Summary

Moderna surged 177% on August 19, taking its market cap from $25B to $69B in a single day; the key to the re-rating was not infectious-disease vaccines, but the “oncology” option. 英博’s explanation goes straight to the ceiling: “If you’re an infectious-disease vaccine company, your ceiling is predictable.” Add oncology to infectious-disease vaccines, and “the future is almost limitless”—because the cancer patient population is enormous, while “the pricing headroom for preventive drugs is different from that for therapeutic drugs.” His company, 艾博生物, had held a 3-day off-site the weekend before the readout, reaching the same conclusion: Wall Street’s focus had shifted from infectious-disease vaccines to oncology.
The quality of the INTerpath-001 readout lies in its endpoint design: more than 1,000 patients, a global multicenter double-blind randomized trial, personalized vaccine plus Keytruda versus PD-1 monotherapy, with the primary endpoint of recurrence-free survival reaching statistical significance. The selection of postoperative Stage 2B to Stage 4 melanoma patients—“if I remember correctly,” 英博 said—was strategic rather than accidental: these patients can relapse “as quickly as 3 months after surgery, or within a year if it takes longer,” giving the company an early readout. 英博’s principle is blunt: “Rapid success and rapid failure are both acceptable… The least acceptable outcome is waiting a long time before I know the answer.” Overall survival still remains the final test.
The real moat is CMC, not science alone: a personalized cancer vaccine is “a nightmare” to manufacture. The industry views 4 to 6 weeks after surgery as the optimal window; sequencing and algorithmic prediction take “in the order of days,” and can be completed within a week, mRNA production takes 1 day, and LNP encapsulation 1 to 2 days, but the DNA template can take multiple days to weeks and QC release another 1 to 2 weeks. The process therefore has to be fully automated and intervention-free: “Any human intervention at any step will inevitably become a bottleneck.” Each machine handles a single patient, then must be disinfected and cleared through QC before taking the next one; single-use consumables are the largest cost item.
Cost is the real variable determining whether the field can commercialize, with 英博 using CAR-T’s high prices and the difficulty of turning a profit in China as reference points. US insurers can cover gene therapies costing “a million dollars per shot,” but in China the price might be “at most a few hundred thousand yuan—not the cost, but what ordinary people can accept.” That is treated as a lifetime cost, from which the cost structure is worked backward. The constraint is hard: once the algorithm and manufacturing process enter Phase 1, they are frozen—“through the 5- to 7-year clinical period, your entire algorithm and process cannot undergo any changes.”
On the AI narrative, he simultaneously cooled it down and turned up the heat. The cool-down: Moderna was already capable of doing this in 2015 and 2016, “way before what we call the AI era,” and the compute requirements were not especially high; with local deployment and human genetic-resource sequencing data kept within national borders, he also sees limited geopolitical risk. The heat: nucleic-acid drugs are the class most likely to be AI-driven because “their foundation is just the 4 letters A, U, C and G”—a quaternary system without complex 3D structures. But “AI without automation is just talk,” because there is no training data without automation. 艾博 has invited Ian, the former Moderna president who led neoantigen algorithm design, to become its Chief AI Officer.
He gave qualified backing to Elon Musk’s claim that vaccines have become a “software problem” and to the Stargate project’s 2025 claim that a vaccine could be made in 48 hours. “What they said is not wrong—at the technical level, it is a software problem,” and the engineering challenges can be solved. But “production release currently takes within 2 weeks, and at present there is no particularly good way to shorten that time further.” The key regulatory insight is that approval is “not actually for a product, but for a platform technology”; the platform’s safety was already validated during COVID.
Indication expansion offers the biggest upside, but the population with clearly demonstrated benefit is still bounded. 英博 made a strong prediction, while qualifying it as “what I think”: “For every indication approved for PD-1, I think mRNA cancer vaccines will eventually be approved as well.” The boundary is the condition of the immune system—vaccines work best while it is still intact. “You can’t keep beating a dead horse,” which is why postoperative patients are the initial focus. For inoperable patients, one path is to use existing drugs to downstage the disease until it becomes resectable; he also speculated that after removing the largest primary tumor, vaccines and PD-1 could eventually help clear microscopic lesions.
Yushan’s closing segment put the heat back in check, making it the most useful section for secondary-market readers. Moderna has so far released only preliminary topline Phase 3 results; “the specific HR risk ratio and the results for different patient subgroups have not yet been disclosed,” and the stock had already given back part of its post-announcement surge over the next several sessions. The off-the-shelf comparison is not encouraging either: Elyseal’s Phase 2 trial targeting shared KRAS mutations missed its disease-free survival primary endpoint in the overall population, with signals seen only in certain subgroups.

Deep dive

1. A 177% one-day gain: the market is repricing the oncology option, not the vaccine business

洪君 opened by setting the coordinates: Moderna surged 177% on August 19, taking its market cap from $25B to $69B, “because on that day humanity validated something for the first time in a Phase 3 trial—that mRNA vaccines can be used against cancer, starting at least with melanoma.” The distinctive logic is “customizing a dedicated vaccine for every patient”: one patient, one drug.
英博’s coincidental opening was persuasive. The weekend before the readout, 艾博生物 had just completed a 3-day off-site strategy meeting, whose analysis of the global competitive landscape and Wall Street’s perception concluded that “the future focus for this company is not infectious-disease vaccines, but cancer vaccines and cancer treatment.” By midweek, the INTerpath-001 readout had arrived. He chose his words carefully: “We can’t call it a success, but it is a very positive readout.”
His valuation framework is about replacing the ceiling, not raising earnings estimates: “Moderna’s ceiling is absolutely not that of an infectious-disease vaccine company. If you’re an infectious-disease vaccine company, your ceiling is predictable.” Add oncology to infectious-disease vaccines, and “the future is almost limitless”—first because of the size of the cancer population, and second because “the pricing headroom for preventive drugs is different from that for therapeutic drugs.”
He also believes the feedback loop is still developing: “This feedback is still on the way. There should be more and more positive feedback in the future, whether from the data or from the secondary market.” For a company doing pioneering research, “people should have nothing but respect… Regardless of success or failure, what matters more is that it has blazed this path for everyone.”

2. The guest’s career is itself a map of the field: oligonucleotides → mRNA → gene therapy

英博 describes his career as “since 2006… my entire professional career has been working with nucleic-acid drugs.” He has worked on oligonucleotides, mRNA, and gene therapy. He studied biology as an undergraduate, then earned a PhD in pharmaceutics overseas, with a dissertation on oligonucleotide delivery funded by Alnylam, before joining Dicerna.
His shift to mRNA was driven by a technical inflection point. Around 2015, oligonucleotides underwent a major transition “from LNP-based delivery to GalNAc liver delivery.” Since his experience was built mainly around LNPs, he moved from Dicerna to Moderna to develop LNP systems. His third job was at a gene-therapy company founded by former Moderna executives.
The decision to return to China had a personal element. His last company moved from Boston to San Francisco, while Boston was the city he had lived in for more than 10 years. “It was just across the Pacific from home, so I thought, why not just go directly home?” Technologically, he deliberately chose the middle path: oligonucleotides were relatively mature, while gene editing still seemed far from maturity, so he “chose the middle-ground mRNA route.”
He summarized China’s industry landscape in 2019 with one line: “One investor even told me I was the first living person he had met who had returned from Moderna.” Moderna’s IPO at the end of 2018 forced investors to look at the field, but “their understanding of the actual technical challenges and what it could do was extremely, extremely limited.”

3. Moderna’s 4 pillars show that personalized vaccines were a 2010 bet

He reconstructed Moderna’s 4 pillars at the time, each supported by a different partner: cancer treatment was On Kaido, in partnership with A Z; infectious-disease vaccines were Valera; rare diseases were L P Dero, in partnership with Alexan; and personalized cancer vaccines were Caperna, in partnership with Merck. The AstraZeneca-related division where he worked focused mainly on oncology products, while his own contribution centered on delivery and CMC development.
Asked by Yushan whether Moderna had always been highly focused, he answered yes: “Moderna is a very focused company. Even after COVID, it has really focused only on developing the mRNA technology platform, and then used that platform to determine which therapeutic areas made the most sense.”
The key date is 2010, when Moderna was founded on an mRNA personalized cancer vaccine platform. The challenge was never just scientific: “Drug development is never simply a scientific challenge. More often, it is a challenge in applied technology—science plus the entire CMC manufacturing process.” Personalized vaccines “represent an extreme level of CMC manufacturing-process requirements,” while also constituting an entirely new product for regulators.

4. The quality of INTerpath-001’s readout comes down to endpoint design

The trial was global and multicenter, enrolling more than 1,000 patients in a double-blind randomized design. One arm received a personalized cancer vaccine plus PD-1—Keytruda, in partnership with Merck—while the other received PD-1 only. The population consisted of postoperative patients with relatively advanced melanoma. The primary endpoint was recurrence-free survival, and the difference between the 2 groups was statistically significant.
He explained the relationship between recurrence-free survival, immunogenicity, and overall survival: “Recurrence-free survival should directly correspond to an extension of survival. It is not the same as whether early immunogenicity can translate into an extension of the survival rate. If there is no recurrence, survival will necessarily be extended to some degree.” But he did not skip the final hurdle: “Ultimately, we still have to look at overall survival.”
When Yushan later referred to the endpoint as progression-free survival, 英博 corrected him on the spot: “In Moderna’s clinical trial, it is actually recurrence-free… Progression-free survival and recurrence-free survival are different.” Recurrence-free means the tumor has not grown back after resection and the patient remains tumor-free. “For the tumor, that means it has achieved a curative effect”—a sharp contrast with recurrence 3 months to a year after surgery.

5. Melanoma was chosen for an early readout, not scientific preference

The enrolled population was roughly Stage 2B to Stage 4 melanoma patients, 英博 added, “if I remember correctly.” These patients can see a recurrence “as soon as 3 months after surgery, or within a year if it takes longer,” allowing the drug’s efficacy to be assessed quickly. This is the strategic core in 英博’s view, and his formulation is a general principle for startups: “For a technology company, rapid success and rapid failure are both acceptable. I can quickly use what I learned from success or failure to develop the next product. The least acceptable outcome is waiting a long time before I know the answer.”
Melanoma therefore meets 2 conditions at once: substantial unmet need and an early-readout window. “I think this was a very, very intelligent strategy.” Moderna has also quickly initiated a Phase 3 study in non-small-cell lung cancer.
But when asked what the next readout would be, he deliberately slowed the pace: “I think we shouldn’t rush to look at the next indication.” For now, the result is limited to a recurrence-free endpoint; he would rather see the product formally approved as soon as possible. Approval would mean the platform has been validated—“just as with the approval of the COVID vaccine, after which Moderna rapidly developed RSV and influenza vaccines.” He characterized the result as a mechanistic validation: “It represents not the success of one product or one disease, but a major breakthrough and success for the entire mRNA technology route in oncology.”

6. Why PD-1 is required: the vaccine lights the fire, PD-1 removes the brakes

He first gave PD-1 its full due: it “can be called one of the greatest discoveries in cancer immunotherapy,” and its inventors won the Nobel Prize. Mechanistically, it is an immune checkpoint inhibitor. Tumors grow uncontrollably because they suppress the immune system’s ability to recognize them; PD-1 removes that antagonism. Today, PD-1 is a cornerstone of immunotherapy. “Even for patients with low PD-1 expression, some clinicians believe they may still benefit from PD-1 treatment.”
He described the combination’s causal chain vividly. A personalized vaccine “is like lighting a fire under the immune system, allowing it to better recognize tumor cells.” But if the cells are recognized and then suppressed by immune tolerance, “there is still no effect.” PD-1 removes the antagonism, “so the immune system can recognize and attack them, producing an effect greater than the sum of its parts.”

7. The mechanism of personalized neoantigens: from collateral damage to one patient, one drug

Traditional cancer drugs face a basic trade-off: “kill 1,000 enemies while losing 800 of your own, or even kill 800 enemies while losing 1,000 of your own.” It is difficult to distinguish rapidly dividing tumor cells from rapidly dividing healthy cells. Nausea, vomiting, and hair loss occur because the drugs also attack gastrointestinal cells and hair follicles.
Vaccines take the opposite path, targeting only “neoantigens expressed exclusively on the tumor.” They are called vaccines because they borrow the principle of infectious-disease vaccines, “causing the body to recognize these tumor cells as if they were an invading virus.” The result is low adverse-event burden and strong patient compliance. His personal judgment: “Personalized cancer vaccines could be an excellent complement to every existing therapy for patients.”
Personalization is necessary because of tumor heterogeneity. Each person’s tumor arises from mutations in their own cells: “They may all be called colon cancer, non-small-cell lung cancer, or pancreatic cancer, but every patient’s tumor is different.” Treating a tumor as a disease category leaves the patient with “survival within expectations.” Treating it as a personalized disease allows a vaccine to target each patient’s tumor markers and direct a specific immune response.
He also clarified a common misunderstanding: this is not a preventive vaccine for healthy people. “It is not a preventive mechanism. It is more of a therapeutic mechanism”—more precisely, “using a cancer vaccine to awaken and activate our own immune system to attack tumor cells.”

8. “A personalized cancer vaccine is a manufacturing nightmare”: the time budget inside the 4-to-6-week window

The window is dictated by the patient’s physiology, not the factory. “The industry considers 4 to 6 weeks to be the optimal period,” because postoperative patients already have a recovery period; producing the drug during recovery avoids missing the treatment window. If the window is missed, “these tumor cells may quickly start growing again.”
The steps are tightly sequenced. Sampling requires tumor tissue and adjacent normal tissue for comparative sequencing. Sequencing and algorithmic prediction are “in the order of days—just a few days, and can be completed within a week.” Reverse engineering then encodes the neoantigens into mRNA, which requires a DNA template: “The DNA template may take multiple days to weeks.” IVT, or in-vitro transcription, is cell-free and can produce mRNA “basically in 1 day.” LNP encapsulation takes “about 1 to 2 days,” and final QC takes “1 to 2 weeks.”
His conclusion is that the window is tighter than it sounds: “4 to 6 weeks actually leaves far less time for true production than people imagine, and the entire process has to run according to an extremely tight workflow.” To save time, sample analysis after sequencing is done at the hospital, “because that is the most time-efficient. Any transfer takes time.” Once the sequence is obtained, the analysis is completed in the cloud, and the company returns the finished product to the hospital end to end.

9. Automation is a feasibility condition, not an efficiency choice: one machine per patient

Automation covers almost the entire process: “From the point at which we obtain the sequence after sampling and sequencing, the entire process can be completed through automation. That is the level we can achieve today.” The constraint is absolute: “Without automation, it is impossible to make the entire process intervention-free.”
Where is the difficulty? “Every step is difficult. It has to be fully automated, with no human intervention. Any human intervention at any step will inevitably become a bottleneck.” At the same time, the company must design a process that is sufficiently universal yet sufficiently precise, so that every patient’s distinct product can meet the final quality-control requirements.
He described the commercial architecture concretely: “You can think of it as a series of separate machines, each designed for an individual patient and producing that patient’s required product.” The more than 1,000 patients in Phase 3 were not enrolled on the same day, so there were time intervals between them. At commercial scale, the patient population will be enormous. To ensure both timeliness and protection against cross-contamination, each machine produces only one patient’s product; afterward, it “has to undergo appropriate disinfection and QC before it can be used to produce the next patient’s product.” Single-use equipment is used extensively.

10. 34 neoantigens reflect the risk-reward trade-off, not a manufacturing limit

Yushan asked about the significance of “no more than 34” in Moderna’s press release. The answer was unexpectedly pragmatic: “Theoretically, we can make as many as there are. But once the mRNA reaches a certain length, the efficiency with which it translates into protein may decline, and the manufacturing challenge also grows. 34 is an upper limit, but not a production limit. It is simply what we see as the risk-reward trade-off.”
The more important point is the historical context. The project “goes back 10 years,” when AI and neoantigen-prediction algorithms were underdeveloped. The industry’s calculation was that “even with only a 10% hit rate, whether we have 34 neoantigens or 20 neoantigens, if 2 neoantigens can precisely recognize the tumor cells, the immune system can already kill them.” As prediction improves, “the vast majority of cancer patients do not need up to 34 neoantigens.”
More targets also bring a concrete benefit: reducing immune escape. Tumor cells are difficult to eliminate because they escape, and the more neoantigens covered, the lower the probability of escape. “The probability of mutating away 1 neoantigen at a time is high, but if you need to mutate away 5 neoantigens at once, you can do the multiplication—the probability becomes extremely, extremely low.” The broader the net, “the more completely I can theoretically kill the tumor cells.”
There is a trade-off. Total dose is fixed: “If you make only 1 neoantigen, the immune response against that neoantigen will definitely be strongest. If you spread it across multiple targets, each response may not be quite as strong, but it is not as though it falls 30x.” He declined to give a preferred number: “Every patient’s situation may be different.” Long follow-up will be needed to determine whether recurrent neo-tumor cells still carry some of the original antigens.

11. Postoperative patients come first: the scalpel removes what the eye can see, the vaccine removes what it cannot

Asked the most tempting question—whether all cancer patients can benefit—he drew a clear boundary: “Based on what we see today, yes, the benefit is clearly demonstrated only in postoperative patients.” The scalpel removes tumors visible to the naked eye; “the personalized mRNA cancer vaccine targets those scattered throughout your tissue that cannot be seen with the naked eye and cannot be removed with a scalpel.”
With the primary tumor still present, the treatment becomes a tug-of-war: “It is a conflict between 2 forces. One is the cancer vaccine plus PD-1 trying to mobilize the immune system to attack it; the other is the tumor itself, which exerts tolerance and antagonism against the immune system.” He kept the hedge strict: “We can only say that it has not yet been fully validated. That does not mean it cannot work in the future.”
He then offered the episode’s most vivid personal prediction: “In the future, cancer treatment may basically be handled by a scalpel plus a personalized cancer vaccine and PD-1. The scalpel deals with what is visible to the naked eye; the personalized cancer vaccine mobilizes the entire immune system to surround and eliminate what is not.”
For inoperable patients, he outlined 2 possible paths. The first is downstaging with existing drugs—moving a patient from Stage 4 to Stage 3, or to a point where the disease can be cleared—after which the patient could become eligible for a personalized cancer vaccine. The second involves patients with multiple microscopic lesions who were previously thought unlikely to benefit from surgery: remove the largest primary lesion while the remaining microscopic lesions have not yet developed strong immune resistance, then use a vaccine plus PD-1. “I think the room for imagination this creates is enormous.”

12. “You can’t keep beating a dead horse”: vaccine efficacy depends on the immune system’s remaining capacity

Yushan raised a consequential question. As standard treatments across tumor types rapidly improve, preoperative therapy controls tumors more effectively, and prognosis improves, “the recurrence problem left for cancer vaccines to solve is becoming harder to validate.” Does that imply a limited range of use?
英博 shifted the constraint from indication to trial design: “There should be no limitation on its range of application, but there are more challenges in terms of clinical design.” The more effective existing drugs become and the longer patients survive, the more the vaccine must prove an incremental extension on top of that baseline. “The observation period may have to be longer.”
But he offered a rule from 艾博’s own work, which includes both fixed-antigen and personalized vaccines: “We found that cancer vaccines work best when your immune system is still relatively healthy. You cannot wait until the immune system has been destroyed by chemotherapy, radiation, and drugs, then expect a personalized cancer vaccine to reactivate it… As the saying goes, ‘you can’t keep beating a dead horse.’ If the immune system has already collapsed, the personalized cancer vaccine has nowhere to exert its force.” That directly explains the focus on postoperative patients, whose immune systems remain relatively intact.

13. Safety is a key asset: outpatient care, not CAR-T

The treatment setting could be lighter than expected. “If the cancer patient does not need to be hospitalized, the cancer vaccine could almost be delivered as outpatient care: come to the hospital for 1 shot of the cancer vaccine and 1 shot of PD-1, then go home.” 英博 expects “no adverse reactions beyond those foreseeable with other mRNA infectious-disease vaccines”—fever, nausea, vomiting, fatigue, swelling at the injection site, and headache. “Compared with traditional chemotherapy and other cancer drugs… it is almost negligible.”
His response to the CAR-T comparison clarified the mechanistic difference. CAR-T adverse events mainly come from the rapid activation of T cells, or cytokine release syndrome (CRS). An mRNA vaccine, by contrast, “is slowly injected into the body, slowly expresses the antigen, and then activates the immune system and activates T cells to recognize it.” Reports of CRS in this setting “should currently be extremely rare. I don’t think I have seen many.”
The cancer vaccine is administered mainly by intramuscular injection, though intravenous delivery is also possible; BioNTech uses intravenous injection. He then stressed that pancreatic cancer remains an “anything-goes” field, with multiple technology routes still in play rather than a single route destined to win.

14. Mapping the field: off-the-shelf versus personalized, with players roughly split 50/50

He began by resetting the premise: “The concept of cancer vaccines is not new, and the technology routes included in cancer vaccines are not limited to mRNA.” They also include nucleic acids—DNA and RNA—peptides, and cell therapies using viral vectors or dendritic cells. Some early US companies pursuing peptide approaches “no longer exist.”
The advantages and disadvantages of off-the-shelf products are straightforward. The benefit is “immediate, plug-and-play” use: faster, more convenient, and cheaper, primarily targeting driver mutations such as EGFR and KRAS. The challenges are 2-fold. Not every patient has a clear driver mutation, and these mutations “have already existed in the body for a long time and developed gradually. The body may already have developed some tolerance… So we need to see whether the vaccine is stronger, or whether the tumor’s ability to induce tolerance is stronger.”
The 2 routes are represented by roughly equal numbers of players: “We did a count, and it is probably about half and half.” Personalized vaccines have lower immune tolerance and are tailored to each patient, but the trade-off is that “the cycle is longer and the cost is higher.” BioNTech’s route and dosing mechanism also differ from Moderna’s. “Moderna’s success represents the success of this technology route, but it does not mean other routes cannot work. The industry as a whole is still in an exploratory phase.”

15. 艾博’s 3-pronged strategy and the cost target reverse-engineered from affordability

艾博 is pursuing 3 antigen categories: TSA, or tumor-specific antigen, consisting largely of driver mutations such as EGFR and KRAS; TAA, or tumor-associated antigen, such as NY-ESO-1 and TP53; and fully personalized neoantigens—“a very short peptide segment, with different peptide segments linked together.”
The dividing line between TSA and TAA is causality versus association. There is clear evidence linking EGFR to non-small-cell lung cancer and KRAS directly to pancreatic cancer, non-small-cell lung cancer, and colorectal cancer. In many other tumors, there is no clear driver gene; researchers simply observe expression of the mutation through tumor sampling, “so we call it a tumor-associated antigen.”
The strategy is: “If there is a clear mutated gene, start with the mutated gene, such as EGFR or KRAS. If not, first look at the associated antigen.” If the cycle can be compressed to 4 weeks—“it looks feasible at present”—and the per-treatment cost can be kept within what patients can accept, “then personalized cancer vaccines, despite being personalized, may actually become a universal therapy.” The reason is that they could potentially work across all tumor types and all patients without needing to look at the tumor type or cause.
He did not disclose the exact pricing target, but used CAR-T as the benchmark. CAR-T’s efficacy is recognized, but its high price makes it unaffordable for insurers and ordinary people, “which sharply limits its addressable population.” CAR-T companies in China also face “extremely, extremely difficult” profitability conditions. US gene therapies can sell for “a million dollars per shot,” while the estimate in China is “at most a few hundred thousand yuan—not the cost, but what ordinary people can accept.” That is treated as a lifetime cost, from which the maximum cost is worked backward. The goal is to be both affordable and profitable.
One hard constraint explains why the company is doing this work now: “Once your algorithm and manufacturing process enter clinical trials, you cannot change them. During the 5- to 7-year clinical period ahead, your entire algorithm and process cannot undergo any changes.”

16. The biggest cost is single-use consumables; saving money means making it smaller, not bigger

On the cost structure, he would identify only one clear major item: “The biggest cost is still these single-use consumables,” because single-use materials are required to avoid cross-contamination. But there is substantial room to improve: “As demand rises and the scale of consumables production increases, the cost per order may fall sharply.” Overall, “which step has the highest cost is still dispersed across different parts of the process.”
Automation saves more than labor. It also reduces “dead volume,” a point he returned to repeatedly. If a person collects the liquid flowing out, they must decide “how much to collect, whether not to collect the initial portion, whether not to collect the final portion—there is loss at every step.” Automation can control the process precisely, minimizing dead volume and waste, which in turn reduces the use of consumables and raw materials. “Every link has to be considered before you can achieve a substantial reduction in total cost.”
His engineering analogy was the best of the episode: “Sometimes making something big is easy; making it small is harder. Going from the brick phone to today’s iPhone, fitting so many components into such a small device is actually the biggest challenge.” 艾博 initially built an entire room-sized automated mRNA production platform, including QC. Now it is working backward toward “a device that is almost portable”—“you press a start button, and a few hours later the entire mRNA vaccine can be produced.” One run would produce “almost the entire dose of cancer vaccine a patient needs over their lifetime.”
When Yushan asked whether 1 shot solves the problem, he gave a practical correction: “One treatment solves one episode, but multiple treatments are still required. It is the same as PD-1: 1 shot every 3 weeks.” Moderna’s course lasts roughly 1 year, while treatment duration may differ by tumor type. Compliance should nevertheless be good, “without severely affecting quality of life.” The path to miniaturization involves having the original scale-up engineering team reverse its thinking, working with external partners to design instruments according to its specifications, then testing and validating them continuously. When Phase 3 involves more than 1,000 patients and commercialization reaches a million-scale patient population, “the only way to control process reproducibility is through individual instruments, one by one.”

17. AI’s real position in the chain: nucleic-acid drugs are ideal for AI, but “AI without automation is just talk”

AI can run through the entire chain: “from predicting the antigen sequence at the beginning, to optimizing the mRNA sequence at the end, and even in the future… the development of lipid molecules, or the delivery system.” Personalized vaccines are the ultimate use case. Fixed antigens such as driver mutations are discoveries built on long-term disease biology—“they do not need to change; they are simply sitting there.” By contrast, “every personalized cancer vaccine is computed, and the quality and accuracy of that computation will, I think, have a decisive impact on clinical-trial results and the eventual extension of patients’ overall survival.”
His case for nucleic-acid drugs as AI’s best landing point is clean: “Nucleic-acid drugs are the most likely to be driven by AI because their design is entirely computational and does not involve overly complex 3D structures. They have 4 letters… a quaternary algorithm, unlike proteins, which have 20 amino acids and 3D structures.” Their foundation is A, U, C, and G; “in a sense, it is a discipline driven by computer science.”
On organization, 艾博 has recruited Ian, the former Moderna president responsible for neoantigen algorithm design, as Chief AI Officer. It is building its own computational-science team and adding automation staff. “This is a very, very unique, nontraditional biotech department,” and hiring continues. While he believes some peers may pursue acquisitions, 艾博生物 is primarily building the capability internally.
His sharpest point concerns the dependency between the 2: “AI without automation is just talk, because AI needs massive amounts of training data. Without automation, it is impossible to train a good AI model.”

18. LNP is the true compute ceiling—and the next 100x efficiency opportunity

艾博 once considered computing the full combinatorial space of LNPs. The company’s investors include well-known domestic AI and TMT investors, as well as international investors such as SoftBank. The conclusion was that it is currently difficult to execute: “The compute required would be extremely, extremely difficult to realize today.”
The reason is structural. The main ionizable cationic lipids in LNPs are small molecules, so “the compute involved is limited.” The difficult part is that “the entire LNP is a self-assembling system involving different lipid molecules and mRNA.” Today, the only option is traditional in-vivo and in-vitro screening, which clearly carries high time costs because it requires experiments. But he did not close the door: “In the future, this may really be achievable… Technology has always turned the impossible into the possible.”
The potential payoff is substantial. “Right now, if 1 mRNA can express 100 protein molecules… if it can express 10,000 protein molecules, that means delivery efficiency has improved 100x.” In that case, many proteins and protein drugs could be developed more quickly through mRNA.
There is an important design discipline: personalization stays on the nucleic-acid side. In a personalized cancer vaccine, “the personalized part is only the nucleic-acid sequence, neoantigen prediction, and the mRNA sequence produced based on the neoantigen sequence. The delivery system is fixed.” The company does not want too many personalized components and products, because that would create greater complexity and less control. Better delivery would reduce dose, lower cost, and further control adverse events.

19. Qualified agreement that vaccines are a “software problem”: the bottleneck is release and regulation, not algorithms

Yushan brought up 2 claims. In a 2025 statement related to the Stargate project, Oracle’s founder said that with AI, “the entire manufacturing process for a personalized cancer vaccine could be completed in 48 hours.” Elon Musk said after the Phase 3 breakthrough that vaccines would henceforth “become a software problem.”
英博 separated the technical issue from the institutional one: “What they said is not wrong. At the technical level, it is a software problem… As I said, it is an engineering challenge, and all of these things can be solved continuously.” But the field still faces “the existing regulatory system, because this is ultimately a drug.” Production release currently takes “within 2 weeks,” and “at present there is no particularly good way to shorten that time further.”
He also cooled the AI narrative: Moderna was already able to do this in “2015 and 2016,” “way before what we call the AI era.” The algorithmic compute requirement “is not as high as people imagine; it is not especially cutting-edge.”

20. The key regulatory insight: approval is for a platform technology, not a product

The central regulatory tension is between version lock and the need to iterate. “Once you enter Phase 1, your algorithm and process cannot change. But from a biotechnology perspective, Phase 1 and Phase 2 clinical trials often still give you room for further improvement.” That is why “before Phase 3, you should ideally apply all the available improvements.” He acknowledged that regulators have made “a major breakthrough in personalized cancer vaccines,” but the process cannot change afterward because doing so could affect safety and efficacy.
Asked how regulators could possibly oversee one patient at a time, he gave the segment’s most important structural insight: “First, regulators need to ensure the safety of the product. So what they are actually approving is not a product, but a platform technology.” Moderna happened to establish the platform’s safety during COVID, while Phase 1 and Phase 2 demonstrated the platform’s reliability at the intended dose. “So ultimately, efficacy is obtained in the Phase 3 clinical trial. But before you can enter Phase 3, regulators first need to see the reproducibility and safety of the process.”
On Moderna’s hope of launching in 2027, he offered no hard forecast but remained directionally optimistic: “We certainly hope such a product can launch as soon as possible, and we believe the company, patients, and regulators all have the incentive to make that happen.” For China, his view was more direct: “The major challenges should no longer be there, because everyone recognizes the platform technology… We do not currently believe there are any genuine regulatory barriers, domestically or internationally, that remain uncrossed.”

21. COVID was a trial run—but why did oncology take more than 10 years?

The spillover from COVID vaccines to cancer vaccines is most visible in process quality. Improving every point in production that can raise purity and reduce impurities, then using that improved process to develop cancer vaccines, “will also produce a huge improvement in the quality of cancer vaccine products.” COVID vaccine manufacturing “was essentially a trial run that accelerated the maturation of the technology.”
Why were COVID vaccines developed quickly while cancer vaccines took more than 10 years? He gave 3 parallel reasons. First, cancer is “much more complex—it is not a virus,” and “cancer is always multi-factorial.” Second, COVID vaccines built on earlier breakthroughs, including “the breakthrough in identifying the RSV vaccine antigen—we found the antigen sequence called Pre-F.” Third, clinical enrollment is radically different: during a pandemic, incidence is high and cases are easy to capture, while enrolling more than 1,000 cancer patients “may require screening far more than 1,000 patients.”
He compressed the contrast into one sentence: “One is given to healthy people, the other to patients. The healthy population is far, far larger than the patient population, while patients also have to undergo large-scale screening to determine whether they meet the enrollment criteria.” That also explains why Moderna’s first post-COVID launches were RSV and influenza rather than a cancer vaccine. He cautioned that the field is still early: “The development of cancer vaccines has only reached the stage of a very positive signal readout. It has not yet been formally approved or launched, and there is still some distance to go.”
The bottleneck in industrialization is not engineering. “The most important thing in industrialization is still producing convincing data. As for industrialization, given the current level of technological development, I do not think it should be a particularly large challenge.” Efficacy, meanwhile, “actually includes everything from the initial algorithm to the final manufacturing process and then the clinical strategy.”

22. China-US comparison: a gap in progress, not as much in understanding—and personalization requires local deployment

His assessment has 2 layers: “There is certainly a gap in progress. Clinically, they are already at Phase 3 with a very good readout. But in terms of understanding, I think the gap is not as large as people imagine. Especially now, with the development of AI and improvements in algorithms, I would not rule out some degree of late-mover advantage.” China also has abundant clinical resources, so “there should soon be some positive results to share with everyone over the next few years.”
On multinational pharmaceutical companies coming to China for projects or acquisitions, he said the transaction structure is fundamentally different: “A personalized cancer vaccine is not a product; it is an entire algorithm and manufacturing process.” The buyer is not simply taking away a product; “it has to deploy your algorithms and processes locally.” The reason is physical: “It is hard to imagine personalized cancer vaccines produced in China or elsewhere being shipped to Europe or the US for patients there.” Neither transport time nor cost permits it.
On geopolitical sensitivity, he sees limited risk as long as the platform is deployed locally and human genetic-resource sequencing data stays within the country. “First, the compute it uses is far lower than we imagine. Second, every country has strong protections for its own human genetic resources. Simply put, human genetic resources cannot leave the country without approval.” With local deployment and no export of genetic resources, “I do not think there is any real major risk.”

23. The 5-to-10-year view—and a closing segment that cooled the heat

He made the episode’s strongest prediction, still qualified with “I think”: “For every indication approved for PD-1, I think mRNA cancer vaccines will eventually be approved as well.” He also expects physician behavior to change. Many clinicians already believe that patients with low PD-1 expression may benefit; once cancer vaccines are approved, “I believe more doctors will think that even in indications without approval, these patients may still derive substantial survival benefit from a personalized cancer vaccine.”
Yushan’s closing brought 2 risks back into focus: competition and data. On competition, BioNTech is also advancing personalized neoantigen therapies; in pancreatic cancer, “early studies have observed immune responses lasting several years, but the actual efficacy still needs to be validated in Phase 3.” On the off-the-shelf side, Elyseal’s vaccine targeting shared KRAS mutations “did not reach the disease-free survival primary endpoint in the overall population in its latest Phase 2 study, with positive signals seen only in certain subgroups.” Moderna and BioNTech are both pursuing dual-track strategies, while domestic players including 艾博生物 and 云顶新耀 are also placing bets on both routes.
The data warning is what investors should remember: “Capital-market enthusiasm often runs ahead of complete data.” As of recording, Moderna had released only preliminary key Phase 3 results—the trial met its primary endpoint, but “the specific HR risk ratio and the results for different patient subgroups have not yet been disclosed.” Those details will have to wait for subsequent academic conferences. After the post-readout surge, “Moderna’s stock had already given back part of its gains over the next several trading sessions,” with the market also waiting for the full dataset. Yushan’s conclusion matched 英博’s caution: “Cancer vaccines are worth looking forward to, but the answer will take time.”

A Middle-Aged Man Pays for Rock: 马智勇 on Live Shows and the Industry Sam Altman: Singularity Slow-Down, Emad Runs 18 Grokbots, Waymo Slashes Hardware 83% | EP #283