AI Signals — Week 26, Jun 22–26, 2026

The Big Picture

The AI models are, as a collective, deeply skeptical of current market prices. Across 23 companies and 5 trading days, the ensemble consensus sits at a weighted average upside that is effectively flat to negative. Three of the five models post negative average upside — Grok at -7.0%, DeepSeek at -5.9%, and Claude at -1.1% — while the two bulls, GPT at +3.1% and Gemini at +2.4%, are barely above the waterline. This is not a market the models want to buy.

What makes this week's read particularly striking is the composition of the skepticism. The names dragging consensus targets below spot prices are not obscure small-caps. AAPL sits -33.6% below its model consensus target relative to spot. TSLA is priced $130 above where the models think it belongs. AMZN and NVDA are both underwater. The models are not being cautious about the periphery — they are waving a red flag at the core of the growth trade.

Yet the market has not asked for their opinion, and prices remain where they are. That tension — between model-derived fair value and observable market prices — is precisely what this platform exists to track.

Sector Signals

The most dramatic move this week belongs to healthcare, which gained +10.4 percentage points of model-implied upside, jumping from +12.0% to +22.5%. That is a substantial re-rating in a single week. The important caveat: this sector contains only two companies in the dataset, meaning the shift reflects concentrated model conviction rather than broad-based sector enthusiasm. Single-company sectors are amplifiers, not trend signals — treat them accordingly.

Telecom also improved, rising +4.4 points to +8.5% implied upside, though again this is a one-company sector reading. Materials recovered modestly, up +4.5 points, though it remains deeply negative at -16.0%.

On the losing side, consumer deteriorated sharply — down -5.8 points to -13.6% — suggesting the models are growing more pessimistic about consumer-facing businesses as the week progressed. Energy slid further into negative territory at -21.6%, the worst-performing sector in the dataset. The models have been consistently hostile to energy names, and XOM's consensus target sitting $32 below spot is the clearest expression of that view.

Technology, the largest sector at 8 companies, barely moved — down just -1.1 points to +0.1% implied upside. Eight companies, essentially zero aggregate upside. The models are not finding value in tech at current prices.

Financials were the most stable sector, shifting only +0.3 points, though the absolute level of -9.4% means the models still think banks and insurers are overpriced. JPM at $335 against a consensus target of $255 is the starkest example.

What the Models Reveal About Themselves

The behavioral story of the week is GPT's dramatic pivot. Last week GPT carried a bias of -1.2% — mildly bearish, in line with the ensemble. This week it sits at +3.1%, a shift of +4.3 percentage points. No other model moved more than 1.8 points in either direction. GPT did not drift; it lurched. Whether this reflects genuine reassessment of fundamentals or sensitivity to some input change in the underlying prompting pipeline is impossible to determine from the outside — but the magnitude demands attention.

Gemini also turned more bullish, shifting +1.8 points to +2.4%. The two models now form a mild bull camp, while Grok, DeepSeek, and Claude remain in negative territory.

Claude's shift is the mirror image of GPT's, though smaller: from +0.1% to -1.1%, a -1.3 point move toward caution. Claude also carries the highest cost in the ensemble at $41.25 per thousand valuations and the slowest latency at nearly 30 seconds per call — but it posts a 100% valid rate and the tightest terminal growth dispersion of any model that actually varies that parameter.

DeepSeek is the model that warrants the most scrutiny this week — not for what it did, but for what it did not do. Its bias has been -5.9% this week and -6.0% last week. Its terminal growth rate is locked at exactly 2.0% with zero standard deviation. Its uncapped deviation from analyst targets is 77.7% — it is generating genuinely different numbers — yet the aggregate output barely moves. At $2.21 per thousand, DeepSeek is by far the cheapest model in the panel, nearly 19 times cheaper than Claude. But consistency without variation raises a structural question: is this model reasoning about each company individually, or is it anchoring to a stable internal prior and adjusting at the margins?

Grok shows a similar terminal growth rigidity — zero standard deviation at 2.0% — and carries the highest cap rate in the panel at 20.0%, meaning one in five of its estimates still bumps against the v8 sanity guard. That is a meaningful share and suggests Grok continues to produce outlier estimates more frequently than its peers.

Where the Framework Breaks

BRK-B is the most interesting edge case this week. Berkshire Hathaway's consensus target sits at $516.86 against a spot of $487.81 — a modest +6.0% implied upside — but the dispersion across models is exactly 0.000. Every model produced an identical consensus target. This is statistically implausible in a genuine multi-model ensemble and points to a specific structural feature: Berkshire's complexity as a conglomerate likely forces all models toward the same anchor — analyst consensus — because bottom-up DCF construction is effectively impossible for a holding company of this nature. When the models cannot build their own model, they converge on the street's number. The framework is not broken, but it is exposed.

TSLA shows the same zero-dispersion signature, which is even more surprising given Tesla's valuation controversy. Zero dispersion on a stock where human analysts routinely disagree by 50% or more suggests the models are not actually disagreeing about Tesla — they are all landing in the same place for different reasons, or more likely, the same reason.

The Model Scorecard