Biotech vs. the Feedback Loop: What Would an AI-Native Future Actually Look Like?
For years, venture firms and startup founders have flirted with the dream of "AI for drug discovery." But if we’re honest, much of biotech today still looks more like 1990s Genentech than it does OpenAI.
Why? Because biotech remains trapped in a structure that resists the core advantages of AI: speed, iteration, and compounding feedback loops.
As investors at the intersection of computation and biology, we’ve started asking a more uncomfortable question: Is biotech even structurally compatible with AI-native acceleration, or is the traditional VC model increasingly out of sync with biological reality?
Biology Doesn’t Want to be Software
AI thrives when feedback loops are tight and data-rich. But in biology, the loop from hypothesis → experiment → insight is long, noisy, and expensive. You can’t A/B test a mouse model like you can a web interface. Wet labs introduce latency. Animal models, by their nature, are slow and complex. Regulatory paths impose hard gating functions.
So while AI models scale exponentially, biology scales in fits and starts. This misalignment undercuts the compound advantages that made the "AI dev stack" so transformative in other sectors.
This isn’t just a speed issue – it’s an epistemology issue. AI learns best when the data generation process is synthetic, standardized, and iterative. Biology offers the opposite: heterogeneity, stochasticity, and often non-reproducible results.
How Biotech Can Work with AI – With Constraints
But this doesn’t mean biotech can’t become AI-native. It just means that we need to recalibrate our expectations — and re-architect how these companies are built.
Here’s what a “biotech on steroids” powered by AI might realistically look like:
1. Synthetic Systems for Closed-Loop Feedback
The most promising AI-biology platforms don’t just analyze data – they generate it, and do so in synthetic, tractable environments. Think programmable cell-free systems, synthetic biology chassis, or engineered cell lines designed for high-throughput iteration. This creates a pseudo-software loop where AI can actually learn and improve.
2. Internal Data Moats, Not External APIs
The best AI-native biotechs are proprietary data engines. Public datasets in biology are often low-resolution or poorly annotated. The companies winning this race aren’t just running models – they’re inventing new instruments, assays, and automation to generate vertically integrated, high-fidelity datasets no one else has.
3. Modality-Constrained, Domain-Specific AI
Rather than general-purpose biology, most AI-native platforms specialize: RNA design, antibody affinity prediction, enzyme optimization, etc. This specialization constrains the data regime and allows for deeper domain modeling – much like AlphaFold’s performance in protein structure, but expanded across other therapeutic modalities.
4. AI as Compiler, Not Replacer
AI doesn’t replace biology – it compiles it. The most effective use cases turn biological programs (e.g., gene circuits, miniproteins, regulatory elements) into design problems that can be modeled, optimized, and built. In this sense, biology becomes the "hardware" to AI’s compiler.
5. New Org Structures: Cloud Biology + Systems Engineering
An AI-native biotech team won’t look like a traditional R&D org. It will be a hybrid of systems engineers, experimental biologists, and computational scientists. The lab will be cloud-managed and modular. Engineering sprints will replace rigid program management. Data will flow across wet and dry lab like packets on a network.
Where the VC Model Still Fits And Where It Doesn’t
This re-architecture has real implications for venture capital. The traditional biotech VC model is predicated on milestones (e.g., IND submission, Phase I readout) and well-understood regulatory gates. But AI-native biotech companies often front-load technical risk while back-loading clinical validation.
In this model:
- The inflection points are fuzzy (e.g., model accuracy, data asset depth, number of wet-dry iterations)
- The capital stack looks more like software (seed to Series B for platform buildout) but ends with biotech-style needs (trials, regulatory approval)
- The exit pathways favor M&A or hybrid licensing over IPOs (especially for platforms without a lead asset)
This doesn’t make the VC model obsolete – but it does require recalibrating timelines, benchmarks, and return expectations. Fund structures need to allow for deep platform investment before clinical derisking. Syndicates must include co-investors who understand both software and translational science. And VCs must be able to underwrite data moats and model fidelity – not just preclinical tox studies.
The Emerging Alternative Playbooks
We’re already seeing new archetypes emerge at the AI–biology frontier:
- "Biology Compiler" Companies: Firms like Generate are treating biology like code – using deep generative models to create de novo therapeutics.
- "Feedback Lab Operators": Companies that tightly couple automated wet lab infrastructure with AI training loops (e.g., DNA synthesis → phenotyping → model update).
- "Platform Licensing Engines": Firms that don’t own products but generate leads and sell platforms into pharma, similar to cloud AI APIs in other sectors.
Each of these approaches rethinks what a biotech company is – and challenges the way we, as investors, define “progress.”
Where We Go From Here
Biotech isn’t broken. But its structure – from wet lab inertia to regulatory drag – demands a more deliberate approach to AI integration. You can’t just bolt on an LLM and call it platform innovation.
The winners in this next wave will be architecture-first: companies that start from the feedback loop and work backward to biology. They’ll build experimental systems that compress time and cost. They’ll generate proprietary datasets that fuel differentiated models. And they’ll demand that we, as VCs, learn to underwrite not just biological risk – but computational potential.
At Tachyon, we call this biology as A.I. Not as a metaphor – but as a roadmap for rethinking how new therapies are discovered, designed, and deployed.
