GSI Technology (GSIT) 2027 會計年度第二季法說會：毛利率壓力與 Plato 進展

GSI Technology (NASDAQ: GSIT) 公布 2027 會計年度第二季營收為 600 萬美元，淨虧損 700 萬美元。製造端的產能限制拖累了出貨與毛利率，而公司仍持續投資於 Gemini-II 商業化、GRINN 軟體平台及其次世代 Plato APU。

重點摘要

• 2027 會計年度第二季營收為 600 萬美元，落在管理層財測區間內，但略低於中位數，主因庫存短缺及製造交期拉長導致部分出貨遞延。
• 毛利率降至 48.5%，低於公司預期的 53% 至 55% 財測區間，主要受到晶圓、委外封裝、預燒（burn-in）與測試成本上升的影響。
• GSI 淨虧損達 700 萬美元，或稀釋後每股虧損 0.18 美元。營業費用增至 1,040 萬美元，其中包括 Plato 開發成本以及非常態性的法律與專業顧問費用。
• 針對 2027 會計年度第三季，管理層預估營收為 500 萬至 600 萬美元，毛利率約為 38% 至 40%。
• 台灣「智慧守護眼」（Smart Guardian Eye）第一期專案的初期階段已於季末完成驗收。該套基於 Gemini-II 的系統正透過約 20 支攝影機提供即時 AI 影像監控與事件回報。
• Plato 仍預計於 2027 年 3 月進行設計定案（tape-out），首批試產晶片目標於 2027 年夏末開始初步送樣。該晶片設計功耗低於 10 瓦即可執行 AI 推論。

重要財務數據

軍事與國防客戶佔第二季出貨量的 46.7%，高於上一季的 31.3% 及去年同期的 28.9%。京元電子 (KYEC) 佔淨營收的 19.7%，高於第一季的 16.1%；而 GSI 在本季未向益華電腦 (Cadence) 出貨。

業務與營運表現

Gemini-II 與「智慧守護眼」

台灣「智慧守護眼」第一期專案已通過驗收的初期階段涵蓋約 20 支攝影機。該平台結合影像判讀、事件偵測與自然語言處理能力，以支援公共安全監控與緊急事件應變。

管理層計劃爭取第二期專案，預計將部署規模擴大至約 80 支攝影機，並加入自動語音辨識功能。後續任何工作仍須視政府公開招標競爭程序以及縣府的最終決定而定。

GSI 打算將已完成的概念驗證與早期部署成果作為參考範例，在類似的監控與智慧基礎設施商機中重複運用其軟體與工程資源。

GRINN 軟體平台

GSI 將其 AI 輔助軟體開發套件命名為 GRINN，即「GSI 神經網路執行階段架構」（GSI Runtime Infrastructure for Neural Networks）的簡稱。管理層表示，GRINN 能在數小時內將來自 Hugging Face 和 GitHub 等平台的現有模型適配至其 APU。

原先需約一人年（person-year）的開發工作現已縮減至不到一個月。部分任務甚至可在不到一週內完成，不過具體時間仍視模型與應用複雜度而異。

該公司預計在 2026 年 12 月底前向首批客戶發布 GRINN，隨後於 2027 年上半年推向商業市場。管理層認為該平台至關重要，能讓客戶與系統整合商無需依賴 GSI 有限的內部工程資源即可自行開發模型。

國防與太空專案

在與 G2 Tech 合作的 Sentinel 專案下，GSI 完成了初期晶片與軟體的交付工作。Gemma-3 12B 模型的實驗室測試顯示，首個 Token 生成時間約為 2.4 秒。由 G2 Tech 主導的現場實測演示可能於 2027 年上半年進行，具體取決於合作夥伴與參與的國防機構。

GSI 也正根據美國陸軍第一期小型企業創新研發（SBIR）獎助計劃，開發一款強化型（ruggedized）Gemini-II 邊緣 AI 節點，預計於 2027 年夏季完成。

針對太空發展署（SDA）的專案，該公司已完成標準 Gemini-II 元件的輻射測試。管理層稱初步的單一事件閂鎖（single-event latch-up）測試結果令人鼓舞，並表示這顯示具備耐輻射能力，而總游離劑量（total ionizing dose）測試結果的最終判讀仍在進行中。

Plato 開發進度

Plato 仍預計於 2027 年 3 月進行設計定案，首批試產晶片目標於 2027 年夏末開始初步送樣。此款 APU 旨在以低於 10 瓦的功耗提供本地端 AI 推論，應用範疇涵蓋電池供電攝影機、小型自主無人機、行動機器人與遠端工業感測器。

管理層表示，潛在策略合作夥伴已對 Plato 展現高度興趣。此類合作夥伴關係可提供硬體與軟體開發資源，並拓展 GSI 目前無法獨立觸及的市場。截至本次電話會議召開時，尚未簽署任何協議。

管理層財測指引

管理層表示，製造成本壓力與產品組合將持續影響第三季毛利率。GSI 計劃立即調漲新訂單價格，並自 1 月 1 日起調漲既有未交貨訂單價格。該公司預期這些舉措加上庫存狀況改善，將有助於自 2027 會計年度第四季起帶動營收與毛利率回升。

與 Plato 相關的外部設計成本預計將持續支出至預定的 2027 年 3 月設計定案為止。單季現金消耗量可能會隨開發里程碑及其他專案進度而有所變動。

風險與關注焦點

• 庫存短缺，以及晶圓、載板、封裝、預燒與測試交期拉長，正導致向客戶的出貨遞延。
• 大型客戶並未持續提早下單建立積壓訂單，加劇了高出貨量產品的不確定性。
• 製造成本壓力預計將持續至 2027 會計年度第三季，進而導致毛利率預估降至 38% 至 40%。
• 「智慧守護眼」第二期專案取決於政府公開招標競爭程序及縣府的最終決標結果。
• Sentinel 現場實測演示取決於 G2 Tech 與參與的國防機構；未來的資金與訂單則有賴後續的專案決策。
• Plato 首批晶片可能需要軟體因應方案（workarounds），管理層表示在進入商業化量產前是否需要再次進行晶片改版（revision）仍不確定。

分析師問答精選

管理層將供應受限歸咎於載板與封裝交期拉長，以及大型客戶提前下單不足等多重因素。小型客戶大致上已提前建立積壓訂單，但高產量產品面臨的風險依然較高。

關於策略合作夥伴關係，管理層表示洽談旨在引導 GSI 進入目前無法單獨拓展的市場。合作夥伴可提供硬體系統與軟體開發支援。該公司指出，協議可能於 2026 年底或 2027 年初敲定，但強調目前尚未簽署任何協議。

管理層將 GRINN 視為推動 Gemini-II 規模化商業應用的關鍵，因為客戶與整合商將能自行開發演算法與模型，而無須依賴 GSI 內部團隊。

針對 Plato，管理層表示各界興趣廣泛，涵蓋電池供電設備、工業電腦、企業伺服器與資料中心。GSI 正在缺乏足夠資源獨立開發應用的較大市場中積極尋求合作夥伴。

法說會電話會議完整逐字稿

完整財報電話會議逐字稿

管理層陳述

Operator

Good day, and welcome to GSI Technology's Second Quarter Fiscal Year 2027 Results Conference Call. [Operator Instructions]

Before we begin today's call, GSI Technology has requested that I read the following safe harbor statement. The matters discussed in this conference call may include forward-looking statements regarding future events and the future performance of GSI Technology that involve risks and uncertainties that could cause actual results to differ materially from those anticipated. These risks and uncertainties are described in the company's Form 10-K filed with the Securities and Exchange Commission.

Additionally, I have also been asked to advise you that this conference call is being recorded today, November 5, 2026, at the request of GSI Technology. Lee-Lean Shu, the company's Chairman, President and Chief Executive Officer, will be hosting the call today.

With him are Douglas Schirle, Chief Financial Officer; and Didier Lasserre, Vice President of Sales.

I would now like to turn the conference over to Mr. Shu. Please go ahead, sir.

Lee-Lean Shu

Thank you, operator, and thank you all for joining us today. In the second quarter of fiscal 2027, we made progress on our projects and plan to bring our APU technology to market. We recently announced that the initial stage of our Smart Guardian Eye Phase I project in Taiwan was accepted. The system is now providing live AI video monitoring and incident reporting, confirming Gemini-II can perform in real-world applications. We plan to leverage these projects into repeatable business and revenue.

In addition to securing design wins for Gemini-II, we are advancing 2 important initiatives to grow our APU business: our AI-assisted software development kit, or SDK, and the Plato, our next-generation APU. Our team continues to test and improve our AI-assisted SDK, and I will update you on that progress.

Plato, our next-generation APU, remains on schedule for tape-out in March 2027. We expect to have the first-spin chips ready for initial sampling by late summer of calendar 2027. Plato is designed to run AI inference at less than 10 watts of power. We believe its low-power capability will open new market opportunities for GSI. We are also seeing interest from potential strategic partners who could help us to expand our capabilities to enter new markets.

For the remainder of fiscal 2027, we have 3 main goals. We plan to pursue Phase II of the Smart Guardian Eye project, finalize our AI-assisted SDK for commercial release in the first half of calendar 2027, and complete Plato's tape-out by seeking strategic partners as we prepare to bring it to market.

Didier will now provide more detail on our business and technology progress, followed by Doug, who will review our financial results and outlook. Afterwards, we will be happy to answer your questions. Didier over to you.

Didier Lasserre

Thank you, Lee-Lean. As we discussed on our last call, our Gemini-II commercialization strategy is to focus our resources on a small number of proof of concepts and early deployment programs. Our goal is to turn these projects into reference applications and apply the software and engineering work to similar opportunities. This approach will allow us to pursue new applications more efficiently without significantly expanding our internal engineering resources.

I'll begin with Smart Guardian Eye in Hsinchu County, Taiwan, where shortly after the quarter end in early October, we achieved our first most visible commercial milestone to date. As Lee-Lean mentioned, the county accepted the initial stage of Phase I, and the system is now providing live AI video monitoring and incident reporting.

As a reminder, the Phase I program covers approximately 20 cameras and demonstrates that Gemini-II can support real-world applications where timely detection and response are crucial to public safety.

Our next opportunity in this project will be Phase II, which is expected to expand the deployment to approximately 80 cameras and introduce automatic speech recognition. We are preparing to submit a proposal once the government issues its quote request. Any additional work we receive on this project will depend on the competitive procurement process and the county's final decision.

The Smart Guardian Eye platform combines image interpretation, event detection, and natural language capabilities that identify relevant activity and help operators respond more rapidly. We are pursuing opportunities to apply these capabilities to other surveillance and smart infrastructure applications, leveraging the software and integration work developed for this project.

This is the model we intend to use: develop a solution for a specific application, demonstrate its performance in the field, and reuse the technology across future deployments.

Let me switch gears now to talk about the AI-assisted SDK that we are now calling GRINN. GRINN. GRINN which stands for GSI Runtime Infrastructure for Neural Networks, is expected to play an important role in our Gemini-II commercialization strategy.

Our software engineers are now using GRINN for model and algorithm development. They can take existing models from open source platforms such as Hugging Face and GitHub and adapt them to run on the APU within hours. GRINN is designed to support a broad range of neural networks, significantly reducing engineering efforts previously required to bring models onto our architecture.

On our last call, we reported that GRINN had reduced development work previously requiring approximately 1 man-year to less than a month. Since then, we have made further improvements, with some development tasks now completed in less than a week. While development times vary depending on the complexity of the model application, these improvements demonstrate GRINN's growing capabilities.

Beyond reducing development time, GRINN has developed an ability to understand the APU architecture and apply that knowledge to optimize how AI models are implemented on our chip. We believe that this hardware-aware approach is an important differentiator from conventional software development tools.

By simplifying model adaptation and reducing engineering effort, GRINN should make it easier and faster for customers and system integrators to evaluate Gemini-II, while allowing GSI to support more simultaneous evaluations. Our development team is incorporating feedback from internal users to further refine GRINN's performance. We expect to release it to initial customers before the end of December 2026, followed by a commercial release in the first half of calendar '27.

Turning to our Sentinel program with G2 Tech. We have completed our chip and software deliverables for the early phase of the project. Laboratory testing demonstrated a fast first to -- I'm sorry, time to first token of approximately 2.4 seconds for the Gemma-3 12B model, meeting the low latency and power efficiency requirements of this drone and perimeter security applications.

The next step is a field demonstration under G2 Tech's leadership. The timing depends on our partner and the participating defense agencies. We currently anticipate that this demonstration could take place in the first half of calendar 2027. A successful demonstration would represent another important validation of Gemini-II, along further funding and commercial orders will depend on subsequent program decisions.

We also continue to advance our government-funded Small Business Innovation Research, or SBIR programs. Under the U.S. Army Phase I SBIR award, we are developing a ruggedized Gemini-II-based Edge AI node designed for demanding military environments. This program is valuable because government funding supports the development of the hardware and software that could potentially be adapted by us for other defense applications.

It also allows us to demonstrate the APU's capabilities and systems that operate reliably under difficult field conditions. We expect this project to be completed in summer 2027. In our work with the Space Development Agency, we completed radiation testing on a standard off-the-shelf Gemini-II device. We received and reviewed the single event latch-up test results and are awaiting the final interpretation of the total ionizing dose results.

The preliminary findings based on the single event latch-up test results are very encouraging and indicate that Gemini-II is radiation tolerant. This implies that an off-the-shelf Gemini-II chip can be suitable for certain space and harsh environment applications. We look forward to sharing the information once the final reports are available.

As Lee mentioned, Plato remains on schedule for tape-out in March 2027. We are targeting our first-spin chips for initial sampling by late summer of calendar '27. Plato is designed to deliver AI inference at less than 10 watts, making it suitable for AI processing that takes place locally with limited power, cooling, or access to electricity.

Examples would include battery-powered cameras, small autonomous drones, mobile robots, and remote industrial sensors. To put that into perspective, Plato utilizes about the same amount of power as an LED bulb in a desk lamp. We believe this combination of low power consumption and fast response times would open up a large market for GSI in autonomous systems, robotics, intelligent cameras, and other edge AI applications.

As we experienced with Gemini-II, initial silicon may be suitable for evaluation with software workarounds while additional testing and refinement continue. However, it remains to be seen whether a subsequent chip revision will be necessary before commercial production.

Plato's preliminary performance and power efficiency targets are also attracting interest from potential strategic partners. Partnerships are an important part of our commercialization strategy for Plato. They could provide the resources, software expertise, and market access to pursue new opportunities that we cannot address independently today.

In summary, our recent progress provides further evidence that Gemini-II can perform in real-world customer applications. Going forward, our focus remains on completing existing programs, turning successful projects into additional commercial opportunities, and delivering GRINN to support a broader ecosystem for customers and system integrators. At the same time, we are keeping Plato's development program on schedule and pursuing strategic partners to support its commercialization.

With that, I'll turn it over to Doug. Doug, go ahead, please.

Douglas Schirle

Thank you, Didier. In my comments, I will provide additional context around the second quarter financial results included in the earnings release we issued today after the market close. Revenue for the second quarter of fiscal 2027 was $6 million, within our guidance range but slightly below the midpoint.

During the quarter, some inventory levels were insufficient to fulfill certain customer orders, and extended manufacturing lead times made it more difficult to replenish inventory. As a result, some customer shipments have been delayed, which is also affecting our revenue outlook for the third quarter.

Overall, our SRAM revenue has remained relatively stable over the past 6 quarters, although customer ordering patterns have become more variable in this fiscal year. As we have discussed previously, these fluctuations in customer order patterns can affect quarterly revenue and gross margin.

For example, we had no shipments to Cadence during the second quarter, while sales to KYEC represented 19.7% of net revenues compared to 16.1% in the preceding quarter. Despite the variability in quarterly orders, we continue to anticipate shipments to our customers supporting AI chip design and semiconductor manufacturing.

Shipments to military and defense customers accounted for 46.7% of total shipments in the second quarter, compared with 28.9% in the year-ago quarter and 31.3% in the preceding quarter. This proportion is consistent with historical levels before the recent increase in sales to customers supporting AI chip design and semiconductor manufacturing.

Second quarter gross margin was 48.5%, below our guidance range of 53% to 55%, primarily reflecting higher manufacturing costs related to constraints affecting wafers and outsourced assembly, burn-in, and testing. These cost pressures are continuing into the third quarter. Together with the anticipated product mix, they are affecting our gross margin outlook.

As stated in today's earnings release, we currently anticipate third quarter fiscal 2027 revenue of $5 million to $6 million with gross margin of approximately 38% to 40%. To address higher manufacturing costs, we plan to implement customer price increases immediately on new orders and effective January 1 on existing backlog. We're also working with our manufacturing partners to improve inventory availability, while the replenishment will take time given current manufacturing lead times. We anticipate that these measures will support a recovery in revenue and gross margin beginning in the fourth quarter of fiscal 2027.

Turning to operating expenses. Total operating expenses increased to $10.4 million compared with $6.7 million in the year-ago quarter and $8.8 million in the preceding quarter. Two factors contributed to this increase. The year-over-year increase in R&D this quarter was primarily driven by external design services related to Plato.

As I've stated on our prior earnings calls, we expect these development costs to continue through the planned tape-out in March 2027. The increase in R&D expenses was partially offset by higher government funding under our SBIR programs. These offsets totaled $556,000 during the quarter compared with $279,000 a year ago and $336,000 in the preceding quarter.

As a reminder, SBIR funding offsets certain development expenses and varies quarterly based on project milestones and the timing of reimbursements. Our software development costs can also be affected by fluctuations in the Israeli shekel relative to the U.S. dollar.

The second factor affecting operating expenses this quarter was an increase in selling, general and administrative expense, primarily related to approximately $2 million of non-routine legal and professional fees associated with an internal assessment of certain vendor matters within our foreign operations. These fees were partially offset by a $600,000 gain from assets received in the vendor settlement, resulting in a net impact of approximately $1.4 million. The internal assessment was concluded during the quarter, and we do not expect significant additional expenses related to these matters.

Further information is included in our earnings release and SEC filings. Operating loss for the second quarter was $7.4 million compared to $5.4 million in the preceding quarter. Our net loss was $7 million, or $0.18 per diluted share.

Turning to the balance sheet, our cash and our cash flow. We ended the quarter with $69.2 million in cash and cash equivalents compared with $77 million at June 30, 2026. Net cash used in operating activities was approximately $8 million during the quarter, reflecting spending on Plato development, Gemini-II commercialization, and the internal assessment discussed earlier. You can find the detail of the cash flow in today's earnings release.

Our operating cash usage this quarter exceeded our previously discussed expectation of approximately $4.0 million per quarter, primarily for the reasons I just reviewed. Despite the higher cash usage this quarter and based on our current operating plan, we continue to anticipate quarterly cash usage of approximately $4 million, or $16 million annually, as we complete Gemini-II software development and advance Plato toward tape-out. Actual cash usage may vary depending on the timing of development milestones and other program activities.

We ended the quarter with no debt and a cash balance of $69.2 million, above the $67.2 million we reported as of the end of fiscal 2026. Our debt-free balance sheet provides the financial resources to continue our planned investments in Gemini-II commercialization, including software development and ongoing development of Plato, while maintaining a disciplined approach to capital allocation. This concludes our prepared remarks.

Operator, at this point, please open the call to Q&A.

Operator

[Operator Instructions] Our first question will come from David Williams of Needham.

分析師問答

David Williams

I guess maybe first, maybe around the constraints that you mentioned and maybe some of the order patterns. And if you can maybe give us a little more color around where you're seeing the constraints and the magnitude of those and where you think you might have an opportunity to overcome some of that in the next couple of quarters?

Douglas Schirle

Well, we have issues with customers, particularly the larger customers, that as much as we tell them that they need to place backlog given the manufacturing constraints that we're seeing and lead times we're seeing, that hasn't happened yet. A lot of our smaller customers have placed backlog, and that's probably okay. But the real issue has been with the larger customers with the higher volume products.

Didier Lasserre

Yes. And that also is coupled with the extension on the lead times, substrate lead times, assembly lead times have all gone out as well. So it's between those 2, it's affected -- that's the constraint.

David Williams

Okay. Great. And then maybe, Didier, I think you had mentioned the strategic partners in your script. And just wonder if you can elaborate a bit more on those strategic partners, what that could look like and maybe when you expect to be able to announce something more solid there.

Didier Lasserre

That's a good question. We can't discuss anything yet because nothing is signed at this point. We are having discussions and as we discussed, it would be to bring us into a market or markets that we are currently not addressing ourselves.

And so it's certainly interesting for us for that to have somebody else do the development, both on the hardware system side along with the software to support those. And so we've been negotiating for some time now. There might be something by the end of this year or into early next year.

David Williams

Okay. Great. And then maybe just any color on the SDK, like the GRINN, I believe, is what you've named that. But maybe just talk about that expectation. Obviously, it's doing better today. You said 1 week versus a month in terms of how quickly that's accelerating. But it seems like that platform opens a lot of opportunities and really gets you close to that commercialization. And one, I guess, is that fair? And then what maybe are you most encouraged by from -- that you've seen from GRINN so far?

Didier Lasserre

Yes, absolutely. We need the software tools. Up until now, we've been completely reliant on our Israeli team for all of the models and algorithm development. And as you guys know, it's -- we have limited resources within the company. And so in that respect, it limits what we can do. And so what we're doing now, and we kind of discussed this, is to try and different POCs.

We've done a drone. We're in a smart city. We're looking at other POCs. And that shows that we're viable for a certain market. But then in order to attract customers and integrators to really address all the customers within those markets, they're going to need the tools.

And so with GRINN, it's going to allow them to develop their own algorithms and models without us helping them. And because, again, we just can't scale that way the way we're doing it now. So GRINN is going to be critical for us to be able to build the ecosystem for all the customers along the markets.

David Williams

Great. Fantastic. And then maybe just lastly, you talked a little bit about seeing some interest in Plato. Can you talk a little bit about the different platforms and where you're seeing the interest between Gemini and Plato and maybe just the size of the magnitude of that interest today? And that's all for me.

Lee-Lean Shu

Yes. We -- we are seeing all kind of use cases for Plato from the battery operator, operation to industrial PC to enterprise server, even to the data center. So we can almost see the whole spectrum of the AI application with Plato. That's why we are talking about we are talking to strategic partners to take part in some of those larger markets because we don't have that much resource to go into each market to develop the application for that, yes. So pretty much that's what we are seeing right now.

Operator

[Operator Instructions] This concludes our question-and-answer session. I'd like to turn the call over to Lee-Lean Shu for any closing remarks.

Lee-Lean Shu

Our goals for the coming quarter -- upcoming quarter are straightforward: execute on the customer programs now underway, continue advancing GRINN development, and deliver the next milestones in the road map while pursuing the strategic partnership that can support its commercialization. We appreciate the continued support of our shareholders and look forward to reporting on our progress.

One final note, I'd like to remind listeners that Didier will attend the Roth Technology Conference on November 18. Please contact your ROTH representative to book a meeting. We are also participating in the annual Needham Growth Conference in January 2027, and we encourage you to reach out to your Needham representative to book one-on-one meetings. Thank you again for joining us today and for your continued interest in GSI Technology.

Operator

The conference has now concluded. Thank you for attending today's presentation, and you may now disconnect.