$347M Is Flowing into AI Protein Design Platforms — Here Is What the Frontier Looks Like

$150 million. That's what it now takes to build a frontier AI model for protein design — and investors are lining up. Nearly $350 million has flowed into three lead startups in less than eighteen months, each applying large-scale language model techniques to biology's hardest design problem. The money is coming from software infrastructure funds, not traditional biotech venture. That alone tells you the thesis has changed.

Protein engineering used to be a process of directed evolution — mutate thousands of variants, test each one, repeat — with no guarantee that hundreds of costly wet-lab experiments would yield a single commercially viable molecule. The last eighteen months have changed that picture entirely. Frontier AI models trained on billions of protein sequences can now generate functional proteins from a text-like prompt, compressing what once took years into a matter of weeks or even days.

Profluent: The $150M Bet on Programmable Biology

Profluent, founded in 2022 and based in Emeryville, California, is the furthest along this new trajectory. The company raised $106 million in November 2025 in a round co-led by Altimeter Capital and Bezos Expeditions, bringing its total funding to $150 million. The backers are not typical biotech investors — Altimeter is a growth-stage software fund that backed Snowflake and Datadog; Bezos Expeditions is Jeff Bezos's personal investment vehicle.

As we wrote in July, Its AI-designed base editors already reached the clinic through a $160 million ARPA-H programme. The company has expanded beyond gene editing into antibodies, enzymes, and industrial proteins, with commercial partnerships spanning Eli Lilly (recombinases for genetic medicine), Corteva Agrisciences (agricultural proteins), and Integrated DNA Technologies (enzyme design). Its Protein Atlas now contains over 115 billion unique sequences — the largest known protein dataset in the world.

"This positions them not just as a frontier science company, but as the foundation of a massive new industry," said Jamin Ball, partner at Altimeter Capital, at the time of the round.

Cradle: The Platform Play at $103M

Cradle, headquartered in Amsterdam, takes a different approach. Rather than developing its own protein candidates, it sells a software platform that lets pharmaceutical and industrial biotechnology companies design their own. The company raised $73 million in Series B funding led by IVP in November 2024, bringing total disclosed funding past $100 million.

Cradle's customers include Novo Nordisk, Johnson & Johnson, and Novonesis. The platform claims to accelerate R&D timelines by up to twelve times — a client improved the activity of a P450 enzyme by four times in three experimental rounds versus the typical ten. CEO Stef van Grieken frames the mission as putting AI protein tools "in the hands of one million scientists."

Latent Labs: DeepMind's AlphaFold DNA Enters the Fray

Latent Labs emerged from stealth in February 2025 with $50 million — $40 million of it a Series A co-led by Radical Ventures and Sofinnova Partners — and a founding team drawn directly from DeepMind's AlphaFold project. CEO Simon Kohl co-developed AlphaFold2 and co-led DeepMind's protein design team.

In July 2025 the company released Latent-X, a generative model that designs protein binders from scratch at the all-atom level — generating structures ten times faster than previous methods. In July 2026 it followed with Latent-X2, which claims drug-like developability and low immunogenicity in generated antibodies, and Latent-Y, an AI agent that designs therapeutic antibodies autonomously from a text prompt. The company is backed by Google Chief Scientist Jeff Dean, Anthropic CEO Dario Amodei, and Cohere CEO Aidan Gomez.

The arc is recognisable. Like early cloud infrastructure, each platform builds on the one before — but instead of compute cycles or storage bytes, the unit is a designed protein optimised for a specific function.

EvolutionaryScale and the Open-Source Question

Not every AI protein platform is chasing a proprietary moat. EvolutionaryScale, founded in 2023 by former Meta AI scientists, raised $142 million in seed funding — one of the largest seed rounds in AI biology — and built ESM3, a 98-billion-parameter model trained on 2.78 billion protein sequences. Its breakthrough: generating esmGFP, a novel fluorescent protein equivalent to simulating 500 million years of evolution, published in Science in January 2025. Chan Zuckerberg Biohub acquired the company in November 2025, taking the technology in-house.

The open-source dynamic creates a tension. Profluent's Protein Atlas of 115 billion sequences is proprietary; ESM3's weights are public. Which strategy wins depends on whether the data moat holds. Foundation models like ESM3 are publicly available, which lowers the barrier for any well-funded biology lab to generate novel proteins. The startups building proprietary platforms argue that what matters is not the base model but the feedback loop — each customer project generates new experimental data that refines the model, creating a data moat that open weights alone cannot replicate.

Scaling Laws Hit Biology

What unites these three companies is a shared bet on an idea borrowed from large language models: that AI model performance improves predictably with scale. Profluent demonstrated this in a NeurIPS 2025 spotlight paper, showing that protein language models follow reliable scaling laws — larger models trained on more sequences generate better proteins with predictable consistency.

The consequence: protein design shifts from a discovery process into an engineering discipline. A pharmaceutical company no longer needs to run ten thousand random mutations and hope; it can specify a target profile — stability at 60°C, binding affinity in the low nanomolar range, low immunogenicity — and ask the model to generate candidates that hit those specifications. The wet lab shifts from a discovery engine to a validator.

The compute requirements are not trivial. A 98-billion-parameter protein model like ESM3 requires thousands of GPU-hours to train, and the inference cost per designed protein — while falling — still exceeds what most academic labs can afford at scale. This creates a natural advantage for well-capitalised platforms and explains why investors with software infrastructure experience, not biotech specialists, are writing the largest cheques.

The Bottleneck That Remains

AI capability has outpaced wet-lab throughput. Each AI-designed protein still needs to be synthesised, expressed, purified, and tested. A single design-build-test-learn cycle can take weeks and cost tens of thousands of dollars, even with AI compression. The companies that solve this — either by building their own automated labs, as Cradle is doing, or by partnering with contract research organisations at scale — will capture the most value.

It has the dataset (115 billion sequences). Cradle has the customer workflow (pharma partnerships across multiple verticals). Latent Labs has the foundational model pedigree. None of them yet has the end-to-end loop closed — the automated lab that feeds design data back into the model at industrial scale. The winner of this category will be the one that does first, and the next twelve months will separate the platforms from the projects.