Introduction
Artificial intelligence has become one of the most heavily funded technologies in modern history. Companies across industries are investing billions of dollars into AI systems, infrastructure, and research. The expectation is simple. AI will transform how we build products, make decisions, and even discover life-saving drugs.
One of the biggest promises came from the healthcare sector. AI-driven drug discovery was expected to reduce research time, cut development costs, and accelerate medical breakthroughs. The vision sounded revolutionary, especially for an industry where timelines often stretch over a decade.
However, the reality tells a more complex story. In 2025, AI drug discovery partnerships reportedly crossed $15 billion in announced deal value. Despite this massive investment, the industry produced zero FDA-approved drugs. This gap between expectation and outcome raises serious questions about how AI actually performs in high-stakes environments.
The $15 Billion AI Drug Discovery Experiment
The pharmaceutical industry entered AI partnerships with strong optimism. Big players announced multi-billion-dollar collaborations with AI startups and research labs. These announcements created a strong narrative that AI was ready to transform medicine.
But there is an important detail behind these numbers. Most of the $15 billion figure came from total deal value, not actual upfront investment. In reality, companies only committed a small portion at the beginning, while the rest depended on future milestones, clinical success, and regulatory approvals.
This structure shows something important. Even the companies investing in AI were cautious. They were not fully convinced that outcomes would match expectations. That caution now appears justified when we look at the results.
After years of experiments and trials, the outcomes remain underwhelming. Several drug candidates designed or optimized using AI failed in later clinical stages. Some well-funded startups struggled to survive after disappointing results. The promise of rapid AI-driven drug discovery has not yet materialized at scale.
Why AI Scaling Is Facing Pressure
For a long time, the AI industry followed a simple belief. Bigger models would lead to better intelligence. More data and more compute would eventually unlock Artificial General Intelligence.
This belief drove massive investment in model training and infrastructure. However, the returns are now becoming harder to justify. Each new improvement requires significantly more computing power than the last, while performance gains are becoming smaller and more expensive.
This creates a clear concern. If scaling alone cannot produce the expected breakthroughs, then the industry may need to rethink its core strategy. The gap between cost and improvement is widening, and this is forcing a deeper evaluation of current AI development approaches.
The $700 Billion Infrastructure Bet
AI infrastructure spending is now at a historic level. Industry projections suggest that companies may spend nearly $700 billion on AI infrastructure by 2026. This includes data centers, GPUs, networking systems, and energy consumption required to support large-scale models.
This level of investment shows strong belief in AI’s future. But it also introduces risk. Infrastructure alone does not guarantee intelligence. If the underlying models have limitations, then increasing compute power will only improve performance up to a point, after which returns may flatten significantly.
This is why many experts are now asking a difficult question. Are we building the foundation of a new intelligence system, or simply scaling an already limited architecture at massive cost?
What Large Language Models Actually Do
Large Language Models, or LLMs, are powerful tools that can write code, generate text, answer questions, and summarize complex information. Their output often feels intelligent and human-like, which has led to high expectations about their capabilities.
But internally, they work very differently from human reasoning. LLMs do not understand meaning in a human sense. Instead, they generate responses by predicting the most likely next word based on patterns learned from large datasets. This makes them extremely effective at language tasks, but not necessarily reliable for factual or causal reasoning.
This is why LLMs can produce accurate answers in one situation and incorrect answers in another. They generate plausible responses, not guaranteed truths. Understanding this distinction is critical for businesses using AI in real production systems.
Causality and the Limits of AI Understanding
Computer scientist Judea Pearl has strongly influenced how researchers think about AI limitations. He argues that statistical models mainly learn correlations, not causation. This distinction becomes extremely important in fields like medicine and drug discovery, where understanding cause and effect is essential.
Pearl’s Ladder of Causation highlights three levels of understanding. The first is association, where systems detect patterns. The second is intervention, where systems understand the impact of actions. The third is counterfactual reasoning, where systems imagine alternative outcomes.
Most AI systems today operate at the first level. They are good at finding relationships in data, but weaker at understanding why those relationships exist. In drug discovery, this becomes a critical limitation because identifying patterns is not enough. Scientists must understand how and why biological systems respond in real conditions.
Jagged Intelligence in Real Systems
AI does not behave like a consistently intelligent system. Instead, it shows what researchers call jagged intelligence. This means AI can perform extremely well in some tasks while failing unexpectedly in others.
A model might generate advanced code but struggle with basic logical reasoning in another scenario. This inconsistency creates challenges for real-world deployment because businesses cannot assume uniform performance across all tasks.
As a result, companies must design systems with validation layers, testing frameworks, and human oversight. AI cannot be treated as a fully autonomous decision-maker in critical environments.
Why Software Engineers Still Matter
There is ongoing debate about whether AI will replace software engineers. Current evidence suggests the opposite. AI has become a productivity tool that helps developers write code faster, generate ideas, and automate repetitive tasks, but it does not replace engineering judgment.
Software systems still require architecture design, performance optimization, security review, and production reliability. These responsibilities depend on human expertise. In fact, demand for engineers who understand AI systems is increasing.
Many organizations now rely on experienced leaders, including a fractional CTO, to guide AI adoption and ensure systems are built correctly and aligned with business goals.
The Shift Toward Practical AI
The most successful companies are not asking how to use AI everywhere. Instead, they are asking where AI actually solves real problems. This shift is important because it moves AI from hype-driven adoption to problem-driven implementation.
Vertical AI systems built for specific industries and workflows are often more effective than general-purpose models. They are trained on relevant data and optimized for specific use cases, which leads to better performance and clearer business outcomes.
The Real Mistake Companies Make
Many organizations start with the wrong question. They ask how to add AI to their business instead of asking what problem needs to be solved. This leads to wasted investment and low-impact projects.
A better approach is to start with a clear business challenge, then evaluate whether AI can solve it efficiently. This ensures that technology serves the business instead of the other way around.
Conclusion
The AI industry is entering a more mature phase. The excitement is still strong, but expectations are becoming more realistic. The $15 billion drug discovery example shows that investment alone does not guarantee success.
AI is powerful, but it is not magic. It has clear strengths and equally clear limitations. Companies that understand this will make better decisions and build stronger systems.
The future of AI will not belong to those chasing artificial general intelligence at any cost. It will belong to those solving real problems with well-engineered systems and practical execution.
Organizations that adopt this mindset will outperform those driven by hype. This is the core idea behind StartupHakk, where technology is always focused on real-world value, execution, and clarity over speculation.
