AI Can Process 15 Million Molecules Daily, Yet Still Can't Cure Alzheimer's

In April 2026, The Next Web (TNW) published an article that challenges the hype surrounding AI in pharmaceuticals: the technology indeed accelerates certain stages of drug development, but according to TNW's assessment, the revolution in drug discovery is radically overstated. Health chatbots represent a documented danger, and the most severe diseases—including Alzheimer's—remain unsolved despite impressive computational power figures.

A Telling Example: 15 Million Molecules for One Disease

The key example TNW cites is work by pharmaceutical company Novartis. In late 2025, a research team working on Huntington's disease used generative AI to computationally design 15 million potential compounds for a specific type of molecule. This is a gigantic volume—several orders of magnitude greater than what a traditional laboratory could screen in comparable time manually or even with classical high-throughput screening.

It is precisely these numbers—millions of AI-generated drug candidate molecules—that are typically cited as proof of revolution in pharmaceuticals. TNW does not deny that generative models can indeed traverse chemical space at scales inaccessible to humans. The problem, in TNW's view, is different: the scale of candidate generation is not the same as the scale of actual clinical success. Between "designing 15 million molecules" and "obtaining a working drug that has passed clinical trials" lies a long path, where the vast majority of candidates are filtered out.

Why the Most Severe Diseases Remain Unsolved

The main thesis of TNW's article: where genuine breakthroughs are needed—in neurodegenerative diseases like Alzheimer's, Huntington's disease, and other complex, poorly understood pathologies at the molecular level—AI has not yet shifted the situation. The reason is not a lack of computational power, but that the biological mechanism itself is poorly understood for such diseases: a model can generate millions of plausible drug candidate molecules, but if science does not yet fully understand which exact target needs to be struck to stop Alzheimer's progression, even the most powerful generative system will optimize the wrong problem.

Key theses from TNW's article:

• The AI revolution in drug discovery is real, but, in TNW's assessment, radically overestimated
• Medical health chatbots are a documented source of risks, not just convenience
• The most severe diseases remain unsolved despite AI progress
• Example: Novartis team in late 2025 generated 15 million potential compounds for Huntington's disease therapy

It is telling that TNW chose not a startup promising revolution, but a large, conservative pharmaceutical company with a long history—Novartis—as its example. This is an important detail: this is not about marketing claims from a small AI company needing to attract investment, but about actual work processes within an organization that by definition is cautious about bold promises in drug development and must substantiate any results with long and expensive clinical trials.

What This Means for the Biotech Industry

The TNW article functions as a counterweight to the wave of enthusiastic headlines about "AI that will solve cancer and Alzheimer's in a few years." In practice, generative AI in pharmaceuticals today works best with the part of the process that lends itself to formalization—generating and initially screening chemical structures, predicting their properties, optimizing known molecular scaffolds. But where fundamental understanding of disease biology is needed, as with Alzheimer's and other neurodegenerative disorders, computational power alone does not substitute for decades of missing scientific knowledge.

For investors and pharmaceutical companies, this means a need to more soberly assess which stages of drug development AI actually accelerates, and which remain as long and uncertain as they were before generative models appeared—and not confuse impressive figures of generated compounds with a guarantee of clinical success.

The second part of TNW's thesis also deserves particular attention—about health chatbots as a documented source of risk. Unlike laboratory AI applications, where an erroneous candidate is simply filtered out at the next testing stage, an error by a chatbot giving medical advice directly to users can cause real harm without any intermediate specialist oversight—and it is precisely this difference in risk level, in the publication's view, that deserves far more public attention than it currently receives.