AI Signals — Week 29, Jul 13–17, 2026

The Big Picture

Before reading anything else in this brief, absorb one fact: on 2026-07-14, the valuation engine upgraded from v8 to v8.1, and the published headline number changed in four simultaneous ways — mid-year discounting applied, targets rolled to true 12-month values, an erroneous +1pp debt premium stripped from the CAPM anchor, and the 30% analyst-consensus blend removed from the headline figure. The net mechanical effect was a +10 to +25 percentage point upward shift in published targets across the board. Every model looks more bullish this week than last. Almost none of that is signal.

With that caveat firmly in place, the aggregate picture is this: five models covering 24 companies over five days produced 594 valid valuations out of 600 attempts. The consensus upside across all models sits near flat to mildly negative in real terms, with the market-cap-weighted universe still looking expensive to the models even after the mechanical lift. The models, stripped of the engine noise, are not enthusiastic about current prices.

Trends

KONE (KNEBV) is the week's most credible momentum story. Three rising days out of four, a +11.98% consensus range, and a +33.6% target-price revision — the largest upward pure estimate change in the dataset. Crucially, the KONE revision appears to have begun before July 14, meaning it carries genuine model conviction rather than mechanical uplift. The models are reassessing KONE's elevator cycle exposure and margin trajectory, not just recalibrating a formula.

Alphabet (GOOGL) moved in the opposite direction: three falling days, a 21.16% consensus range (the widest trending stock this week), and a current upside of only +14.5% despite the engine tailwind. Wide range plus falling trend equals model disagreement compounding into pessimism. The models are not aligned on Alphabet's AI monetisation path, and the trend suggests that disagreement is resolving bearishly.

Apple (AAPL) tells a starker story. Three falling days, a -20.5% pure target-price cut — the second-largest downward revision of the week — and a consensus upside of -54.5% against a spot price of $333.26. The models collectively think Apple is trading at roughly double fair value. That view has been hardening all week, independent of the engine change.

Wärtsilä (WRT1V) rose on three of four days with a tight 3.09% range — low dispersion, steady upward drift. Not dramatic, but the kind of quiet conviction that tends to be more durable than a single-day spike.

Sector Signals

Industrials made the sharpest move in model favour, swinging +14.8 percentage points to a +9.3% average upside from -5.4% last week. Three companies drive this, and the engine change contributes, but the KONE revision suggests genuine reassessment too. Healthcare also gained ground, up +12.7pp to +27.8% — though with only two companies (ORNBV and JNJ), this is a thin sample. Note that JNJ still sits at -9.0% upside individually; Orion is doing the heavy lifting at +75.6%.

Energy lost the most ground, falling -8.1pp to -36.0% average upside. With XOM at -40.5% and Neste (NESTE) at -23.9%, the models are consistently and increasingly hostile to energy valuations. XOM's target was cut by a pure -16.3% this week. The models see energy cash flows as structurally impaired at current prices.

Telecom is a single-company sector (Elisa, ELISA) at +20.9% upside — treat it as a stock call, not a sector signal. Materials similarly rests entirely on UPM (UPM) at -16.9%.

Technology, the largest sector at eight companies, barely moved: +1.3pp to +2.7%. Beneath that placid surface lies enormous dispersion — TIETO at +116.6% and TSLA at -56.5% are in the same bucket. The sector average is nearly meaningless here.

What the Models Reveal About Themselves

The engine change is the dominant story in model evolution this week, and intellectual honesty demands we say so plainly. All five models shifted bullish. Gemini moved the most: from -9.4% to +3.2%, a +12.5pp swing. Grok followed at +11.5pp, DeepSeek at +8.1pp, Claude at +5.9pp, and GPT at a modest +1.2pp. The ordering roughly tracks each model's sensitivity to the CAPM anchor correction and the removal of the analyst-consensus blend — models that previously leaned harder on those inputs moved more.

What is genuinely interesting is what the engine change *reveals* about prior behaviour. GPT's tiny +1.2pp shift suggests it was already producing estimates closer to pure DCF logic, less anchored to the consensus blend. Claude's shift of +5.9pp is moderate despite its high token consumption — it appears to have been partially self-correcting the CAPM issue through its reasoning process. DeepSeek, despite the largest bear bias at -12.4%, still shifted +8.1pp, implying its pre-change estimates were being pulled down by the now-removed constraints.

DeepSeek's 5% failure rate (6 invalid valuations out of 120) deserves a flag. At $3.23 per thousand valuations, it is by far the cheapest model — 13x cheaper than Claude at $41.69. But a model that cannot parse roughly one in twenty inputs introduces a systematic gap in coverage that compounds over time.

Gemini's stddev_cagr of 8.31 — more than double GPT's 3.63 — confirms what prior weeks have suggested: Gemini is the most internally volatile model, capable of both the highest conviction calls and the widest swings. It is the analyst who writes the most interesting research and also the one most likely to contradict themselves.

Where the Framework Breaks

TIETO (TIETO) sits at +116.6% consensus upside with a spot price of €17.86 against a target of €38.69 — and a dispersion of exactly 0.0. Every model agrees perfectly. That unanimity is not reassuring; it is suspicious. When five independent models converge on identical outputs for a small-cap Finnish IT services company, the most likely explanation is that they are all drawing from the same thin pool of public information and applying identical heuristics, not that they have independently reached the same conclusion through different analytical paths. Perfect agreement in a high-uncertainty name is a sign of correlated ignorance, not correlated insight.

TSLA and BRK-B also show zero dispersion — but for different reasons. Tesla's 0.0 dispersion at -56.5% upside, combined with a -30.9% pure target cut this week, suggests the models have converged on a shared bearish narrative. Berkshire's 0.0 at +4.7% reflects the opposite: a conglomerate so diversified that the models default to a mean-reversion anchor. Neither zero is informative in the same way.

The Model Scorecard