AI Signals — Week 34, Aug 17–21, 2026

The Big Picture

The AI valuation panel enters the back half of August in a state of quiet disagreement. Aggregate consensus upside across 24 companies sits in low single digits for most models, but that headline masks a portfolio that is deeply bifurcated. On one side: TIETO (+100.1%), ORNBV (+94.0%), and META (+51.6%) — names the models collectively believe the market has dramatically mispriced. On the other: TSLA (-54.2%), FUM1V (-54.2%), AAPL (-41.5%), and AMZN (-36.1%) — household names the models think are priced for perfection that will not arrive.

The panel is not bearish on the market. It is bearish on specific, high-profile, high-multiple names. That distinction matters. When AI models systematically discount the most widely-held US mega-caps while finding value in mid-cap Finnish industrials and European pharma, the question is not whether the models are right — it is what structural feature of their training is producing this pattern.

Trends

Two names generated enough directional consistency to register as trends this week, and they point in opposite directions.

NVIDIA fell on 3 of 4 trading days in model consensus, with a range of 15.06pp across the week. At -20.5% implied downside from a spot of $216.85, the models are not merely cautious — they are increasingly unified in their caution. NVIDIA's dispersion of 0.274 is the second-highest in the panel, meaning individual models disagree sharply on how overvalued it is, but they agree on the direction. Three consecutive down-days in consensus is a signal worth watching: it suggests models are incorporating new information (likely earnings revisions or competitive data on Blackwell ramp economics) and arriving at the same uncomfortable conclusion.

Nokia, by contrast, rose on 3 of 4 days, with a tighter range of 4.98pp. Its consensus target of €9.64 against a spot of €8.78 gives only +9.8% upside — modest by this panel's standards — but the directional consistency suggests building conviction rather than noise. The +5.6% pure estimate revision in tp_changes this week reinforces that this is genuine model upgrading, not price drift.

Sector Signals

Technology is the week's standout gainer in model sentiment, with implied upside rising +5.5pp to +12.3% across 8 companies. This is a meaningful move, though it should be read carefully: sector-level shifts blend market price movements with model revisions, so part of this improvement may reflect a sector that sold off while model targets held steady. Either way, the models are more constructive on tech than they were seven days ago.

Healthcare also improved, up +4.0pp to +37.8% — the highest absolute upside of any multi-company sector. The models have been persistently bullish on healthcare for weeks; this week's move reinforces rather than initiates that view.

Telecom deserves a note. It shed -7.9pp to land at +26.1% — still the highest single-company sector reading, but the direction is worth flagging. With only Elisa in this bucket, the shift is entirely stock-specific rather than a macro call on the sector.

Energy deteriorated further, now at -27.8% implied downside across 3 companies including XOM and FUM1V. The models have been structurally bearish on energy for months; this week's -3.2pp drift suggests no change of heart.

What the Models Reveal About Themselves

The most important story this week is DeepSeek's +6.4pp bias shift — from -0.4% last week to +6.0% this week. On its face, this looks like a dramatic change of view. It is not. As noted in the methodology, DeepSeek's serving infrastructure switched to a reasoning-mode variant around July 31, which measurably reduced the model's structural bearishness. The bias was trending toward zero from that date; this week's reading confirms the new regime has fully settled. Pre- and post-July 31 DeepSeek comparisons span a series break and should not be treated as evidence of market insight.

Stripping out DeepSeek, the remaining four models show more modest evolution. Gemini and GPT both shifted +2.1pp, suggesting mild convergence toward the bullish end of the panel. Claude nudged up +0.5pp to +6.1% — the most stable model in the panel week-to-week. Grok was perfectly flat at -1.9%, which is itself informative: Grok's terminal growth assumptions (2.04%, the lowest in the panel) and its highest cap rate (18.6%) produce a consistently conservative output that does not move much regardless of what the underlying companies do.

The GPT/Grok bear camp versus the Claude/DeepSeek/Gemini bull camp is now a structural feature of this panel, not a weekly fluctuation. GPT's -1.8% average upside comes with the lowest confidence score (0.60) and the tightest CAGR dispersion (3.49%) — suggesting a model that is not just bearish but certain about its bearishness in a way that should make sophisticated users suspicious.

Where the Framework Breaks

TIETO and BRK-B both show dispersion of exactly 0.0 — meaning every model in the panel produced an identical consensus target. For BRK-B at +7.2% upside, this is plausible; Berkshire is a well-understood, heavily-covered conglomerate. For TIETO at +100.1% upside against a spot of €19.12, zero dispersion is a red flag. When five independent models agree that a stock is worth exactly twice its current price, the most likely explanation is not analytical convergence — it is that all models are anchoring to the same analyst consensus target and the v8.1 sanity guard is not wide enough to force genuine independent estimation. A 100% implied upside with 0.0 dispersion is not a signal. It is a data artifact masquerading as one.

TSLA presents the mirror-image problem: -54.2% implied downside, also with 0.0 dispersion. Two names at the extreme ends of the upside distribution, both with perfect model agreement. The framework's inability to generate genuine disagreement on its most controversial outputs is its most important current limitation.

The Model Scorecard

*DeepSeek bias shift reflects provider-side reasoning-mode change (July 31), not a genuine revision.