Dr. Hamid Noori is the CEO of The Cultivated B. As a scholar, he conducted his work at Princeton, MIT, Max Planck Society and IHES.
getty
Talk to anyone who has spent time on a biomanufacturing floor of a pharmaceutical company, and you will eventually hear some variation of the same story. A run went sideways, but not because of a lack of skill or oversight. By the time the offline assay confirmed what the cells were doing, hours had passed, and the batch was unrecoverable. Several members of my own team have lived through that exact scenario inside large biopharma and CDMO operations.
Based on my observations across the industry, compound manufacturers report batch failure rates of roughly 5% to 10% on average for biologics and ingredient production, costing billions of dollars annually. Even individual failed batches can sometimes carry losses in the tens of millions. And the consequences ripple beyond any single manufacturer: In 2020, the FDA reported that 62% of all drug shortages were caused by manufacturing and product quality problems, resulting in supply disruptions.
These numbers point to something the industry has tolerated for too long. The bioreactor—the core technology powering pharmaceuticals, biologics, food and advanced materials—still operates, in most facilities, as a black box.
For decades, the industry's instinct in response to variability and yield pressure has been to scale up: bigger facilities, bigger tanks and more redundant capacity. That playbook produced real gains, but it is reaching its limits. Adding another 20,000-liter tank does not tell you what is happening inside the one you already have. The next decade of biomanufacturing will not be won by the operators with the largest footprint. It will be won by the operators with the most intelligent one.
Operators set empirical inputs. They monitor a handful of process parameters, such as pH, dissolved oxygen and temperature. They wait for offline samples. The biology itself, the metabolic state that determines whether a run succeeds or fails, has historically been only partially observable. Critical biochemical variables such as amino acids, glucose, lactate and ammonia are often measured as snapshots of the process dynamics, outside the process environment, introducing delays that make real-time decision-making nearly impossible.
The result is process control that is reactive by design. If your most informative measurements arrive on a four-hour lag, the best you can do is correct after the fact. Batch variability, wasted feed, prolonged development cycles and the occasional catastrophic loss are not operator mistakes. They are the predictable output of running a complex biological system on lagging indicators.
What is shifting now is the sensing layer. New in-line measurement technologies enable tracking of substrates, by-products and metabolites continuously, inside the reactor, without pulling samples. The data is no longer a snapshot taken every few hours. It is a continuous multidimensional time series. When you can observe metabolic dynamics as they unfold rather than reconstruct them after the fact, the meaning of "control" changes.
Bioprocess data is time-dependent and governed by nonlinear interactions that don't behave the way intuition suggests. A subtle change in one variable may be meaningless on its own. Even excellent operators cannot reliably hold those relationships in their heads in real time, especially across a fleet of reactors running different products at different scales.
This is where AI moves from buzzword to load-bearing infrastructure. A model continuously analyzing high-dimensional time-series data can pick up patterns that don't show up on any single chart. Done well, it doesn't just flag deviations after they happen—it predicts where the run is heading. That is a fundamentally different capability than dashboards have ever offered.
The economic case for getting this right is substantial. BCG estimated that "integrating digital capabilities such as AI, digital twins, and advanced automation into a traditional production system can unlock an incremental 10% to 25% savings on conversion costs." Such savings can directly translate into broader patient access and stronger margins.
When real-time sensing and continuous AI interpretation are wired together, the process can begin to adjust itself. Feed strategies, environmental setpoints and control parameters can be tuned during the run, based on where the biology is actually going rather than where the SOP assumed it would go.
The implications extend well beyond a single tank. Historically, every bioreactor has been an island. Insights stayed local. A learning that emerged at one site rarely made it to a sister site running a similar process, let alone to a CDMO partner across an ocean. A unified software layer changes that. When sensing data, models and process knowledge live in the same system, you can monitor and optimize across reactors, sites and modalities. Each run makes the next one smarter, not metaphorically, but literally, because the model has more ground truth to learn from.
It does not mean removing humans from the process. It does not mean a fully hands-off facility within five years. The regulatory, validation and safety considerations in pharma alone make that timeline unrealistic, and frankly, undesirable.
What it does mean is a steady migration of decisions that today require human judgment toward systems that can support—and, in well-defined cases, execute—those decisions in real time. It means processes that are continuously understood rather than retrospectively explained. It means a meaningful reduction in batch failures, faster tech transfer and development cycles measured in weeks rather than quarters.
For founders and operators in this space, the strategic implication is straightforward. The companies that will win the next decade won't be the ones with the most capacity, the largest facilities or the cleverest individual model. They will be the ones that build integrated systems—sensing, software and intelligence—that get better the more they are used. Scale will still matter, but it will be a function of intelligence, not a substitute for it.
The black box era of biomanufacturing is ending. The question every operator and executive should be asking right now isn't whether this transition is happening. It is how prepared their organization will be when it arrives, because, in my experience, it arrives faster than anyone plans for.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。