the hardware-software gap

For each (date, ticker) row in our backtest history, compute the 10-day forward return relative to QQQ. Group tickers into two clusters and compare distributions.

HARDWARE / infrastructure / power (16):

NVDA · MU · AMD · AVGO · TSM · INTC · ARM · ASML · MRVL · SMCI · ANET · NBIS · VRT · DELL · VST · CEG

SOFTWARE / platform / cloud (13):

MSFT · META · GOOGL · PLTR · ADBE · CRM · DDOG · NFLX · SNOW · TTD · CRWV · AMZN · ORCL

Excluded: AAPL (consumer tech), TSLA (EV), MP (materials). Sample: 3,161 row-level forward-return observations from 2025-11-17 through 2026-05-08.

Three things to notice.

The full-period gap is large enough to be meaningful (4.9 percentage points of 10-day forward excess return) and statistically clean (t-statistic above 13). This is not a couple of outlier weeks doing the work.

The gap has widened in the last 90 days. Whatever is driving the bifurcation is not fading; it is getting more pronounced.

Hardware as a cluster is positive. Software as a cluster is slightly negative. This is the kind of asymmetry most market commentary doesn’t capture because it lumps the two together as “AI exposure.”

1. Software CONFIRMED is the worst software setup.

When a software name shows the topicspace state CONFIRMED — meaning price is moving with the narrative — the 10-day forward excess return is −5.07pp. Not neutral. Negative.

By the time price confirms a software narrative, the trade is already over. This contradicts the natural intuition that confirmation is the safer entry. In software, confirmation has been the worst signal in our sample.

Compare to hardware CONFIRMED at +2.25pp. Same state label; opposite direction of return. The validity matrix had already told us this implicitly — Growth Software has zero eligible states — but seeing the magnitude is different.

2. CRWV behaves like hardware, not software.

We classify CRWV as “Growth Software” in our sector map. Its mean 10-day forward excess is +4.83pp — well inside the hardware distribution. CRWV is GPU cloud: it sells AI compute capacity. The classification was wrong; the data corrected it.

GOOGL (+1.11pp) and AMZN (+0.68pp) sit at the boundary as hyperscalers. The clean partition is not “hardware vs. software” — it’s “capacity providers vs. interpretation layer.” CRWV is a capacity provider with software-shaped accounting.

3. The AI-celebrity hardware names are the cluster’s laggards.

Inside the hardware cluster, the names everyone talks about — NVDA, SMCI, VST, CEG — are at the bottom of the distribution.

The contested-narrative hardware names (INTC fighting a turnaround story, MU fighting an HBM-glut narrative, MRVL trying to convince the market about ASIC revenue) deliver. The names whose narratives are uniformly bullish across analyst desks (NVDA, SMCI as the “AI server” poster child, VST and CEG as “AI power”) do not.

The bifurcation is not uniform across topicspace states. It concentrates in the contrarian states — places where narrative and price are misaligned.

The largest gap, by far, is NEG_CONFIRMATION — a setup where price is selling alongside a bearish narrative. In hardware, this has historically resolved upward by +11.6pp on average. In software, it just slightly underperforms.

Same pattern in DISAGREEMENT and DIVERGENCE: hardware names whose narrative and price are fighting each other have produced clean forward returns; software names in the same setup have not.

We don’t make recommendations from this kind of analysis. But three implications fall out of the data.

The validity matrix should explicitly weight what we now know. Software CONFIRMED at −5pp is not just “not eligible” — it is a negative-EV cell. Hardware NEG_CONFIRMATION at +11.6pp is not just “eligible” — it is the highest-EV cell in the matrix. The current eligibility logic captures the direction; it doesn’t capture the magnitude.

Sector classification should reflect return signature, not industry label.CRWV behaves like hardware. The current taxonomy buckets it as software. That mismatch weakens every downstream signal that depends on the bucket.

How a name is being narrated matters more than what sector it is in. Within the same hardware cluster, the contested-narrative names outperformed the consensus-bullish names by 8-10 percentage points. The crowd is the price; the alpha is in what the crowd isn’t yet sure of.

5.7 months is a meaningful sample but not a long one. We don’t know if the gap holds across a different market backdrop — these months covered a particular phase of the AI buildout cycle.

The 10-day horizon was chosen for consistency with the rest of topicspace’s research; a longer horizon could either confirm or wash out the gap.

Cluster definitions are not neutral. Reclassifying CRWV (and possibly GOOGL, AMZN) tightens the gap further; that’s on the work list.

Most importantly: this is one finding from one slice of the data. It changes how topicspace weights setups going forward, but it doesn’t replace per-name reading. Some software names will work; some hardware names will fail. The cluster-level gap is the pattern underneath.

That’s the kind of disagreement worth measuring.