AI Signals — Week 39, Sep 21–25, 2026

The Big Picture

Something quiet happened this week that deserves more attention than it will get. Across all five models in the panel — claude, deepseek, gemini, gpt, and grok — the average implied upside fell. Not one model moved in the other direction. The shifts are modest individually: -1.8pp for Claude, -2.0pp for DeepSeek, -0.7pp for Gemini, -1.2pp for GPT, -0.6pp for Grok. But unanimity is rare, and unanimity without an obvious catalyst is rarer still.

Remember that bias_shift figures blend model revisions with market price movements — if prices rose this week, part of the apparent bearish drift is simply the market running ahead of static estimates. Still, the direction is consistent enough to note. The panel's center of gravity sits at roughly -1.6% average upside when you weight models equally, compared to a modestly positive reading last week. That is not a crash signal. It is a whisper.

The broader picture these models are painting is one of a market that has, in aggregate, outrun fundamental value on the names that matter most. AAPL at -48.4% implied downside. TSLA at -58.0%. NVDA at -41.6%. These are not rounding errors. They represent a structural disagreement between DCF-anchored AI reasoning and whatever narrative is currently sustaining those price levels.

Trends

The trends array is empty this week — no company generated consecutive multi-day directional consensus moves sufficient to register. In a five-day window, that absence is itself informative: model estimates are not building momentum in either direction on any individual name. The panel is recalibrating, not trending.

Sector Signals

Healthcare led the week's losers in relative terms, with the sector's implied upside falling 5.7pp to +31.1% — still the most bullish sector reading in the panel, but the direction of travel matters. The two-company sector (ORNBV and JNJ) is pulling in opposite directions: Orion sits at +77.1% implied upside while Johnson & Johnson languishes at -15.5%. Averaging them flatters neither. Treat healthcare's headline number as a tale of two very different situations.

Industrials shed 2.4pp to land at +11.4%, a meaningful move for a three-company sector. Technology, the largest group at eight companies, slipped just -1.1pp to +2.2% — thin consensus upside for a sector that contains both GOOGL at +31.7% and NVDA at -41.6%. The average conceals a canyon.

On the improving side, financials edged up +1.6pp to -4.0% — still negative, but less so. The four-company financials group includes JPM at -11.1%, which the models have consistently viewed as expensive. Telecom's single representative, ELISA, improved +1.5pp to +26.3% implied upside — though single-company sectors should be read as stock calls, not sector calls.

Energy remains the panel's most convicted bearish sector at -28.3% implied upside, barely changed week-on-week. XOM at -29.6% and NESTE at -9.9% anchor that view. The models appear to be applying discount rates and terminal growth assumptions that simply cannot justify current energy valuations at prevailing commodity price assumptions.

What the Models Reveal About Themselves

The cost-versus-performance table this week tells a story that should unsettle anyone paying premium rates. Claude costs $46.13 per thousand valuations and delivers +1.7% average upside — the most expensive model, with middling bullishness and a 12.0% cap rate suggesting it occasionally hits the sanity guard. DeepSeek costs $2.62 and lands at -1.3% — a 3pp difference in output for a 17x difference in price.

Gemini is the panel's most confident model at 0.82 average confidence, and also the most bullish at +2.8%. High confidence and high upside is a combination worth watching: it can mean genuine analytical conviction, or it can mean the model is not sufficiently uncertain about things it should be uncertain about. Gemini's stddev_cagr of 6.47 — the widest in the panel — suggests it is at least generating genuine dispersion across names rather than anchoring to a single growth assumption.

GPT is the outlier that demands explanation. At -7.6% average upside and 0.60 confidence — the lowest in the panel — it is simultaneously the most bearish and the least sure of itself. A model that thinks stocks are expensive but cannot say so with conviction is a model in tension. Its stddev_cagr of just 3.17 is the tightest in the panel, suggesting GPT is applying a narrow band of growth assumptions regardless of company-specific circumstances. That is a form of model rigidity worth monitoring.

DeepSeek's stddev_terminal_growth of 0.0 is striking — it applies exactly the same terminal growth rate to every company it values. That is not analysis; that is a parameter that has been fixed. It produces fast, cheap, consistent output. Whether it produces right output is a different question.

Where the Framework Breaks

TIETO presents the most interesting edge case this week. The stock trades at €18.45 with a consensus model target of €38.00 — +106% implied upside — and a dispersion of 0.0. Zero dispersion means every model produced an identical target price, which is statistically implausible unless the models are converging on the same analyst consensus anchor or the same capped output. With the v8 sanity guard in place (0.40x–2.0x of analyst consensus), a +106% upside reading that shows zero dispersion suggests the models may all be hitting the upper bound of the sanity guard rather than independently arriving at the same intrinsic value. The framework is producing agreement where there should be uncertainty. That is not consensus — it is a ceiling.

TSLA also shows zero dispersion at -58% downside. Here the interpretation is different: the models may genuinely agree that Tesla's current price is indefensible on a DCF basis. But zero dispersion at an extreme is still worth flagging — it could reflect genuine analytical convergence, or it could reflect that all models are anchoring to similar near-term earnings estimates that embed the same set of analyst assumptions.

The Model Scorecard