AI Signals — Week 25, Jun 15–19, 2026

The Big Picture

This was a week of quiet but unmistakable capitulation. Not in the market — in the models. Every one of the five AI systems running valuations on this platform shifted its average upside estimate downward over the five-day period. That kind of unanimity is worth pausing on. Individual models drift bearish for idiosyncratic reasons all the time. All five moving in the same direction simultaneously suggests something more systematic: either the underlying fundamental data flowing into these models deteriorated across the board, or the models are picking up on a shared macro signal — stretched valuations, rate sensitivity, margin pressure — and converging on a more cautious read of it.

The aggregate picture is stark. Across 23 companies and 575 valuations, the median model now sees either flat or negative upside. Only TIETO, ORNBV, MSFT, META, KNEBV, GOOGL, ELISA, and BRK-B carry positive consensus targets. The rest — fifteen names — are priced above where the models think they should be. That's not a market call. It's a signal about how AI valuation engines are currently calibrated against current prices.

Trends

The trend data this week tells two very different stories about model conviction.

MSFT logged rising consensus targets on 3 of 4 days, with a weekly range of 10.49% — meaningful movement for a mega-cap. The models are not just bullish on Microsoft; they are consistently, repeatedly bullish. That kind of directional persistence is what separates genuine conviction from noise. When a model revises a target upward three days running, it is incorporating new information each time and arriving at the same conclusion: the stock is undervalued relative to fundamentals.

NVDA is the mirror image. Three consecutive days of falling consensus targets, a 13.16% weekly range, and a spot price of $210.69 against a consensus target of just $176.04. The models are not merely cautious on NVIDIA — they are actively retreating from prior estimates. Combined with a -20.9% target price change on the week (the largest move in the entire dataset), this is the clearest expression of model-level conviction reversal in the current period.

KONE fell for all 4 tracked days, though with a narrow 2.58% range. Persistent but shallow — the models are grinding lower on the Finnish elevator giant without drama. Procter & Gamble rose 3 of 4 days within a tight 1.08% band, suggesting stable, low-volatility bullish drift rather than a re-rating event.

Sector Signals

Healthcare's collapse is the headline rotation story. The sector's model-implied upside fell from +23.3% last week to +12.0% this week — a -11.2 percentage point swing. With only two companies in this sector (JNJ and ORNBV), that move is concentrated and meaningful rather than diversified noise. Orion's target dropped -17% on the week; Johnson & Johnson already sits in negative upside territory. The models appear to be repricing the defensive premium that healthcare commanded last week.

Energy is the lone bright spot in relative terms, improving +5.3 points to -18.9% implied upside. That it remains deeply negative tells you the improvement is from a very low base — XOM and NESTE are still well above model fair value — but the direction of travel is at least constructive.

Financials deteriorated -3.4 points to -9.7% across four companies. With JPM trading at $325 against a consensus target of $254.87, the models are making a pointed statement about bank valuations at current levels. Materials (single company: UPM) slipped further to -20.5%. Technology, the largest sector at eight names, barely moved — -0.5 points to +1.2% — which masks enormous internal dispersion between names like MSFT and NVDA pulling in opposite directions.

What the Models Reveal About Themselves

The model evolution data this week is genuinely unusual. All five models shifted bearish simultaneously, but the magnitude varies in ways that reveal character.

GPT is the outlier. Its bias swung -4.4 points — from +3.2% to -1.2% — in a single week. No other model came close to that velocity of change. GPT also carries the lowest terminal growth rate dispersion (stddev of 0.0), meaning it applies a uniform 2.0% terminal growth assumption across all companies. A model that doesn't differentiate terminal growth by company is more sensitive to changes in near-term cash flow assumptions, which may explain why its aggregate bias can shift so sharply when earnings-related inputs move.

Claude dropped -0.6 points to a near-zero +0.1% average upside — essentially flat on the market. What's notable is Claude's terminal growth standard deviation of 0.43, the only model showing meaningful variation in this parameter. Claude is the only model that appears to genuinely customize long-run growth assumptions by company. That intellectual effort costs money: at $39.79 per thousand valuations, Claude is nearly 18x more expensive than DeepSeek's $2.21. Whether that premium buys better outcomes remains the central unanswered question of this platform.

Grok is the structural bear, sitting at -7.3% average upside with a -1.7 point weekly shift. Its uncapped terminal value deviation of 86.2% — highest of any model — suggests it is generating the most aggressive DCF assumptions even as its price targets remain the most pessimistic. That combination is internally coherent: Grok may be applying high discount rates that overwhelm its growth assumptions.

Gemini stands out for a different reason: its CAGR standard deviation of 15.94 dwarfs every other model (the next highest is DeepSeek at 6.17). Gemini is not just bullish on some stocks and bearish on others — it is wildly differentiated in its growth forecasts. High dispersion can mean genuine analytical nuance or unstable outputs. At $11.14 per thousand, it sits in the middle of the cost range, making it neither the cheap option nor the premium one.

Where the Framework Breaks

TSLA and AAPL both show zero dispersion — every model agrees exactly on the consensus target price, producing $243.92 and $181.38 respectively against spot prices of $400.49 and $298.01. Both carry -0.4 upside scores, the most negative in the dataset.

Zero dispersion is a red flag, not a sign of confidence. When five independent models with different architectures, training data, and cost structures arrive at identical numbers, the most likely explanation is not that they have all independently reached the same correct answer. It is that they are drawing on the same anchoring data — analyst consensus targets, perhaps — and failing to diverge from it. For the two most widely-followed consumer technology stocks in the world, that convergence is precisely where you'd expect the most analytical disagreement. The framework is producing false precision on the names that deserve the most uncertainty.

The Model Scorecard