Published December 8, 2025 · Updated December 8, 2025
Google DeepMind CEO Demis Hassabis stated in a recent Business Insider interview that massive scaling of compute, data and model architecture remains the most direct route toward artificial general intelligence (AGI). But he warned that the industry is now hitting hard boundaries: spiraling compute costs, environmental pressures and diminishing returns.
The remarks reignite one of the most important debates in AI today — whether scaling alone is enough to reach AGI, and what happens when the industry reaches the physical and economic limits of exponential growth.
Key Takeaways
- DeepMind CEO says “scaling remains the most promising path” toward AGI.
- Larger multimodal models continue to show emergent abilities.
- But scaling faces limits: soaring compute costs, hardware bottlenecks and energy consumption.
- Hassabis says the industry must innovate beyond brute-force scaling while keeping safety in focus.
- The debate reflects deeper questions about AGI timelines, sustainability and governance.
Explore More
Learn more about AGI, future AI capabilities and scaling laws in Arti-Trends’ core hubs:
- AI Guides Hub — technical breakdowns, scaling laws & model evolution
- AI Tools Hub — how scaling shapes real-world model performance
- AI News Hub — fast updates on frontier labs & breakthroughs
- AI Investing Hub — compute, hardware and infrastructure impact on markets
These hubs help readers understand both the opportunities and constraints of modern AI.
Why scaling still matters
Hassabis explained that the most powerful breakthroughs at DeepMind and Google Research in the past five years — multimodal reasoning, improved planning capability, scientific problem-solving — came primarily from:
- larger datasets
- more powerful accelerator hardware
- smarter architectures (e.g., Mixture-of-Experts)
- huge training runs at unprecedented scale
These scaling runs continue to reveal emergent intelligence: abilities that smaller models cannot express.
He emphasized that DeepMind still sees scaling as a reliable research direction — but not a guaranteed road to safe AGI.
The limits of brute-force scaling
Hassabis pointed to three key bottlenecks slowing the pace of exponential growth:
1. Compute costs exploding
Frontier training runs are projected to require billions of dollars by 2027–2028.
Even Big Tech may struggle to sustain this trajectory.
2. Hardware & supply chain bottlenecks
Nvidia GPUs, HBM3e memory and advanced packaging remain scarce worldwide.
This constrains training schedules and model release cycles.
3. Environmental & sustainability concerns
Training frontier LLMs requires massive energy consumption — raising political and ethical pressure.
Governments may eventually regulate extreme compute usage.
Together, these factors make pure scaling less sustainable long-term, forcing labs to explore hybrid approaches.
Beyond scaling: what comes next?
Hassabis says the next phase of AI progress will likely require:
- algorithmic breakthroughs
- more efficient architectures
- better reasoning mechanisms
- improved memory & planning
- alignment-focused research
- hybrid symbolic-neural approaches
- governments setting guardrails around compute thresholds
AGI may still come from scaling, but breakthroughs in efficiency and safety will decide whether large models remain viable.
Ethical & economic implications
The scaling debate raises profound questions:
For society
- Will AGI be accessible to everyone, or only to companies with billion-dollar compute clusters?
For the environment
- Can the world sustain the energy demands of trillion-parameter models?
For governance
- Should there be global limits or transparency rules for compute usage?
For businesses
- Scaling may reshape markets as smaller players lose access to training resources.
What happens next
DeepMind is expected to continue training larger multimodal models while simultaneously pursuing more efficient architectures.
Regulators in the US, EU and UK are also evaluating policies that may directly impact frontier scaling — including compute monitoring and resource transparency.
The question is no longer whether scaling leads toward AGI…
But how far the world is willing — or able — to scale.
Source
Business Insider — DeepMind CEO says scaling is the path to AGI, but the limits are becoming impossible to ignore
(December 2025)
Explore more
For deeper insights into AGI, scaling laws and compute evolution, you can explore:
