Gemini-II Delivers 3-Second Edge LLM Inference at 30 W Versus 100 W GPUs

GSITGSIT

GSI Technology’s Gemini-II processor achieved a 3-second time-to-first-token for multimodal LLM inference at approximately 30 watts in third-party tests, compared to 12 seconds on Qualcomm’s Snapdragon X Elite (30 W) and 3 seconds on NVIDIA’s Jetson Thor (>100 W). This result highlights Gemini-II’s superior power efficiency for edge AI deployments.

1. Q3 Fiscal 2026 Financial Highlights

GSI Technology reported third quarter net revenues of $6.1 million, a 12 percent increase year-over-year driven by sustained demand for its high-performance SRAM solutions. Gross margin narrowed to 52.7 percent from 54.0 percent in the prior-year quarter, reflecting a shift in product mix. Sales to key accounts included $1.1 million to KYEC (17.9 percent of revenues), $675,000 to Nokia (11.1 percent) and $233,000 to Cadence Design Systems (3.8 percent). Military and defense applications accounted for 28.5 percent of shipments, while SigmaQuad products comprised 41.7 percent. The company’s operating loss widened to $6.9 million, primarily due to higher R&D spend of $7.5 million, up from $4.0 million a year earlier, as GSI invested in IP acquisitions and consulting for its next-generation Plato and Gemini-II platforms.

2. Balance Sheet and Cash Flow Strength

GSI ended the quarter with $70.7 million in cash and cash equivalents, up from $13.4 million at the close of fiscal Q4 2025. The increase reflects net proceeds of $46.9 million from a Registered Direct Offering in October 2025. Operating activities used $7.9 million, primarily for continued development and commercialization of Gemini-II and Plato. Investing activities consumed $296,000, while financing activities generated $53.5 million. Working capital expanded to $71.7 million, and stockholders’ equity reached $83.6 million, underscoring a fortified liquidity position to support R&D milestones and potential design-win pursuits in defense and edge deployments.

3. Gemini-II Benchmark and Commercialization Roadmap

In independent third-party testing, GSI’s Gemini-II APU achieved a time-to-first-token of approximately three seconds on a multimodal 12-billion-parameter vision-language model while consuming 30 watts at the subsystem level. This result compares favorably against competitive embedded platforms reporting similar latency at materially higher power budgets. CEO Lee-Lean Shu highlighted that this performance validates Gemini-II’s energy-efficient, low-latency profile for power-constrained edge applications such as autonomous perimeter security using drones. The company has secured a proof-of-concept agreement with G2 Tech and two government agencies, expecting roughly $1 million in funded engagements, and is targeting initial design wins for defense-oriented and select commercial edge AI programs.

4. Q4 Outlook and Strategic Priorities

For fiscal Q4, GSI anticipates net revenues between $5.7 million and $6.5 million with gross margins of approximately 54 percent to 56 percent. Strategic priorities include advancing the Plato hardware development kickoff funded by recent equity raises, expanding the Gemini-II software stack, and leveraging third-party validation to secure early design-win commitments. The company plans to deepen engagements in autonomous and defense markets, capitalize on governmental proof-of-concept funding, and broaden its edge AI footprint through targeted partnerships and system integrations.

Sources

SGG