AI Signals — Week 27, Jun 29–Jul 03, 2026

The Big Picture

The AI model panel covering 24 companies across five trading days this week delivered a verdict that is, in aggregate, mildly bearish — but the headline number conceals a story of divergence that is far more interesting than the average suggests. Across 580 total valuations, the consensus upside sits at roughly -8% to -9% when you weight by model count, meaning the panel collectively believes the market is pricing most names ahead of fundamental value. That is not a new finding. What is new is the velocity of change inside individual models, and the sharp sector reordering that happened quietly beneath the surface.

The v8 engine change from June 10 has now fully settled into the data. With the old tiered caps gone and the wide sanity guard (0.40x–2.0x of analyst consensus) in place, terminal value dispersion is genuine rather than mechanically compressed. This week's uncapped deviation figures — ranging from 64% for Gemini to 92% for Grok — reflect real model disagreement about long-run cash flows, not an artifact of ceiling constraints. Investors reading these numbers should treat them as a signal of model confidence, not model error.

Trends

The trends array is empty this week — no company in the coverage universe generated multi-day consecutive movement in the same direction across the model panel. That absence is itself a data point. In a week where individual target prices moved sharply (see NVDA +21.8%, GOOGL +15.5%), the lack of sustained directional momentum suggests the revisions were episodic rather than the product of building conviction. Models are repricing, but they are not yet telling a coherent multi-day story about any single name.

Sector Signals

The most consequential rotation this week was the technology sector's +6.7-point improvement in model-consensus upside, lifting it to +6.7% from a near-flat +0.1% the prior week. Eight companies drive that figure, so it is a broad-based shift rather than one outlier dragging the average. The mega-cap US names — NVDA, GOOGL, META, MSFT, AMZN — all saw target price increases ranging from +9.3% to +21.8%, suggesting the models collectively absorbed some positive signal about AI-era earnings trajectories. Whether that signal was new data or simply mean-reversion after prior conservatism is harder to say.

The mirror image was healthcare, which fell -9.5 points to a still-positive +13.0% consensus upside. With only two companies in that bucket, the move is statistically fragile — a single name's revision can swing the sector average dramatically. Treat single-company sectors (materials, telecom) with the same caution: UPM alone defines materials at -11.8%, and Elisa defines telecom at +11.1%.

Energy remains the panel's most consistently unloved sector at -22.5% average upside, barely changed from -21.6% last week. Three companies — XOM, NESTE, FUM1V — share that burden, and the models show no sign of warming to the sector. Fortum's consensus target of €12.46 against a spot of €20.05 is the starkest single-name disconnect in the entire universe: a -38% implied downside that has been persistent for weeks.

What the Models Reveal About Themselves

The most behaviorally significant development this week is Claude's +3.6-point bias shift, moving from -8.2% to -4.6% average upside. That is not a small twitch — it represents a meaningful recalibration of how Claude is weighting near-term earnings against its discount rate assumptions. Claude also carries the highest per-valuation cost at $39.91 per 1,000, the slowest average latency at 28 seconds, and the widest uncapped terminal value deviation at 86.5%. The pattern is consistent: Claude is doing more work, spending more tokens, and arriving at more differentiated conclusions. This week, that differentiation resolved bullishly.

GPT moved modestly in the same direction (+1.2 points to -0.9%), making it the least bearish model on the panel — essentially flat to fair value on average. Its 96.6% validity rate (four failed valuations out of 116) is the only blemish on an otherwise clean week operationally.

DeepSeek and Grok are a study in anchoring. DeepSeek sits at -15.7% average upside with a -0.7-point drift; Grok at -16.8% with a -0.2-point drift. Both have near-zero terminal growth dispersion (stddev of 0.0), meaning they are applying identical terminal growth assumptions across every company in the universe. That is not a feature — it is a limitation. A model that assigns the same long-run growth rate to TSLA and Berkshire Hathaway is not modeling; it is templating. The cost efficiency is real ($2.22 per 1,000 for DeepSeek), but so is the intellectual flatness.

Gemini occupies an interesting middle ground: the highest average confidence score at 0.76, the highest average CAGR estimate at 10.5%, but also the widest CAGR standard deviation at 14.62 — nearly double any other model. Gemini is simultaneously the most self-assured and the most internally inconsistent model on the panel. It is confident, but not always about the same thing.

Where the Framework Breaks

TSLA and BRK-B both show zero dispersion this week — every model in the panel converged on identical consensus targets. For Berkshire at $516.86 against a spot of $507.78, that near-zero upside with perfect agreement is almost philosophically appropriate: a conglomerate designed to be fairly valued at any point in time. But for Tesla at $244.27 against a spot of $393.45 — a -38% implied downside — unanimous agreement is alarming rather than reassuring. When five structurally different models with different training data and different architectures all arrive at exactly the same number for the most narratively contested stock in the market, the most likely explanation is not convergent insight. It is convergent training data. The models have all read the same analyst reports, absorbed the same bearish DCF frameworks for TSLA, and reproduced them faithfully. That is the framework breaking: consensus masquerading as conviction.

The Model Scorecard