THE OPEN-PREM
INFLECTION POINT V4
The evolution of self-hosted enterprise infrastructure since spring, alongside the strategic impact of agent-generated software architectures.
The Open-Prem Inflection Point is a framework created by David Borish that tracks when on-premises AI becomes a practical alternative to cloud AI for enterprises. The V4 edition (October 2026) finds that the best open-weight model, Xiaomi’s MiMo-V2.6-Pro, scores 46 on the Artificial Analysis Intelligence Index against 58 for the top proprietary model, and that open-weight models have carried most OpenRouter tokens since June. It also extends the thesis to the application layer: coding agents now let a company build its own software on permissive open-source bases and keep the result private. The paper covers models, licensing, agents, hardware economics, compliance, and build-versus-buy. Against low-cost open-model APIs, the case for on-premises deployment rests on control rather than token savings.
WHAT CHANGED IN V4
-
Four released open-weight models now exceed one trillion parameters: MiMo-V2.6-Pro (1.02T, MIT), DeepSeek V4-Pro (1.6T, MIT), Qwen3.8 (2.4T), and Kimi K3 (2.8T). The top four open-weight models on the Artificial Analysis leaderboard are Chinese. Meta’s Muse Spark Open (39) is the top US open-weight model. GLM-5.3, Kimi K3, Step 5, MiniMax M3, Qwen3.8-Max, and Muse Spark Open carry custom license conditions, so license review is required.
-
Open-weight models passed closed models in OpenRouter token volume in June and held about 60% by mid-June. This measures hosted API traffic, not self-hosting. On Vercel’s AI Gateway, open-weight models ran 29% of June tokens on under 4% of spend.
-
Payback depends on the comparison. For a $40,000 deployment at 20 million tokens per day, it is 3.3 months against GPT-6 Astra, 13.2 months against a $5 blended price, and about 10 years against MiMo-V2.6-Pro via hosted API, before power and staff. Hardware prices rose: RTX 5090 cards sold for $4,930 to $5,090 in September against a $1,999 list price.
-
AI-built software is the new layer. In October, Brandon Thomas used Claude Opus 5.5 to build early-alpha, permissively licensed reimplementations of Adobe and Office apps (ArtCraft). McKinsey’s 2026 survey found 32% of organizations declined to buy a software product or feature because they could build it in-house. Systems of record stay with vendors.
-
OpenClaw runs agents locally, with a cloud model supervising and local open-weight models executing. ClawHub grew from 13,700+ to about 52,700 skills by June with no mandatory security review, and RankClaw found 1,103 malicious skills among 14,706 scanned. Run only bundled or internally built skills from a private registry.
-
NemoClaw remains in early preview (version 0.0.117 and later) and now works as a sandbox layer for other agent frameworks. Nemotron 3.5 Super VL runs on a single DGX Spark, so a fully local deployment fits on a desktop-class system.
-
The EU Digital Omnibus moved Annex III high-risk obligations to December 2, 2027 and Annex I obligations to August 2, 2028. Article 50 transparency obligations took effect August 2, 2026. IBM’s 2026 report puts the average US breach cost at $11.5 million and finds shadow AI in 43% of breaches. On-premises deployment eases auditing but does not remove shadow AI, which still requires policy, access controls, and monitoring.
ABOUT THE AUTHOR
David Borish is an Enterprise AI Strategist with 25 years of experience across technology (AI), CPG, sports tech, and finance. He created the Open-Prem Inflection Point framework and delivers the Open-Prem Strategy Accelerator workshop in partnership with IBM. He is the author of "The Tony Hawk Paradox: When Video Games Predict Reality," a book on simulation theory and the pattern of video games predicting real-world capability
Last Updated: October 25th, 2026
Related: Link to Open-Prem | Link to Open-Prem V2 | Link to Open-Prem V2 Update | Link to The Instruction Gap Link to All Papers | Speaking