广船国际技术 2025年度报告

Fiscal 2026 Annual Reportand Proxy Statement July 16, 2026 To Our Stockholders: Over the past year, GSI Technology has made meaningful progress in transforming the company from aspecialized memory supplier into an emerging, differentiated edge AI platform company. This transition remains anchored by the strength of our SRAM business. In fiscal 2026, net revenue grewmore than 22% to $25.1 million, while gross margins expanded to 54.5%. We continue to see durable SRAMmarket demand in high-value applications including AI infrastructure, semiconductor testing, defensesystems, and an emerging radiation-hardened memory opportunity. Our higher-density products cancommand premium pricing, strong margins, and potentially recurring long-duration program revenue. Thisbusiness provides both financial stability and strategic relevance as the memory foundation for our AIarchitecture. What has changed most significantly over the past year is our progress in moving the APU platform fromdevelopment into commercialization. A year ago, the central question was whether our compute-in-memory architecture could deliver meaningfuladvantages over conventional AI processors. Today, through live customer proof-of-concepts and early-stage deployments, we believe our offering is increasingly being validated. Gemini-II, which remains in pre-production rather than full commercial production, is now being evaluatedin real-world environments where performance requirements are unforgiving and conventional GPU-basedarchitectures often fail to meet the necessary power and latency thresholds. While revenue from our APUproducts, including Gemini-II, has not been material to date, we believe these early engagements arepositioning the platform for future commercialization. In one active defense deployment for a drone security application, Gemini-II demonstrated sub-3-second time-to-first-token performance while operating at approximately 30 watts — well below competitive GPUsolutions and materially faster than other low-power alternatives. This is an important proof point for thecommercial viability of our architecture, validating its ability to address a growing class of AI workloadswhere speed of response and energy efficiency matter more than maximum throughput. We believe this distinction is becoming increasingly important as AI inference moves from the data center tothe edge. Across defense, smart infrastructure, physical AI, industrial systems, and future autonomous platforms,customers are expressing growing demand for inference solutions that can operate within constrained power,limited footprints, and extreme environments. In these applications, efficiency is emerging as a primarybuying criterion, and this is where GSI is highly differentiated in the field. Our active proof-of-concept programs are helping validate not only the performance of Gemini-II, but also therepeatability of our deployment model. In Taiwan, our Smart City initiative is now in Phase-1 deployment, beginning with 20 cameras, and asuccessful Phase-1 could scale materially into future, larger deployments. The architecture of thisdeployment — where one Gemini-II APU supports multiple camera inputs — creates a potentially attractivemodel for broader rollout and introduces a complementary software and recurring revenue componentalongside silicon sales. At the same time, our growing pipeline of Department of War SBIR programs continues to provide animportant source of non-dilutive funding while accelerating product validation in high-value markets. Theseprograms help offset development costs while placing our technology into mission-critical operationalenvironments where performance and reliability are essential. While these programs have not yet convertedinto follow-on production contracts or material commercial revenue, we believe they position us to pursuethose opportunities as our technology matures. An important area of progress over the past year has also been in software enablement. As with any new computing architecture, the long-term success of our platform depends not only on siliconperformance, but on the quality and accessibility of the software tools that support customer adoption.We continue to make meaningful progress strengthening our compiler stack, improving deployment flexibilityand expanding the range of applications we can support. Just as important, these deployments are helping us establish reusable software and systems frameworks.This reduces future customer onboarding complexity and moves GSI closer to a repeatable commercializationmodel that can eventually scale through system integrator partnerships across multiple vertical markets. Looking ahead, Plato represents the next major step in our roadmap. While Gemini-II has established our position in high-performance edge inference, Plato is being designed tobroaden that opportunity significantly. With a target power envelope of approximately 10 watts, a packagefootprint one-quarter