Great Divergence: Proprietary AI Owns Verticals, Open-Weight Wins Horizontal

IsoDDE: Proprietary Drug Discovery at 2x AlphaFold Accuracy

Isomorphic Labs announced IsoDDE, doubling AlphaFold 3 accuracy on the Runs N' Poses protein-ligand benchmark, outperforming it by 2.3x on antibody-antigen prediction, and surpassing gold-standard physics-based methods (Free Energy Perturbation). The critical characteristic: the system is fully closed. Nature's assessment was stark: "scant insight into how to achieve similar results".

Isomorphic has $600M+ in funding and nearly $3B in pharma partnerships with Eli Lilly, Novartis, and J&J. CEO Demis Hassabis targets first AI-designed clinical trials by end of 2026. The proprietary approach is not altruistic — it is strategic. Those pharma partnerships provide training data that academic labs cannot access. The more proprietary data Isomorphic trains on, the wider the performance gap, the more proprietary data partners share. The advantage compounds through exclusive access.

Claude's Insurance Vertical: 94% Production-Grade Accuracy

Claude Sonnet 4.6 achieves 94% accuracy on the Pace insurance computer use benchmark — the highest score any model has achieved for real insurance workflows (intake, FNOL, spreadsheets, legacy applications). The Vercept acquisition ($50M) integrates computer-use perception and routing capabilities in-house. Anthropic is explicitly building a proprietary insurance automation stack.

UiPath lost 3.6% in market value on the Vercept announcement — the market pricing in that Anthropic is building a proprietary computer-use stack that will displace enterprise RPA in regulated verticals. Like IsoDDE, Anthropic's advantage is deepening through enterprise integration and proprietary perception infrastructure.

Mistral: Open-Weight Distribution Over Vertical Depth

Mistral's model family — Large 3 (675B total parameters, 41B active via sparse MoE) and Ministral dense variants (3B/8B/14B with 256K context) — are released under Apache 2.0. Fully open, commercially unrestricted. Ministral 14B scores 85% on AIME 2025 — technically strong but does not dominate any single vertical the way IsoDDE dominates drug discovery or Claude dominates insurance.

Instead, Mistral wins on distribution: the Accenture partnership gives access to 500,000+ consultants across regulated industries globally. The Apache 2.0 license means enterprises can run models on-premises, fine-tune without restrictions, and avoid vendor lock-in. EU AI Act compliance is a distribution advantage, not a capability advantage.

The Economics of Vertical Proprietary vs. Horizontal Open

This divergence reflects fundamental market economics. In verticals (drug discovery, insurance automation), the value comes from domain-specific training data, proprietary benchmarks, and regulatory moats. Isomorphic's pharma partnerships provide training data that academic labs cannot access. Anthropic's deep insurance integration provides real-world workflows that emerge only from production deployment. These advantages compound.

In horizontal markets (general enterprise AI), the value comes from distribution reach, deployment flexibility, and cost. Mistral's Apache 2.0 license eliminates vendor lock-in concerns. EU positioning eliminates regulatory friction. The models themselves are technically competitive but do not dominate any single vertical.

The implication for market structure is a barbell: proprietary AI labs capture outsized value in deep verticals (drug discovery, regulated-industry automation, specialized scientific computing), while open-weight providers capture volume in broad horizontal enterprise deployment. The middle — closed-source general-purpose APIs without vertical depth (and without open-source cost advantage) — becomes increasingly squeezed from both sides.

The Open-Source Catch-Up Question

The contrarian case: vertical proprietary dominance may be temporary. Boltz-2 (open-source AlphaFold alternative) already exists, and while IsoDDE outperforms it by 19.8x today, the open-source drug discovery community has historically closed capability gaps within 12-18 months of proprietary announcements (as happened with AlphaFold 2 to OpenFold). Similarly, open-source computer-use frameworks could emerge to replicate Vercept's capabilities.

But the compounding advantage matters. Each month that IsoDDE maintains its 19.8x lead over Boltz-2, it compounds access to proprietary pharma data and customer lock-in. The question is whether the vertical lead compounds faster than open-source can catch up.

The Deeper Concern: Scientific Progress

AlphaFold's open release democratized protein structure prediction and earned Hassabis a Nobel Prize. IsoDDE's proprietary release — while commercially rational — creates a two-tier scientific ecosystem where pharmaceutical companies with partnership access have capabilities that academic researchers cannot even evaluate, let alone replicate. This risks fragmenting the global scientific commons that has historically accelerated discovery.

What This Means for Practitioners

ML teams deciding between building on proprietary APIs vs. open-weight models should base the decision on vertical depth requirements. For domain-specific applications (drug discovery, insurance, legal) where proprietary training data and benchmarks matter, closed-source vertical leaders will maintain widening performance gaps. The advantage is structural: exclusive data access compounds over time.

For general enterprise deployment where cost, deployment flexibility, and regulatory compliance matter, open-weight models via distribution partners offer better economics. The Accenture-Mistral partnership demonstrates that distribution scale can compete with capability depth in horizontal markets.

Adoption timeline: IsoDDE-powered drug design is targeting clinical trials by end of 2026. Claude insurance automation is production-ready now. Mistral via Accenture is rolling out across regulated industries over the next 6-12 months. Winners are clear: Isomorphic Labs (vertical drug discovery), Anthropic (vertical regulated-industry automation), Mistral (horizontal open-weight distribution). Losers: open-source drug discovery tools (Boltz-2 falling further behind), general-purpose closed-source APIs without vertical depth, and enterprise RPA incumbents (UiPath already down 3.6%).