AI-native biotech in 2026: companies building the new stack

Two documents published within a week of each other in April 2026 describe the same shift from opposite ends of the industry. Bessemer Venture Partners, mapping startups, argues that the next generation of biotech winners will be defined by their data infrastructure, not their science alone. AstraZeneca’s R&D leadership, writing in Cancer Discovery, reaches the same conclusion from inside a top-10 pharma: “it is the unique data fabric of each organization and not the models themselves that will increasingly define its edge.”

The shared premise is uncomfortable arithmetic. Nearly 90 percent of drugs entering clinical trials fail, and R&D cost per approved therapy has doubled roughly every nine years. Between 2012 and 2022, some 200 AI-for-drug-discovery companies raised $18 billion, and by 2024 more than 350 biological AI models had been published; models alone did not fix the failure rate. What is different in 2026 is a coherent stack emerging underneath the models. This post maps that stack layer by layer (the VC view), walks through how an established pharma actually absorbs it (the AstraZeneca view), and closes with what the two frameworks agree on.

Layers after Bessemer Venture Partners (2026). Selection and placement are our reading of public materials; illustrative, not exhaustive.

Bessemer’s market map organizes the space into three layers. At the top, biology-native data collection: companies generating multimodal biological measurements at scale, purpose-built for models rather than repurposed from legacy assays. In the middle, workflow automation: agentic platforms that orchestrate tools and unify R&D processes. At the bottom, lab automation: the physical infrastructure compressing experimental timelines.

One proof point Bessemer cites for the whole thesis: every protein-targeted small-molecule cancer drug the FDA approved between 2019 and 2023 relied on structural data from the Protein Data Bank. Public data infrastructure already made one generation of drugs possible; the bet is that proprietary, purpose-built data infrastructure makes the next one.

Layer 1: biology-native data collection

These companies span the drug development continuum, from target discovery and molecular modeling through molecular design to biomarker discovery and clinical prediction. Sixteen worth knowing (stage placement is our reading of what each company does, verified against public materials in August 2026):

Layer 2: workflow automation

This is the youngest layer of the stack, and its entrants approach it from very different angles: Potato structures and automates lab protocols, Convoke runs agentic competitive overviews and ideation, Phylo builds a collaborative research workspace. Rather than catalogue a category whose differentiation is still settling, we will show the layer through the example we know best.

Hydra is PharosBio’s autonomous analysis platform. Give it a research direction and it plans the analysis, executes it across ~100 scientific databases and 200+ codified skills, and validates every result before you see it. That is the workflow layer’s job description: data in, validated decisions out. And because the layer sits directly above the physical one, Hydra’s integration skills can hand designed experiments to lab automation partners and pull the measured data back, as its Adaptyv Bio skill does for protein designs. Worked examples are in our case studies.

Layer 3: lab automation

We covered this layer in depth in our guide to lab software and, more directly, in Lab automation meets computational biology (the 8 automation levels, four buying models, and ten companies including Medra, Lila Sciences, Automata, and Trilobio). Bessemer’s map adds names on the software-and-connectivity edge of the layer:

The startup map shows what is for sale. The harder question is how an established R&D organization actually absorbs it. In their Cancer Discovery article, AstraZeneca’s R&D leadership (Goodwin, Barry, Weatherall, Platz, and Reis-Filho) describe AI in pharmaceutical R&D as three tiers with different goals and timelines.

After Goodwin et al., AstraZeneca R&D, Cancer Discovery (2026).

Tier 1 removes repetitive work for individual scientists: automated literature review, natural-language protocol generation, and exploratory data analysis beyond fixed workflows. Tier 2 builds AI into processes so it augments domain experts continuously: systems that monitor preclinical and clinical data streams, competitive intelligence engines tracking scientific and patent activity, and decision support that integrates historical project data into go/no-go calls. Tier 3 is the ambitious one: AI-driven hypothesis generation and end-to-end optimization through interconnected agentic frameworks that combine specialized models, including omics models paired with language models.

Three nuances make the framework more than a maturity ladder. First, the epistemics: tiers 1 and 2 answer known unknowns, questions you already know to ask; only tier 3 explores unknown unknowns, and it demands curated datasets, traceability to source, semantic layers, and sandbox environments before it can be trusted. Second, the economics: AI is “durational, not permanent.” Tools will be replaced, so the authors argue for annual AI budgeting, like reagents, rather than one-off transformation projects. Third, the proof: quantitative continuous scoring of TROP2 became the first computational-pathology predictive biomarker used to select patients for antibody-drug-conjugate therapy, evidence that the upper tiers are reachable, not aspirational.

A venture market map and a pharma adoption framework are different instruments, but they are pointed at the same object. Five points of agreement stand out:

• Data before models. Bessemer's first principle is mechanism-informed multimodal datasets; AstraZeneca says the data fabric, not the models, defines the edge. Both treat models as increasingly commoditized.
• Agentic AI is the operating layer. Bessemer's agentic frameworks across R&D workflows are AstraZeneca's tier-3 mechanism: interconnected agents combining specialized models.
• The loop must close physically. Bessemer's lab automation principle mirrors AstraZeneca's design-make-test-learn loop and lab-in-the-loop experimentation.
• Fit for purpose beats maximal complexity. The more-plex fallacy and mechanism-of-action-informed data collection are the same warning against measuring everything and hoping.
• Infrastructure is the moat. Both locate durable advantage in the system that generates evidence, not in any single molecule or model.

Where they differ is geometry. The map is spatial: all three layers are being built simultaneously by different companies. The tiers are temporal: an organization climbs them in sequence, and each tier raises the bar for data quality and governance. Startups sell layers; pharma buys tiers. The practical question for any R&D organization is which companies from the map serve the tier you are actually on:

Whichever tier your organization is on, the workflow layer is the entry point that pays for itself first. Tier 1 is a scientist asking Hydra an exploratory question today. Tier 2 is Hydra monitoring a portfolio continuously, as in our portfolio monitoring case study. Tier 3 is Hydra handing validated designs to automated labs and learning from what comes back.