AI Signals — Week 31, Jul 27–31, 2026

The Big Picture

Five models, 24 companies, a full week of data — and the most interesting story has nothing to do with any stock. Week 31 is defined by infrastructure noise masquerading as signal. The v8.1 engine change on July 14 already shifted published targets mechanically upward by roughly 10–25 percentage points; now, on the final day of the week, a provider-side switch on DeepSeek's serving layer injected a second discontinuity. Readers who see a suddenly more optimistic AI panel should know that at least two of the forces driving that appearance are plumbing, not judgment.

Stripping away the mechanical effects, the underlying picture is one of a panel that is modestly bullish on aggregate (consensus upside of +9.9% for technology, flat-to-negative for most other sectors) but deeply divided on individual names. The spread between TIETO's +104% implied upside and AAPL's -52% is not a sign of a coherent market view — it is a sign that the models are working from very different priors about what growth is worth in 2026.

Trends

The trend data this week tells two distinct stories about model conviction. JPM built a rising consensus over three of four trading days, with a 14.9% intra-week range — the widest of any trending name. That is not a gentle drift; that is a panel actively revising upward, and the pure estimate revision of +7.7% in tp_changes confirms the move is real, not just a price-deflation artifact. The models appear to be warming to financials selectively, even as the sector aggregate barely moved.

The counterpoint: META and GOOGL both fell on three of four days, with ranges of 10.0% and 7.1% respectively. This is the models expressing doubt about two of the most widely-held AI-beneficiary narratives in the market. META's estimate was also cut by -8.3% in pure revision terms. When the models that are most enthusiastic about technology broadly are simultaneously trimming their two largest ad-platform names, that divergence is worth watching. JNJ, meanwhile, rose on all three available days — a quiet but consistent signal in a name the panel already views as -14.3% overvalued.

Sector Signals

The sector rotation table rewards careful reading. Technology is the only sector to gain model favor this week, with a +4.0pp shift to a consensus upside of +9.9%. But this is an eight-company sector, and the gains are concentrated — the GOOGL and META trend data suggest the lift is coming from elsewhere in the cohort, likely MSFT (+18.3% estimate revision) and GOOGL's still-elevated absolute upside of +40.1%.

Energy recovered the most in absolute terms, with a +10.8pp shift — but it remains deeply negative at -26.3% consensus upside across three companies. A sector climbing from -37% to -26% is not a recovery; it is a slightly less dire verdict. Note that energy's shift reflects both model revisions and market price moves; XOM at -40.6% and FUM1V at -41.5% are not names the models are rushing to endorse.

Materials is a single-company sector (UPM), so its -13.5pp shift is entirely that one name's story — and UPM's estimate was cut by -17.8% in pure revision terms, the sharpest downgrade in the panel this week. One company does not make a sector signal, but it does make a pointed model statement about Finnish forestry.

What the Models Reveal About Themselves

The DeepSeek story is the week's most structurally important behavioral note. The model's median gap moved from roughly -8% in prior weeks toward approximately -3.4% this week — and on July 31 specifically, a provider-side switch to a reasoning-mode variant under the same model name caused a one-day outage before being repaired. The reasoning variant is measurably less bearish. Readers should treat any DeepSeek comparison spanning July 31 as crossing a series break. The +1.1pp bias shift reported for DeepSeek this week understates the discontinuity because it averages across days before and after the switch.

GPT is the more interesting puzzle. While claude (+1.3pp), gemini (+1.1pp), grok (+2.0pp), and even deepseek (+1.1pp) all moved toward more bullish territory, GPT's bias fell -2.6pp to -4.5% — making it the panel's most bearish model alongside deepseek. GPT also carries the highest cap rate at 16.7%, meaning it is hitting the sanity guard ceiling more often than its peers. A high cap rate combined with a falling bias suggests GPT is not just being conservative — it is being conservative in a way that the engine's own guardrails are partially overriding. That is a model under tension with its own framework.

On cost and efficiency: deepseek at $3.01 per thousand valuations remains the panel's price-performance champion, running at roughly 1/14th the cost of claude ($41.85) while generating comparable token volumes. GPT's 4.2-second median latency is the fastest by a factor of nearly three versus the next competitor — speed and cost efficiency do not always travel together.

Where the Framework Breaks

TIETO at +104% implied upside with zero dispersion is the week's sharpest edge case. Zero dispersion means every model in the panel produced an identical or near-identical estimate — which for a stock trading at €18.73 against a consensus target of €38.25 is not reassuring. It suggests the models are anchoring on the same input (almost certainly analyst consensus targets, which are themselves elevated) rather than independently deriving value. When five different architectures agree perfectly on a deeply out-of-the-money call, the agreement is not evidence of conviction — it is evidence of a shared prior. BRK-B also shows zero dispersion at near-flat upside, which is more defensible: the models may simply find Berkshire genuinely hard to disagree about.

TSLA at -48.4% with zero dispersion is the mirror image: unanimous bearishness, no debate. For a stock this volatile and this narratively contested, a dispersion of zero is itself a red flag about model independence.

The Model Scorecard

*DeepSeek bias shift spans the July 31 provider-side series break; pre/post comparisons are not clean.