A new TD Cowen survey of 80 biopharma executives puts a hard number on what AI has done to early drug research: preclinical costs and timelines are down by up to 70 percent. The same survey estimates that new drug development programs could grow more than 10 percent over the next three to five years, with roughly $1 billion in incremental spending flowing into software, sequencing tools and modeling platforms. Neither figure has translated into a single FDA-approved drug.
That gap is the story. Preclinical work, the phase before a compound ever reaches a human being, is where biotech burns most of its budget chasing candidates that will eventually fail. If AI models can screen out weak candidates before a wet lab technician ever touches them, the industry gets more attempts within the same R&D spend. Brendan Smith, TD Cowen’s director of life sciences equity research, described the goal as wanting to “create more shots on goal,” generating enough data to keep improving the prediction models themselves.
That framing is optimistic by design. It measures efficiency inside a stage of drug development that has never been the bottleneck determining whether a drug reaches patients. Clinical trials are. A compound can clear every computational filter, get flagged by an “in silico” platform as stable and non-toxic, and still fail once it meets the biological variability of actual human bodies. Axios, which first reported the TD Cowen findings on July 20, 2026, notes that some investors already question whether AI’s preclinical gains will show up in patient outcomes at all.
The 70 percent cost reduction and the zero-approval record are not contradictory. They describe two different problems. AI is demonstrably good at compressing the search space: fewer synthesis cycles, faster toxicity simulation, quicker iteration on dosing models for populations like newborns or pregnant patients that are hard to study directly. AI has not yet demonstrated that the candidates it prioritizes survive contact with human trial data any better than the ones a traditional pipeline would have selected. Historically, roughly 90 percent of drug candidates that enter clinical trials still fail, and skeptics interviewed for the survey say that rate could hold even as AI reshapes the front end of the pipeline.
Two structural forces will keep pressure on this bet regardless of the outcome. The Trump administration’s push to reduce animal testing is steering more biomedical work toward computational tools and 3D tissue models by default, not because the science has proven itself, but because the regulatory door is opening. Separately, China’s biotech sector is scaling faster on cheaper labor and shorter turnaround times, pulling in new investment that raises the cost of the U.S. sitting out an unproven technology.
For anyone allocating capital or roadmap time to AI-biotech, the number to track is not the preclinical savings percentage vendors will keep quoting. It is the FDA approval count for AI-originated compounds, currently zero. Until that number moves, treat AI drug discovery platforms as productivity tools for the research phase, not as a proxy for clinical success, and size any investment thesis around the 90 percent trial failure rate rather than the cost curve improving in silico.
Reported by Axios on July 20, 2026.
