ARKA-MEM · ANALYTICS
Arka AI Memory Tax
// WEEKLY SNAPSHOT · 2026-06-26 → 2026-07-02 (5 trading days)
A losing week — almost all of the -11.9% move came from our semiconductor tilt, not stock-picking.
The book fell about -11.9% this week as the Semiconductors (SEMI) factor alone subtracted roughly -15.8%, swamping small gains elsewhere. That lines up with the Korea chip-deleveraging shock and Meta's reported move to sell cloud compute, both of which intensified AI-capex overcapacity fears. Stock-picking was a minor part of the story, making up only about 6% of the total move — this was mostly a factor-driven week, not a stock-picking one.
Events this week
- 2026-06-29 South Korea unveiled a major AI/semiconductor industrial strategy anchored by Samsung Electronics and SK Hynix; market read-through was mixed as investors weighed strategic support against future memory-supply-glut risk.[1][2][3]
- 2026-06-29 Asia risk tape was choppy after the prior Korea-led chip deleveraging shock; South Korea's KOSPI pared losses to close only slightly lower as the Korea chip plan and U.S.-Iran truce/talks steadied sentiment.[1][2][3]
- 2026-07-01 Meta reportedly began building a cloud business to sell excess AI compute capacity; this intensified AI capex ROI / overcapacity fears and pressured AI infrastructure and neocloud-exposed names.[1][2][3]
- 2026-07-01 Fed Chair Kevin Warsh, speaking at the ECB forum in Sintra, reiterated commitment to the Fed's 2% inflation target, emphasized independence, and pushed back against giving explicit forward guidance.[1][2][3]
- 2026-07-02 U.S. nonfarm payrolls slowed sharply to 57k jobs in June versus roughly 110k expected; this reduced near-term Fed-hike pressure but added to growth-slowdown concerns.[1][2][3]
- 2026-07-02 U.S. semiconductor selloff intensified: the SOX index fell roughly 5.4% for a second straight down day, with weakness across AI and memory-linked names; Nasdaq underperformed even as the Dow closed at a record.[1][2][3]
- 2026-07-02 Fed's Daly said U.S. policy was “slightly restrictive” and the next move remained uncertain; she highlighted both persistent-inflation risk and the risk that AI investment slows if returns remain unclear.[1][2][3]
- 2026-07-02 Korea/Asia chip deleveraging remained a key global signal: reports showed KOSPI down sharply again, with SK Hynix and Samsung Electronics under heavy pressure as leveraged ETF selling and concerns about large new memory investments amplified volatility.[1][2][3]
- 2026-07-03 Reuters Morning Bid framed the macro tape as “soft jobs, hard choices”: weaker payrolls eased dollar/rate pressure, but chip stocks remained under pressure as investors rotated away from AI-heavy exposure.[1][2][3]
- 2026-07-04 U.S. cash equity markets were closed for the Independence Day holiday; no new U.S. market price signal for the weekly snapshot.[1][2][3]
Weekly factor scorecard
exactSemiconductors (SEMI) subtracted about -15.8% this week; Market (MKT) added about +3.0%; most other factors were small. Stock-picking added about +0.8%. Total book move: about -11.9%.
Factor tilts vs what actually happened
exactWhat our factor tilts predicted for the week (about -12.6%) was close to what actually happened (about -11.9%); stock-picking explained only about 5.6% of the gap.
Semis factor tracked the broader chip selloff
estimateOur Semiconductors (SEMI) factor moved closely with the broader chip-market proxy this week (correlation about 0.95 over 5 trading days), similar to its usual relationship (about 0.94 over the past 20 days) — a short-sample read that may not hold.
Defense barely offset the semis pain
exactAerospace & Defense (DEF) contributed close to 0% for the week while Semiconductors (SEMI) subtracted about -15.8%; on the days semis fell, defense added almost nothing to offset it.
Basket tracked the volatility spike closely
estimateThe basket moved closely with Equity Volatility (VIX) this week (beta about 1.8, correlation about 0.89 over 5 trading days), higher than its usual relationship (beta about 0.8 over the past 20 days) — a short-sample read that may not hold.
AI-capex names: no clean read this week
estimateWe had 0 usable trading days to measure NVDA, AMD, COHR, LITE and WOLF's specific reaction to the Meta cloud-capacity headline this week — no reading available.
Growth factor moved more than usual around Fed comments
estimateOn July 1-2, Growth (GROWTH) factor daily moves were roughly 1.0 to 1.9 times their typical size versus the past 20 days, while Interest Rates (RATE) stayed close to normal — just 2 days, too few to call this a pattern.
How to read this
- Stock-picking
- Name-specific stock moves the factor model doesn't explain — i.e. selection, not the book's factor tilts.
- Factor contribution
- How much each factor exposure (e.g. Market, Semis) added to or subtracted from the week's return.
- Predicted vs Actual
- What the book's factor tilts implied for the week vs what it actually returned; the gap is stock-picking.
- Estimate (short-sample)
- A co-movement read over only ~5 trading days — directional, not statistically reliable, and may not hold.
All event-factor links below are correlations observed in a 5-trading-day window; most are not statistically significant and may reverse. Tier-1 figures are exact accounting identities.
Narrative generated by claude-opus-4-8 (effort high) on 2026-07-05T01:32:15Z — all figures machine-verified against this run’s computed results.