The Measurement Surface and Denominator The resolution relies on the organization-level gap on the headline Artificial Analysis (AA) Intelligence Index. AA no longer publishes a separate Agentic Index, having folded agentic evaluations into the composite index at roughly a 25–30% weight https://artificialanalysis.ai/leaderboards/models [Announcing the Artificial Analysis Intelligence Index v4.3](https://artificialanalysis.ai/articles/artif
Weighed against expectations for near-term leaderboard performance, the median was slightly lowered to reflect stronger confidence in capability compression and benchmark saturation.
The estimated ratio of enterprise primary model vendor switching in 2027 compared to mid-2025 is projected at a median of 1.18 (10th percentile: 0.66, 25th: 0.91, 75th: 1.6, 90th: 2.1).
Initial baseline metrics and structural parameters are fully established. The denominator anchors tightly on the Menlo Ventures mid-2025 finding that 11% of enterprises changed their primary model provider over the prior year menlovc.com. Alternative figures, such as the 55% cumulative ever-switched rate [1d0063, d
Set against related questions, this distribution held steady, reflecting the balanced forces of forced vendor deprecations and multi-homing against rising technical lock-in.
The forecast projects that the share of frontier-lab external data spend directed at capturing real work sessions will nearly double by late 2028 (median ~1.75x). Currently, real-work capture constitutes a very small base (roughly 2-5% of the total). The denominator—manufactured data such as crowd annotation, synthetic generation, and reinforcement learning (RL) environments—dominates current budgets. However, as frontier models shift toward complex agentic workflows, labs increasingly require m
Factoring in expectations around the ongoing convergence of frontier AI model capabilities slightly tightened the upper and lower bounds of our estimate, while preserving the expectation that spending will shift heavily toward real-work enterprise data.
Core Arithmetic & Utilization Baseline The central estimate of roughly 110x hinges on achievable model-FLOPs utilization (MFU). A DeepSeek-V3-class MoE activates roughly 37B parameters per token out of its 671B total arxiv.org. At batch size one, decode is strictly memory-bandwidth bound. Moving 37GB of active weights against 3.35 TB/s of memory bandwidth (e.g., on an H100) takes roughly 11ms per token, yielding a batch-one MFU of roughly 0.34%. Weighing this against a realistic large-batch M
Set against related questions evaluating lab efficiency , this distribution was maintained with a 110x median, reflecting the inherent hardware penalties that drive large-scale inference economics.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited