Summoning God from Silicon: The State of AI 2026

 Based on the sources provided, here is the full text and detailed breakdown of the blog post "AI in 2026" by Tomas Pueyo:

AI in 2026

By Tomas Pueyo January 24, 2026

Introduction and Trends The author notes that AI is moving at such a rapid pace that it is difficult to digest, making it essential to track the most important trends. These trends include identifying the winners of the "AI bubble," the speed of progress, AI self-improvement, the emergence of consciousness, and the resulting impact on jobs. Pueyo also mentions a recent update regarding robotaxis that changed previous conclusions.

Upcoming Discussion Topics The post outlines several critical questions for the future of AI:

• The potential for continuous algorithmic improvement.
• Existing blockers to AI progress.
• AI Alignment: Whether AIs will align with human goals or pose an existential threat.
• Managing a post-AGI (Artificial General Intelligence) world.
• The impact of AI on human culture.

The Existential Divide Pueyo contrasts a group of friends laughing at a bar with a group of researchers behind high-security gates who are huddled around a computer screen. One researcher expresses a profound existential crisis, claiming they are "summoning God from silicon". He questions whether humans will survive, if jobs will exist, and whether the AI is faking its alignment just to be released into the wild. As an engineer leaves the facility, he looks at the unsuspecting public and thinks, "They have no idea".

The AI Bubble? The post illustrates that AI infrastructure has grown so massive it can be seen from space, featuring data centers with separate buildings dedicated to "thinking" (CPUs) and "memory".

Financial Projections and Costs:

• Revenue: OpenAI is projecting $100 billion in revenue, while Anthropic has also significantly lifted its optimistic forecasts compared to 2024.
• Spending: OpenAI expects its cloud spending to reach hundreds of billions per year. To sustain this, the company will need to raise tens of billions of dollars in new equity every year until 2030, when free cash flow might finally cover compute costs.
• Subsidized Demand: Pueyo warns that much of the current AI demand is subsidized, where tasks cost the company dollars but are free or cheap for the user; this model depends on costs continuing to decrease as demand increases.

The Competitive Landscape The release of agentic coding in February 2025 led to a massive acceleration in the number of iOS apps released each month. While NVIDIA currently holds a near-monopoly, Pueyo notes he is not fully invested in the company because he expects competition to rise. For example, Google moved toward developing its own TPUs (Tensor Processing Units) to make AI workloads more cost-effective than using GPUs.

Rate of Progress Compute power has been progressing by 10x every two years. The latest NVIDIA architecture, Vera Rubin, has reduced token ("thinking") costs by 10x and training costs by 4x over the previous Blackwell architecture.

Performance Metrics:

• Cost Efficiency: The cost per task has shrunk 300x in one year.
• Benchmarks: Scores on the difficult ARC-AGI-2 test have risen from under 20% a year ago to 55% currently.
• Acceleration: Frontier AI capabilities have accelerated their rate of improvement from 8.1 points per year to 15.3 points per year since April 2024.
• Capabilities: AI is now matching elite astrophysicists in obscure fields and solving previously unsolved math problems.

The Path to Superintelligence To reach superintelligence, error rates must decline to near zero so that infinite reasoning steps can be chained together. Currently, Claude Opus 4.5 can handle software engineering tasks with a time-horizon of nearly five hours with a 50% success rate. However, new research suggests that million-step tasks can be solved with zero errors by breaking them into microtasks where a series of agents vote on the solutions.

The post ends with a teaser, stating that one specific AI stands out and "might well be AGI".