# Stacked S-Curves of AI Development

> By Scott Covert · August 3, 2026 · AI Stock Market Impacts
> Source: https://aistockmarketimpacts.com/special-reports/stacked-s-curves-of-ai.html

**The verdict, first:** Raw pretraining scale really is flattening — even the field's own scaling-law pioneer says so. Ilya Sutskever, accepting NeurIPS 2024's "Test of Time" award: **"Pre-training as we know it will unquestionably end."** That's a genuine, citable data point, not hype-cycle spin.

But that's one curve out of at least seven running at once. Reasoning/inference-time compute started its own steep climb almost exactly as pretraining's curve began to bend. Agentic orchestration started its climb the year after that. Meanwhile, four much slower curves — vertical adoption, physical robotics, scientific validation, and economic restructuring — are each just beginning multi-year-to-multi-decade climbs of their own, bottlenecked by things that have never moved at software speed: trust, biology, hardware, and institutions.

## Timeline overview (2012–2040+)

| # | Curve | Started | Status |
|---|---|---|---|
| 1 | Base model capability | ~2012 | Flattening — declared so by the field itself in 2024 |
| 2 | Reasoning / inference-time compute | ~2024 | Still steep |
| 3 | Agentic orchestration | ~2025 | Adoption climbing fast |
| 4 | Vertical / domain integration | Fragmented per industry | Coding earliest (~2022), medicine/finance lag |
| 5 | Physical / robotics embodiment | ~2024-25 | "Validation year," not mainstream |
| 6 | Scientific validation (biotech) | Design ~2020, validation still flat | Design transformed, approval untouched |
| 7 | Economic / organizational restructuring | ~2023 | Barely rising — electricity's own lag ran ~40 years |

## What each curve actually is, and what's verified

**1. Base model capability (pretraining scale).** Started roughly 2012, climbed steeply through the GPT-3 era, and is now the one curve in this stack the field's own leadership has explicitly called flattening. Sutskever, NeurIPS 2024, Vancouver: "Pre-training as we know it will unquestionably end." "The 2010s were the age of scaling, now we're back in the age of wonder and discovery." *Source: techmeme.com/241213/p33*

**2. Reasoning / inference-time compute.** Started almost exactly as curve 1 began to bend — OpenAI's o1 (Sept 2024), then DeepSeek-R1 (Jan 20, 2025) proved a second, independent lab could reach frontier reasoning via reinforcement learning rather than more parameters. (The pure-RL characterization applies to the R1-Zero research variant specifically; the shipped R1 used a cold-start SFT phase before RL.) McKinsey's real projection: inference becomes more than half of AI compute demand by 2030 — not "75%," a figure that doesn't trace to McKinsey. Separately, McKinsey does project a genuine $7 trillion in AI data-center infrastructure investment, from a different report. *Sources: fireworks.ai/blog/deepseek-r1-deepdive; mckinsey.com (Future of AI Workloads); mckinsey.com ($7T Data Center Build-Out)*

**3. Agentic orchestration / tooling.** Gartner, verbatim: "40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025" — an 8x jump in one year. McKinsey Global Institute separately projects AI agents and robots could generate roughly $2.9 trillion/year in US economic value by 2030, at ~27% average automation of current work hours. *Sources: gartner.com (Aug 26 2025 press release); mckinsey.com/mgi (Nov 2025)*

**4. Vertical / domain integration.** No single clean statistic covers "verticals" as one category — by nature it fragments per industry. Coding is furthest along (digital-native, low regulatory friction, ~2022 onward); medicine and finance lag hardest, gated by liability, compliance, and legacy-system integration. Curves 5 and 6 below are this curve's two most measurable sub-cases.

**5. Physical world / robotics (embodiment).** 2026 is explicitly called a "validation year, not proof of mainstream adoption" in current trade coverage. Roughly $6B went into world-model companies in Q1 2026 alone. China's National Venture Capital Guidance Fund targets roughly $138B for AI/robotics — as a 20-year commitment, not a lump sum. The honest counterweight: most humanoid robots run 90 minutes to 8 hours per charge (not one universal figure), and policies hitting 95% success in the lab commonly drop to roughly 60% in the real world. *Sources: techtimes.com (June 2026); theaiinsider.tech*

**6. Empirical / scientific validation (the biotech case).** The design phase is genuinely transformed; the validation phase is untouched. AlphaFold's real, verified effect: it raised proteome structural coverage from roughly 48% to 76% — not the commonly repeated (and peer-review-contradicted) claim that it cut experimental structure-determination work by 60-70%. As of this writing, zero FDA approvals exist for AI-discovered drugs, and Phase II success rates for AI-derived candidates (~40%) are running at historical norms, not above them. The bright spot: Insilico Medicine's rentosertib — first molecule combining an AI-discovered target with an AI-designed drug to show a Phase IIa efficacy signal (Nature Medicine, June 2025) — and AI-native programs running ~80-90% Phase I success versus a historical baseline near 52%. *Sources: pnas.org/doi/10.1073/pnas.2315002121; insilico.com/news*

**7. Economic / organizational restructuring.** Framed explicitly in current commentary as Solow's productivity paradox with "computer age" swapped for "artificial intelligence." Global AI investment exceeds $500B/year (Goldman Sachs, hyperscaler capex); OECD total factor productivity growth remains flat (~0.4% in 2024) — two separate, real findings. The historical precedent: Edison's Pearl Street Station opened Sept 4, 1882; the productivity impact of electrification didn't show up in US manufacturing data until the 1920s — a documented ~40-year lag (Paul David, *The Dynamo and the Computer*, American Economic Review, 1990). The named framework for why: the "Productivity J-Curve" (Brynjolfsson, Rock & Syverson, NBER 2018, published *American Economic Journal: Macroeconomics* 2021) — value capture requires expensive complementary investment that initially masks a technology's true potential. *Sources: ideas.repec.org (David 1990); nber.org/papers/w25148; oecd.org Compendium 2026*

## What ties it together

Every curve in the fast software stack (1→2→3) is capability moving at software speed. Every curve in the slow diffusion stack (4 through 7) is capability running into something that has never moved at software speed — biology, hardware, or institutions — and it won't start now just because the underlying model got smarter. The mistake the doom cycle and the hype cycle both keep making, in opposite directions, is measuring the whole phenomenon by whichever single curve happens to be loudest that quarter. Pretraining flattening is real. It's also just one row in a seven-row chart.

---

*This page is an analysis of publicly reported research, earnings, and academic findings (current as of August 2026), not investment advice and not a prediction of any specific market event, date, or security. Do your own diligence.*

*AI Stock Market Impacts · https://aistockmarketimpacts.com*