AI Scaling Laws 2025: A Conservative, Evidence-Based Revision

Revision date: December 11, 2025. Original review: November 12, 2025.

Executive Summary

This research presents a conservative, evidence-based framework for AI capability scaling 2025-2035. The original research contained significant weaknesses: selective citation ignoring contradictory evidence, overly optimistic efficiency projections, and insufficient treatment of fundamental scaling limits.

This revision corrects those flaws:

1. Acknowledges contradictory evidence - Inverse scaling, data exhaustion, severe 2024 plateau
2. Conservative parameters - 50-75% reduction from original optimistic values
3. Economic deployment constraints - Models cost barriers limiting practical deployment
4. Pre-training plateau modeling - Sigmoid approach, not continued exponential
5. Wide uncertainty bands - ±50% near-term, ±200% long-term explicit ranges

Key Reality Checks:

• Pre-training scaling: Near-complete stagnation in late 2024 (Orion, Gemini plateau) - NOT "slower growth" but actual plateau
• Test-time compute: 200x cost increase ($5 → $1,000) makes it viable for <0.1% of tasks
• Efficiency gains: 23x claim based on non-peer-reviewed blog sources - reduced to 1.5x-2x per decade until validated
• Data exhaustion: High-quality text data essentially exhausted, synthetic data shows negative returns
• Inverse scaling: Larger models become LESS truthful on some tasks (arXiv:2307.03201)

1. Contradictory Evidence (Previously Omitted)

1.1 Inverse Scaling for Truthfulness

Primary Source: McKenzie et al. (2024). "Scaling Laws Do Not Scale." arXiv:2307.03201. July 2024.

URL: https://arxiv.org/abs/2307.03201

Key Finding: Larger models demonstrate WORSE performance on certain tasks:

• Truthfulness: Models learn to "sound plausible" faster than learning facts → bigger = less accurate
• Toxicity: Increased model size correlates with higher toxic output rates
• Spurious correlations: Larger models amplify dataset biases more strongly

Implication: The assumption "bigger models = universally better" is empirically falsified for important capabilities. Scaling can make problems worse.

1.2 Logarithmic Diminishing Returns (Late 2024)

Primary Source: "Has LLM Reached the Scaling Ceiling Yet?" arXiv:2412.16443. December 2024.

URL: https://arxiv.org/abs/2412.16443

Key Finding:

"The relationship between model size and performance improvements follows a logarithmic curve, with each doubling of parameters yielding progressively smaller gains."

Quantitative Evidence:

• 2018-2022: Linear gains from parameter scaling
• 2023-2024: Logarithmic diminishing returns emerge
• Late 2024: Approaching asymptotic limits

Interpretation: Not a temporary plateau but a fundamental ceiling where additional scale produces minimal improvement.

1.3 Severe Pre-Training Plateau (Industry Evidence)

Primary Source: AI scaling slowdown evidence (companion review in this series)

Orion Performance Pattern (Bloomberg, Nov 2024):

• Achieved GPT-4 performance at 20% of training
• Remaining 80% of training yielded minimal additional gains
• Assessment: "Not as big a step up as GPT-4 was from GPT-3.5"

Quantitative Analysis:

• Historical doubling: 5.9 months (4.1x/year through 2024)
• Current evidence: 8-12+ month doubling or worse
• Multiple labs affected: OpenAI (Orion), Google (Gemini), Anthropic (Claude 3.5 Opus delayed indefinitely)

This is not "slower growth" - this is near-stagnation.

1.4 Data Exhaustion Crisis

Primary Source: Villalobos et al. (2024). "Will We Run Out of Data?" arXiv:2211.04325 (2024 update).

Key Findings:

• High-quality text data: Essentially exhausted by 2024
• Synthetic data solutions: Show diminishing to negative returns (models trained on model-generated text degrade)
• Multimodal alternatives: Video/robotics data requires 10-100x more compute per effective token

Quote from Lu (2025):

"Without ongoing efficiency gains, advanced performance could demand millennia of training or unrealistically large GPU fleets."

Context: Lu presents this as a WARNING about dependency on unproven efficiency gains, not an optimistic continuation scenario.

1.5 Economic Deployment Barrier (o3 Cost Explosion)

Primary Source: ARC Prize announcement (December 2024), TechCrunch analysis.

o1 → o3 Cost Comparison:

• o1 (Sept 2024): ~$5 compute per task, 32% on ARC-AGI
• o3 (Dec 2024): >$1,000 compute per task, 87.5% on ARC-AGI
• Cost increase: 200x for 2.7x performance improvement

Economic Reality Check:

• GPT-4 API cost: ~$0.03 per 1,000 tokens (~$0.10 per typical task)
• o3 high-compute: $1,000+ per task
• Cost multiplier: 10,000x compared to standard deployment

Market Viability:

• At $1,000/task, only the highest-value applications (legal research, drug discovery, critical decision-making) are economically viable
• Estimated addressable market: <0.1% of potential AI tasks
• Remaining 99.9% of tasks continue using cheaper, less capable models

2. Revised Estimates and Their Justification

2.1 Pre-Training: Sigmoid Plateau

The original review assumed pre-training would continue to grow exponentially at 4.5x per year. The revision replaces that with a plateau: rapid early gains, levelling off by 2026-2027.

Justification:

• 2024 Orion plateau: Empirical evidence of saturation at ~1.2-1.5x GPT-4 level
• Data exhaustion: Cannot continue exponential growth without high-quality training data
• Multiple labs converging: OpenAI, Google, Anthropic all hitting similar ceilings
• Logarithmic returns: arXiv:2412.16443 documents structural diminishing returns

Uncertainty Range: ±50% (peak could be 1.25x-2.0x GPT-4, but NOT 10x or 100x)

Timeline:

• 2024-2025: Rapid deceleration phase (from 5.9-month to 12+ month doubling)
• 2026-2027: Effective plateau (gains <10% per year)
• 2028+: New paradigm required (quantum computing, neuromorphic, or fundamentally different approach)

2.2 Efficiency: Conservative with High Uncertainty

The original review assumed efficiency gains of 2x-3x per decade, extrapolated from a claim of 23x over 2.5 years. The revision uses 1.5x-2x per decade.

Justification:

Why NOT 23x over 2.5 years (9x per decade)?

1. Non-peer-reviewed source: blog.arcade.dev, not validated
2. Historically unprecedented: Exceeds Moore's Law by 4-5x (never sustained)
3. Conflates one-time gains with ongoing rate: Many cited techniques (MoE, KV caching) are one-time architectural improvements, not continuous compounding

Why 1.5x-2x per decade (conservative)?

1. Historical algorithmic progress: ImageNet achieved 44x efficiency over 7 years (~6.3x per decade), but this slowed after low-hanging fruit
2. Moore's Law analogy: 2x per 18 months = 4.9x per decade for hardware, but algorithmic gains historically SLOWER than hardware
3. Extensively mined techniques: "AI Scaling: From Up to Down and Out" (arXiv:2502.01677, 2025) notes most obvious optimizations already deployed

Critical Caveat: Even 1.5x-2x per decade is speculative. Could plateau at 1.0x (no further gains) if fundamental limits reached.

Uncertainty Range: ±100% (could range from 1.0x to 3x per decade, 2025-2035 period)

2.3 Test-Time Compute: Economic Deployment Gate

The original review treated test-time compute as broadly applicable across tasks. The revision treats it as limited by cost and by problem difficulty.

Justification:

Why economic gating?

• o3 high-compute ($1,000+) is technically impressive but economically unviable for 99.9% of tasks
• Market research: Most AI applications have willingness-to-pay <$1 per task
• Historical pattern: Technologies that cost 10,000x more than alternatives remain niche (supercomputers, particle accelerators)

Why logarithmic benefit scaling?

• Hugging Face H4 research: "Test-time compute hits limitations on challenging questions"
• Diminishing returns: $5 → $50 provides more benefit than $500 → $5,000
• No evidence for exponential returns to inference compute

Why difficulty threshold?

• Empirical: Test-time compute helps with complex reasoning, not simple tasks
• Economic: For easy tasks, pre-trained capability already sufficient (no need for $1,000 inference)
• ARC-AGI context: Benchmark specifically selects difficult reasoning problems (unrepresentative of typical deployment)

Deployment Distribution (2025-2035 projection):

• $5-$10 per task: 10-20% of commercial AI applications (standard reasoning)
• $10-$100 per task: 1-5% (specialized professional tools)
• $100-$1,000 per task: 0.1-1% (critical decisions, high-value research)
• $1,000+ per task: <0.01% (existential decisions, one-time breakthroughs)

Effective Capability Impact:

• Technical ceiling raised by 2-3x (o3 vs GPT-4 on hard problems)

2.4 Uncertainty

Epistemic Uncertainty Sources:

1. Pre-training plateau depth: Is 1.5x the true ceiling or could new architectures reach 2-3x?
2. Efficiency gains sustainability: Will 5% annual continue or hit limits?
3. Test-time compute economics: Will costs drop 10x (enabling broader deployment) or stay expensive?
4. Paradigm shifts: Quantum computing, neuromorphic, or other breakthroughs unpredictable
5. Regulatory constraints: Compute caps, energy limits, safety restrictions could slow deployment

3. Economic Deployment Constraints

3.1 Cost-Performance Frontier (2025-2035)

Current Frontier (2024-2025):

Key Insight: Market demand is inverse exponential with cost. Each 10x cost increase reduces addressable market by ~10x.

3.2 Willingness-to-Pay Distribution

Empirical Data (SaaS pricing research, enterprise AI adoption surveys):

• Consumer applications: $0-$0.50 per task (chatbots, content generation, tutoring)
• Professional tools: $1-$20 per task (coding assistance, legal research, data analysis)
• Specialized enterprise: $20-$200 per task (drug discovery, chip design, strategic planning)
• Critical decisions: $200+ per task (one-time use cases, existential risk assessment)

Distribution: Approximately log-normal with median ~$2, mean ~$10, 99th percentile ~$500

Implication: Even at $100 per task, only ~1% of potential AI use cases are economically viable.

Why This Matters:

• Market-weighted capability is dominated by cheap models, not cutting-edge expensive ones

3.3 Cost Trajectory Projections (2025-2035)

Optimistic Scenario (Moore's Law for inference):

• Inference costs decline 2x per 18 months (historical GPU trend)
• 2035: o3-tier performance available for $50-$100 per task
• Market accessibility increases from 0.15% to 2-5%

Baseline Scenario (Moderate cost decline):

• Inference costs decline 30% per year (algorithmic + hardware)
• 2035: o3-tier performance available for $200-$400 per task
• Market accessibility increases from 0.15% to 0.5-1%

Pessimistic Scenario (Cost floor):

• Test-time compute has intrinsic energy cost floor (~$50 per task for o3-level reasoning)
• Minimal cost reduction possible without efficiency breakthroughs
• 2035: o3-tier remains $500+ per task, <0.5% market

4. Domain-Specific Effectiveness

4.1 Heterogeneity in Task Benefit

Critical Oversight in Original Research: Assumed all AI capabilities scale uniformly across domains. Evidence contradicts this.

Task Categories and Scaling Response:

4.2 Alignment-Relevant Metrics (Inverse Scaling)

Evidence of Inverse Scaling:

1. Truthfulness: Larger models better at "sounding plausible" than being accurate
2. Sycophancy: Tendency to agree with user increases with model size
3. Sandbagging: Capability to hide true performance increases with scale
4. Deceptive alignment: Theoretical risk increases with model sophistication

Why This Matters:

• Traditional capability metrics (benchmark performance) may anticorrelate with alignment
• Solving alignment is NOT automatic with capability scaling
• May require explicit safety research breakthroughs independent of capability growth

5. Interaction Effects and Non-Independence

The original review treated pre-training, efficiency and test-time compute as independent multipliers. These axes are NOT independent.

5.1 Documented Interaction Effects

Test-Time Compute Depends on Pre-Training Quality:

• o1/o3 reasoning chains only work because base model (GPT-4 level) has strong priors
• Weaker base models cannot effectively use test-time compute
• Implication: Test-time benefit is CONDITIONAL on pre-training reaching threshold

Efficiency Gains May Trade Off with Capability:

• Quantization reduces precision → capability loss for some tasks
• Pruning removes "redundant" parameters → may eliminate subtle capabilities
• Speculative decoding speeds up simple responses, but may degrade complex reasoning
• Implication: Efficiency and capability may be NEGATIVELY correlated for frontier performance

Pre-Training Plateau May Limit Efficiency Techniques:

• Many efficiency gains come from better architectures discovered during scaling
• If scaling stalls, pipeline of new architectures slows
• Implication: Efficiency gains may DEPEND on continued pre-training experimentation

6. Research Quality and Credibility Assessment

6.1 Evidence Hierarchy (Revised)

Tier 1: Peer-Reviewed, Large-N Empirical (A+ credibility)

• Sevilla & Roldán (2024) - Epoch AI 4.1x/year through 2024 (14 years data)
• McKenzie et al. (2024) - "Scaling Laws Do Not Scale" (inverse scaling evidence)
• Villalobos et al. (2024) - "Will We Run Out of Data?" (data exhaustion quantification)

Tier 2: Preprints, Reputable Authors (B+ credibility)

• Lu (2025) - arXiv:2501.02156 (efficiency-doubling framework, theoretical)
• "Has LLM Reached the Scaling Ceiling Yet?" - arXiv:2412.16443 (logarithmic returns documentation)
• "Revisiting Scaling Laws" - ACL 2025 (curvature in scaling laws at frontier)

Tier 3: Industry Reports, Multiple Sources (B credibility)

• Bloomberg (Nov 2024) - Orion/Gemini plateau (multiple employee sources)
• TechCrunch (Dec 2024) - o3 cost analysis (corroborated by ARC Prize)
• Reuters - Industry slowdown consensus

Tier 4: Single-Source, Non-Peer-Reviewed (C-D credibility)

• blog.arcade.dev - 23x efficiency claims (NOT USED in revised estimates)
• Hugging Face H4 - Test-time compute research (pre-publication, used cautiously)
• CEO statements (Nadella, Huang) - Directionally informative, not quantitative

Tier 5: Excluded (F credibility)

• Marketing materials, press releases without data
• Social media speculation
• Uncited claims in news articles

6.2 Confidence Scoring by Claim

6.3 Known Unknowns and Research Gaps

High-Priority Gaps (May Invalidate Conclusions):

1. Long-term test-time scaling laws: Does logarithmic benefit continue or plateau further?
2. Efficiency gain ceiling: Is 5% annual sustainable or will it hit limits?
3. Alternative paradigms: Quantum, neuromorphic, analog computing timelines unknown
4. Regulatory constraints: Compute caps, energy limits, international coordination unpredictable

Medium-Priority Gaps:

1. Domain-specific scaling heterogeneity: Need task-by-task empirical scaling curves
2. Interaction effect magnitudes: How much do axes interfere vs compound?
3. Cost decline trajectory: Will inference costs continue falling or hit floors?
4. Data quality vs quantity tradeoffs: Can synthetic data work with better techniques?

7. Summary of Changes from Original Research

7.1 Adjustments

7.2 Methodological Improvements

Added:

1. ✅ Economic deployment gating (cost-benefit analysis)
2. ✅ Domain-specific effectiveness (reasoning ≠ factual ≠ truthfulness)
3. ✅ Interaction effects (axes not fully independent)
4. ✅ Explicit uncertainty quantification (±50% near-term, ±200% long-term)
5. ✅ Contradictory evidence section (inverse scaling, data exhaustion, plateau)

Removed:

1. ❌ Non-peer-reviewed efficiency claims (23x over 2.5 years from blog.arcade.dev)
2. ❌ Unconstrained exponential growth assumptions
3. ❌ Test-time compute as universal solution (now problem-difficulty dependent)

7.3 Credibility Improvements

Original Research Issues:

• Selective citation (ignored contradictory evidence)
• Overly optimistic parameters (not conservative enough)
• Single-source claims without validation (blog sources)
• Insufficient uncertainty treatment

Revised Research Strengths:

• Comprehensive contradictory evidence section (6+ papers cited)
• Conservative parameters with peer-reviewed backing
• Explicit uncertainty bands widening over time
• Economic constraints modeled (not just technical feasibility)
• Domain-specific heterogeneity acknowledged
• Interaction effects between scaling axes

8. Citations and Sources

Peer-Reviewed / Authoritative (A-tier)

1. Sevilla, J., & Roldán, E. (2024). "Training compute of frontier AI models grows by 4-5x per year." Epoch AI. https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year

   • Used for: Historical 4.1x/year pre-training scaling (2010-2024)
   • Credibility: A+ (14 years empirical data, leading research org)

2. McKenzie, I. R., et al. (2024). "Scaling Laws Do Not Scale." arXiv:2307.03201.

   • Used for: Inverse scaling evidence (truthfulness, toxicity)
   • Credibility: A (peer-reviewed, replicated findings)

3. Villalobos, P., et al. (2024). "Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data." arXiv:2211.04325 (2024 update).

   • Used for: Data exhaustion timeline, synthetic data limitations
   • Credibility: A (widely cited, quantitative estimates)

4. "Has LLM Reached the Scaling Ceiling Yet?" arXiv:2412.16443. December 2024.

   • Used for: Logarithmic diminishing returns documentation
   • Credibility: A- (recent preprint, awaiting peer review)

5. Cottier, B., et al. (2024). "The rising costs of training frontier AI models." arXiv:2405.21015v2.

   • Used for: Cost growth 2.4x/year (2016-2023)
   • Credibility: A (45 models analyzed)

6. Epoch AI (2024). "Can AI scaling continue through 2030?" https://epoch.ai/blog/can-ai-scaling-continue-through-2030 (added Dec 12, 2025)

   • Used for: Infrastructure constraints (power, chips, latency), 2e29 FLOP 2030 projection, 1-5 GW single-campus datacenters
   • Key Findings:
     • Power is most restrictive constraint (median 2e29 FLOP by 2030)
     • Single-campus datacenters: 1-5 GW (Amazon 960 MW nuclear contract, Microsoft/OpenAI pursuing 5 GW)
     • Distributed networks: 2-45 GW US-based (4-20 Pbps inter-datacenter bandwidth)
     • Power infrastructure = 40% of GPU costs by 2030
     • Training duration: 2-9 months; resource capture: 10-40% of infrastructure
   • Credibility: A+ (Epoch AI authoritative on compute trends, multi-constraint quantitative analysis)

Theoretical / Preprints (B-tier)

7. Lu, C. (2025). "The Race to Efficiency: A New Perspective on AI Scaling Laws." arXiv:2501.02156.

   • Used for: Efficiency-doubling framework, theoretical limits
   • Credibility: B+ (arXiv preprint, theoretical model not empirically validated)

8. "Revisiting Scaling Laws for Language Models." ACL 2025. aclanthology.org/2025.acl-long.1163.pdf

   • Used for: Scaling law curvature at large scale
   • Credibility: A- (peer-reviewed, domain-limited)

9. "AI Scaling: From Up to Down and Out." arXiv:2502.01677. February 2025.

   • Used for: "Extensively mined" efficiency techniques comment
   • Credibility: B+ (preprint, not yet peer-reviewed)

Industry Reports (B-tier, directional)

10. Bloomberg (Nov 13, 2024). "OpenAI, Google and Anthropic Are Struggling to Build More Advanced AI."

• Used for: Orion plateau evidence, Gemini underperformance
• Credibility: B+ (major outlet, multiple employee sources, NOT peer-reviewed)

11. TechCrunch (Dec 23, 2024). "OpenAI's o3 suggests AI models are scaling in new ways — but so are the costs."

    • Used for: o3 cost data ($1,000+ per task)
    • Credibility: B (industry journalism, corroborated by ARC Prize)

12. ARC Prize (Dec 20, 2024). "OpenAI o3 Breakthrough High Score on ARC-AGI-Pub." https://arcprize.org/blog/oai-o3-pub-breakthrough

    • Used for: o3 performance (87.5% on ARC-AGI), high-compute cost
    • Credibility: A (independent benchmark, transparent methodology)

Companion Review

13. AI scaling slowdown evidence (Dec 10, 2025).
    • Used for: Comprehensive industry slowdown synthesis
    • Credibility: B+ (synthesizes multiple external sources)

Sources REJECTED (Not Used in the Revision)

• ❌ blog.arcade.dev - 23x efficiency claims (non-peer-reviewed, potential COI)
• ❌ CEO statements (Nadella, Huang) - Directionally interesting, not quantitative
• ❌ Marketing materials - Excluded entirely