Ten people caught the same infection from one lunch. The closest tables weren’t the ones that got it.

In January 2020, ten people across three families were infected in a single lunch at a windowless, air-conditioned restaurant in Guangzhou, China. One diner, recently arrived from Wuhan, sat at a table with four family members. Within two weeks, nine more diners at two nearby tables tested positive. Other diners in the same room — some sitting just as close to the index patient, some closer — never caught it.

The difference wasn’t distance. It was air. Investigators from the Guangzhou Center for Disease Control reconstructed the seating plan and the restaurant’s single wall-mounted split-unit air conditioner, then modelled the room’s airflow. The three affected tables sat in a row directly in the unit’s recirculation path. The AC pulled room air in, cooled it, and blew it back out in a loop that carried exhaled droplets across the two downstream tables — and only those tables. No fresh air was being introduced; the unit only recirculated what was already in the room.

This is the failure mode CFD exists to catch: a space that looks uniform to the eye is rarely uniform in airflow. Two tables sat closer to the index diner than the ones that got sick, tucked just outside the recirculating band, and stayed clean for the whole exposure window. A seating chart tells you proximity. It tells you nothing about which cubic metre of the room the air in front of your face came from thirty seconds ago — or where it’s going next.

Zephyr’s simulation of this layout traces the same mechanism: an aerosol released at the index table follows the AC’s recirculation loop and reaches exactly the two downstream tables the outbreak did, while the two off-axis tables — modelled at their real, closer distance — see none of it. The fix isn’t a bigger air conditioner or more distance between chairs. It’s breaking the loop: one dedicated extract that continuously removes room air, so what leaves a diner’s mouth has somewhere to go besides the next table over. Re-run with that single extract active, the same simulation shows the plume reaching zero of the three tables it previously filled — not because it was diluted, but because it was removed.

The lesson generalises past restaurants. Any enclosed room that recirculates air without a real exhaust path — a windowless meeting room, a basement gym, a waiting area with a wall-mounted split unit — can build the same kind of loop. Ventilation standards are usually written in air changes per hour, a room-average number. This outbreak happened in a room that likely met a reasonable ACH figure on paper. The number that mattered wasn’t the average — it was where, specifically, the recirculated air went, and that only shows up in a flow simulation, not a spec sheet.