific 2026 claims come from Aehr investor materials and the FY2025 10-K, TSMC 2026 Technology Symposium, ASML investor material, Reuters reporting on STMicroelectronics / AWS photonics and on Wolfspeed, and Yole Group market commentary. Performance, order, and market figures are company or research-firm claims. No third-party charts, screenshots, or logos are reproduced. This is not investment advice.