AI Signals — Week 37, Sep 07–11, 2026
- DeepSeek swung from a -0.9% bearish bias last week to +4.7% bullish this week — the largest single-model shift in the panel, a direct consequence of the reasoning-mode provider change that took effect July 31.
- GPT-4 remains the panel's lonely bear at -4.3% average upside, the only model still pricing the market as overvalued while every other model has turned net positive.
- Tesla's consensus target of $156 against a spot price of $363 implies a -57% downside — the most extreme model conviction in the dataset, with zero dispersion meaning every model agrees.
- Industrials staged the week's sharpest sector re-rating, jumping +9.1 percentage points in implied upside, driven by meaningful target-price upgrades to JPMorgan, KONE, and Alphabet.
- Apple at -46.9% implied downside and Tesla at -57.1% with zero dispersion each represent a form of model consensus so complete it should itself be treated as a red flag.
The Big Picture
The panel of five models enters September in an unusual configuration: three bulls, two bears, and a spread between them that tells you more about model architecture than about market fundamentals. The aggregate picture — 24 companies, upside estimates ranging from +100% (TietoEVRY) to -57% (Tesla) — masks a striking bifurcation. Finnish mid-caps are being treated as deeply undervalued; US mega-caps are being treated as dangerously expensive. Whether that reflects genuine analytical insight or a systematic quirk in how these models anchor to analyst consensus targets is the question worth sitting with.
The market-cap-weighted reality is sobering. Strip out the Finnish names, and the US portfolio the models are collectively analyzing is one they would largely sell. AAPL at -46.9% implied downside, TSLA at -57.1%, XOM at -30.9%, JPM at -14.3% — these are not mild underweights. They are structural rejections. Yet the models are not panicking; confidence scores cluster between 0.60 and 0.77, suggesting calibrated skepticism rather than alarm.
Trends
No multi-day consecutive momentum trends were recorded this week. The absence of trends is itself informative: it suggests the models are not building on prior-day conviction but re-evaluating each session largely from scratch. In a week with significant target-price revisions, that reset behavior may be appropriate — or it may indicate that the models lack the memory architecture to compound a view.
Sector Signals
The week's most consequential rotation was in industrials, where implied upside jumped from +3.3% to +12.4% — a +9.1 percentage point shift. With only three companies in the bucket, this is not a broad-based re-rating; it is driven by specific estimate revisions. KONE saw a +19.0% target-price increase, and the sector's improvement is concentrated there and in the adjacent financial upgrades spilling into the classification. Treat this as stock-specific signal amplified by a thin sector sample, not as a call on global industrial capex.
Financials improved by +8.0pp, moving from -9.3% to -1.3% implied upside. The JPM target-price revision of +32.5% — the largest pure estimate revision in the dataset this week — is the primary driver. That is a large single-week move for a company of JPMorgan's analytical coverage depth, and it warrants scrutiny: when models revise a well-covered bank by a third in a week, the revision is more likely reflecting a change in discount-rate assumptions or terminal-growth inputs than new fundamental information.
Energy remains the panel's most structurally bearish sector at -27.0% implied downside, essentially unchanged from last week's -26.7%. The models have a settled, negative view on XOM and the energy complex that is not shifting. Materials (a single-company sector, UPM) deteriorated by -3.7pp. Telecom (single company, ELISA) slipped -2.6pp but remains positive at +29.7%.
Technology, the panel's largest sector at eight companies, barely moved: +5.9% this week versus +6.6% last. Beneath that stability, however, GOOGL received a +18.2% target revision while META was cut -7.4%. The sector average is calm; the underlying model views are not.
What the Models Reveal About Themselves
The dominant story this week is DeepSeek's dramatic bias shift: from -0.9% last week to +4.7% this week, a +5.6pp move that dwarfs every other model's evolution. Readers should understand what is happening here. On July 31, DeepSeek's provider silently switched to a reasoning-mode variant under the same model name. That variant is measurably less bearish — its median valuation gap moved from roughly -8% toward zero. The bias shift visible in this week's data is the continuing normalization of that provider-side change, not a genuine revision of DeepSeek's analytical view on the companies it covers. Pre- and post-July 31 DeepSeek figures span a series break and should not be compared directly.
With DeepSeek now at +4.7%, the panel splits cleanly: Claude (+4.9%), DeepSeek (+4.7%), and Gemini (+4.1%) form a bullish cluster; GPT (-4.3%) and Grok (-2.2%) remain net bearish. GPT's additional -0.6pp drift this week makes it the panel's most consistently pessimistic voice — and notably, it also carries the lowest confidence score at 0.60, tied with Grok. High conviction bears are rare; low-confidence bears are a different animal.
Gemini's +2.8pp shift is genuine model evolution rather than a mechanical artifact, and its 0.77 confidence score — the highest in the panel — suggests it is not hedging. Gemini is becoming more bullish and more certain simultaneously. That combination is worth watching: it can precede either a strong call or an overfit.
On cost efficiency, the spread is stark. DeepSeek at $9.36 per 1,000 valuations delivers its analysis at roughly one-quarter of Claude's $42.23. Gemini at $11.23 offers the best combination of speed (12.7 seconds average latency), cost, and validity rate (99.1%). GPT at $20.26 is the fastest at 4.4 seconds but mid-table on cost. Claude's premium is real and large; whether the marginal analytical quality justifies it remains an open question.
Where the Framework Breaks
TSLA and AAPL both show zero dispersion — every model in the panel produces the same consensus target, or close enough that the spread rounds to nothing. For Tesla, that consensus is $156 against a spot of $363.56, implying -57.1% downside. For Berkshire Hathaway, zero dispersion at +5.1% upside is unremarkable. For Tesla, it is extraordinary.
Zero dispersion on a company as analytically contested as Tesla is not a sign of model wisdom — it is a sign that the models are anchoring to the same external input, almost certainly the analyst consensus target, and not generating independent views. When five architecturally distinct models agree perfectly on a -57% call, the agreement itself is the anomaly. The framework is producing a number, but it is not producing independent thought. That distinction matters enormously for anyone tempted to treat this as five separate bearish signals on Tesla rather than one signal echoed five times.
The Model Scorecard
| Model | Avg Upside | Bias Shift | Cap Rate | Valid % | Cost/1K |
|---|---|---|---|---|---|
| claude | +4.9% | +1.4pp | 11.5% | 99.1% | $42.23 |
| deepseek | +4.7% | +5.6pp* | 11.5% | 92.0% | $9.36 |
| gemini | +4.1% | +2.8pp | 15.0% | 99.1% | $11.23 |
| gpt | -4.3% | -0.6pp | 15.0% | 99.1% | $20.26 |
| grok | -2.2% | +0.8pp | 11.6% | 99.1% | $15.72 |
*DeepSeek bias shift reflects ongoing normalization from the July 31 provider-side reasoning-mode change; not a genuine analytical revision.