Scaling Walls, Building Bridges: State of AI 2025 - Year In Review
Trends in performance, governance, economy & research from 100+ papers.
A personal note before we begin: This consolidated presentation was built using Google NotebookLM—my favorite GenAI tool, which I now recommend to any executive struggling to keep pace with global movements across the AI landscape.
This isn't official RBD research, but a personal workflow and work product I want to share with you.
To keep the pulse on areas you care about in 2026, I recommend using ChatGPT 5.2 in agent mode to scrape papers and research of your choice, quarterly. Upload to GoogleNotebook LM and use the mind map, slide deck and concept summaries to inject information quickly + reclaim your time.
Now, to the findings.
[You can download the slide deck here.]
The dominant AI strategy of the past decade—scale compute, scaledata, scale parameters—is hitting a wall. And the implications for enterprise leaders are significant.
The Costs Have Become Astronomical
Training compute is doubling every five months. The numbers tell the story: training the original Transformer in 2017 cost roughly $670. GPT-4 in 2023 cost approximately $79 million. Llama 3.1 in 2024 reached $170 million. Training datasets are doubling every eight months, with Meta's Llama 3.3 trained on 15 trillion tokens.
The carbon footprint follows the same curve. Training Llama 3.1 produced 8,930 tons of CO2—equivalent to the annual emissions of nearly 500 Americans. Microsoft, Google, and Amazon are now securing nuclear energy agreements to power AI operations.
The Frontier Is Tightening
Despite these investments, differentiation through scale alone is becoming difficult. The performance gap between #1 and #10 models on Chatbot Arena fell from 11.9% to 5.4% in one year. The gap between top US and Chinese models on MMLU narrowed from 17.5 to 0.3 percentage points. Open-weight models closed the gap with closed models from 8% to under 2%.
Scale is yielding diminishing returns.
A New Path: Smarter Architectures
This is where the research gets interesting. New architectures like Titans and the MIRAS framework represent a fundamental shift—from static models requiring costly retraining to dynamic systems capable of "test-time memorization": learning and adapting while running, without offline retraining.
The core insight borrows from neuroscience: effective learning requires distinct modules for short-term and long-term memory, with a "surprise metric" that prioritizes novel, pattern-breaking information. The result is models that achieve higher accuracy with linear (not quadratic) inference speed.
What This Means for Leaders
The next frontier is not simply building bigger models, but building smarter ones. The strategic implications are threefold: accessibility (efficient models lower barriers to entry), personalization (real-time adaptation without new versions), and sustainability (an escape from the diminishing returns of brute-force scaling).
The organizations that thrive will be those that understand the scalingwall is real—and that the bridge to the future runs through architectural innovation, not just capital expenditure.
All this, at a glance below:
Warmest,
Megan
RBD helps complex organizations embed AI as a core capability. Take our un-gated 3 minute enterprise capability assessment for customized recommendations here: https://rbdco.ai/ai-tools
