From Google to ByteDance, every major tech giant will eventually own a pharmaceutical company.

On September 16, Reuters broke the news that ByteDance spun off its internal AI pharmaceutical R&D team to establish a new company named Anew Labs. The company completed a $290 million first round of external financing, with a post-money valuation of approximately $1.5 billion, and ByteDance retains a 56% controlling stake.

The investor list is almost a who's who of China's primary market: HSG (formerly Sequoia China), IDG, and GL Ventures led the round, with 5Y Capital as a co-lead investor, followed by Gaorong Capital, Chunhua Capital, and Boyu Capital. Two distinct names also appear on the list: China Biopharmaceutical, and Shanghai Future Industry Fund.

$1.5 billion is roughly equivalent to the current market capitalization of Schrödinger, the decades-old listed company that has been developing computational chemistry software for 30 years and is regarded as a standard configuration by global pharmaceutical enterprises.

A new company established for three years, with a core team of 50 people and all four pipelines remaining at the pre-clinical stage, is valued roughly on par with that 30-year-old industry veteran in the eyes of the capital market.

Schrödinger, the famous thought experiment: before the box is opened, the cat is both alive and dead at the same time.

And Anew Labs is in exactly such a box right now: valued at $1.5 billion based on its platform, and worth at most $500 million based on the residual value of its pipelines. The two valuation results hold valid at the same time, until something opens the box.

From a recruitment notice to a $1.5 billion company

The story of Anew Labs dates back nearly six years before this spin-off news.

At the end of 2020, ByteDance AI Lab began to quietly recruit talents for AI pharmaceutical R&D. That was the heyday of ByteDance, when Douyin, TikTok, and Toutiao achieved overwhelming success with their recommendation algorithms. However, inside this recommendation machine, a small group of people were working on something with a completely different rhythm: developing drugs with AI.

In 2021, Liu Kai took the lead in forming a formal drug discovery team. He did not start as an algorithm scientist, but had seven years of experience in venture capital before that. In hindsight, letting an investor build this team was an accurate choice: half of the work of drug development is science, and the other half is capital operation and resource integration. By the way, his former employer is now on the list of Anew's investors.

The team is not large, with around 50 people. The official website listed 36 core members in May, combining AI algorithm talents and senior pharmaceutical experts. In June this year, ByteDance completed the spin-off, injecting the team, algorithm platform, and research pipelines into the new entity as a whole. Three months later, the first round of financing was finalized.

Why is it necessary to spin off the team? Industry insiders can easily understand: pharmaceutical R&D follows the double-ten rule, which means a 10-year cycle and $1 billion investment, with verification measured in years; while the rhythm of internet business is measured in weeks. When the two kinds of clocks are installed in the same organization, there will be misaligned incentives, unfocused assessment, and patience will be worn away little by little. Google made exactly the same choice for DeepMind's drug team, and spun off Isomorphic Labs in 2021. The two giants from China and the United States share the same insight on organizational issues.

Spin-off does not mean complete separation. Anew's computing power continues to be supplied by Volcano Engine, and the company has been connected to the computing power pipeline of a large technology enterprise since its birth. Its offices are located in Shanghai, San Francisco and Singapore, and its US entity is named Anew Therapeutics.

In terms of assets, public information can outline a general picture.

Five platforms: AnewFold is used for biomolecular structure prediction, while AnewSampling, AnewOmni, AnewDesign, and AnewMind cover other links of molecular design. International media reports mention that these platforms include Protenix and PXDesign. People familiar with the open source community are no strangers to Protenix, a protein structure prediction model open sourced by ByteDance earlier, which was once regarded as a competitor to AlphaFold3. The technical foundation of this company is the continuation of ByteDance's public achievements in AI for Science in recent years.

Four pipelines: covering small molecules, antibodies and cyclic peptides. The most noteworthy one is called "the world's first" by the company: a small molecule inhibitor that can simultaneously target all three dimers of the IL-17 family.

This line needs further explanation:

IL-17 is one of the most successful target families in the field of immunology. Novartis's Cosentyx and Eli Lilly's Taltz have annual sales of billions of US dollars, and are the main treatments for psoriasis and ankylosing spondylitis. There is only one huge regret in this market: all existing drugs are biological agents that must be injected; and for the three dimers of the IL-17 family (A/A, A/F, F/F), even the strongest dual-target antibody currently only covers two of them.

Why has no one ever succeeded in targeting all three with a small molecule? Because the target itself is a forbidden zone. The binding surface between IL-17 and its receptor is large and flat, which is a famously "undruggable" protein-protein interaction. Large molecules like antibodies can spread their arms to wrap the whole flat surface, while small molecules have to be like a nail wedged into a smooth wall.

If it is successfully developed, the situation will be completely different: patients can switch from lifelong injection to taking a pill every day, no cold chain is required, and manufacturing costs will be greatly reduced. Earlier this year, Anew's biology leader publicly reported on this pipeline at an immunology conference in Boston.

Another pipeline targeting IL-4R is also worth mentioning: it is the core target for atopic dermatitis, which also represents a market of tens of billions of dollars, and has the potential for oral administration.

Of course, to fully understand this company, there is one point that must be remembered: none of the four pipelines has entered the clinical stage.

The end of large technology enterprises leads to pharmaceutical companies

Anew is not an isolated case.

If we spread out the global AI pharmaceutical R&D map in 2026, several forces are almost in place at the same time.

Nvidia provides computing power. In January, it and Eli Lilly announced the joint construction of an AI joint innovation laboratory, with an investment of more than $1 billion over five years, aiming to achieve a dry-wet closed loop that "completes one round of iteration in two hours" — traditional medicinal chemists take weeks to complete one round of design, synthesis and testing.

OpenAI and Anthropic provide models. OpenAI released GPT-Rosalind, a dedicated model for life sciences, in April. Amgen and Moderna have integrated it into their R&D processes, and Novo Nordisk immediately announced a strategic cooperation. Dario Amodei, CEO of Anthropic, publicly said that life sciences is one of the company's highest strategic priorities; a more radical move also took place in April, as the company acquired Coefficient Bio, which was established only eight months ago, for about $400 million, shifting from selling tools to directly holding pipelines. Bristol Myers Squibb is promoting Claude to more than 30,000 employees.

Anew's most direct counterpart is Google-backed Isomorphic Labs. The company spun off from DeepMind and led by Nobel laureate Demis Hassabis completed a $2.1 billion Series B financing on May 12, led by Thrive Capital, with Alphabet, GV, Temasek, and the UK's sovereign AI fund all participating, bringing its total accumulated financing to about $2.6 billion. Its engine IsoDDE covers protein-ligand modeling, affinity prediction, and pocket identification, and it has signed multi-billion-dollar cooperation agreements with Johnson & Johnson, Eli Lilly, and Novartis. It set itself the goal of putting AI-designed drugs into human trials by the end of 2026, and this schedule has been postponed once from the end of 2025.

Capital is also pouring into the transaction segment. The total value of global pharmaceutical transactions in the first quarter of this year reached $88 billion, a year-on-year increase of 30%, and institutions predict that the full-year figure will hit $150 billion. AstraZeneca's $18.5 billion deal with CSPC, Eli Lilly's $8.85 billion deal with Innovent Biologics, and Eli Lilly's deal with Profluent worth up to $2.25 billion. Large pharmaceutical companies are using real money to lock AI and original innovation into their own pipelines.

Looking at ByteDance's move on this chessboard: the computing power is connected to its own Volcano Engine, the model is connected to its own accumulated AI4S capabilities, the capital side is connected to top-tier VCs, state-owned assets and pharmaceutical companies, and the pipelines directly target global-level targets.

Anew Labs' ambition is hardly hidden.

$1.5 billion and 56% equity

Let's first answer the most concerned question in the market: why is a 50-person company with zero clinical pipelines valued at $1.5 billion?

According to the traditional pipeline valuation method, the residual value of each pre-clinical pipeline is tens of millions to $200 million, and the total value of four pipelines is at most $500 million. The extra $1 billion is priced by the market for other things.

The first factor is the possibility of the spillover of ByteDance's capabilities. Essentially, Anew translates ByteDance's capabilities in model training, engineering implementation and computing power scheduling accumulated in the era of recommendation systems to a new battlefield. This makes mathematical sense: the core of recommendation algorithms is to quickly search for the optimal solution in a huge continuous solution space, which is mathematically similar to drug molecular design. Investors are betting that the machine trained in short video scenarios can also perform well when using protein corpus.

The second factor is the machine that produces pipelines, rather than the pipelines themselves. A judgment from CITIC Securities and Haitong Securities can be used as a reference: the competitive unit of AI pharmaceutical R&D is upgrading from a single molecule to the systematic capability of "data × model × pipeline". Molecules may fail, but the production line will not. Isomorphic can raise $2.1 billion, while a biotech holding a specific drug may not be able to raise a fraction of that amount. That means the market has decided to price the factory, not the product.

The third factor is time. The last wave of Chinese AI pharmaceutical companies was established in batches from 2018 to 2021, and most of them are still stuck in the positioning dilemma of "whether AI is a tool or a drug". Giving a $1.5 billion valuation at the angel round is a head start in the primary market — once the clinical data comes out, the valuation may rise by a large margin or drop to zero. Options are bidirectional by nature.

What is more thought-provoking than the $1.5 billion valuation is the 56% equity ratio.

ByteDance neither fully owns the company nor completely sells it. The problem of full ownership has been mentioned before: the conflict between the double-ten rule and the rhythm of the internet business is almost unsolvable; a complete sell-off means giving up a strategic card, and may even cultivate a future competitor. Controlling spin-off is a balance between the two: the parent company retains control and upside financial potential, while the new company gains independent capital access, independent compensation system, and the rhythm that belongs to the pharmaceutical industry.

This strategy is becoming a global paradigm. Alphabet adopts spin-off for Isomorphic while maintaining deep participation; Anthropic chooses to acquire Coefficient Bio directly; Nvidia does not hold equity and only forms alliances. The level of investment is a measure of the confidence of each party. It can be seen that ByteDance and Google are the most committed: they contribute funds, computing power, and hold controlling stakes.

For large Chinese internet enterprises, the spillover value of this matter may exceed the financing itself. Those internal AI teams working on materials, batteries, and climate research now have a referable path: spin off, connect to the parent company's computing power, introduce industrial capital, and grow at the rhythm of the industry. It is predictable that the strategic departments of many large enterprises will study this 56% equity structure.

The bottleneck has shifted

Where is the next bottleneck for AI pharmaceutical R&D? A judgment from Industrial Securities points out the most counter-intuitive point of this industry: The rate-determining step is not how fast the model runs, but how fast wet experimental data can be generated.

The arms race on the model side is already overheated. Tools like AlphaFold have pushed the marginal cost of structure prediction to almost zero, and generative models spit out millions of candidate molecules every day. However, when the output speed of the model exceeds the experimental verification speed by 10,000 times, the bottleneck shifts completely: you can design 100 million molecules, but the laboratory can only synthesize 50 of them in a week.

This also explains the shift of the narrative protagonist, from algorithm companies to companies with wet experimental capabilities. The selling point of the joint laboratory between Nvidia and Eli Lilly is the two-hour iteration cycle, which does not compete for how fast the model runs, but how fast the closed loop rotates; the capital premium of XtalPi is based on 300 robots and 200,000 pieces of self-owned experimental data per month.

For Anew, this is the part that most needs to be supplemented in its business map. According to public information, it is a company with heavy computing and light experimental layout. All five platforms are dry experiment tools, and its wet experimental capabilities have not been disclosed. There are not many options: self-construction, cooperation, or outsourcing to Chinese CROs.

Conversely, this just explains the intention of China Biopharmaceutical's strategic investment — the pharmaceutical company's experimental system, clinical resources and regulatory experience are exactly the missing half of Anew. This is not a financial investment, but a complementary cooperation.

There is another layer of relationship that pushes this matter deeper.

I have written before that global payers are forming a joint encirclement: the United States uses the MFN clause to anchor drug prices at the international lowest level, and China uses the dual catalog of basic medical insurance and commercial insurance to divert high-value drugs. The ceiling of single drug price is systematically suppressed, and excess profits can only come from more patients multiplied by longer medication duration.

Putting Anew's pipelines into this framework, you will see a perfect coincidence. The number of patients with autoimmune diseases is counted in tens of millions, which means more patients; the diseases require lifelong medication, which means longer medication duration; the oral small molecule form does not require a cold chain, and the manufacturing cost is much lower than that of antibodies, which means the price is affordable. The same calculation in reverse is almost fatal to biological agents: once the oral small molecule achieves the same curative effect as Cosentyx, those injectable drugs that cost tens of thousands of dollars per year will face both patent cliff and dosage form replacement at the same time.

AI reduces R&D costs, and payers push down sales prices. The two forces exert pressure from both ends of the industrial chain, squeezing the old model of "high unit price, small patient group, biological agent" in the middle. Anew's preference for targets — IL-17 and IL-4R, both are large markets with oral administration potential — shows that it understands the endgame rules better than anyone else.

The economic form of drugs is approaching consumer goods: large market, low unit price, long cycle. Whoever accepts this setting first will update their pipeline model first.

The box has not been opened yet

Of course, there are three questions with no answers at the moment.

The first is clinical verification. All four of Anew's pipelines are at the pre-clinical stage, Isomorphic has postponed the human trial schedule, and so far there is no drug designed entirely by AI that has completed phase III trials across the industry. CITIC Securities and Haitong Securities said that the clinical verification wave of AI-designed drugs is coming. The word "coming" currently supports hundreds of billions of dollars in valuations. From 2027 to 2029, these companies will submit their results one after another, and that will be the coming-of-age ceremony for this track.

The second is ByteDance's patience. This company has launched multiple business lines in its history, and has also shrunk its business many times, including the education and gaming sectors. The 56% controlling stake can be a long-term commitment, and it can also be the first asset to be realized in the next strategic contraction. To judge the fate of a company, the equity structure does not make the decision, but how long the parent company is willing to wait makes the decision. The answer is not in the announcement.

The third is the convergence speed of valuation. The market is currently pricing based on the platform, but the value of the platform must eventually be realized through pipelines. If the platform cannot produce molecules that enter clinical trials within three to five years, the platform valuation will shrink to the pipeline valuation. Schrödinger's market capitalization fell from $8 billion to $1.5 billion, which is a record of the platform narrative being corrected by the pipeline reality.

Anew now has the