Shou Zi Chew, CEO of TikTok, puts forward a new judgment: AI will rewrite the development of science within three years, what are ByteDance's bargaining chips?

Shou Zi Chew rarely talks specifically about science in public.

On October 12, at the inaugural DBS Asian Future Forward Forum in Singapore, the TikTok CEO took the stage in a dark suit and held a dialogue with Xu Zhikun, Head of Retail Banking at DBS Group, with the theme "Unleashing Human Potential in the AI Era".

He did not talk about short videos, e-commerce, or geopolitics. What he discussed was physics, biotechnology, chemistry and materials science.

"In the next two to three years, we will make real scientific breakthroughs in these fields, and all of this will be enabled by AI. This is very, very, very likely to happen."

This is not an offhand prediction. As a veteran from ByteDance, almost every word he says can find corresponding evidence in ByteDance's organizational structure and capital flow over the past two years.

ByteDance's AI for Science team has been spun off for independent financing. Shou Zi Chew revealed at the forum that ByteDance is "exploring how to cooperate with a large number of scientists to make some progress in the scientific field". He did not elaborate on the details.

But ByteDance's actions in this track are far more than what he mentioned.

In June 2026, ByteDance fully spun off its internal AI for Science drug discovery team that had been operating for 5 years. The core team, algorithm platform, and pipeline assets under research were all injected into a new entity, branded as Anew Labs. The team has about 50 members, led by Liu Kai, with offices in Shanghai, San Francisco and Singapore.

Three months later, Anew Labs completed its first round of independent financing: 290 million U.S. dollars, with a post-money valuation of 1.5 billion U.S. dollars. The lead investors are HSG (formerly HSG China), IDG Capital and GL Ventures, 5Y Capital co-led the investment, and China Biopharmaceuticals and Shanghai Future Industry Fund followed as strategic investors. ByteDance retains 56% of the shares and continues to hold controlling stake.

The $290 million Series A financing refreshed the record for a single round of domestic AI pharmaceutical financing this year. It is not common to see a lineup of HSG, IDG and GL Ventures betting on the same company in the same round in the 2026 venture capital market.

Why did ByteDance spin off this business? Industry insiders gave a straightforward explanation: the industry logic and management methods followed by AI drug discovery are completely different from those of the Internet main business. Pharmaceutical R&D follows the "dual 10 law" — it takes 10 years and 1 billion U.S. dollars of investment, and the verification cycle is calculated in years. Internet business is good at rapid iteration. If the two rhythms are placed in the same organization, the incentives will be misaligned and the evaluation will lose focus. Only by spinning out independently can the company access industrial capital and attract pharmaceutical talents.

After the spin-off, Anew Labs' core pipeline is an IL-17 small molecule inhibitor. According to the company, this is the world's first small molecule inhibitor that can simultaneously inhibit the three dimers of IL-17 family AA, AF and FF. The IL-17 pathway is a key target for autoimmune diseases such as psoriasis and ankylosing spondylitis. The competition on this target is already very fierce, but there is no second public case that uses AI to achieve comprehensive trimer blockade.

Protenix-v2: An Underestimated Open-Source Model If Anew Labs is ByteDance's "external interface" in AI4S, then Protenix-v2 is its underlying capability.

In April 2026, ByteDance's Seed team released Protenix-v2, a model for biomolecular structure prediction and design. With a parameter size of 464 million, its performance is on par with or even surpasses DeepMind's AlphaFold 3 in a number of core benchmark tests. It is one of the few open-source foundations in the world that can currently perform high-precision end-to-end prediction of protein, DNA, RNA and small molecule ligand complexes.

The key lies in the word "open source".

The weights of AlphaFold 3 are not fully open and cannot be privately deployed. Protenix-v2 adopts the Apache 2.0 license, with all codes and model parameters open, available for both academic research and commercial use. This means that any pharmaceutical company or any laboratory can deploy Protenix-v2 on its own server, without having to transmit data to ByteDance, and without worrying about intellectual property ownership.

This choice itself is a business judgment. ByteDance has no shortage of model capabilities, but lacks an ecological niche. In a field where DeepMind has defined technical standards, open source is the fastest way to penetrate — making Protenix-v2 the default tool for academia and small and medium-sized pharmaceutical companies, so that ByteDance has the opportunity to master the de facto interface in the field of structure prediction.

The performance improvement of Protenix-v2 is also supported by specific data. In antibody-antigen prediction, it only takes 5 random seeds to surpass the results of 1000 seeds of the previous generation. Antibody-antigen prediction is one of the hardest nuts to crack in AI pharmaceuticals, because the complementarity-determining regions of antibodies are highly variable, and traditional methods are difficult to achieve high accuracy. The progress of Protenix-v2 in this direction directly corresponds to the most time-consuming link in drug design.

29.6 Billion U.S. Dollars: ByteDance's AI Ammo Depot Shou Zi Chew dares to say "two to three years", and his confidence comes not only from models and teams, but also from capital.

In September 2026, ByteDance obtained a three-year loan of 29.6 billion U.S. dollars from nearly 30 banks, coordinated and arranged by Citigroup and JPMorgan Chase. The money will be used for artificial intelligence projects outside China, including data centers under construction in Southeast Asia. The transaction has no collateral, and lenders said the arrangement reflects ByteDance's credibility.

29.6 billion U.S. dollars is close to a quarter of ByteDance's total revenue in 2025. In the financing history of Chinese technology companies, this is one of the largest bank loans. And there is no mortgage — banks look at ByteDance's cash flow and balance sheet, not a specific asset.

ByteDance has not disclosed how much of this money will flow to AI4S. But judging from the spin-off financing of Anew Labs, the continuous open source of Protenix, and the investment density of the Seed team in biomolecular structure prediction, the weight of AI4S in ByteDance's AI territory is rising, not falling.

Industrialization of AI4S is Stuck at "Verification" Rather Than "Discovery"

Shou Zi Chew said AI will bring scientific breakthroughs within three years, but the real bottleneck in the AI4S field may not be on the discovery side.

Rentosertib from Insilico Medicine is a good observation sample. This oral small molecule inhibitor, whose target discovery and molecular design are completed by AI, targets idiopathic pulmonary fibrosis. In September 2026, the first patient administration in Phase III clinical trial was completed at Peking Union Medical College Hospital. From target identification to candidate molecule determination, all rely on the AI platform, which took 18 months and synthesized 78 molecules to obtain the candidate compound.

However, the Phase III clinical trial involves 47 research centers, 320 patients, and a 52-week treatment period, and key data is expected to be read out in 2029. AI has compressed the discovery phase from four or five years to one and a half years, but the time of the verification phase has not been shortened proportionally.

Ren Feng, Co-CEO of Insilico Medicine, made it very clear: "AI can generate a large number of R&D ideas, but decision-making and evaluation still rely on human experts. At present, there is a shortage of compound talents who understand both AI and pharmaceuticals, and the cooperation needs of many enterprises at home and abroad cannot be met due to insufficient talent supply."

An unnamed clinical researcher once said a sentence to the media, which is worth placing here: "AI technology itself does not constitute evidence of clinical benefit, and ultimately it depends on the hard endpoint data of Phase III trials."

This actually points to the core tension of AI4S at present. The speed at which models generate candidate molecules is already faster than the experimental verification capacity. The faster dry experiments run, the more prominent the capacity bottleneck of wet experiments becomes. The research report of CITIC Securities summed this up as: the value anchor is shifting from "model capability" to "scientific research productivity, data assets and systematic biological design".

In other words: the laboratory that can keep up with the rhythm of AI is qualified to talk about industrialization.

Bifurcation of Two Routes

ByteDance takes the "self-developed pipeline" route. Anew Labs holds the drug pipeline by itself, bears the risk of failure by itself, and advances the clinical process by itself. The logic of this route is: AI is mature enough to directly lead drug discovery.

The other route is "providing platforms". Nvidia and Eli Lilly jointly established an AI joint innovation laboratory in January 2026, with an investment of more than 1 billion U.S. dollars over five years, to build a dry-wet closed-loop R&D process that can operate in 2 hours. OpenAI released GPT-Rosalind, a dedicated life science model, in April, which was trained for 50 biological workflows, and Amgen and Moderna have imported it into their R&D processes.

The divergence between the two routes is not in technology, but in judgment. The self-developed pipeline route bets that "AI can already make drugs", while the platform route bets that "the greatest value of AI is to lower the threshold of the entire industry". Dario Amodei, CEO of Anthropic, regards life sciences as one of the company's "most strategically significant" fields. In April 2026, he acquired Coefficient Bio, an AI pharmaceutical startup that was only 8 months old, for about 400 million U.S. dollars, and directly entered the drug development field. Isomorphic Labs of Google was spun off from DeepMind, completed a 2.1 billion U.S. dollars Series B financing in May, and is in talks for a new round of financing according to Bloomberg, with a valuation of at least 40 billion U.S. dollars and a maximum possible of 50 billion U.S. dollars. Isomorphic not only cooperates with pharmaceutical companies, but also holds its own pipelines, walking on two legs.

On October 7, Biohub, a non-profit scientific research institution supported by Meta, announced the expansion of the "Virtual Biology Program" with a total investment of about 1.8 billion U.S. dollars. Participants include the U.S. Department of Energy, the National Institutes of Health, Google DeepMind, Isomorphic Labs, Meta and Nvidia. Google DeepMind, Meta and Isomorphic Labs jointly invested 300 million U.S. dollars, aiming to build an artificial intelligence model that can predict cell behavior.

This is a signal. American tech giants have chosen the path of jointly building infrastructure in AI4S — data sharing, model sharing, and standard sharing. ByteDance takes another path: self-developed models, spin-off incubation, independent financing, and vertical integration.

Shou Zi Chew's "Balance Technique"

Back to Shou Zi Chew's speech. He talked not only about opportunities, but also about risks.

He quoted the thought experiment of the "paperclip maximizer": a super-intelligent AI set to maximize the output of paperclips might convert all matter on Earth and even human beings themselves into raw materials for making paperclips. He said that when AI excessively pursues a single goal without appropriate restrictions, it may produce results that are inconsistent with the real intentions of human beings.

He advocates a development model where "humans are involved", especially in high-risk work where automation should not be completely relied on. Taking TikTok itself as an example: AI has been used for content moderation and platform security, which can identify voices and images in real time during live stream management, and assist in detecting violations that human moderators may miss. But he also admitted that users' viewing and interaction performance on AI-generated videos is "still relatively limited".

This admission is quite interesting. TikTok is the world's largest application of AI recommendation algorithms, but in the matter of AI-generated content, the result of users voting with their feet is not optimistic. There is a long distance between technical capability and user acceptance.

The structure of Shou Zi Chew's speech is actually very clear: the long-term value of AI is underestimated, and short-term expectations are overestimated; scientific breakthroughs will occur in two to three years; but humans need to maintain judgment and creativity to harness it. This is a safe statement, which not only gives confidence to the capital market, but also does not make unfulfillable promises.

But what he is really saying may be more specific than these words: ByteDance has bet enough money and people on this track, and it is time to deliver results.

This article is from the WeChat official account "Tech Business", author: Tech Business, authorized for release by 36Kr.