GSI Technology says its Gemini-II AI processor is now in production and uses 98% less power than NVIDIA GPUs while delivering comparable performance, according to a Cornell University comparison in a retrieval-augmented generation test.
The company is targeting early production by the end of 2027, supported by two proof-of-concept programs backed by the U.S. Defense Department and a Taiwan municipality, and plans to release an alpha version of its AI-assisted software development kit this fall.
With $77 million in cash and no debt, GSI believes it has sufficient resources to meet near-term milestones, though execution on proof-of-concept completion, software rollout, and next-generation chip design will be critical.
What happened
GSI Technology's Vice President of Sales Didier Lasserre said Gemini-II hardware is now in production and that Cornell University found it consumed 98% less power than an NVIDIA GPU in a retrieval-augmented generation test at comparable performance. The company is working toward early production by the end of 2027 while expanding software tools and completing two proof-of-concept programs—one Defense Department-funded drone surveillance project and one Smart City project with a Taiwan municipality.
Why it matters
GSI's compute-in-memory architecture addresses a fundamental inefficiency in conventional processors by performing calculations within on-chip memory rather than moving data repeatedly between processors and external memory, reducing power consumption significantly. This approach is particularly suited to search, high-performance computing, and edge applications where power constraints are critical—a key advantage in markets like defense and autonomous systems where Lasserre said government entities have shown early interest and made financial commitments.
What to watch
GSI has $77 million in cash, no debt, and is burning roughly $4 million per quarter. Key milestones include completion of the drone and Smart City proof-of-concept projects, the software development kit rollout this fall, and the Plato chip's tape-out in spring. The Taiwan Smart City project could expand to between 2,000 and 6,000 cameras by the end of 2027 if Phase III production deployment proceeds.
GSI Technology, a fabless semiconductor company headquartered in Sunnyvale, California, is positioning its Gemini-II artificial intelligence processor as a near-production-ready platform while working toward early production by the end of 2027. Vice President of Sales Didier Lasserre outlined the company's progress at a Canaccord session following the June-quarter earnings report, emphasizing that Gemini-II hardware is already in production and has secured third-party validation and government-backed engagements.
The key technical advantage of Gemini-II is its compute-in-memory architecture, which loads a model or database into on-chip memory and performs calculations directly within the memory bit line, eliminating repeated data movement between processors and external storage—a bottleneck in conventional CPU and GPU designs. A Cornell University comparison of a GSI board with an NVIDIA GPU in a retrieval-augmented generation test found that GSI consumed 98% less power at comparable performance. Lasserre described this approach as particularly suited to search and high-performance computing applications. The company's next-generation product, Plato, is designed to target large language models, vision-language models, and edge applications by increasing data-entry bandwidth, with expected near data-center performance within a power budget of 2 to 10 watts; design completion and tape-out are expected in spring.
GSI is currently executing two announced proof-of-concept programs. The Department of Defense-funded drone surveillance project required a customer specification of time-to-first-token below three seconds and power consumption below 50 watts; in a June laboratory demonstration for the Defense Department, GSI and drone partner G2 Tech achieved 2.5 seconds at 30 watts. The Taiwan Smart City project begins with video analysis from 20 cameras to identify events in real time rather than recording for later review. A potential Phase II would expand to 80 cameras with audio; a potential Phase III production deployment could involve between 2,000 and 6,000 cameras by the end of 2027, with one Gemini-II chip supporting every four cameras alongside an annual recurring software license. Lasserre noted that an application for the drone proof of concept took approximately one person-year to develop before GSI's AI-assisted software development kit; early internal use of the new tool indicates development time could be reduced to weeks. The company plans to release an alpha version of the SDK this fall to select customers, followed by a broader release next year.
Defense interest has translated into concrete funding. GSI has won two Small Business Innovation Research awards with the Air Force Research Laboratory, two awards with the Space Development Agency, and one award with the U.S. Army for a ruggedized edge server concept. One active Space Development Agency award is funding radiation testing of a commercial Gemini-II part for potential space use; testing completed in June showed zero single-event latch-ups, though the company was still awaiting the third-party report as of the time of Lasserre's remarks, with total ionizing dose testing expected to begin at the end of August or early September. The Army-funded ruggedized edge server could support object detection or synthetic aperture radar imagery processing in field environments and may become a future product and revenue source.
Financially, GSI has $77 million in cash and no debt. The company is burning roughly $4 million per quarter, with an expected additional expense of several million dollars in spring related to Plato tape-out. Lasserre said the company believes its cash position is sufficient to meet near-term milestones. Looking ahead, he identified progress in completing the drone and Smart City proof-of-concept projects, the SDK rollout, and Plato's tape-out as key items for investors to monitor, framing the challenge ahead as primarily one of execution rather than technology validation.
GSI Technology is positioning Gemini-II as a power-efficient alternative to conventional AI processors by leveraging a compute-in-memory architecture that moves calculations into on-chip memory rather than shuttling data between separate processors and external storage. This design principle directly addresses a longstanding inefficiency in CPU and GPU designs, and Lasserre's mention of Cornell University validation—98% less power consumption versus NVIDIA at comparable performance—suggests the company has third-party backing for its efficiency claims.
The company's progression from introducing the chip roughly 18 months ago to securing production-ready hardware and government-backed engagements reflects a deliberate path toward market validation. Two proof-of-concept programs illustrate how GSI is targeting near-term defense and smart-city use cases where power constraints and edge deployment are paramount. The drone surveillance project's specification (sub-3-second time-to-first-token at under 50 watts) and GSI's June demonstration (2.5 seconds at 30 watts) suggest the hardware is already close to operational requirements, while the Taiwan Smart City project's phased expansion—from 20 cameras to potentially 2,000–6,000 by end of 2027—points to a concrete revenue pathway if Phase III proceeds.
Software is emerging as a critical enabler. The company's alpha release of an AI-assisted software development kit this fall, preceded by internal tests showing development time dropping from roughly one person-year to weeks, signals that GSI is removing a key barrier to adoption. With $77 million in cash, no debt, and quarterly burn of roughly $4 million, the company's runway appears sufficient to execute on these milestones, though Lasserre's framing of the challenge as "execution" rather than "technology validation" suggests that delivering on proof-of-concept timelines and software tooling will determine whether the 2027 production target is met.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Hygon Information posted a 66.5% increase in first-half 2026 revenue, driven by accelerating AI deployment and…
An analyst projects that Taiwan Semiconductor Manufacturing Company (TSMC) and Broadcom will each reach a $4 t…
Vanguard's July Market Perspectives report argues that as AI productivity gains spread beyond the tech sector…
Airbnb, Booking Holdings, and Expedia have climbed nearly 40% at the median since May 19, while hotel chains C…
AppLovin Corporation (NASDAQ:APP) reported second-quarter revenue of $1.92 billion, missing analyst estimates…
Self-driving laboratory systems—which combine AI agents, simulations, and robotic hardware in a closed loop—ar…
The AI news that matters, in one minute each morning.
Sign up free