AI on Deterministic Rails

Narrative

The dominant finding across McKinsey, Gartner, Bain, and a16z research for 2025–2026 is structural: AI adoption is near-universal but durable enterprise value remains concentrated in a small minority. McKinsey's November 2025 State of AI survey (1,993 respondents) found 88% of organisations using AI in at least one function, yet only 39% reporting EBIT impact and just 5.5% qualifying as high performers with more than 5% EBIT attributable to AI. Gartner's successive forecasts — 30% GenAI PoC abandonment by end of 2025, upgraded to 50% in its 2026 analysis, and a further prediction that over 40% of agentic AI projects will be cancelled by 2027 — provide the clearest quantitative frame for the so-called PoC-opalypse. S&P Global data adds a sharp data point: the share of enterprises scrapping most AI initiatives jumped from 17% in 2024 to 42% in 2025.

The token-cost crisis arrived sharply in early 2026. Per-token prices fell roughly 98% from late 2022 levels, yet enterprise AI bills rose an estimated 320% over the same period, with the average annual enterprise AI budget growing from $1.2 million in 2024 to $7 million in 2026. The mechanism is the agentic multiplier: Gartner's March 2026 analysis found agentic models require 5–30 times more tokens per task than standard chatbots, because each step in a multi-agent loop resends the full context window. Uber exhausting its entire 2026 AI coding budget by April became the emblematic case; the FinOps Foundation reported companies calling in April saying they were already 3x over their full-year token budgets. Goldman Sachs projects global token consumption will multiply 24x by 2030. Bain's parallel finding — that the top 5% of users consume more tokens than the other 95% combined — locates the cost problem precisely in the highest-value, hardest-to-throttle engineers.

The open-weight model story changed the cost calculus for private inference. DeepSeek's V3 release in December 2024 and R1 in January 2025 demonstrated frontier-comparable reasoning trained for approximately $6 million in compute, with inference costs reported at 12.5x cheaper than Claude 3.5 Sonnet. By H1 2026, DeepSeek V4 reduced cost-to-serve by over 10x versus V3.2 through hybrid sparse attention, and the broader open-weight stack (Qwen 3, Llama 4, Mistral) closed to within 5% of closed-frontier models on coding and reasoning benchmarks. Sovereign-cloud deployment patterns consolidated around vLLM and air-gapped clusters running quantised DeepSeek V4 or Qwen 3.6, with switching to self-hosted open models offering 70–90% cost savings for organisations processing millions of tokens daily.

On the investment side, CB Insights documented $66.6 billion in AI funding in Q1 2025 alone — nearly two-thirds of all 2024 AI investment — with agentic AI platforms and multi-agent orchestration recording the highest CB Insights Mosaic health scores across all industries (705+). Sequoia's 2026 AI Ascent named Claude Code and long-horizon agentic coding as the first genuinely discontinuous capability shift, distinct from the incremental gains of 2023–2025. A16z's successive CIO surveys tracked the transition from model interchangeability to harness and workflow lock-in: as one enterprise leader told a16z researchers, all prompts have been tuned for a specific provider, and switching now risks breaking downstream agent dependencies. The symbiotic-stack thesis — AI layered onto deterministic workflow infrastructure rather than replacing it — found clearest expression in a16z's adoption analysis, which identified verifiable outputs and well-defined standard operating procedures as the primary selection criteria for successful deployments.