Compute (AI Capabilities): Research Report

📋Page Status

Quality:3 (Stub)⚠️

Words:3.6k

Structure:

📊 15📈 0🔗 0📚 56•14%Score: 10/15

Executive Summary

Finding	Key Data	Implication
Training costs escalating	2.4x/year growth; GPT-4 cost $78M, Gemini Ultra $191M	Billion-dollar training runs by 2027; only well-funded orgs can compete
Supply chain concentration	TSMC: 92% of advanced chips; ASML: 100% of EUV machines	Single points of failure create governance opportunities and geopolitical risk
Hardware bottlenecks easing	H100 lead times dropped from 11 months (2023) to 8-12 weeks (2024)	Memory (HBM) now the binding constraint through 2027
Compute governance emerging	US: 10²⁶ FLOPs; EU: 10²⁵ FLOPs reporting thresholds	Compute is measurable, concentrated, physical—ideal governance lever
Energy demand doubling	Data centers: 415 TWh (2024) → 945 TWh (2030); AI: 35-50% of DC load	Infrastructure growth outpacing electricity supply; nuclear partnerships emerging

---

Research Summary

Compute has emerged as the most governable input to AI development because it is measurable, concentrated, and physical. Training costs for frontier AI models have grown at 2.4× per year since 2016, with GPT-4 costing approximately $78 million and Gemini Ultra around $191 million—projecting to billion-dollar training runs by 2027. The supply chain exhibits extreme concentration: TSMC manufactures 92% of advanced chips, ASML holds a monopoly on EUV lithography equipment, and NVIDIA controls roughly 80% of the AI accelerator market.

These chokepoints create natural governance leverage. The US, EU, and California have all implemented compute-based regulatory thresholds (10²⁵-10²⁶ FLOPs) that trigger reporting requirements. Hardware bottlenecks have shifted from GPU availability to high-bandwidth memory (HBM), with lead times dropping from 11 months to 8-12 weeks. However, energy constraints are intensifying: AI-driven data center demand is projected to consume 945 TWh by 2030, prompting major tech companies to pursue nuclear partnerships. The concentration of advanced chip manufacturing in Taiwan (TSMC) and lithography equipment in the Netherlands (ASML) creates significant geopolitical risk, driving US efforts to reshore semiconductor production through the CHIPS Act.

---

Background

Compute—the hardware resources required to train and run AI systems—has emerged as perhaps the most tractable lever for AI governance. Unlike algorithms (which can be shared instantly) or data (which is hard to track), compute is measurable (FLOPs, GPU-hours), concentrated (few chokepoints like TSMC, ASML, NVIDIA), and physical (can be tracked, controlled, and metered).

Why Compute Matters for AI Safety

The compute required to train frontier AI models has grown exponentially—doubling roughly every 6 months since 2012. This growth rate far exceeds Moore’s Law (doubling every 18-24 months), creating a “compute overhang” where the hardware investment for cutting-edge AI becomes the primary bottleneck. This makes compute a critical variable for both capability forecasting and governance design.

Training frontier models now costs tens to hundreds of millions of dollars in compute alone. Anthropic CEO Dario Amodei has stated that frontier AI developers are likely to spend close to a billion dollars on a single training run in 2025, with up to ten billion-dollar training runs expected in the next two years. This concentration of resources creates natural governance chokepoints.

---

Key Findings

Training Compute Costs: Exponential Growth

The cost trajectory for training frontier AI models has followed a remarkably consistent exponential trend:

Model	Year	Training Compute Cost	Notes
GPT-3	2020	$4-12M	Established LLM scaling paradigm
GPT-4	2023	$78M	Per Stanford AI Index 2024
Gemini Ultra	2024	$191M	Google’s flagship model
Projected (2027)	2027	>$1B	If 2.4x/year growth continues

Cost Breakdown

For frontier models like GPT-4 and Gemini, the most significant expenses are AI accelerator chips and staff costs, each costing tens of millions of dollars. Other notable costs include server components (15-22%), cluster-level interconnect (9-13%), and energy consumption (2-6%). Source: Epoch AI (2024)

Epoch AI’s analysis found that the amortized hardware and energy cost for the final training run of frontier models has grown at a rate of 2.4x per year since 2016 (95% CI: 2.0x to 3.1x). This rate significantly exceeds Moore’s Law and suggests that “given that total model development costs at the frontier are already over $100 million, these advances may only be accessible to the largest companies and government institutions.”

OpenAI’s 2024 compute spending illustrates the scale: $3 billion on training compute, $1.8 billion on inference compute, and $1 billion on research compute amortized over multiple years (Epoch AI, 2024).

Scaling Laws: From Training to Inference

While traditional scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022/Chinchilla) focused on training compute, recent research has expanded to inference scaling laws:

Scaling Dimension	Key Finding	Implication
Training	Performance ∝ compute^α (α ≈ 0.5-0.7)	Predictable capability growth with compute
Inference	Test-time compute can be more efficient than parameter scaling	Smaller models + advanced inference may be Pareto-optimal
Architecture	MLP-to-attention ratio, GQA affect inference cost	Conditional scaling laws needed for deployment
Efficiency	”Densing law”: capability density doubles every 3.5 months	Same capability with exponentially fewer parameters over time

The Inference Revolution

Research on inference scaling laws (arXiv:2408.00724) found that scaling inference compute with strategies like tree search, best-of-n, and weighted voting can be more computationally efficient than scaling model parameters. This suggests a shift from “bigger models” to “smarter inference” as a cost-optimization strategy.

The “densing law” published in Nature Machine Intelligence states that capability density (capability per parameter) doubles approximately every 3.5 months, indicating that equivalent model performance can be achieved with exponentially fewer parameters over time. This has significant implications for compute efficiency and deployment costs.

Hardware Supply Chain: Critical Chokepoints

The AI compute supply chain exhibits extreme concentration at multiple levels:

TSMC: The Chip Manufacturing Bottleneck

Metric	Value	Source
Market share (advanced chips)	92%	AILAB Blog (2025)
Geographic concentration	Single island (Taiwan), 13,976 sq mi	Global Taiwan Institute
Economic impact of disruption	$10 trillion (10% of global GDP)	Verdantix (2025)
Key customers	Apple, NVIDIA, Qualcomm, Samsung, AMD	Industry analysis

The Taiwan Chokepoint

Taiwan manufactures over 60% of the world’s semiconductors and almost 90% of the most sophisticated chips. If China were to invade Taiwan and destroy TSMC in a best-case scenario, or capture the company in a worst-case scenario, China would effectively neutralize global access to advanced semiconductors and secure a decisive advantage in AI development. Source: Tandfonline (2025)

TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) packaging technology has emerged as the specific bottleneck for AI chips. The production bottleneck lies primarily in TSMC’s CoWoS packaging process, which cannot scale fast enough. Additionally, the TRX5090 substrate—a critical component binding the GPU core to its high-bandwidth memory—is in extremely limited supply, with only a handful of manufacturers in Japan and Taiwan able to produce it at required precision and volume.

TSMC is diversifying with plans for six semiconductor fabs in Arizona, plus expansion in Japan and Germany. However, diversification is not feasible in the short term due to high reshoring costs and talent acquisition challenges (NPR, 2025).

ASML: The EUV Lithography Monopoly

Metric	Value	Source
EUV market share	100%	CNBC (2022)
Machine cost	$150-400M (High-NA: $370M+)	Industry reports
R&D investment	$9B over 30 years	xLight analysis
Parts per machine	>100,000	Technical specifications
Suppliers	800+ globally	Supply chain analysis

ASML builds 100% of the world’s extreme ultraviolet lithography machines, without which cutting-edge chips are simply impossible to make. The Dutch company is the sole supplier of EUV machines, winning a 30-year race that granted a monopoly on selling the tool essential for fabricating leading-edge semiconductors (Strange VC (2025)).

Technical Barriers to Entry

ASML’s monopoly stems from extraordinary technical requirements: EUV light wavelength of 13.5 nanometers (1/5000th width of human hair), mirrors polished smooth to within a single atom, and operation in vacuum. The whole system requires components from over 800 suppliers across the globe, including world’s most advanced optics from Germany’s Zeiss, which took 15 years to develop. Source: Mazi Asset Management

Potential competition: Pat Gelsinger (ousted Intel CEO) as executive chairman of xLight—a startup founded in 2024—is developing free-electron lasers driven by compact particle accelerators as an alternative to ASML’s laser-produced plasma. The Trump administration injected up to $150 million into xLight from the 2022 CHIPS and Science Act, though this represents early-stage R&D with uncertain timeline (24/7 Wall St., 2025).

NVIDIA and Memory Bottlenecks

GPU shortages dominated 2022-2023, but the situation has evolved:

Period	H100 Lead Time	Binding Constraint
2023 (peak shortage)	11 months	GPU chip supply
Early 2024	3-4 months	CoWoS packaging
Late 2024	8-12 weeks	High-bandwidth memory (HBM)
2025-2027	Variable	Memory supply (SK Hynix, Samsung, Micron)

Memory Becomes the Bottleneck

A deeper supply crisis emerged as memory chips became the binding constraint on AI infrastructure globally. High-bandwidth memory (HBM) hit severe shortages as manufacturers Samsung, SK Hynix, and Micron operated near full capacity while reporting six-to twelve-month lead times. SK Hynix told analysts that shortages may persist until late 2027, reporting that all memory scheduled for 2026 production is already sold out. Source: AI News (2025)

NVIDIA’s market dominance is reflected in financials: Q3 fiscal 2026 total revenue reached $57.0 billion, with data center operations accounting for $51.2 billion—90% of the company’s entire business. An NVIDIA H100 AI accelerator sells for $25,000-40,000, giving the company unusual pricing power during shortage periods (BattleforgePC (2025)).

Compute Governance: Regulatory Frameworks

Compute thresholds have emerged as a central mechanism for AI governance across multiple jurisdictions:

Jurisdiction	Threshold	Reporting Requirements	Status
US (EO 14,110)	10²⁶ FLOPs (10²³ for bio)	Notify government, report security measures, share red-team results	Revoked by Trump EO 14,148; rules proposed
EU AI Act	10²⁵ FLOPs	Notify Commission, perform evaluations, assess systemic risks, report incidents	Active; affects 5-15 companies
California (SB 53)	10²⁶ FLOPs + $500M revenue	Disclose “frontier AI framework,” report catastrophic risk assessments quarterly	Enacted 2024
New York (RAISE Act)	Revenue-based (removed compute threshold)	Report critical incidents within 72 hours	Signed into law 2025

Why Compute Thresholds Work

Compute thresholds are currently the best tool available for identifying potentially risky GPAI models because they are: (1) based on a risk-correlated, quantifiable metric; (2) difficult to circumvent; and (3) measurable before model development and deployment. However, they serve as a proxy for risk rather than a precise measure, requiring complementary capability evaluations. Source: Institute for Law & AI (2024)

Rationale: Training compute can serve as a proxy for the capabilities of AI models. A compute threshold operates as a regulatory trigger, identifying which models might possess more powerful and dangerous capabilities that warrant greater scrutiny (GovAI Research Paper).

Limitations: The debate has centered mostly around a single training compute threshold, but governments could adopt a pluralistic and risk-adjusted approach by introducing multiple compute thresholds that trigger different measures according to degree or nature of risk. Some proposals recommend a tiered approach that would create fewer obligations for models trained on less compute (Institute for Law & AI, 2024).

Export Controls and Geopolitics

The US has implemented increasingly strict export controls on advanced computing chips to China, with significant implications:

Policy Action	Target	Impact	Assessment
October 2022	Advanced chips (A100, H100) to China	Created incentive for compute-efficient algorithms	Accelerated Chinese innovation
October 2023	Expanded chip restrictions	Cloud computing loopholes remain	Limited effectiveness
December 2024	High-bandwidth memory (HBM) restrictions	Targets deployment compute for reasoning models	Addresses inference scaling
AI Diffusion Framework	Three-tier country system	Byzantine limits for ~150 middle-tier countries	Critiqued as overreach

The Efficiency Paradox

DeepSeek’s models are a stark illustration of why U.S. export controls on advanced computing chips, instead of impeding China’s AI progress, may actually be accelerating it. As a direct result of U.S. controls on advanced chips, companies in China are creating new AI training approaches that use computing power very efficiently. When China develops the ability to produce its own leading-edge chips, it will have a powerful combination of both computing capacity and efficient algorithms. Source: Brookings (2025)

Hardware-Enabled Governance Mechanisms (HEMs): RAND researchers introduced the concept of HEMs, which could be installed on chips otherwise prohibited from export. HEMs could provide “some level of confidence that they could not be misused” through technical enforcement (RAND (2024)). However, HEMs face significant threat vectors and would require robust protection measures.

Cloud Computing Alternative: Providing compute as a service offers superior governance opportunities compared to chip export controls. Unlike AI chips accumulated over time, cloud access provides point-in-time computing power that can be restricted or shut off as needed—making it a more precise tool for oversight (Brookings (2024)).

Energy Consumption: The Infrastructure Challenge

AI’s compute demands translate directly into energy consumption at unprecedented scale:

Metric	2024	2030 (Projected)	Growth Rate
Global data center electricity	415 TWh (1.5% of global)	945 TWh (3% of global)	15%/year
US data center electricity	183 TWh (4% of US total)	~320 TWh (8.6% of US total by 2035)	Doubling by 2035
AI share of DC power	5-15%	35-50%	AI-specific servers: 30%/year
AI-specific servers	53-76 TWh	165-326 TWh	4.3x growth

Energy Growth Context

From 2024 to 2030, data centre electricity consumption grows by around 15% per year, more than four times faster than the growth of total electricity consumption from all other sectors. BloombergNEF forecasts US data-center power demand will more than double by 2035, with average hourly electricity demand nearly tripling from 16 GWh (2024) to 49 GWh (2035). [Sources: IEA (2024), BloombergNEF]

Energy sources (US, 2024): Natural gas supplied over 40% of electricity for data centers, renewables (wind/solar) about 24%, nuclear around 20%, and coal around 15% (Pew Research, 2025).

Environmental footprint: Company-wide metrics from environmental disclosures suggest that AI systems may have a carbon footprint equivalent to that of New York City in 2025. In 2023, US data centers directly consumed about 17 billion gallons of water, with hyperscale facilities expected to consume 16-33 billion gallons annually by 2028 (ScienceDirect (2025)).

Infrastructure investment: The data center real estate build-out has reached record levels based on select major hyperscalers’ capital expenditure, trending at roughly $200 billion as of 2024 and estimated to exceed $220B by 2025 (Deloitte (2025)).

---

Causal Factors

The following factors influence AI compute availability, cost, and governance effectiveness. This analysis is designed to inform future cause-effect diagram creation for the AI Transition Model.

Primary Factors (Strong Influence)

Factor	Direction	Type	Evidence	Confidence
Scaling Laws	↑ Compute Demand	cause	2.4x/year cost growth; predictable capability returns	High
Supply Chain Concentration	↑ Governance Tractability	intermediate	TSMC 92%, ASML 100% create chokepoints	High
Training Cost Escalation	↓ Actor Diversity	intermediate	$100M+ limits to well-funded orgs; billion-dollar runs by 2027	High
Hardware Bottlenecks	↓ Capability Growth Rate	leaf	Memory (HBM) shortages through 2027	High
Energy Infrastructure	↓ Deployment Speed	leaf	15%/year DC growth exceeds grid capacity in some regions	High

Secondary Factors (Medium Influence)

Factor	Direction	Type	Evidence	Confidence
Compute Thresholds	↑ Regulatory Oversight	intermediate	US/EU/CA frameworks at 10²⁵-10²⁶ FLOPs	Medium
Export Controls	Mixed Effect	leaf	May accelerate efficient algorithms (DeepSeek example)	Medium
Algorithmic Efficiency	↓ Compute Demand	cause	”Densing law”: 2x capability density every 3.5 months	Medium
Inference Scaling	↑ Deployment Compute	cause	Test-time compute increasingly important (o1, r1 models)	Medium
Geopolitical Tensions	↑ Supply Risk	leaf	Taiwan Strait conflict would disrupt 92% of advanced chips	Medium

Minor Factors (Weak Influence)

Factor	Direction	Type	Evidence	Confidence
Cloud Governance	↑ Oversight Precision	intermediate	Point-in-time control vs. chip accumulation	Low
Hardware-Enabled Mechanisms	↑ Export Flexibility	intermediate	RAND proposal; unproven threat model	Low
ASML Competition	↓ Monopoly Risk	leaf	xLight startup; early stage ($150M funding)	Low
TSMC Diversification	↓ Geographic Risk	leaf	Arizona/Japan/Germany fabs; long timeline	Low

---

Scenario Variants

Compute dynamics could evolve along several distinct pathways with different implications for AI safety:

Scenario	Mechanism	Timeline	Warning Signs	Governance Implications
Compute Overhang	Algorithmic breakthroughs make existing compute far more capable	2-5 years	Efficiency gains exceed hardware scaling; DeepSeek-style innovations	Thresholds become unreliable proxies
Hardware Plateau	Physical limits (energy, memory, lithography) constrain scaling	5-10 years	Slowing Moore’s Law; energy grid bottlenecks	Increased focus on algorithmic safety
Geopolitical Disruption	Taiwan conflict disrupts TSMC; China controls advanced chips	1-10 years	Escalating Taiwan Strait tensions	Western AI capabilities severely degraded
Decentralized Compute	Distributed training across many small GPUs becomes viable	3-8 years	Successful federated learning for frontier models	Governance via chokepoints becomes infeasible
Energy Bottleneck	Grid capacity limits data center expansion before compute saturation	3-7 years	Brownouts near mega-clusters; nuclear partnerships stall	Natural brake on capability growth

Scenario Interactions

These scenarios are not mutually exclusive. For example, geopolitical disruption could accelerate both compute overhang (drive for efficiency) and hardware plateau (supply constraints). The energy bottleneck scenario may be the most likely near-term constraint based on current infrastructure timelines.

---

Open Questions

Key Uncertainties

These questions represent the highest-value areas for follow-up research and have significant implications for AI governance strategy.

Question	Why It Matters	Current State
How robust are compute thresholds to algorithmic progress?	DeepSeek achieved competitive results with less than 10²⁵ FLOPs	Thresholds are static; efficiency gains accelerating
What is the energy ceiling for AI?	May be binding constraint before chip supply	Projections vary widely; grid capacity unclear
Can TSMC diversification succeed in time?	Arizona fabs won’t reach volume until late 2020s	Geopolitical risk timeline uncertain
Will inference scaling change governance calculus?	Shifts compute from training (one-time) to deployment (ongoing)	Reasoning models (o1, r1) show importance; December 2024 HBM controls respond
How effective are export controls?	May accelerate rather than impede Chinese AI progress	DeepSeek case study suggests efficiency paradox
What are the limits of hardware-enabled governance?	Could enable export of otherwise-restricted chips	Threat model unproven; RAND early-stage research
Will memory (HBM) remain the bottleneck?	Determines whether GPU shortages return	SK Hynix: shortages through late 2027
Can cloud-based governance scale globally?	More precise than chip controls but requires infrastructure	Loopholes exist; university research concerns

---

Sources

Academic Papers & Preprints

• Epoch AI (2024). “The rising costs of training frontier AI models” - Comprehensive cost analysis; 2.4x/year growth rate
• arXiv (2024). “Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference” - Test-time compute optimization
• arXiv (2024). “Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs” - Conditional scaling laws
• Nature Machine Intelligence (2025). “Densing law of LLMs” - Capability density doubling every 3.5 months
• arXiv (2024). “Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models” - MoE efficiency metrics

Policy & Governance

• Institute for Law & AI (2024). “The Role of Compute Thresholds for AI Governance” - Comprehensive threshold analysis
• RAND (2024). “Hardware-Enabled Governance Mechanisms” - Technical solutions for export control
• GovAI (2024). “Training Compute Thresholds: Features and Functions in AI Regulation” - Regulatory framework analysis
• Brookings (2024). “The new AI diffusion export control rule will undermine US AI leadership” - Critique of three-tier system
• Brookings (2025). “DeepSeek shows the limits of US export controls on AI chips” - Efficiency paradox analysis
• RAND (2025). “Understanding the Artificial Intelligence Diffusion Framework” - Supply chain controls

Industry Analysis & Data

• Epoch AI (2024). “How much does it cost to train frontier AI models?” - GPT-4: $78M, Gemini Ultra: $191M
• Epoch AI (2024). “Most of OpenAI’s 2024 compute went to experiments” - $3B training, $1.8B inference
• Tom’s Hardware (2024). “NVIDIA’s H100 AI GPU shortages ease” - Lead time evolution
• AI News (2025). “AI Chip Shortage 2025: What CTOs Learned the Hard Way” - HBM bottleneck through 2027
• CFR (2025). “China’s AI Chip Deficit: Why Huawei Can’t Catch Nvidia” - Huawei 5% of NVIDIA capacity in 2025

Semiconductor Supply Chain

• AILAB Blog (2025). “The $10 Trillion Chokepoint: How One Company Powers the AI Revolution” - TSMC 92% market share
• Tandfonline (2025). “China’s semiconductor conundrum: understanding US export controls” - Taiwan invasion scenarios
• ScienceDirect (2025). “From vulnerabilities to resilience: Taiwan’s semiconductor industry” - $10T disruption cost
• NPR (2025). “As political winds shift, top chipmaker TSMC looks beyond Taiwan” - Arizona/Japan/Germany expansion
• CNBC (2022). “Inside ASML, the company advanced chipmakers use for EUV lithography” - 100% EUV monopoly
• Strange VC (2025). “ASML’s 30-Year Monopoly: The Moonshot Bet No One Can Replicate” - $9B R&D, 30-year race
• 24/7 Wall St. (2025). “Monopoly No More? ASML May Suddenly Have a New Competitor” - xLight startup, $150M CHIPS Act funding

Energy & Infrastructure

• IEA (2024). “Energy demand from AI” - 415 TWh (2024) → 945 TWh (2030)
• Pew Research (2025). “What we know about energy use at U.S. data centers amid the AI boom” - 183 TWh US (4% of total)
• BloombergNEF (2025). “Power for AI: Easier Said Than Built” - 16 GWh → 49 GWh by 2035
• Deloitte (2025). “As generative AI asks for more power, data centers seek cleaner energy” - $220B capex in 2025
• ScienceDirect (2025). “The carbon and water footprints of data centers and AI” - 17B gallons water (2023), 16-33B (2028)

---

AI Transition Model Context

Model Integration

This research directly informs the Compute (AI Capabilities) factor in the AI Transition Model. Compute is a fundamental driver of AI capabilities and exhibits unique governance properties: it is measurable, concentrated in few chokepoints, and physical (unlike algorithms or data). These properties make compute a critical leverage point for influencing the AI transition trajectory.

Connections to Other Model Elements

Model Element	Relationship	Key Insights
AI Capabilities (Algorithms)	Complementary	Algorithmic efficiency (densing law) reduces compute requirements; may undermine threshold-based governance
AI Capabilities (Adoption)	Enabling	Training costs ($100M+) create barriers to entry; energy infrastructure limits deployment speed
AI Ownership (Companies)	Concentrating	Only well-funded organizations can afford frontier models; drives consolidation
AI Ownership (Countries)	Geopolitical lever	Export controls and supply chain chokepoints (TSMC, ASML) create interstate competition
Misalignment Potential (AI Governance)	Primary intervention point	Compute thresholds enable reporting, evaluations, audits before deployment
Misuse Potential	Limiting factor	High training costs reduce rogue actor capabilities (though inference compute different)
Transition Turbulence (Racing)	Accelerator	Shortages and strategic competition increase pressure for rapid deployment
Civilizational Competence (Governance)	Test case	Effectiveness of compute governance indicates broader governance capacity

Strategic Implications

The research reveals several strategic considerations for the AI transition:

1. Governance window closing: If algorithmic efficiency continues to double capability density every 3.5 months (densing law), compute thresholds will become less reliable proxies for risk within 2-3 years. This suggests urgency in establishing complementary governance mechanisms.

2. Energy as natural brake: Infrastructure constraints (15%/year data center growth vs. slower grid expansion) may limit capability growth independent of policy choices. This could provide additional time for governance development but may also increase racing incentives.

3. Geopolitical fragility: 92% concentration in Taiwan creates catastrophic downside risk. TSMC diversification timeline (late 2020s for volume production) may not align with AI capabilities timeline (potentially transformative AI by 2027-2030).

4. Efficiency paradox: Export controls may accelerate development of compute-efficient algorithms, potentially making frontier capabilities accessible to a wider range of actors. This suggests limits to supply-side interventions.

5. Inference shift: Growing importance of deployment/inference compute (o1, r1 reasoning models) changes governance focus from one-time training runs to ongoing operational compute. Cloud-based governance may be more effective for this regime.

The compute landscape is evolving rapidly, with multiple uncertainty dimensions (algorithmic efficiency, hardware bottlenecks, energy constraints, geopolitical shocks) that could significantly alter the AI transition trajectory. Adaptive governance mechanisms that can respond to these shifts will be critical.