$95
Research Brief

How Compute Scaling Forces New Efficiency Designs

How physical limits on compute scaling are forcing design shifts in model design, deployment, and enterprise cost structure.

Megan C. Starkey | 2025 | RBD. Intelligence Center
$170M
Cost to Train
Llama 3.1-405B
8,930t
CO₂ Emissions
One Frontier Model
0.7%
Elo Gap Between
#1 and #2 Models
15
Sources Cited
Across 3 Sectors
5
Emerging Designs
Profiled
Governing Insight

Make stronger AI infrastructure decisions before the cost commitments lock in.

Compute scaling costs are doubling every five months while performance gaps compress to fractions of a percent. This research maps the inflection point, profiles the efficiency designs emerging from the frontier, and provides the vendor evaluation framework and TCO diagnostic your team needs to avoid locking into the wrong cost structure.

Executive Summary

Brute-force compute scaling is hitting three walls simultaneously: training costs doubling every five months, energy demands equivalent to the annual carbon output of nearly 500 people per model, and performance differentiation collapsing to fractions of a percent. These walls are not arriving sequentially. Their simultaneous convergence creates an inflection point that no amount of capital can overcome through scaling alone. New designs from Google Research provide the first empirical evidence that memory-efficient, continually learning systems can match or exceed scale-dependent transformers at a fraction of the cost.

15
Sources Cited
3
Sectors Covered
6
Exhibits
24 mo
Decision Horizon
Key Findings

Key Findings

Three Scaling Walls Are Converging Simultaneously
Cost, energy, and diminishing performance returns are tightening from every direction at once, creating a forcing function that incremental optimization cannot resolve.
New Designs Break the Efficiency-Performance Trade-Off
Memory-efficient systems from Google Research demonstrate that continual learning and selective memory can match frontier transformer accuracy at a fraction of the computational cost.
The Competitive Advantage Shifts from Scale to Adaptation Speed
When every organization has access to models of approximately equivalent quality, deployment velocity and domain specificity become the decisive differentiators.
Inside This Brief
Everything a leadership team needs to make stronger AI infrastructure decisions before the cost commitments lock in.
  • 6 original exhibits
  • 2 decision tools
  • 15 sources across 3 sectors
What's Inside

Table of Contents

Research Brief
$95
Instant Access — Individual License
  • Full 12-section research brief with complete analysis
  • 6 original exhibits with data visualizations
  • Vendor Design Evaluation Scorecard (decision tool)
  • 5-Year TCO Variable Checklist (decision tool)
  • Complete source bibliography (15 cited sources)
A custom research engagement on AI infrastructure economics runs $25,000–$75,000. This brief delivers the analysis, vendor evaluation framework, and TCO diagnostic for $95.
Purchase Research Brief

Looking for ongoing access? The IC Subscription includes all research, frameworks, and strategic insights.

Next Step

Make more informed AI infrastructure decisions before the cost commitments lock in.

The decisions your organization makes about AI design over the next 24 months will determine your cost structure, competitive position, and regulatory exposure for the next decade. RBD. helps leadership teams evaluate these decisions with clarity.

Schedule a Conversation