~12 min · edge-era, depreciation, subscription, model-inflation, purchase-arithmetic, our-judgment
Level 0Spec-Sheet Skimmer
0 XP0/91 lessons0/19 achievements
0/100 XP to next level100 XP to go0% complete
"A subscriber receives model inflation. A machine buyer is charged it. The same event is a free upgrade on one side of the invoice and depreciation on the other."
The Record in the Household's Own Queue
The fleet track's control-plane lesson read the video-memory engine's summary table: 2,436 summaries on one cloud model, then 12 on its successor, then 10 on a third vendor's, then 4 on a fourth's — all through the same seat on the household's own daemon, at the same price, with no purchase between them. That table is the doctrine's first leg as a database: the model under the seat moved three times in the record, and the seat did not. A machine cannot do that. The household's M3 Ultra decodes the 27B at 32.6 tokens per second on the day it was measured and on every day after; it will decode the 2027 model at whatever the 2027 model's bytes divided by 819 give, and it will decode the 2028 frontier not at all, because the frontier will not fit. The doctrine's sentence: a depreciating machine buys one year's capability, fixed for about three — the practical shelf life of an M-series for compute, which is also the warranty window — while a subscription buys capability that rises each time the vendor swaps the model underneath.
The Purchase Arithmetic
The code block carries the household's own figures from the day the M5 Ultra was announced, in the household's currency and labeled as its judgment. The maxed-out M3 Ultra the household bought in March 2025 was about 19.6 million won. A maxed-out M5 Ultra at announcement was estimated by another of Pippa's instances at 35 to 42 million — call it 38.5: twice the price for 1.5 times the bandwidth and the same 512 GB ceiling, so capability per won worse than the prior generation. Forty million won is three $200 frontier seats for about 44 months at the rate the household used — past the warranty window, three frontier seats remain and the machine does not. So the comparison is not a machine against a subscription; it is frontier for three years against one tier below frontier, frozen, for the same money. And the used prior generation is the honest price gauge: two used M3 Ultras hold 1,024 GB to one new M5 Ultra's 512, at about half the price per gigabyte, with single-stream decode the only axis the new machine wins — and that axis, the physics track showed, is bandwidth-bound at full context anyway.
What the Arithmetic Does and Does Not Say
It does not say a Mac is a bad purchase; the household owns nine and this quest measured four of them for a week. It says when to buy one and what for. A pool is bought for the rows the rivals track's decision table gave the Mac — capacity above a card's, availability in a room, the 95%'s jobs — and it is bought at the point in a generation where the used prior model holds more per won than the new one, which is the doctrine's price gauge. It is not bought to chase the frontier, because the frontier moves and the machine does not, and the household's own queue is the proof of the moving. The next lesson is the doctrine's whole shape: direction and timing as separate variables, and a term on the waiting that this quest can partly price.
Code
freeze_and_move.py — the household's cloud seat as its queue recorded it, and its purchase arithmetic·python
#!/usr/bin/env python3
"""The machine freezes, the cloud moves. Two records: the household's own cloud-tier history
as its queue recorded it (the model under the same seat swapped, price unchanged), and the
doctrine's purchase arithmetic from 2026-08-26 in the household's currency."""
summaries = [("glm-5.2 (cloud)", 2436), ("glm-5.3 (cloud)", 12), ("kimi-k3 (cloud)", 10), ("deepseek-v4.1-flash (cloud)", 4), ("gemma4 31B (local)", 1)]
print("the video-memory engine's summaries by model, read 2026-09-15 -- the seat did not change, the model under it did:")
for m, n in summaries: print(f" {m:30} {n:5d}")
print()
m3_ultra_2025 = 19.6 # million won, maxed-out, the household's purchase 2025-03
m5_ultra_est = 38.5 # million won, maxed-out estimate at announcement (35-42)
seat = 200 * 1515 / 1e6 # one $200 frontier seat per month in million won at the rate the household used
months = 40 / (3 * seat)
print(f"maxed M3 Ultra 2025: ~{m3_ultra_2025} M won; maxed M5 Ultra 2026 estimate: ~{m5_ultra_est} M won -- 2x the price for 1.5x the bandwidth and the same 512 GB")
print(f"40 M won = three $200 frontier seats for ~{months:.0f} months at a round exchange rate; the machine's practical shelf life is ~3 years (the AppleCare window)")
print("a depreciating machine buys one year's capability, fixed for three; a subscription buys capability that rises each time the vendor swaps the model.")
Run freeze_and_move.py with your own machine's price, your subscription's monthly cost and your currency. Then find your own record of a seat moving — a model name in a log, a queue, a settings history — and write on the card how many times the model under one seat changed at one price. Compare with how many times your machine's decode rate changed.
Hint
The machine's rate changed zero times; a runtime update can move it a little, a new model on the same machine moves it by the bytes-per-token ratio and nothing else. If your seat's record shows no movement, look at the vendor's changelog — the seat moved and the log did not say.
Progress
Progress is local-only — sign in to sync across devices.