The Big Picture

Five AI models. Twenty-three companies. Five trading days. The aggregate picture that emerges from Week 18 is one of a panel caught between two gravitational pulls: a mild but real recovery in model optimism on one side, and a stubborn cluster of overvalued mega-caps on the other.

The consensus across all models points to a market that is, at best, fairly priced. Average upside estimates range from GPT's +4.5% at the bullish extreme to DeepSeek's -4.6% at the bearish end — a spread of nearly nine percentage points between models analyzing the same stocks on the same days. That spread is not noise. It is a structural feature of how differently these systems weight growth expectations versus discount rates, and it deserves more attention than it typically gets.

What shifted this week is the direction of travel. Four of five models moved toward optimism. The exception — DeepSeek, down another 0.2 points — barely moved at all, which is its own kind of statement.

Trends

The trend data surfaces a genuinely interesting tension. Nokia (NOKIA) posted rising consensus on 3 of 4 days, with a weekly target price range of 11.4% — the widest of any trending name. Models are not just nudging Nokia upward; they are actively revising their view of it. The +18.4% jump in consensus target price is the largest single-week TP revision in the entire dataset. Yet the stock trades at €10.61 against a consensus target of €7.53 — meaning even after all that upward revision, models still think Nokia is 29% overvalued. The revision tells you models are less pessimistic. The gap tells you they are not optimistic.

KONE (KNEBV) shows a cleaner story: 3 rising days, a +5.3% TP revision, and a consensus that now sits modestly above spot. Industrial momentum, quiet but consistent.

Microsoft (MSFT) is the week's notable deterioration — 3 falling days, a -2.9% TP cut, and a model community that appears to be digesting something in the earnings or macro backdrop that it does not like. When the world's second-largest company by market cap starts accumulating downward revisions, it is worth watching whether that becomes a trend or a blip.

Sector Signals

The sector rotation data tells a story of defensive repositioning with a growth footnote.

Healthcare leads all sectors with +22.5% average upside and a +5.9 point improvement week-on-week — the largest positive shift in the table. With only two companies in this bucket (ORNBV and JNJ, though JNJ itself sits in negative upside territory), the aggregate is being pulled heavily by Orion's +46% implied upside. Single-company sector readings require a discount, but the directional signal is real.

Energy is the panel's most unloved sector at -23.5% upside, deteriorating by another 5.3 points from last week's already-negative -18.2%. With XOM trading at $154 against a consensus target of $125, models are making a clear call on commodity valuations that runs counter to much of the sell-side consensus.

The quiet improvers are worth noting: financials recovered 3.2 points to reach near-flat (-1.0%), and industrials gained 2.5 points. Neither is a ringing endorsement, but the direction suggests models are becoming less alarmed about rate-sensitive and cyclical names.

Telecom is the sole sector that moved the wrong way, slipping 2.4 points — driven entirely by Nokia's persistent structural discount despite the TP revision.

What the Models Reveal About Themselves

The model evolution data this week is the most interesting it has been in months. GPT executed a +4.2 point bias shift — from +0.3% last week to +4.5% this week. That is not a drift; that is a recalibration. GPT is now the panel's most bullish voice by a significant margin, and the speed of that shift suggests it is responding to something in the earnings data or macro inputs rather than gradually updating priors.

Claude also moved meaningfully, flipping from -1.1% to +0.2% — a +1.4 point swing that crosses the neutral line. Two weeks ago Claude was a mild bear. This week it is, just barely, a bull.

Grok improved by +1.3 points but remains in negative territory at -2.0%. Gemini gained +1.0 point.

Then there is DeepSeek. Unchanged in direction, barely changed in magnitude, running at a terminal growth assumption of exactly 2.0% with a standard deviation of zero — meaning it applies the same terminal growth rate to every single company it values. That is not analysis; that is a prior masquerading as analysis. It may still produce useful relative rankings, but its absolute valuations should be treated as a systematic screen, not a DCF.

The cost dimension adds another layer. DeepSeek processes a valuation for $2.18 per thousand versus Claude's $41.24. Claude is nearly 19x more expensive and takes 28.9 seconds per call versus DeepSeek's 8.1 seconds. If the outputs were meaningfully different in quality, that premium might be justified. The data does not yet confirm it is.

Where the Framework Breaks

The zero-dispersion cases are where the framework becomes philosophically uncomfortable. Apple (AAPL), Tesla (TSLA), NVIDIA (NVDA), Nokia, and DeepSeek's terminal growth all show dispersion of zero or near-zero. For AAPL and TSLA, every model agrees: both are overvalued, with AAPL's consensus target at $225 against a spot of $271, and TSLA at $313 against $382.

When five independent models converge completely, one of two things is true: either the answer is genuinely obvious, or the models are drawing from sufficiently similar training data that their "independence" is illusory. For stocks as widely covered and narratively complex as Apple and Tesla, the latter explanation deserves serious weight. Consensus among AI models is not the same as consensus among independent minds.

The Model Scorecard