AI DEMAND THERMOMETER · AS OF Q2 2026 · PROTOTYPE

ZekiRex™ AI Demand Thermometer

Last updated 2026-10-04 21:18 PDT

Q2 2026: money 97 (Overheated) · chips 90 (Overheated) · power 30 (Cooling). Money and chips are overheated; power is the one part not keeping up.

We read the three gauges separately rather than as one number: averaging lets one gauge's extreme be buried by the others (the average, for reference, is 72).

What this thermometer is for

How hot AI demand is right now, split into three gauges on the same 0–100 scale: money (investment), chips (AI chip and cloud revenue) and power (data-center construction).
Which part cools first. Whether all three rise together, or one lags, shows where demand is bottlenecked or cracking.
Not missing the turn in the chip and memory cycle. Cooling signals are recorded every quarter under the same rules, and whether they were right is scored in public.

What it does not do: forecast next quarter, or name stocks or when to buy or sell. Over the past 10 quarters the reading did not call next-quarter memory prices or stocks in advance.

Money big-tech capex

97 ▲

Overheated
prev. qtr 93

Chips AI chip & cloud revenue

90 ▲

Overheated
prev. qtr 70

Power data-center build

30 ▲

Cooling
prev. qtr 10

Efficiency shock
No alert now
Last alert: DeepSeek, Jan 2025

Quarterly money, chips and power (0–100 percentile) · 2024Q1–2026Q2

In 2024 power (data-center construction) ran hottest. From the second half of 2025, money and chips climbed while power cooled: more money committed, slower building. The power series is in real (inflation-adjusted) dollars, so it reads lower than the nominal figure.

Same time axis, 2004–2026 · blue = cool zone, orange = hot zone · bottom: whether semis beat the market

Track record — when each gauge was at an extreme (long history, proxy data)

The AI-specific numbers start in 2023. We rebuilt the same structure from official statistics going back to the 2000s and counted how semiconductor stocks did against the market after each gauge hit an extreme. The clear patterns: semis were strong after money cooled and after chips cooled. But the 11 money-cool quarters are only three cycles (2009, 2020, 2023); chips about nine. "Weak after hot" did not show up; after money ran hot, the beat rate (64%) was about the baseline.

Gauge	Zone	Quarters	Semis (SMH) vs S&P 500 4 qtrs later, mean	Median	Beat rate	Now (pctile)
Money (information-processing equipment investment)	bottom 20% (cool)	11	+23.4%p	+19.0%p	100%	98
Money (information-processing equipment investment)	top 20% (hot)	14	+20.7%p	+4.2%p	64%	98
Chips (semiconductor industrial production)	bottom 20% (cool)	32	+14.3%p	+10.3%p	84%	57
Chips (semiconductor industrial production)	top 20% (hot)	3	+3.4%p	+3.5%p	100%	57
Power (power construction spending)	bottom 20% (cool)	14	+4.8%p	+3.1%p	64%	54
Power (power construction spending)	top 20% (hot)	9	+12.8%p	+5.1%p	89%	54

2004–2026, one-quarter publication lag, percentiles from data known at the time. Baseline (all quarters): mean +11.2%p, median +5.1%p, beat rate 67% — semis beat the market over this period anyway, so what matters is beating the baseline. These zones were chosen after seeing the results, so only results from Q4 2026 on count as evidence. Four-quarter returns overlap, so independent observations are fewer. Not a buy or sell recommendation.

All zone records — from the end of each zone's first quarter, excess return vs the S&P 500

Money

Zone	Period	Value	SMH +2Q	SMH +4Q	Micron +2Q	Micron +4Q
hot	2008Q1	80	-2%p	+5%p	-20%p	+8%p
cool	2009Q1–2009Q4	7	+4%p	+2%p	+69%p	+108%p
hot	2010Q2–2010Q4	98	+4%p	+4%p	-27%p	-40%p
hot	2018Q1–2018Q2	85	-8%p	-5%p	-24%p	-28%p
cool	2019Q4–2020Q3	12	+12%p	+38%p	-0%p	+24%p
hot	2021Q2–2021Q3	99	+7%p	-10%p	-1%p	-23%p
hot	2022Q1–2022Q2	89	-10%p	+7%p	-15%p	-13%p
hot	2022Q4	93	+34%p	+48%p	+10%p	+46%p
cool	2023Q2–2023Q4	15	+8%p	+48%p	+28%p	+86%p
hot	2024Q4	91	+10%p	+32%p	+41%p	+223%p
hot	2025Q2–2026Q3	100	+19%p	+114%p	+121%p	+816%p

Chips

Zone	Period	Value	SMH +2Q	SMH +4Q	Micron +2Q	Micron +4Q
cool	2006Q3–2007Q1	13	-9%p	-3%p	-37%p	-50%p
cool	2009Q1–2009Q4	2	+4%p	+2%p	+69%p	+108%p
hot	2010Q2–2010Q3	93	+4%p	+4%p	-27%p	-40%p
hot	2011Q1	81	-3%p	+3%p	-41%p	-36%p
cool	2012Q1–2012Q2	13	-13%p	-11%p	-28%p	+12%p
cool	2012Q4–2013Q3	14	+4%p	+2%p	+113%p	+213%p
cool	2015Q2–2016Q3	8	-1%p	+3%p	-24%p	-29%p
cool	2018Q3–2019Q4	6	+3%p	+10%p	-6%p	-7%p
cool	2020Q2–2020Q3	15	+22%p	+33%p	+25%p	+26%p
cool	2021Q4–2022Q1	13	-13%p	-15%p	-20%p	-27%p
cool	2023Q2–2023Q4	11	+8%p	+48%p	+28%p	+86%p

Power

Zone	Period	Value	SMH +2Q	SMH +4Q	Micron +2Q	Micron +4Q
hot	2008Q1–2008Q2	95	-2%p	+5%p	-20%p	+8%p
cool	2010Q1–2010Q4	3	+2%p	+11%p	-28%p	-3%p
hot	2012Q2–2012Q3	92	-3%p	+1%p	-4%p	+109%p
cool	2013Q2–2013Q3	10	-3%p	+9%p	+37%p	+108%p
hot	2014Q2	96	+5%p	+5%p	+1%p	-48%p
cool	2015Q1–2015Q2	10	-3%p	+0%p	-38%p	-61%p
cool	2016Q2	19	+19%p	+28%p	+53%p	+102%p
cool	2017Q3–2018Q1	13	+7%p	-2%p	+28%p	-1%p
hot	2019Q4–2020Q1	93	+12%p	+38%p	-0%p	+24%p
cool	2020Q4–2021Q1	12	+6%p	+14%p	-1%p	-3%p
hot	2023Q4–2024Q1	87	+35%p	+15%p	+40%p	-25%p

Average (ref.)

Zone	Period	Value	SMH +2Q	SMH +4Q	Micron +2Q	Micron +4Q
cool	2005Q1	28	+9%p	+2%p	+24%p	+33%p
hot	2008Q1	82	-2%p	+5%p	-20%p	+8%p
cool	2008Q4–2009Q4	14	+20%p	+35%p	+90%p	+277%p
hot	2011Q1	70	-3%p	+3%p	-41%p	-36%p
cool	2012Q1	25	-13%p	-11%p	-28%p	+12%p
cool	2013Q2–2013Q3	22	-3%p	+9%p	+37%p	+108%p
cool	2015Q1–2015Q3	23	-3%p	+0%p	-38%p	-61%p
cool	2016Q1–2016Q3	21	+21%p	+30%p	+65%p	+161%p
cool	2019Q1–2019Q2	18	+7%p	+19%p	-1%p	+11%p
cool	2020Q3–2020Q4	19	+21%p	+19%p	+69%p	+23%p

Correlation — looking at the extremes too

Linear correlations are all near zero: there is no "higher is better" relationship; meaning shows up only at the extremes. So we also measured being in the cool zone, the hot zone, and extremeness. Money's extremeness is the largest (+0.39 to +0.42), but shorten the period and the hot side disappears (+0.00 through 2024) while the cool side stays (+0.33). Chips, power and the average show no signal.

Gauge	Measure	SMH +2Q	SMH +4Q	Micron +2Q	Micron +4Q
Money	Linear	-0.00	-0.03	+0.08	+0.13
	Rank	-0.20	-0.25	-0.17	-0.25
	In cool zone	+0.24	+0.21	+0.12	-0.03
	In hot zone	+0.21	+0.19	+0.24	+0.22
	Extremeness \|v−50\|	+0.42	+0.39	+0.38	+0.27
Chips	Linear	-0.03	-0.09	-0.02	+0.00
	Rank	-0.10	-0.20	-0.13	-0.20
	In cool zone	+0.07	+0.11	+0.05	-0.01
	In hot zone	-0.03	-0.07	-0.09	-0.11
	Extremeness \|v−50\|	-0.10	-0.01	-0.14	-0.13
Power	Linear	+0.07	+0.13	-0.01	+0.02
	Rank	+0.12	+0.20	+0.06	+0.16
	In cool zone	-0.18	-0.24	-0.17	-0.19
	In hot zone	-0.03	-0.05	-0.14	-0.06
	Extremeness \|v−50\|	-0.15	-0.20	-0.23	-0.26
Average (ref.)	Linear	-0.01	-0.01	+0.02	+0.08
	Rank	-0.14	-0.15	-0.15	-0.21
	In cool zone	+0.04	+0.05	+0.09	+0.05
	In hot zone	-0.08	-0.05	-0.14	-0.07
	Extremeness \|v−50\|	+0.09	+0.04	+0.15	+0.08

Threshold: beyond ±0.30 for +2Q and ±0.42 for +4Q to rule out chance (overlapping returns mean fewer independent observations). Blue = positive, orange = negative.

Money — does the result rest on the recent boom alone?

Period	Cool · SMH +4Q	Cool · Micron +4Q	Hot · SMH +4Q	Hot · Micron +4Q
All, 2004–2026	+0.21	-0.03	+0.19	+0.22
Through 2024 (recent boom excluded)	+0.33	+0.05	+0.00	-0.09
Through 2021	+0.32	+0.06	-0.18	-0.22

The numbers in Q2 2026

Gauge	Indicator	Value (YoY)
Money	Big-5 (MSFT·GOOGL·AMZN·META·ORCL) capex growth	+86.5%
Money	Capex ÷ operating cash flow	97.4%
Money	Backlog (RPO) growth, MSFT·AMZN	+108.5%
Chips	NVIDIA data-center revenue growth	+116.6%
Chips	Average growth, AWS · Google Cloud · Azure	+54.0%
Power	US data-center construction spending (real, quarterly sum)	+31.7%

Rules (fixed)

Money, chips and power are each shown on their own 0–100 percentile scale. Their average is given for reference only.
Each indicator uses year-over-year growth, converted to a percentile against a fixed base window, Q1 2024–Q2 2026. The base window will not be changed later.
Efficiency shock (doing the same work with fewer resources) is shown as an alert, not a number.
Bands: overheated 80+ · hot 60–80 · normal 40–60 · cooling 20–40 · cold under 20. Arrows compare with the previous quarter.
Watched alongside: memory stocks' strength relative to the S&P 500. In all four past DRAM cycles it turned at or before the price peak.

Scoring — the average over the last 10 quarters

Over the past 10 quarters, correlations between the reading and next-quarter memory measures were all near zero or weakly negative. Ten points cannot tell signal from chance.

Comparison	Correlation
Reading → next-quarter DRAM contract price change	−0.28
Reading → next-quarter average return, three memory makers	−0.21
Change in reading → next-quarter DRAM contract price	0.08
Change in reading → next-quarter memory stocks	−0.05

How to read this: with 10 data points, a correlation must exceed ±0.63 before it can be told apart from chance (5% significance). All four values are less than half that. The data so far could not confirm a relationship — which is not proof that none exists.

Of 16 quarters of DRAM contract prices, only 2 are actuals; the rest are TrendForce forecast midpoints. The whole period was an upcycle in which demand never cooled. The first public scoring comes after all Q3 2026 results are out (early November), on this page and in the weekend edition.

Limits

This is a prototype. Some figures come from secondary sources (stockanalysis.com, Our World in Data) and are being checked against company releases. Anything that changes will be corrected, and the correction noted.
Power rests on a single indicator.
Much of the backlog growth comes from a few large contracts; demand may lean on single customers.

Sources: company earnings releases · stockanalysis.com quarterly cash-flow statements · US Census Bureau construction spending (via Our World in Data) · TrendForce DRAM contract prices. For information only; not investment advice. Past results do not guarantee future returns.

Get the quarterly scoring — subscribe free