Historical Baseline The forecast depends on the organization-level point gap between the #1 and #2 organizations on the Artificial Analysis Intelligence Index on December 31, 2028, divided by the 2024–2026 average gap. Reconstructing the historical baseline yields a mean gap of roughly 2.5 to 3.0 points. While today's gap sits at a temporarily elevated 4.6 points following Claude Fable 5.2's quick overtake of Gemini 4 Carbon [Gemini 4 Carbon is the new #1 on the Artificial Analysis ...](http
Set against related questions, this was adjusted slightly upward to reflect a stable expectation that the top two labs will maintain historically normal margins without collapsing into a dead heat.
The Measurement Baseline The denominator for this ratio is 11%, the share of enterprises that changed their primary model vendor in the 12 months prior to Menlo Ventures' mid-2025 survey 2025 Mid-Year LLM Market Update: Foundation Model ... 2025 Mid-Year LLM Market Update: Foundation Model .... It is critical to distinguish this annual rate from cumulative "
Evaluating this estimate alongside enterprise API spending trends and the potential costs of dedicated deployment instances led to a slight narrowing of the uncertainty interval, while preserving the median near 1.16 to balance rising multi-homing against architectural switching friction.
What is Being Measured and the 2026 Baseline
The core metric is the relative shift in the share of external data-acquisition spend directed toward capturing practitioners' actual paid work, comparing end-2028 to end-2026. The 2026 baseline for this "real-work" share is likely in the low single digits to low teens. It includes programs exchanging training rights for tokens or discounts, such as OpenAI's Data Sharing Program [Sharing feedback, evaluation and fine-tuning data, and API ...](htt
Set against related questions, this estimate was slightly adjusted upward to reflect that as AI automates synthetic environment generation , labs will likely reallocate a greater share of their data budgets toward capturing real practitioner workflows.
The baseline estimate centers on the severe model-FLOPs utilization (MFU) penalty of batch-one decoding for a sparse mixture-of-experts model. DeepSeek-V3 activates 8 of 256 routed experts plus one shared expert per token, giving a sparsity of 32 (roughly 37B active out of 671B total parameters) [arXiv 2412.19437]. On an H100, streaming 37GB of active weights per token against 3.35 TB/s of bandwidth caps batch-one MFU near 0.34%. Compared to an achievable 35% MFU at the critical batch size (roug
Evaluating this alongside expected declines in overall serving and hardware rental costs kept our median estimate stable, as the core physical divergence between batch-one and large-batch throughput remains governed by rigid hardware bottlenecks.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited