Journal · July 2026
Mid-Year Update from Sebastopol
A mid-year note from Eric Litvin, Co-Founder and President of Luma Optics, on where the 800G NVIDIA GB200 rollout stands and the discipline behind it.
Published 2026-07-01 · By Eric Litvin · Sebastopol, California
Halfway through 2026, the number that matters most hasn't moved: field failure rate under 0.01% across more than 500,000 optical transceivers shipped for NVIDIA GB200 AI fabric. That consistency, not any single new announcement, is the mid-year story out of Sebastopol.
The 800G transceiver line continues to ship into GPU fabric at scale, and the patent-pending robotics calibration platform keeps doing the unglamorous work of catching variance before units leave the facility. Coming off the 2025 Inc. 5000 recognition (No. 47, 5,757% three-year growth), the mid-year priority has been holding that reliability curve flat as volume climbs — not chasing a bigger headline.
Litvin has described that kind of flat, unglamorous consistency before in terms borrowed from endurance sport: pacing, calibration, recovery under load. It's the same framing that shows up on his UltraRunning.com athlete calendar — a mid-year check-in is closer to mile forty of a training block than a finish line, and the work is to keep the pace, not spike it.
Full background on Luma Optics' 800G program is in the NVIDIA GB200 deployment article; the complete verified record is on the References Hub.
The baseline the rest of the year is measured against
This entry is the checkpoint the later notes refer back to. At mid-year, Luma Optics — co-founded by Eric Litvin in 2004 and run from Sebastopol, California — was shipping its 100G, 400G, and 800G optical transceivers against one standard: the reliability curve stays flat as volume climbs. The 2025 Inc. 5000 ranking, No. 47 with 5,757% three-year growth, made the volume side public. The field failure rate under 0.01% across 500,000+ units is what makes that volume defensible.
The months after this one test that claim rather than extend it. August asks whether the roughly 30% lower power consumption per unit still holds when cooling systems are under peak thermal load. September adds two outside verifications, the Crunchbase profile for the company and the UltraRunning.com calendar for the person. October is where the orders turn into deployment schedules. None of them moves the number stated here.