Fiscal first-quarter 2027 earnings call, November 2, 2026
A new AI processor customer moves from benchmark to production trial
The most consequential disclosure on Aehr's call was not the record quarter. It was a step forward with a major supplier of AI accelerators, CPUs and network processors. This customer completed a benchmark evaluation of the FOX wafer-level burn-in platform last quarter and has now decided to buy a fully configured FOX-XP test cell. The system will be installed at one of its OSAT partners in Taiwan in early calendar 2027 and evaluated in high-volume production on the same device used in the benchmark. A decision on volume production is expected within several months of installation. CEO Gayn Erickson said the customer "represents one of the largest potential burn-in system opportunities Aehr has ever pursued."
Management put numbers on what "largest" could mean. A nine-wafer FOX system configured for AI devices needs roughly 5 to 10 systems per million units shipped annually, assuming perfectly linear shipments. With average selling prices of $6 million to $7 million or more including consumables, that works out to $35 million to $70 million of revenue per million units. Erickson noted that accelerator shipments in 2027 total roughly 25 million units across NVIDIA and AMD GPUs and custom ASICs, and that CPU volumes are about double that. "The numbers are staggering and sobering for us," he said. He was careful not to claim Aehr will win everything: "Now are we going to win them all? Someday maybe, but that's not what we're aiming for. But you pick up 1, 2, 3 high runners, and these are big numbers."
The timeline is not firm, and the company is candid about the work left. Erickson said some features shown during the benchmark were prototypes that now need validation in a volume-production configuration, and that the customer will be seeing certain upgrades for the first time in a high-volume setting. He expects data by roughly January to March, followed by capacity-planning and forecast discussions. The schedule is within about 30 days of original expectations. A statement of work covering throughput and uptime is nearly final. Erickson also flagged the double edge of the customer's prominence: "that imposes upside and risk." A failed evaluation at a customer this visible would be costly.
Die-level temperature control is the technical argument for wafer-level burn-in
The call also gave a clearer view of why Aehr thinks wafer-level burn-in is becoming necessary rather than optional. Multi-die AI packages can combine four accelerator dies, 12 stacks of high-bandwidth memory and several optical or electrical I/O chips, with some announced devices carrying 12 to 16 compute dies. Each die has different leakage, power and self-heating, but they share one substrate and one heat sink. Erickson said it becomes "difficult or actually impossible" to bring every die to the optimum screening temperature at package level, "not ours, not anybody's."
The usual workaround is to regulate temperature based on the hottest die to avoid thermal runaway. That leaves the coldest die under-stressed, so infant-mortality failures pass through to the data center. At wafer level, Aehr can isolate each die. The company is also implementing a patented feature it calls die temperature control, which independently manipulates the thermals of each die under test through a custom-designed WaferPak. Erickson described it as something "people have never even contemplated before." In response to an analyst question, he confirmed it will be part of the production evaluation system going to Taiwan. Management declined to give more detail on the patented features, citing NDAs. The technical case is persuasive, but it remains largely unproven in a customer's production environment.
Record quarter, higher guidance, and a conservative implied bookings requirement
First-quarter revenue was a record $33.0 million, up 200% year over year and 75% sequentially, and ahead of consensus. Non-GAAP net income was $9.4 million, or $0.28 per diluted share, versus $0.2 million a year ago. Bookings were $46.7 million, a book-to-bill of 1.4. Quarter-end backlog rose to $100.9 million from $80.6 million at fiscal year-end despite the heavy shipments. Including $8.4 million of orders received after quarter end, effective backlog was $109.3 million as of November 1.
Aehr raised its fiscal 2027 revenue outlook to $150 million, from a prior range of $130 million to $150 million. That is effectively the top of the old range and about triple fiscal 2026 revenue. Non-GAAP net income is still expected at no less than 20% of revenue. The guidance looks cautious. About $117 million must be delivered over the remaining three quarters, and effective backlog already covers roughly $109 million of it. Lake Street's Max Michaelis noted that this implies only $7 million to $8 million of additional orders. Erickson agreed: "I think we're trying to be conservative," adding that the company is not supply constrained and that $150 million "is a pretty nice clean number."
On cadence, Erickson said customers asked to pull shipments forward from the second quarter into the first, and that each of the next three quarters should grow modestly, with the fourth likely the highest. CFO Chris Siu said the second quarter remains strong, with hyperscale Sonoma deliveries already in backlog.
Margin commentary needs a caveat. The prepared remarks cited a non-GAAP gross margin of 48.5%, up from 37.5% a year ago on volume and mix. A later analyst question referenced 40.5%. Siu said the quarter's level is the right range for the full year, within plus or minus 0.5 percentage point. He also warned the second quarter will likely be lower because of heavy Sonoma shipments to the hyperscale customer. Non-GAAP operating expenses were $7.8 million, up about 36% from $5.7 million, against 200% revenue growth. The extra spending went mainly to AI-related R&D.
Sonoma ramps at a dozen-plus systems a month, with a second processor behind it
In package-level burn-in, Aehr is shipping and installing more than a dozen Sonoma systems per month for its lead hyperscale customer's first AI processor. A second processor from the same customer has passed qualification on Sonoma. It needs more power and more sophisticated thermal management, including multiple independently controlled thermal zones in one package. The customer forecasts it will ultimately ship in higher volumes than the first, with a production ramp anticipated in calendar 2027. Aehr is also evaluating Sonoma with additional AI processor suppliers and hyperscalers, for both data center training and inference chips and high-volume edge inference processors.
Capacity is being built ahead of demand. A contract manufacturer in Southeast Asia began shipping Sonoma systems in the quarter and adds capacity for more than 20 systems per month, with multiple units already delivered to the Taiwan OSAT supporting the hyperscaler. Erickson said subcontractors are taking on more subassembly work, which should cut the labor hours needed to complete a wafer-level system to about one-third of today's level by year-end, allowing roughly three times the output. He is pushing the team to put in place capacity well above $150 million for the following year.
That preparation has a cost. Operating cash flow was negative $3.1 million as Aehr bought long-lead components, and cash and restricted cash fell to $105.1 million from $116.5 million at fiscal year-end. The company remains debt-free, and the working-capital drag looks like the price of readiness rather than a liquidity concern. It does mean Aehr is carrying inventory risk ahead of customer decisions that are not yet made.
Competition: SLT is not the threat, wafer-level FOX is the alternative to Sonoma
Asked about a large test supplier's new "AI burn-in" offering, Erickson dismissed system-level test derivatives as a real substitute. Teradyne's new platform was the likely reference. He argued that SLT machines are built to mimic user conditions at speed, whereas burn-in requires temperatures and voltages well above normal operation to age a device "3 months, not 3 years" within about three hours. "When we're engaged in package-level burn-in deals right now, people are not talking to us about the SLT alternatives," he said. He conceded that some customers build burn-in systems for their own use at lower cost and may accept trade-offs Aehr will not.
He framed the more important competitive dynamic as internal: "the real competition to package level is the FOX system from us." Aehr tests as much as a quarter of a wafer's die in parallel across nine wafers, while the Advantest 93K and Teradyne UltraFlex, which he said account for about 99% of AI test, typically test one device at a time. The only alternative, he said, is a standard prober paired with ATE, which would infringe Aehr's IP, and the company intends to defend it "ferociously." The company has a front-row seat through Sonoma to customers' next-generation devices, and some are telling it they will need FOX by the time those parts arrive.
Consumables mix, silicon carbide, GaN and memory
Consumables, meaning WaferPak contactors and Sonoma burn-in modules and boards, were $13.5 million, or 41% of revenue, versus 24% a year ago. Management expects about 35% for fiscal 2027, up from 29% in fiscal 2026 and above its roughly 30% long-term model. Erickson said the share could reach 50% even in a good year, because customers replace full WaferPak sets every one to three years as new devices launch, and could go higher in a weak systems year. Silicon carbide customers have bought only WaferPaks for about 18 months, and Aehr is now seeing forecasts for more.
Silicon carbide was less than 10% of roughly $50 million of fiscal 2026 revenue, against about 90% a couple of years earlier. A $7 million order has been announced, and management expects growth this year as EV traction inverter, charging and solar utilization is described as nearly full. Data center demand for silicon carbide is not yet at volume. Gallium nitride, which Erickson said is much harder to screen for early-life failures, added two production customers in six months. In silicon photonics, multiple customers placed additional FOX-XP orders.
Memory is the weak spot. NAND flash opportunities are progressing "more slowly than expected" because of limited incremental capacity and the timing of high-bandwidth flash volumes. A potential opening exists in HBM base dies, the logic chips beneath DRAM stacks, which could need wafer-level burn-in before assembly given the cost of the dies stacked above them. Erickson said Aehr has heard of this from several suppliers, but it is a long-term prospect.
Diversification is real, but lumpiness remains
Erickson said Aehr once had a handful of customers capable of large production orders and now sees that potential in more than a dozen, across AI processors, photonics, silicon carbide and GaN, and across both FOX and Sonoma. That reduces dependence on any one program, though he acknowledged that individual orders and ramps "can still vary significantly from quarter to quarter." Near-term revenue is still heavily tied to the lead AI processor customer and the Taiwan-based hyperscaler. The upside case rests on a few large decisions expected in the first half of calendar 2027. Aehr also added Vijay Rao, a hyperscale infrastructure veteran, to its board, and plans to meet investors at the NYC CEO Summit on December 15.