Why Is Suleyman's Exponential Argument So Seductive?

In his April 8, 2026 essay for MIT Technology Review, Suleyman uses a simple analogy: humans evolved for a linear world, but AI is exponential. He notes that training compute for frontier models has grown 10x every 18 months since 2012, citing data from Epoch AI. This is true—but it's also a story of diminishing returns. The marginal gains from adding more compute are declining, as shown by the fact that GPT-4 used 10x more compute than GPT-3 but didn't deliver 10x the capability. Suleyman's argument works as a rhetorical device, but it's a strategic map drawn with too broad a brush.

What Physical Limits Does Suleyman Ignore?

Here's the hard truth: training a single frontier model now consumes roughly 50 GWh of electricity—equivalent to the annual consumption of 5,000 US homes. The world's total data center power capacity is projected to hit 100 GW by 2027, up from 20 GW in 2023, according to the International Energy Agency. But semiconductor fabrication capacity for the most advanced chips (like NVIDIA's H200 and B200) is not growing exponentially—it's constrained by ASML's EUV lithography machine output, which is increasing at only 15-20% per year. Suleyman's essay never mentions these physical bottlenecks. He treats compute as a frictionless resource, which is the same mistake that led to the 2024-2025 AI winter scare.

Who Actually Benefits If Suleyman Is Wrong?

The winners are companies that optimize for inference efficiency rather than brute-force training. Groq, with its LPU architecture, claims 10x lower latency and 5x lower energy per token than NVIDIA GPUs for inference. Edge AI startups like Syntiant and Hailo are building chips that run models at milliwatts, not kilowatts. If the exponential training curve flattens, these companies will capture the market for real-time, on-device AI. The losers are hyperscalers like Microsoft and Google, which have committed over $100 billion combined to build massive training clusters. If those clusters become underutilized because the next generation of models doesn't deliver proportional gains, they will face billions in stranded assets.

What Does This Mean for AI Strategy in 2026-2027?

The practical implication is clear: companies should stop investing in massive training runs for models that are already commoditizing. Instead, they should focus on fine-tuning smaller models for specific tasks—a strategy that yields 80% of the performance at 10% of the cost. This is already happening: Mistral AI's Mixtral 8x7B, which uses a mixture-of-experts architecture, achieves GPT-3.5-level performance with 5x less compute. The market for general-purpose foundation models is consolidating around OpenAI, Anthropic, and Google—everyone else should pivot to vertical applications.

1. Microsoft will announce a $5-10 billion write-down on AI training infrastructure by Q1 2027, citing lower-than-expected utilization of its H200 clusters for next-generation model training.
2. Groq will IPO by Q4 2027 at a valuation exceeding $15 billion, driven by demand for inference-efficient chips that undercut NVIDIA's power consumption per token by 5x.
3. The EU AI Office will mandate energy-efficiency reporting for all foundation model training runs by Q3 2027, forcing companies to disclose power consumption and carbon footprint, accelerating the shift to inference optimization.

• April 2026: Mustafa Suleyman publishes essay in MIT Technology Review arguing AI won't hit a wall due to exponential trends.
• June 2026: International Energy Agency releases report showing data center power demand will reach 100 GW by 2027, up from 20 GW in 2023.
• Q1 2027: Predicted hyperscaler write-down on training infrastructure due to underutilization.
• Q3 2027: Predicted EU AI Office mandate for energy-efficiency reporting.
• Q4 2027: Predicted Groq IPO.