AI Signals — Week 28, Jul 06–10, 2026

The Big Picture

The headline number this week is deceptively simple: every model in the panel moved in the same direction. All five ended the week more bearish than they started it. That kind of unanimity is rare enough to demand an explanation — and the explanation matters enormously for how you interpret the data.

Remember that bias_shift figures blend two things: genuine model estimate revisions and the mechanical effect of rising market prices compressing the gap between model targets and spot prices. When markets rally, upside shrinks arithmetically even if models don't change a single assumption. So the -6.5 pp swing at Gemini, the -4.8 pp at DeepSeek, and the -3.7 pp at GPT almost certainly reflect a market that ran ahead of model valuations this week, not a sudden collective crisis of confidence. The pure estimate revision data — the tp_changes — tell a more surgical story, and that story is concentrated in a handful of names.

The aggregate consensus across 24 companies paints a market that AI models, taken together, regard as modestly overvalued. Only eight names carry positive implied upside. The median sits in negative territory. That is not a crash signal — it is a valuation regime where the models are saying "priced for perfection" rather than "priced for disaster."

Trends

Four names built multi-day rising consensus momentum this week: NOKIA, NESTE, JNJ, and ORNBV. The pairing of Nokia and Neste is worth noting — both are Finnish names that have been deeply out of favour, and both saw their target prices revised upward (+10.0% and +7.7% respectively in pure estimate terms). That is not coincidence; it suggests the models may be responding to the same underlying data signal, possibly currency or Nordic macro inputs.

The conviction behind these trends is moderate rather than emphatic. Nokia's 8.94% intra-week range is the widest of the four, implying the models are still debating the magnitude of recovery even as they agree on direction. Orion's 1.68% range tells the opposite story: quiet, steady upward drift with minimal internal disagreement. When a trend is narrow-ranged and persistent, it tends to be more durable. Watch ORNBV.

Sector Signals

Three sectors lost meaningful ground this week: consumer (-5.5 pp shift), energy (-5.4 pp), and technology (-5.3 pp). That is an unusual combination — normally consumer defensives and growth tech move in opposite directions. When they fall together in model-implied upside, the most likely culprit is a broad market price increase compressing valuation gaps across the board, which is consistent with the unanimous bearish bias shift noted above.

Technology deserves a closer look. The sector still carries a positive aggregate upside of +1.4%, but that is down from +6.7% last week. With eight companies in the bucket, the average masks enormous dispersion: TIETO at +85% implied upside sits in the same sector category as AAPL at -39.7%. The sector number is almost meaningless at this level of internal spread.

The two sectors that gained favour — telecom (+3.0 pp) and healthcare (+2.0 pp) — are both single- or dual-company readings, so treat them as stock-specific signals rather than genuine rotation. Telecom's improvement is essentially the ELISA story; healthcare's is ORNBV and JNJ trending upward simultaneously.

What the Models Reveal About Themselves

Gemini's -6.5 pp bias shift is the week's most interesting behavioral data point — but not for the reason you might think. Gemini also carries the lowest cap rate in the panel at 8.6%, meaning fewer of its terminal value estimates are hitting the sanity guard ceiling. That combination — low cap rate, high volatility in weekly bias — suggests Gemini is generating genuinely dispersed estimates that respond more sensitively to input changes. Its CAGR standard deviation of 12.76 is nearly double any other model's, confirming this is a model that swings wide.

DeepSeek remains the structural outlier on pessimism. At -20.5% average upside and a 22.1% cap rate, it is simultaneously the most bearish and the most constrained model in the panel. Its terminal growth rate is locked at exactly 2.0% with zero standard deviation — a tell that DeepSeek is applying a fixed assumption rather than deriving growth from company fundamentals. That rigidity has a cost: it produces systematically lower valuations that may reflect model architecture rather than genuine analytical judgment.

Claude's avg_uncapped_deviation of 71.5% — meaning its raw DCF terminal values deviate from analyst consensus by an average of 71.5% before any capping — is a reminder that these models are not simply parroting sell-side targets. They are running independent calculations that happen to land in a similar neighbourhood, most of the time.

Where the Framework Breaks

TSLA is the cleanest failure case in the dataset. Its dispersion score is exactly 0.0 — five models producing identical consensus output — while its implied upside sits at -39.4%. Perfect agreement on a deeply bearish number for the market's most sentiment-driven stock. This is not analytical convergence; it is model herding. When every model reads the same training data about Tesla's valuation history and analyst targets, they converge on the same answer. The framework cannot distinguish between "five independent analysts agree" and "five models memorised the same prior." For a stock where narrative drives price as much as fundamentals, that distinction is everything.

AAPL at -39.7% with a 0.043 dispersion tells a similar story: near-unanimous bearishness on the world's largest company by market cap. Either the models are right and Apple is dramatically overvalued, or they are all anchored to the same stale fundamental framework that cannot price the optionality embedded in the Apple ecosystem. The honest answer is probably some of both.

The Model Scorecard