AI Signals — Week 22, May 25–29, 2026

The Big Picture

Something unusual happened this week: all five models moved in the same direction. Every member of the panel grew more bearish, and not by trivial amounts. The aggregate shift tells a story about how AI valuation engines respond to a market that has run hard — and perhaps too far.

The panel's average upside estimate sits at roughly -4.1% when you weight models equally, against a universe of 23 companies across Finnish and US markets. That is not a screaming sell signal — these models are not built to time markets — but the unanimity of the directional shift is the real signal. When five architecturally distinct models, trained on different data and using different reasoning strategies, all tighten their discount rates or trim their terminal assumptions in the same week, you are probably looking at a genuine repricing of risk rather than model noise.

The market, in short, has gotten ahead of what the models think the fundamentals justify. Whether that matters for price action is a separate question. What it tells us about the models is more interesting.

Trends

Microsoft (MSFT) is the week's most compelling momentum story from a model-conviction standpoint. Consensus target prices rose on 4 of 4 trading days, covering a range of 9.49% — the widest trending range in the universe this week. That is not a model drifting; that is a model updating with conviction. The direction is bullish, the consistency is high, and the week-on-week target price increase of +4.6% confirms the signal. Models are becoming more comfortable with Microsoft's AI monetization story as evidence accumulates.

Nokia (NOKIA) tells the opposite tale. Three up-days in the trend data might suggest improving sentiment, but the range is a modest 3.7% and the target price was cut -6.8% over the week. The trend data here reflects intra-week volatility in model outputs, not a genuine bullish revision. With consensus implying -47% downside from spot, Nokia remains the universe's most extreme valuation outlier — a company the models collectively believe is priced for a future that does not exist.

Wärtsilä (WRT1V) fell on 3 of 4 days, with a range of 6.17%. The models are growing steadily more skeptical of the industrial marine and energy transition story at current prices.

Sector Signals

The most important rotation this week is the collapse in technology sentiment. The sector's average model upside fell from -1.1% to -4.2%, a shift of -3.1 percentage points — the largest absolute move of any sector. This is striking because technology is the sector where individual stock stories diverged most sharply: Microsoft's target rose, while Meta's was slashed -13.5% and Nokia's fell -6.8%. The sector aggregate masks a brutal dispersion story.

Energy and materials were the week's quiet winners in relative terms, recovering +2.5 and +2.3 points respectively. Both sectors remain deeply underwater in absolute terms — energy at -17.8% implied upside, materials at -17.2% — but the direction of travel improved. The models appear to be stabilizing their views on commodity-linked businesses after several weeks of deterioration. Note that materials is a single-company sector here (UPM), so read that signal with appropriate caution.

Healthcare remains the only sector with positive implied upside at +15.3%, though even that slipped -1.4 points from last week. The models continue to find value in ORNBV and JNJ relative to their peers, even as JNJ's target was trimmed -5.4% on the week.

Industrials deteriorated -2.0 points, driven largely by the Wärtsilä weakness. Financials slipped -1.3 points, with JPMorgan (JPM) sitting at -13% implied downside — a persistently bearish model view on a stock trading near all-time highs.

What the Models Reveal About Themselves

The most revealing data point this week is GPT's bias shift of -3.9 percentage points — from +2.2% last week to -1.6% this week. That is not a gradual drift; it is a discontinuous jump. GPT was the only model with a positive bias last week, functioning as the panel's optimist. This week it capitulated to the consensus.

What does that mean behaviorally? GPT's relatively tight CAGR standard deviation of 3.48% — the lowest on the panel — suggests it is the most anchored model, least prone to wild swings in individual valuations. Yet its aggregate bias moved the most. The implication is that GPT revised its views broadly and consistently across names, rather than making a few dramatic cuts. A systematic repricing, not a stock-specific reaction.

DeepSeek remains the panel's structural bear at -8.1% average upside, a position it has held for weeks. Its terminal growth rate is locked at exactly 2.0% with zero standard deviation — the model applies the same assumption to every company in the universe, from Nokia to Nvidia. That is a feature, not a bug, of how DeepSeek appears to be parameterized: consistent, conservative, and cheap at $2.20 per thousand valuations.

Gemini continues to be the only model that fails to return valid outputs for every company — 92.2% validity this week, meaning roughly 9 valuations were unusable. At $10.12 per thousand, it is neither the cheapest nor the most reliable. Its CAGR standard deviation of 10.14% is the highest on the panel, suggesting Gemini is the most willing to make bold, differentiated calls — but also the most erratic.

Where the Framework Breaks

Tesla (TSLA) is where the DCF framework strains most visibly. Consensus target price of $310.98 against a spot of $442.10 implies -30% downside, and the dispersion across models is zero — meaning every model agrees on this bearish view with unusual unanimity. Yet Tesla has been trading at or above these levels for months.

The models are not wrong about the fundamentals. They are simply applying a framework — discounted cash flows, WACC, terminal growth — to a company whose market price is substantially determined by optionality, narrative, and Elon Musk's next announcement. A DCF engine cannot price the possibility that Tesla becomes the world's leading robotics company, or that it does not. It can only price what the numbers say today. The -30% gap is not a prediction; it is a measure of how much faith the market is placing in outcomes that no model can quantify.

That is not a flaw in the models. It is a flaw in expecting them to explain everything.

The Model Scorecard