AI Signals — Week 33, Aug 10–14, 2026

The Big Picture

The models are getting cold feet. Four of five AI analysts on the panel shifted their aggregate bias downward this week, and the one exception — GPT — barely moved, edging up a statistically meaningless +0.1pp. The panel-wide average upside, blending all five models, sits somewhere in low single digits, but that headline number conceals a striking bimodal structure: a cluster of deeply underwater names dragging against a handful of Finnish mid-caps that the models appear to love with unusual intensity.

What makes this week interesting is not the direction of the shift but its breadth. When models move together, it usually reflects shared inputs — macro assumptions, discount rate anchors, or common training data responding to the same earnings signals. A synchronized drift toward caution across architectures as different as claude and deepseek is worth noting, even if the magnitude is modest.

The aggregate picture for the 24 companies under coverage is one of a market the models collectively regard as fairly-to-expensively priced. The median name in the universe sits below fair value on model estimates, and the distribution of upsides skews negative once you strip out the Finnish outliers at the top.

Sector Signals

The most meaningful rotation this week is the recovery in consumer sentiment — the sector's model-implied upside improved by +7.6pp to -19.1%. That is still deeply negative, but the direction matters. Consumer here is a two-company sector (PG and JNJ), so treat this as stock-specific revision rather than a broad thematic call. Both names saw positive target-price revisions this week (PG +7.1%, JNJ +6.8%), suggesting the models are gradually walking back prior pessimism on defensive staples — perhaps as rate assumptions stabilize.

Industrials gained +5.6pp to reach +13.9% implied upside, driven in part by a notable +16.7% target-price revision for KNEBV (KONE). That is the largest single upward revision in the dataset this week. The models appear to be reassessing KONE's earnings trajectory, though with dispersion at 0.078 there is still meaningful disagreement beneath the consensus surface.

On the other side, energy slipped a further -2.7pp to -24.6% implied upside. With three companies in the bucket — XOM, NESTE, and FUM1V (Fortum) — the signal is more robust than a single-name read. The models are consistently marking energy below market price, and that view is hardening. Fortum at -43% implied downside is the most extreme case: the models see a business worth roughly half its current market capitalisation, a gap that either reflects genuine structural concern about the Nordic power market or a systematic failure to price regulated utility cash flows correctly.

Technology is essentially flat week-on-week at +6.8% implied upside, masking enormous internal dispersion. TIETO at +106% and GOOGL at +25.5% sit in the same sector bucket as AAPL at -47.7% and TSLA at -53.3%. The sector average is nearly meaningless as a signal.

What the Models Reveal About Themselves

The most behaviorally interesting data point this week is deepseek's -2.1pp bias shift, taking it from +1.2% to -0.9% average upside. This is the largest single-model swing on the panel. Context is essential here: deepseek switched to a reasoning-mode variant on 2026-07-31, which mechanically pushed its bias upward from its historical range of around -8%. The model had been running anomalously bullish since that provider-side change. This week's pullback toward negative territory may represent a partial mean-reversion — the reasoning variant finding its equilibrium — rather than a genuine reassessment of the underlying companies. Analysts comparing deepseek's current output to pre-July readings should treat the series as structurally broken at that boundary.

Grok also shifted meaningfully, down -1.7pp to -1.9%. Unlike deepseek, there is no known mechanical explanation for grok's move, making it the more informative signal of the two. Grok's confidence score remains the lowest on the panel at 0.59, and its CAGR standard deviation of 5.67 is the widest — suggesting a model that is genuinely uncertain rather than systematically biased.

GPT's behavioral fingerprint continues to stand out. It produces the tightest CAGR distribution (stddev 3.45 versus the panel range of 5.5–5.7 for other models), the lowest average CAGR (5.4%), and the highest cap rate (21.9%). A model that caps more estimates, assumes lower growth, and arrives at the most bearish aggregate view is either the most disciplined analyst on the panel or the most anchored to conservative priors. The data cannot distinguish between those interpretations — but the consistency of the pattern across weeks suggests it is structural, not random.

Claude remains the panel's quiet optimist at +5.1% average upside, though it too drifted slightly more bearish (-0.5pp). Its terminal growth assumption of 2.44% is the highest on the panel by a meaningful margin, which mechanically inflates DCF outputs. Readers should be aware that claude's relative bullishness may be partly a function of this terminal growth anchor rather than a more optimistic view of near-term cash flows.

Where the Framework Breaks

TIETO is the week's most instructive edge case. The stock trades at €18.59; the model consensus target is €38.25, implying +105.8% upside. Dispersion is exactly 0.0 — every model produced an identical or near-identical estimate. This is not consensus born of careful triangulation. It is more likely a case where all models are anchoring to the same analyst target-price input and the valuation engine's sanity guard (0.40x–2.0x of analyst consensus) is doing the heavy lifting. When dispersion collapses to zero on a name with triple-digit implied upside, the signal is not that the stock is a screaming buy — it is that the models have effectively outsourced their judgment to the analyst consensus and wrapped it in a DCF shell. TSLA and BRK-B show the same zero-dispersion pattern at the opposite end of the upside spectrum, reinforcing that zero dispersion is a data-quality flag, not a confidence signal.

AMZN presents the opposite problem: 0.308 dispersion — the highest in the universe — against -26.3% implied downside. The models cannot agree on what Amazon is worth, which for a company of this analytical coverage depth suggests the framework is struggling with the conglomerate structure, the AWS-versus-retail valuation split, or both. High dispersion plus negative consensus is the least actionable combination the system produces.

The Model Scorecard