diff --git a/gaze/README.md b/gaze/README.md
index 8143f51..3346530 100644
--- a/gaze/README.md
+++ b/gaze/README.md
@@ -63,7 +63,7 @@ Ground rules, for anyone changing it:
 - **Clean room.** Nothing of Valve's goes in: we don't decompile, disassemble, or patch the `eyetracking` binary or its network weights, and we don't copy their code or weights. Its public output (eye-server.mmap, read-only) is fair game as a baseline and as labels, and so are published papers and openly licensed pupil detectors (check each one's license: PuRe, PuReST, ElSe, and ExCuSe are non-commercial only).
 - **Root only reads.** ft-eyegrab never writes to, stops, or signals the `eyetracking` process, vrserver, or vrcompositor, never opens `/dev/adsp`, `/dev/cdsp`, or `/dev/spidev0.1`, and never writes to `/dev/shm/eye-server.mmap` (it also carries calibration clicks into SteamVR's tracker), `/opt`, or `/persist`.
 - **Eye images are biometric data.** Recordings live outside the repo, in `~/.local/share/frametop/eyes/captures` (0700), and `.gitignore` catches stray frame dumps. They go nowhere but the machine that runs your offline jobs.
-- **Mind the headset's budget.** Finding a pupil takes about 0.4 ms a frame while ft-eyes follows it, and 1.4-2.1 ms when it searches the whole frame. Replays, scoring, and training go to a PC.
+- **Mind the headset's budget.** Finding a pupil takes about 0.4 ms a frame while ft-eyes follows it, and 1.4-2.1 ms when it searches the whole frame. ft-eyes keeps OpenCV and numpy to one thread: their pools of one per core spun idle workers at about a quarter of a core, for frames this small. Replays, scoring, and training go to a PC.
 
 ## Headset fit
 
diff --git a/gaze/tracker/eyes_pupil.py b/gaze/tracker/eyes_pupil.py
index 8884363..cf3672d 100644
--- a/gaze/tracker/eyes_pupil.py
+++ b/gaze/tracker/eyes_pupil.py
@@ -9,7 +9,12 @@ the search runs again with a smaller closing.
 import cv2
 import numpy as np
 
-DARK = 30          # pupil pixels are below this (the face around it is 40-180)
+# One thread: OpenCV's pool of one per core costs more than it saves on a frame this small
+# (a 140-240 px window while it follows the pupil). Its idle workers spun and yielded about
+# 14,000 times a second each, a quarter of a core, beside SteamVR's compositor.
+cv2.setNumThreads(1)
+
+DARK = 30         # pupil pixels are below this (the face around it is 40-180)
 MIN_AREA = 150     # pupil area range in pixels
 MAX_AREA = 20000
 MIN_FILL = 0.75    # blob area / fitted-ellipse area
diff --git a/gaze/tracker/ft-eyes b/gaze/tracker/ft-eyes
index 5e25799..66bbff7 100755
--- a/gaze/tracker/ft-eyes
+++ b/gaze/tracker/ft-eyes
@@ -58,7 +58,12 @@ import time
 from collections import deque
 from pathlib import Path
 
-import numpy as np
+# One thread for numpy's BLAS and OpenMP, set before numpy loads: it would start one per core
+# (8 here) for small arrays that never need them. eyes_pupil keeps OpenCV to one as well.
+for _var in ("OPENBLAS_NUM_THREADS", "OMP_NUM_THREADS"):
+    os.environ.setdefault(_var, "1")
+
+import numpy as np  # noqa: E402
 
 sys.path.insert(0, str(Path(__file__).resolve().parent))
 import eyes_model  # noqa: E402