At WIRED Health in London on April 16, surgeon Ara Darzi — director of the Institute of Global Health Innovation at Imperial College London — described what he called "the first genuine inflection point" in the antibiotic-resistance crisis. Drug-resistant infections cause more than a million global deaths annually and contribute to nearly 5 million more, according to figures cited by Wired's coverage. A 2024 Lancet report projected drug-resistant infections could cause 40 million deaths by 2050. Traditional diagnostics — culturing bacteria from a sample to determine which antibiotic will work — take two to three days, which is time sepsis patients do not have: every hour of delayed treatment increases the risk of death by 4-9%. While waiting, doctors are guessing which antibiotics to prescribe.

AI changes that calculus along three vectors. First, diagnostics. Darzi cited AI-based systems achieving accuracy above 99% without additional laboratory infrastructure — the without-additional-lab-infrastructure clause is the part that matters for deployment in low-resource settings, where the WHO estimates one in three reported infections is already resistant in southeast Asia and the eastern Mediterranean, and one in five in Africa. Second, drug discovery. The UK National Health Service is working with Google DeepMind on an AI system that, in one demonstration, identified previously unknown mechanisms of resistance in 48 hours, cracking a mystery Imperial College London researchers had spent a decade trying to solve. Third, at-scale experimentation: paired with an automated laboratory, deep learning models can run hundreds of parallel experiments around the clock and screen billions of molecular structures in days, while generative AI is being used in candidate-drug design.

The economic story is the harder one. The reason there are few new antibiotics in the pipeline is that the market punishes the only successful version of the product: a new antibiotic that works should be used as little as possible to delay resistance, which means low volume sales, which means low ROI for pharma, which means underinvestment in R&D. Darzi's framing — "a lack of incentives means innovation may not reach patients" — is the polite version of saying the discovery half of the problem is being solved by AI faster than the deployment-economics half is being solved by anyone. AI can find candidate molecules; AI cannot fix the fact that drug companies do not make money on rarely-prescribed antibiotics, that hospitals do not pay enough for rapid-diagnostic kits to amortize their development, or that low-income countries with the highest resistance rates cannot afford the lab equipment AI depends on.

For builders, three concrete things. First, the AI-diagnostics-without-extra-lab-infrastructure claim is exactly the deployment frontier worth watching. If you build in medical AI, the products that reach patients will be the ones that work on existing hardware (smartphones, basic microscopy, point-of-care kits) — not the ones that require a $200,000 sequencer per clinic. Second, the DeepMind/NHS 48-hours-vs-10-years comparison is the durable framing. AI compresses certain research timelines by orders of magnitude, but only on problems where the bottleneck was exhaustive search rather than experimental validation. Validation is still slow, expensive, and human-bound; budget your AI-driven research projects accordingly. Third, the incentive critique applies broadly: any AI tool whose value is highest precisely when it is used least (rare-disease diagnostics, security-incident response, infrastructure monitoring) faces the same misaligned-payment problem. Build that conversation into your business model from day one — government-purchase, subscription-license, outcomes-based pricing — because the tool's effectiveness alone will not unlock the budget.