$2.1bn for AI drug design: where smaller firms should start with AI

On 4 October 2026, Sina Finance, citing PharmaCube, reported that Isomorphic Labs — an AI drug discovery company spun out of Google DeepMind — had closed a $2.1bn funding round, a record for the sector. Around the same time, the company disclosed IsoDDE, its in-house integrated AI drug design engine.

For ordinary operators, the sum is secondary. Where the money went is the point — and it answers a question many are weighing: should AI be bought as ready-made capability, or built from scratch?

What a $2.1bn round has to do with a factory owner

The money did not buy a model. It bought a process.

Isomorphic Labs was founded in London in 2021 by Demis Hassabis and Max Jaderberg. Hassabis founded DeepMind and shared the 2024 Nobel Prize in Chemistry for the AlphaFold 2 work; Jaderberg led development of drug design models including AlphaFold 3.

When AlphaFold2 was released in 2020, the field briefly assumed protein structure prediction was the finish line. It soon became clear that reading a protein structure is not the same as making a drug that works. Between a static structure and a viable new molecule lies a wide gap.

In late September this year, the company published full research on IsoDDE, formally moving from "parsing known structures" to "identifying uncharacterised protein pockets and generating new chemical entities."

In short: AI went from looking at pictures to using its hands.

This is worth a look from every industry. Chatting does not generate revenue. Being wired into a process that must produce a finished good does.

So must we train our own model?

The logic of this round runs the other way.

Both founders have repeatedly made the same argument: drug development should not depend on luck but should be a computable, reason-driven, scalable engineering science. The $2.1bn backs that judgment, not a model with more parameters.

Consider another set of numbers from the same period. A weekly manufacturing investment report showed that in the week of 28 September to 4 October, four Chinese manufacturing firms listed, three were acquired and six raised funding, for a total of about RMB 3.006bn.

Capital is picking scenarios, not parameters.

For a company of a few dozen people, building a model in-house is hard to justify on the numbers: training is only the start, and data, compute, maintenance and iteration follow. Those costs are not what customers pay for.

Where should the first step land?

On the link you already have people doing — but doing slowest.

Three tests apply: does this link have clear right-and-wrong standards, does it involve large amounts of repetitive work, and is there existing data to draw on. If all three hold, start there. In pharma it is molecule design; in trade, customer acquisition and quoting; in restaurants, scheduling and inventory; in wholesale, order entry and reconciliation.

This is not a niche view. IMF Managing Director Kristalina Georgieva, in a speech titled "Europe and the Global AI Race," broke Europe's need for faster policymaking into five areas: finance, energy, business, labour and government. Even policymakers are advancing by domain rather than speaking of AI in the abstract. Companies have even less reason to generalise.

Does the logic change when going overseas?

The logic holds; the difficulty moves.

An analysis by 36Kr Global noted that Chinese firms going overseas have moved from early product trade and cross-border M&A to a new stage centred on deep localisation. The 2026 government work report explicitly called for "supporting the building of open-source AI communities" and "accelerating the promotion of next-generation intelligent terminals and agents."

For firms going overseas, AI saves more than labour. The real difficulty is replicating domestic operating standards in a second or third market — products can be shipped in containers, but judgment cannot, unless it is written into tools.

Rules are shifting too. A 4 October 2026 report noted that the United States, led by the Director of National Intelligence, has formed a "superintelligence" task force to submit a report within 120 days assessing the risks and opportunities of advanced AI and clarifying the federal government's role in AI safety and industry regulation. Anyone doing business across markets should factor such signals into annual planning.

How to put this into practice

If your question is where AI should land first, start with customer acquisition content — it sits closest to cash flow and shows results fastest.

SHEYU AGENT (舍予AI智能体) puts 16 industry advisers and 34 zero-threshold tools in one place. Graphics, AI posters and copywriting — work usually outsourced or shouldered in-house — can be produced there. One login across desktop, mobile and web means no separate setup for each channel.

It does not answer whether to build your own model. It answers whether you can produce something today without adding headcount.

Models will keep getting cheaper. The ability to write an industry's judgment into tools will keep getting more expensive.