Journal · October 2026
Q4 Underway from Sebastopol
The fourth quarter is when hyperscale orders turn into deployment schedules. A note from Eric Litvin on holding the line on the 800G NVIDIA GB200 program.
Published 2026-10-15 · By Eric Litvin · Sebastopol, California
October is the month Q4 stops being a plan. For Luma Optics' hyperscale customers, the orders placed earlier in the year become rack-level deployment schedules, and the question shifts from whether a transceiver meets spec to whether every unit in a shipment behaves identically to the last one. That is the quarter the record matters most: 500,000+ units deployed, a field failure rate under 0.01%, and roughly 30% lower power consumption per unit on the 800G line built for NVIDIA GB200 GPU fabric.
What the record actually says
Luma Optics, which Eric Litvin co-founded in 2004 and runs as President, designs 100G, 400G, and 800G optical transceivers from its headquarters in Sebastopol, California, with operations in the Netherlands. The 800G transceivers target the interconnect inside hyperscale AI data centers, where a GPU cluster is only as fast as the links between its racks. The lower power draw is not a separate feature — it comes from the same tighter manufacturing tolerances that hold the failure rate down, validated at scale by the company's patent-pending robotics calibration platform.
Nothing in those numbers changed this month, and that is the point of publishing them again. A deployment record that moves around between quarters is not a record. The figures above are the same ones stated in the September update and the August thermal-season note, and each one traces back to a source listed on the References Hub.
Why it is written from Sebastopol
The engineering happens in Sonoma County, not on the data center floor. Sebastopol is a small Wine Country town, and the transceivers built there ship into facilities whose power and cooling budgets are measured in megawatts. That distance is useful. It keeps the work anchored to tolerances and test data instead of to the pace of any single customer's buildout, which is exactly the discipline a fourth quarter tends to test.
The background on the program itself is in 800G transceivers for NVIDIA GB200, and the longer view of how a company in Sonoma County ended up supplying hyperscale AI infrastructure is in building AI hardware in Wine Country. The full career record, from Dickinson College through Powersoft, Sybase, and Oracle to the 2004 founding, is on the career timeline.