The Big Picture

Five models. Twenty-three companies. Five days. The headline from Week 20 is deceptively simple: the AI consensus is bearish, and it is getting more so. Four of five models now assign negative average upside to the universe of stocks they cover, with only GPT holding above the waterline at +2.8%. The aggregate picture is one of a market that has run ahead of what discounted cash flow discipline can justify — at least according to the machines.

But aggregate numbers flatten the interesting texture. The more revealing story is how the models disagree, and why that disagreement is itself informative. When models diverge, they expose the seams in their training data, their assumptions about terminal growth, and their tolerance for uncertainty. This week, those seams are showing.

NOKIA sits at the bottom of the consensus table with a -45.5% implied downside (spot €11.92, target €6.49), while ORNBV (Orion Oyj B) leads with a modest positive consensus. The spread between the most and least favored names is enormous — which tells us less about Nokia or Orion than it does about how differently models handle mature-but-struggling industrials versus stable pharmaceutical cash flows.

Trends

KONE Oyj (KNEBV) is the sole name in the trends data with a confirmed directional signal: three rising days out of four tracked, with a range of 1.97%. That is not a dramatic move, but the consistency matters more than the magnitude. When models revise a target upward on consecutive days — rather than oscillating — it suggests the underlying inputs (earnings revisions, macro assumptions, peer comparisons) are moving in one direction rather than creating noise. KONE's +3.6% target price increase this week reinforces the picture. The models are building conviction on Finnish industrials, slowly but without contradiction.

Sector Signals

The sector rotation data this week tells a story of two diverging worlds. Healthcare leads all sectors with a +23.6% consensus upside, up from +20.5% the prior week — a +3.1 point improvement. This is a two-company sector (dominated by JNJ and ORNBV), so treat the number with appropriate skepticism about sample size. But the direction is consistent with a broader model preference for predictable cash flows over growth narratives.

Technology is the week's most significant loser. Eight companies, the largest sector cohort in the dataset, saw consensus upside fall from +2.2% to -1.1% — a -3.2 point swing. This is not noise. With names like META (target price down -13.3% week-on-week), MSFT (down -5.7%), and GOOGL (down -4.5%), the models are collectively marking down the sector's most prominent constituents. The irony that AI models are growing more skeptical of the companies most exposed to AI infrastructure spending is not lost on this analyst.

Materials improved the most in absolute terms (+4.9 points), though it remains deeply negative at -15.8% — a single-company sector, so the signal is thin. Energy stays firmly in the red at -19.9%, with XOM carrying a consensus target of $125.33 against a spot of $152.78. The models have been consistently bearish on energy for weeks; at some point, persistent bearishness becomes its own kind of information.

What the Models Reveal About Themselves

The model evolution data this week is where the real behavioral story lives. Gemini made the largest single-week sentiment shift of any model: from +0.9% bias last week to -0.6% this week, a -1.5 point reversal. Gemini also carries the highest CAGR assumption of the group at 9.6% with a standard deviation of 11.9% — by far the widest dispersion. High variance in growth assumptions combined with sharp week-to-week sentiment swings suggests Gemini is the most reactive model in the panel. It updates aggressively when inputs change, which can look like responsiveness or instability depending on your prior.

GPT moved in the opposite direction, becoming the week's most notable bull with a +1.7 point bias shift (from +1.1% to +2.8%). GPT also has the highest uncapped terminal value deviation at 71.1% — meaning its DCF math is generating intrinsic values that diverge most dramatically from current prices before the cap is applied. GPT appears to be the model most willing to extrapolate optimistic long-run scenarios.

Claude and DeepSeek both drifted more bearish, by -0.5 and -0.8 points respectively. Claude remains the most expensive model at $41.07 per 1,000 valuations and the slowest at 29 seconds per call, but it achieved 100% validity across all 115 valuations — a consistency that has value in production environments where failed outputs create downstream problems.

DeepSeek at $2.19 per 1,000 is the cost story of the week. It processes comparable token volumes to its peers in 6 seconds average latency, at a price that makes the others look artisanal. The question is not whether DeepSeek is cheap — it obviously is — but whether its 47.8% cap rate (the share of valuations where the terminal value required capping) signals a systematic tendency toward extreme outputs that the cap is quietly correcting.

Where the Framework Breaks

TSLA is the edge case that exposes the limits of DCF-based AI valuation most starkly. Spot price: $443.30. Consensus target: $311.25. Implied downside: -29.8%. Dispersion: 0.0 — meaning all five models converged on nearly identical bearish targets, with no disagreement whatsoever.

Perfect model consensus on a stock famous for defying fundamental analysis is not a signal of analytical clarity. It is a signal of shared training data. Every model has absorbed the same analyst reports, the same DCF frameworks, the same skepticism about Tesla's non-automotive revenue streams. The zero dispersion tells us the models are not independently analyzing Tesla — they are independently recalling the same consensus view. When the crowd is unanimous, the framework has stopped being a tool and started being a mirror.

The Model Scorecard