AI Signals — Week 38, Sep 14–18, 2026

The Big Picture

The headline number this week is deceptively calm. Aggregate consensus upside across 24 companies sits at a modest positive, but beneath that surface, the current is running hard in one direction: every model in the panel moved bearish relative to last week. Not four out of five. All five. That kind of unanimity is rare and worth interrogating.

The most natural explanation is not that the models independently revised their fundamental views downward — it is that market prices moved up faster than model targets did. Recall that bias_shift measures the gap between model estimates and current market prices, so a falling bias can reflect a rising denominator as much as a falling numerator. With equity markets continuing their late-summer grind higher, models anchored to DCF fundamentals will mechanically look more cautious each week prices advance without a corresponding earnings catalyst.

Still, the direction of pure estimate revisions — NVDA down -15.2%, AMZN down -8.1%, GOOGL down -3.5%, META down -3.1% — confirms that at least part of the bearish shift is genuine model-level reassessment, concentrated in US mega-cap technology. The models are not just being left behind by a rally; they are actively trimming their views on the sector's most prominent names.

Trends

NVIDIA is the only company with a statistically meaningful directional trend this week: three down-days against one up-day across a 13.2% intra-week range. That range figure is the real story. A swing of that magnitude in a single week, combined with a -15.2% target price cut, tells you the models are genuinely uncertain about where NVDA belongs — not drifting, but actively repricing. The falling trend in consensus, set against a spot price already -30% below that consensus, creates a peculiar situation: the models are bearish on their own prior bullishness. They are cutting targets while still implying significant upside. That is not a contradiction; it is a model in the process of mean-reverting toward a market price it previously dismissed.

Sector Signals

The most interesting rotation this week is what *didn't* happen in technology. With 8 companies, it is the largest sector in the panel, and its aggregate upside slipped from +5.9% to +3.3% — a -2.6 point shift that looks modest until you remember it is dragged down by some of the most heavily-weighted names in global indices. Models are quietly losing conviction on tech even as the sector remains nominally in positive territory.

Consumer staples staged the week's most notable recovery, with the two-company cohort (PG and one other) moving from -20.3% to -15.2% implied upside — a +5.0 point shift driven partly by a +6.4% upward revision to PG's target price. The models are not turning bullish on consumer staples; they are simply becoming less bearish, which in the current environment is its own signal.

Financials deteriorated by -4.3 points to -5.6% aggregate upside, with JPM sitting at -13.3% implied downside. The models see US banking as priced for perfection at current levels. Healthcare, meanwhile, improved to +36.7% aggregate upside — though with only two companies, ORNBV and JNJ, the sector average is dominated by ORNBV's extraordinary +85.6% implied upside, a number that warrants its own scrutiny.

Note that telecom (+24.7%) and materials (-21.9%) each contain a single company — ELISA and UPM respectively — so their sector figures are individual stock readings wearing a sector label.

What the Models Reveal About Themselves

DeepSeek's -4.0 point bias shift is the week's most significant behavioral data point. Its average upside dropped from +4.7% to +0.7% — nearly flat. For context, deepseek's post-July reasoning variant was already measurably less bearish than its predecessor; a sharp single-week move toward neutrality could indicate the model is still calibrating its new persona, or it could simply reflect the market-price effect described above. The honest answer is: we cannot cleanly separate the two without holding prices constant. What we can say is that deepseek's terminal growth assumption is locked at exactly 2.0% with zero standard deviation — a hardcoded prior that no other model replicates. That rigidity is a feature of the reasoning variant's output format, and it means deepseek's valuation dispersion comes entirely from WACC and growth assumptions, not from terminal growth flexibility.

GPT's -6.4% average upside and 0.60 confidence score make it the panel's structural bear and its least confident participant simultaneously. Low confidence paired with negative upside is a coherent stance — the model is saying prices are too high and it is not sure by how much. Claude, by contrast, posts +3.6% upside with 0.72 confidence and the highest cost per thousand valuations at $47.23 — nearly 18x deepseek's $2.62. Whether that premium buys meaningfully better signal remains the central unanswered question of this platform.

Gemini's 0.82 confidence score — highest in the panel — combined with a 6.37% standard deviation in CAGR estimates (also highest) is a paradox worth noting. High confidence, high internal dispersion. The model is sure of itself stock by stock, but those individual certainties do not cohere into a consistent macro view.

Where the Framework Breaks

TIETO sits at +99.6% implied upside with zero dispersion across models. Every model agrees the stock is worth roughly twice its current price, and none of them disagrees. Zero dispersion on a 99.6% upside call is not a sign of model confidence — it is a sign of model anchoring. When five independent systems converge on the same extreme number with no variance, the most likely explanation is that they are all reading from the same analyst consensus target of €38.00 against a spot of €19.04 and reverse-engineering similar DCFs to justify it. The framework is supposed to produce independent valuations; what it may be producing here is a sophisticated consensus-laundering operation. Similarly, TSLA shows zero dispersion on a -56.6% implied downside — the models agree Tesla is dramatically overvalued, but their agreement is suspiciously perfect.

The Model Scorecard