Category: 3. Business

Data source

We used data from the Stockholm Creatinine Measurements (SCREAM) project, which contains healthcare utilization data from all residents of the region of Stockholm, Sweden [13]. A single healthcare provider in the Stockholm region provides universal and tax-funded healthcare to 20–25% of the population of Sweden. SCREAM contains complete information on demographics, healthcare utilization, laboratory tests, dispensed drugs, diagnoses (captured by International Classification of Diseases 10th Revision (ICD-10) codes from primary care, specialist care and inpatient care records) and vital status [13].

Study population

We included adults aged ≥ 18 years residing and accessing health care in Stockholm during 2016 to 2021 (last date available currently in SCREAM), with a diagnosis of HF and available subsequent hemoglobin tests. We excluded hemoglobin tests taken during an inpatient stay, or hemoglobin tests performed within 30 days after a bleeding event or a transfusion code and hemoglobin tests within 30 days from a hospitalization discharge, which could relate to the monitoring and/or resolution of an event. After these exclusions, we selected the date of the first hemoglobin test per patient as the index date for our study and the timepoint where baseline covariates were assessed, and follow-up began. At this point we further excluded patients who had conditions affecting the interpretation of hemoglobin values or of potential incident anemia-related outcomes, including recent pregnancy (within 2 years prior), ongoing or recent history of cancer (a clinical diagnosis of cancer in the previous 3 years, excluding diagnoses of melanomas), a hematologic disease, and chronic infections (e.g., hepatitis, tuberculosis, and human immunodeficiency virus) (Additional file 1: Table S1).

Once the study population had been defined, we identified and excluded individuals with anemia at the index date (i.e. prevalent anemia cases, defined by having a low hemoglobin value according to the World Health Organization (WHO) definition: < 12 g/dL for females or < 13 g/dL for males), or having received an anemia diagnosis (ICD-10 codes D50–D64) in the year prior to the index date or having received anemia treatment (erythropoietin stimulating agents (ESA) or iron) in the year prior to the index date (Additional file 1: Table S1). The resulting cohort was then a cohort of prevalent HF cases free from (recent) anemia and with a baseline hemoglobin value.

Incidence of anemia and processes of care

This study comprised two analyses. The first analysis aimed to quantify the population with HF developing anemia and the processes of care around that anemia event. The primary outcome was thus the first-detected anemia event in each patient, defined as a new onset of low hemoglobin measurement (< 12 g/dL for females or < 13 g/dL for males) followed by a diagnosis of anemia, or initiation of anemia treatment (ESA or iron) within 3 months, or a subsequent hemoglobin measurement with similar magnitude between 3 and 6 months apart (i.e., anemia sustained for at least 3 months) (Additional file 1: Table S2). The event date was the date of the first low-detected hemoglobin. Anemia events were categorized by their severity: severe anemia was defined as hemoglobin < 10 g/dL regardless of sex followed by a diagnosis of anemia, treatment initiation, or a second hemoglobin measurement < 10 g/dL; mild/moderate anemia was defined as a hemoglobin measurement < 12 g/dL for females or < 13 g/dL for males but ≥ 10 g/dL, followed by a diagnosis of anemia, treatment initiation, or a second hemoglobin measurement of similar magnitude. Outcomes based on anemia severity were not mutually exclusive, and a given patient may have developed more than one anemia event of varying severity during follow-up.

The clinical work-up of anemia was evaluated in terms of (i) anemia recognition, defined by the establishment of a clinical diagnosis of anemia within 6 months from the anemia event; (ii) testing for iron stores, defined by the presence of at least one laboratory test of ferritin or TSAT within 6 months from the anemia event; (iii) other laboratory testing, including liver enzymes (alanine aminotransferase (ALT) and aspartate aminotransferase (AST), kidney function (serum/plasma creatinine) and inflammation (C-reactive protein (CRP)); (iv) procedures for ruling out bleeding or cancer, including colonoscopy, urinalysis, cystoscopy, and esophagogastroduodenoscopy; (v) initiation of treatments, including recorded blood transfusions, infusions of intravenous iron, and filled prescriptions of oral iron or ESAs within 6 months from the anemia event (definitions detailed in Additional file 1: Table S3).

Clinical outcomes associated with incident anemia

The second analysis aimed to investigate the association between developing anemia and the subsequent risk of all-cause death, major adverse cardiovascular events (MACE) and hospitalization for HF, and new-onset cancer. Here, anemia was considered a time-varying exposure. Deaths were ascertained by linkage with the Swedish population register, which has no losses to follow up [14]. MACE was defined as the composite of stroke, myocardial infarction, and cardiovascular death. Cancer was defined as any new diagnosis falling under the International Classification of Diseases 10th Revision (ICD-10) C category (Additional file 1: Table S2).

Study covariates and stratifiers

Study covariates were derived at index date and updated again at the time of anemia occurrence. Covariates were selected on the basis of biological plausibility and included demographic information (age, sex, highest level of education attained and calendar year); HF type, categorized as HF with preserved ejection fraction (HFpEF:EF ≥ 50%) or reduced ejection fraction (HFrEF: EF < 50%) by using a prediction model derived in a Swedish cohort and validated in a Dutch cohort [15] (Additional file 1: Method S1); history of comorbidities (diabetes mellitus, hypertension, ischemic heart disease, cerebrovascular disease (CVD), peripheral vascular disease, atrial fibrillation, valve disease, chronic obstructive pulmonary disease, rheumatoid diseases, dementia, liver disease, peptic ulcer disease and melanoma); laboratory tests (hemoglobin and estimated glomerular filtration rate (eGFR) [16]; and use of implantable cardioverter defibrillator or cardiac resynchronization therapy, use of CVD-related medications (renin–angiotensin system (RAS) inhibitors (angiotensin-converting enzyme (ACE) inhibitors/angiotensin II receptor blockers (ARBs)), beta-blockers, calcium channel blockers, loop diuretics, mineralocorticoid receptor antagonists (MRA), digoxin, statins, immunosuppressants, platelet aggregation inhibitors, anticoagulants except heparin, non-steroidal anti-inflammatory drugs, and other blood pressure medications. Detailed definitions of these covariates are presented in Additional file 1: Table S4.

We stratified our analyses by the presence/absence of iron deficiency and the type of care identifying the event (setting). Iron deficiency was defined as ferritin < 100 µg/L, or ferritin between 100 µg/L and 299 µg/L if TSAT was < 20% [17]. We considered that anemia was detected and managed in primary care if the hemoglobin test that defined the event was ordered by a primary care unit and there were no records of a cardiology department visit within 6 months; we considered that anemia was detected and managed in cardiology care if the hemoglobin test that defined the event was ordered by a cardiology department or if there was a recorded visit in a cardiology department within 6 months. For cases not fitting these definitions, we considered that anemia was detected and managed in other sources of care.

Statistical analysis

Continuous variables with normal distribution are presented as mean and standard deviation (SD), whereas those with non-normal distribution are expressed as median and interquartile range (IQR). Categorical variables are presented as counts and proportions (%).

Incidence of anemia and baseline predictors

We first calculated anemia incidence rates by dividing the number of events by the person-time, following patients until the first anemia event detected. Then, we assessed time to anemia event, identifying baseline predictors (all those listed in Additional file 1: Table S5), through multivariable Cox regression models, reported as hazard ratios (HRs) and 95% confidence intervals (CIs). For these analyses, patients were followed until the occurrence of anemia, emigration from Stockholm, death, or end of follow-up, whichever occurred first. Continuous variables were standardized as per SD increase, and the relative importance of each predictor was assessed by the estimated explained variance of the outcomes (R²) and the proportion of overall explainable log-likelihood (Χ²) attributable to each predictor in the analysis of variance.

Clinical work-up of anemia

We described the clinical reactions upon anemia occurrence overall, stratified by anemia severity, by the absence/presence of iron deficiency, and by setting of management. Furthermore, we modelled these processes of care across calendar years to evaluate time trends since the instauration of the ESC guidelines in 2016. In addition, we reevaluated the clinical work-up of anemia after excluding patients who died within the first 6 to 12 months after incident anemia to account for the possibility that some of these patients may have been under palliative care, thereby limiting further diagnostic investigations and anemia treatment.

Adverse outcomes following incident anemia

Finally, we estimated the associations between developing anemia and subsequent outcomes through multivariable-adjusted time-dependent Cox proportional hazards regression. Patients were followed from the index date until the occurrence of adverse outcomes, emigration from Stockholm, or the end of follow-up, whichever occurred first.

To evaluate the consistency of our findings, we performed subgroup analyses by sex, HF type, and diabetes status. To evaluate the robustness of our findings, we performed various sensitivity analyses. First, we reanalyzed associations with all-cause death, MACE, hospitalized HF, and cancer after excluding events within the first 180 days after incident anemia to assess the impact of reverse causation bias (e.g., that anemia may have been identified during the workout and/or clinical investigations following another complication). Second, we reevaluated the associations between developing anemia and subsequent risk of MACE, hospitalized HF, and cancer, considering all-cause death as a competing risk.

Missing data approaches

There were no missing data for covariates except for educational level and baseline eGFR, with a missing rate of 3.4% and 1.9%, respectively (Additional file 1: Table S4). Multiple imputation by chained equations using classification and regression trees was employed to impute complete data sets. The imputation model included the exposure variable, all covariates, the event indicator for the outcome, and the Nelson-Aalen estimate of the baseline cumulative hazard.

Statistical analyses were performed using R Version 4.2.1 (R Foundation for Statistical Computing). Two-sided P < 0.05 was considered statistically significant.

A flurry of bets on the football Club World Cup and summer tennis tournaments helped the UK gambling company Entain to better-than-expected results, helping it brace for looming tax rises in its domestic market.

The owner of brands including Ladbrokes and Coral reported an 11% rise in underlying profits to £583m in the first half of the year, on revenues that rose 3% to nearly £2.6bn.

At BetMGM, Entain’s 50%-owned US joint venture, revenues rose by more than a third. US sports betting continues to grow rapidly following the supreme court’s overturning in 2018 of a decades-old ban on the practice.

But Entain also performed strongly in its home market, with online gambling revenues up 21% in the UK and Ireland.

Entain said the Club World Cup, held in the US earlier this summer, had delivered “fantastic results” for the company, despite a muted response to the tournament in the UK.

Entain describes Brazil as one of its “must win” markets, alongside the UK and US, and it has a partnership with the São Paulo team – and Club World Cup participant – Palmeiras. The company said this helped boost local interest in betting on the tournament.

The Club World Cup final, between Chelsea and Paris Saint-Germain, was the year’s most popular event by the number of bets taken. The French Open and Wimbledon tennis tournaments also delivered new records for the company, amid surging interest from gamblers in a sport whose design inherently creates opportunities to place rapid-fire wagers.

Female tennis players have spoken out this year about the abuse they receive, often from gamblers frustrated by having lost bets.

Entain’s chief executive, Stella David, appointed earlier this year after her predecessor, Gavin Isaacs, departed suddenly after five months, said the results showed the company was “getting stronger, fitter and faster”.

Entain’s improved fortunes in the UK could cushion the blow if the Treasury goes ahead with plans to target the gambling industry with higher taxes, as the chancellor, Rachel Reeves, casts around for ways to plug a fiscal hole.

skip past newsletter promotion

after newsletter promotion

Gordon Brown, the former prime minister who also served a decade as chancellor, recently called for gambling duties to increase by £3bn, from about £2.5bn at present, to pay for lifting the two-child cap on benefits.

The Treasury is expected to increase taxes levied on the sector, albeit by a smaller amount than Brown has suggested.

ADI IGNATIUS: I am Adi Ignatius.

ALISON BEARD: I’m Alison Beard, and this is the HBR IdeaCast.

ADI IGNATIUS: All right, so Alison, today we have a topic that’s really interesting to me and that is the parallel development of AI ecosystems in the West and in China. So you think of how the two pursued divergent strategies with the internet, that really is a separate Chinese internet. The question is whether that’s going to happen with AI and the likelihood at this point is that there isn’t. That can exist multiple ecosystems and that companies can tap into them.

ALISON BEARD: This sounds very interesting. I know that you spent a great deal of time in China earlier in your career, but I imagine it’s changed dramatically because of the advance in the internet and now AI. So what’d you learn?

ADI IGNATIUS: Well, when OpenAI debuted in 2022 with generative AI, and we all suddenly had that at our fingertips. China was way behind. They’re not now. So they’ve caught up and they have competitive advantage in some really clear areas.

So my guest today is Amit Joshi, who is a professor at IMD, and he’s suggesting that rather than this being a winner-take-all, or you have to go with one or the other of these ecosystems that global companies can really work with both that there are advantages to Western AI, to Chinese AI, and if you run a complex business, you might want to be engaging with both.

ALISON BEARD: I imagine though that a lot of companies are maybe hesitant because of concerns about data privacy, security that government policy might change and prevent them from continuing that strategy.

ADI IGNATIUS: Absolutely. Look, I mean, we have these concerns when we deal with any AI company, what are the biases, etc. I think with Western companies dealing with China, absolutely there’s a whole new level of security concern, but there are ways to protect your data, and again, Chinese AI companies are doing some things really well. So at the very least, it’s worth educating yourself. So here’s my conversation with IMD, professor Amit Joshi, co-author of the HBR article, How Savvy Companies Are Using Chinese AI.

One of the big questions for global business obviously is the technological race between China and the US, particularly in the development of AI. There are questions of course, about who’s winning, but also whether the two countries will end up pursuing separate parallel paths. So I want to set some context though. So let’s go back to 2022 when most of us first became aware of generative AI’s remarkable capabilities, OpenAI, an American company suddenly makes ChatGPT available to everyone. It’s an unforgettable moment for many of us. What is happening in the field at that time in China?

AMIT JOSHI: So in 2022, when the US market suddenly pivoted to generative AI, China was still all in on traditional AI, machine learning as we call it, right? I mean, at that time, the real war, it seemed at least on the surface was between American machine learning and Chinese machine learning. And then we had the huge Chinese giants, the Alibabas, the Huaweis of the world, JD.com, Tencent obviously, investing massively in traditional machine learning, getting more data sets, putting in more infrastructure for that, and they were completely almost oblivious to this huge development in the GenAI space.

ADI IGNATIUS: Ok so, fast-forward a couple of years and now it’s turn to shake up the world with the emergence suddenly of DeepSeek, at least to those of us who didn’t know. So people in the West are amazed, they’re skeptical even about some of DeepSeek’s claims. Talk about what happened in those intervening years and the significance of DeepSeek’s emergence.

AMIT JOSHI: So the first thing that happened as soon as OpenAI launched ChatGPT, was shock and awe in China. They were completely taken aback. This was a race where they were supposed to be equal partners, if not winning it on some aspects. And now all of a sudden they seem to be laggers, they’re nowhere in the race, they have nothing that looks like it even on the horizons.

However, credit to them, and this is something they’re amazing at, they pivoted very, very quickly. So if you fast-forward just a couple of years from there onwards, which is January of this year, I believe it was, is when DeepSeek actually launched, DeepSeek was launched, and in the interim, they managed to do this faster than what OpenAI had done. They managed to do it for cheaper than what OpenAI had done, and at least at the time it launched, it rocketed right up on the AI leaderboard. They evidently managed to do it better than what OpenAI had at that time.

ADI IGNATIUS: That’s great context. So now we have two essentially parallel but distinct AI ecosystems, and I want to talk about how they differ in terms that are understandable to people like myself, both in terms of how they’re configured and how they perform. So let’s start with how Chinese AI companies are, how they’ve developed the product and how they’re taking AI solutions to the market.

AMIT JOSHI: So if you will, I’d just like to rewind to how the U.S. systems were built. I mean when the U.S. systems were built, remember OpenAI was this tiny little startup with a couple of hundred people in San Francisco. They essentially used the infrastructure, the chips, the storage, the data that was already available at the time. So they used Azure platform, they use the NVIDIA chips obviously, which were not, let’s face it, they were not designed specifically for generative AI. They were designed for other purposes, well, originally for graphics, but then eventually for traditional machine learning. And they built it off of that.

The first things the Chinese did is they said, “Hey, since we know what we are doing now, how about we actually create the infrastructure that’s kind of designed to do these kinds of things? So let’s take something that kind of vertically integrates this infrastructure, let’s put it all together and then make sure that the stack that we are building is meant for this.” So the customization and infrastructure that they did, it was not about saying, “Let me try and build the best general purpose AI tool that I can get away with.” What they said is, “Can I customize my storage, my chips, my data, my training for this one particular purpose?” So this was one of the first things that the Chinese did very successfully obviously.

ADI IGNATIUS: I think this probably points to a comparison maybe more generally between Western approach to business and a Chinese approach to business where one, as you say, is maybe leading the technology wave and at a certain altitude and where the Chinese maybe are super focused on consumer segments and application. So talk more about that and maybe even examples of what Chinese companies are doing with their AI products.

AMIT JOSHI: So again, the contrast between what’s happening in the West versus what’s happening in China is pretty amazing in this sense. I mean, let’s take the usual Western models, whether it’s ChatGPT, whether it’s Gemini or even open source models like Llama. What we’re trying to do is we’re trying to build the best model that we can do using the highest of the best quality chips, the best storage, the best data or the most amount of data that we have. Contrast that with what the Ant Group did. The Ant Group, they said, “We want to build an AI medical app that’s going to be available to people who have the Alipay app. So it’s an AI doctor that’s available in Alipay. I could use a general purpose GenAI tool, but what if I create a healthcare specific model that uses data specifically from hospitals, that uses infrastructure, that allows this model to make quick inferences faster than a general purpose model and then put that on the app.” Nothing contrasts what we are trying to do in the West versus what they’re trying to do better than this example, in my opinion.

ADI IGNATIUS: So the American president, Donald Trump, his AI plan, at least to the extent that he’s laid it out, is set at trying to achieve U.S. dominance in the field. And I’m interested in your view, is this a winner take all as sort of Sony versus Betamax or can more than one of these ecosystems survive?

AMIT JOSHI: I do not think this is a winner-takes-all battle. I think if we fight it as a winner-takes-all battle, we are completely missing the point. I have a feeling that this is a space where we will have multiple foundation models that will probably be focused on certain areas that’ll be better for one versus the other. But I do think this is something where multiple different models, multiple different technologies will coexist. So in that sense, this is less like social media where we have one dominant social media platform and one chat platform and one search, etc. But this is more akin to mass production where we’ve got about a dozen or more car companies in the world, all of them more or less successful, all of them differentiated in some way.

ADI IGNATIUS: So let me just push on that a little bit because we do sort of have separate and distinct internets. I think in the early days of the internet we didn’t even imagine that was possible, but there really is a distinct Chinese internet and it is almost a parallel and unbridgeable kind of system. But you think AI is different. Explain why.

AMIT JOSHI: First of all, specifically if you stick to social media, a lot of the power for social media comes from network effects. The facts that the value that one person gets from using, for example Facebook or WhatsApp, is heavily dependent on how many total users there are on this. This is not easily transferable. This concept to OpenAI mean as long as OpenAI is providing me with quality answers, I don’t really care if a billion other people are using it. Now there are some technical things, reinforcement learning with human feedback that’s happening in the back end that might marginally make OpenAI’s answers better than, for example, Llama, which is lesser users, but the quality difference is going to be relatively minor.

So because this is not driven primarily by network effects, at least as of now, my sense is that this is going to be akin to an economies of scale kind of a business rather than a network effects business. Now, a network effects business, we know it’s winner takes most if not winner takes all, but in the economies of scale business, we know for the last 80 years that multiple businesses can survive in parallel, they can coexist.

ADI IGNATIUS: So this starts to get to the crux of the HBR article that you co-wrote and that is that global companies have to figure out how to navigate these two very important parallel AI universes. So let’s start with a basic question. Can global companies engage in both of these?

AMIT JOSHI: I think they can. We’ve already started seeing a few examples of companies that have started exploring both ecosystems and figuring out which can work for both. So let us remember one thing. The leading Chinese model that’s out there right now, which is DeepSeek is an open source model, which basically means I can take it, I can modify the weights and I can put it on my servers with minimal cybersecurity risks or with minimal privacy risks, et cetera. It’ll never be zero, but they are minimal. Knowing this, it then becomes imperative for Western companies to try and understand saying, “Okay, I have access obviously to the Western models, but I also potentially have this Chinese model.” And by the way, we were discussing previously, the Chinese models have been trained on the Chinese internet. In addition to the internet, which the Western models are trained on, they can do different things.

They do specialize in different aspects. So is it possible that moving forward, Nestle or Airbnb or Siemens or GE, they discover that if I really want to do high-end innovation, if I really want to do cutting-edge stuff, if I want to do new drug discovery if I’m Pfizer, then it’s much better for me to use some of the cutting-edge Western models. On the other hand, if I want to do customer service, if I want to do supply chain optimization, maybe the Chinese models are better and then I set up my system in a way that depending on the problem, the AI can triage what’s happening,

ADI IGNATIUS: Talk in more detail about one or two of these companies and how they’re engaging with Chinese AI.

AMIT JOSHI: One is BMW. What BMW is planning to do is, especially for cars that are for the Chinese market, but in general cars in the Asian market, it’s integrating DeepSeek, which is obviously the Chinese large language model into its vehicles. So a lot of the AI in BMW vehicles is going to be done by DeepSeek, staying still in the automotive field. Bosch, which is obviously a massive tier one supplier to automotive companies, they’ve just finalized a high-performance computer for AI-enabled vehicles. So they’re going to make massive amounts of these based on the Chinese AI.

And then of course in the other space, in the consumer packaged goods space, we’ve got folks like Procter & Gamble. And what P&G is doing is it’s using Chinese AI along with Western AI to hyper-personalize the messages that is going out. So it’s partnering with Douyin in China to do what’s known as interest-based e-commerce. And what this essentially does is it combines short videos. The short video e-commerce is huge in China. It’s not huge outside of China, but combine that with discovery and understanding which products to recommend in an e-commerce purchase. So really, really we’ve got companies all the way from consumer packaged goods to durables like automotive to industrial goods like Bosch that are kind of already stepping into this dual paradigm.

ADI IGNATIUS: Are the Chinese better at, let’s call it, hyper-personalization with AI? Are they more advanced or more focused on that than Western counterparts?

AMIT JOSHI: So what the Chinese are better at is industry-specific hyperpersonalization. So if you just take general hyperpersonalization, I don’t know if they’re better, but if you ask me, are the Chinese better at hyperpersonalizing in healthcare? My answer to that would be probably yes. Are the Chinese better at hyperpersonalizing in B2B marketing? My answer to that will probably be yes. So I had a conversation with Chris Tung who’s the CMO of Alibaba, and what he said is what they’re using their AI for is to allow the vendors on Alibaba. So they’ve got a couple of hundred million vendors on Alibaba, which is basically the most dominant B2B wending site essentially in Asia.

And what they essentially do is, I mean the classic thing, I mean if I’m a vendor and I’m selling 50 different types of mugs to buyers, to retailers all over the world, I will never ever have the resources to do the marketing materials, to do all the kinds of pictures, the descriptions, etc. For all the 50 types of mugs that I’m selling, I’ll probably go to an agency and have it made for the five best-selling mugs that I have on my portfolio. What I can now do is I can use Alibaba’s AI to create marketing material, images, descriptions, etc., for the remaining 45. And now all of a sudden I have this scaled up amazing personalization that’s possible, which still yesterday was simply not doable.

ADI IGNATIUS: Then what about the opposite? To what extent are Chinese companies operating in the US AI ecosystem?

AMIT JOSHI: I think that to the best of my knowledge is extremely limited so far, primarily driven by restrictions, but also because I think that they are currently more focused on their own internal market. Their own internal market is growing so rapidly, including the government sector. By the way, the government sector has now turned into a massive user of the Chinese ecosystem of the Chinese AI ecosystem for everything from governance to city management to taxes, to you name it. So I think the growth in the pie that they’re seeing in their own backyard is just incredible. They’re not currently actively pursuing the US market to the best of my knowledge, but I bet that’s going to change soon enough.

ADI IGNATIUS: So now let’s talk super practically. All right, so if people are listening to this and thinking, okay, that’s interesting. I need to think about if I’m a global company, I need to think about whether and how I want to be in the Chinese AI ecosystem in addition to the Western ecosystem, where do you start mean? How do leaders, what exactly should they be thinking about or looking at if they’re trying to answer that question?

AMIT JOSHI: There’s a few things I would recommend. I mean, if you are an executive, if you’re an executive listening to this, go play with DeepSeek right now, right? Figure out a way, a safe way, a sandbox, download DeepSeek, make sure it’s cut off from any private secure data if that’s a problem with you, but play with it. Understand how it is similar, but also how it is different than the Western models, for example. What works better on DeepSeek versus what works worse? How does it do on hallucinations? How does it do on answering factual questions, et cetera? Because it’s not a level playing field. It’s better at some worse for some.

Secondly, and more importantly in my opinion, keep a broader eye on what’s happening in China, not just in the LLM market, but also in the larger business models market. To me, the largest story that will probably come out of China eventually, I think two years from now, we won’t be talking so much about how the Chinese models might be better or worse than the Western models.

I think what we might be talking about two years from now is look at this cool business model that they have built based on AI, because this is exactly what we saw with the mobile revolution, right? Were there mobile phones better than ours? I don’t know, maybe it depends who you ask. Did they have significantly different business models that were based on them? 100% yes. So I think this is something executives at businesses need to keep a very close eye on, which is what kind of different business models, what kind of different use cases are the Chinese building based on this technology that they now have?

If you are an American company that’s based in America and using American models so far, I think it’s in your best interests to actually look at these models. It’s quite likely that these models are not only going to be better, but they’re going to be cheaper to do some of your tasks. Why would you want to cede that advantage to competition? Simply because competition did the due diligence and looked at the Chinese models.

ADI IGNATIUS: So China is a particular challenge for Western businesses. My wife is in the academic world with China. When she visits China, like everyone else in her world, she gets a burner phone, she gets a burner laptop, she doesn’t risk her data. I don’t know if that’s excessive paranoia or if it’s proper paranoia. So now we’re suggesting that Western companies engage in China’s AI ecosystem won’t some of those same concerns about data privacy, data integrity – how do we think about that?

AMIT JOSHI: And as if that was not enough to think about already, let us also throw in regulations because the regulations are probably going to be significantly different across these geographies and how western companies need to handle that. I do not want to underplay the security, privacy and even ethical issues that come with using these systems. They exist. Okay? For now, DeepSeek is open source, so you can actually take it and you can put it on your own servers with reasonable amount of safety, a reasonable amount of confidence that nothing crazy is happening in the back end. Not a hundred percent, obviously nothing is a hundred percent on this planet, but reasonable.

But moving forward, there are going to be probably different models that are not open source that will need to be based off of their complete infrastructure stacks. And in that case, Western companies will need to make an informed decision on whether or not they want to play in that ecosystem. Absolutely, yes. Both from a privacy security perspective, but also from a regulations. And then finally from an ethics perspective.

ADI IGNATIUS: And then you’ve got just the uncertainty about relations between Washington and Beijing. As we record this, they’re better in some ways and they have been. Donald Trump is probably going to meet with Xi Jinping before the end of 2025, and that could improve relations. But you don’t know. I mean the latest headline is that President Trump is going after Intel because its CEO has had ties with “Chinese communists.” Well, anybody in authority in China as a communist.

There can be problems in that area. This is not necessarily one that Intel has already said that this is ridiculous, but it is a volatile relationship and I wonder if politics could force companies to choose at some point, and is it responsible to do scenario planning? As you’re thinking about engagement in the ecosystem, I guess to what degree do companies need to think about these kind of worst case scenarios?

AMIT JOSHI: They have to absolutely factor that in irrespective of what’s happening currently politically in the United States. I think we can all agree that we’ve generally across the world, entered an era of more uncertainty of greater volatility for a variety of reasons, whether it’s political, whether it’s environmental, whether it’s around regulations or whatever that is, this is just more volatility. Companies will need to hedge their bets. I mean, it’s not going to happen that just because a couple of presidents don’t like each other or putting tariffs on different parts of the world that you’re going to stop doing business in that part of the world.

And if you then want to continue doing business in that part of the world, you need to be integrated into that ecosystem, into that particular technological stack. So if nothing else, from a pure scenario planning perspective, from a pure hedging perspective, organizations do need to look at this and relate it aside. You know what tool is really great for doing scenario planning? It’s actually DeepSeek and you know why? Funnily enough, it’s because you can actually drill down and see its chain of thought. So there’s a little button that you click and then it actually shows you how it thought about the answer that it actually gave you on scenario planning, which can be tremendously useful for executives. So for folks listening, I encourage you to try it out.

ADI IGNATIUS: Let’s say that some of these risks are manageable and get back to where we were and not worst case. So again, there’s this idea that companies may be running parallel GenAI models and applications. I just want to understand in really practical terms, what does that look like? What does that mean? To what extent is that simple and everyone’s already doing it, or to what extent is that going to force people to change how they think about business?

AMIT JOSHI: I think at some level, most savvy organizations already doing it, but they could be doing it between Gemini and ChatGPT. So most organizations that have thought this through and have a reasonable technological architecture, what they’re doing is if a query goes into the company GPT, for example, they’ve got a little engine in there that says, “Hey, for this particular query, going to Llama might actually be cheaper, faster, use less tokens and give a more accurate answer than going to ChatGPT.” So just as an example, there’s a bank that I’m working with in Asia that has had the system in place pretty much since the early days of ChatGPT, and then they keep adding the different LLMs that come in and they keep optimizing their triage engine that does this. So at some level, savvy organizations already working on this. Having said that, adding Chinese large language models, it’s not just about adding one more LLM to this. They will have to think this through a little bit deeper because of the reasons that we spoke about previously, which was around cybersecurity, privacy, ethics as well as regulations.

ADI IGNATIUS: Well and bias. We certainly know the Chinese internet is scrubbed of or the censors certainly try to successfully scrub the internet of things that the Chinese ruling Communist party thinks are damaging to their image or to China’s image. I assume we would have to have the same concerns about Chinese trained large language models that they may be scrubbed of certain pertinent relevant material. That would be a bias that could be tricky for Western companies.

AMIT JOSHI: Without a question. And obviously, again, I encourage listeners to try this out for themselves. I mean, if you are going to use DeepSeek, go ask it what happened in Tienanmen Square and see what answer you get.

ADI IGNATIUS: What do you get? Have you done that?

AMIT JOSHI: I have, multiple. This is the first thing I did when I went to DeepSeek and it says basically, I mean I’m paraphrasing here, but it said nothing happened. It was a nice sunny day and you can go on and on about controversial topics and it’ll essentially just back out, right? It’ll essentially say, “Hey, look, I don’t have the answer to this question,” or “I’m sorry, I cannot provide an answer to this question,” which is fine. I mean this is something, but you’re absolutely right. I mean there are some aspects, I mean especially for organizations that are going to be dealing with things like this that are going to be dealing with kind of ethical outcomes. These are things that they will need to consider, which is not to say that the Western models are completely free and clear and completely above board. They have their own sets of scrubbing that goes on, which also we have to account for.

ADI IGNATIUS: Okay, so we’ve talked about some of the risks, so I want to throw it back to you. Your research has suggested that what Chinese AI companies are doing in some ways is so extraordinary that Western companies would be foolish not to at least experiment with them. So make the case – what is it that the Chinese companies do so well that western companies really need to see if there’s a place for them in that ecosystem?

AMIT JOSHI: So what the Chinese have done really well, and which in our paper we call the three Cs, but which in simple words is they have created an ecosystem that from the ground up is customized for GenAI. They have taken that in a low-cost model, and then finally they have built it such that all their applications are calibrated for the real world. They’re not calibrated to win competitions. That’s an aside, but they’re calibrated for healthcare, they’re calibrated for pharma, they’re calibrated for supply chain, they’re calibrated for CPG. This is something that western companies should not ignore.

I’m not saying that we cannot do it in the West. I’m not saying that the Western models are not capable of that, but this is something that the Chinese are trying differently. They’ve been very, very successful. There’s a lot that we can learn just as they learned from looking at our models. There’s a lot that we can learn from looking at theirs, and it would be silly of us not to do that. And then of course, the big one, in my opinion, what business models these folks build based on these tools, that’s going to be the real kicker.

ADI IGNATIUS: Yeah. This will be fascinating to watch how this unfolds and there’s an option for Western companies to form, not just to plug into a DeepSeek or another LLM, but to actually form deeper partnerships. What’s your advice in terms of how Western companies might think about that?

AMIT JOSHI: Most Western companies, most companies around the world are now looking at agents and agentic AI, et cetera. Agentic AI is in many ways ripe for these kinds of partnerships. So please do a look at how some of these tools can fit in to your overall AI portfolio, overall AI strategy. It’s unlikely, almost impossible that you’re going to be sticking to just one tool, one type of AI or one infrastructure. You will be using a variety of these anyway, inside, most medium to large organizations look at these, look at tools like Manus, which was a Chinese company which recently moved to Singapore, which is building agents on top of these tools, and it’s using a combination of Western models and Chinese models already. This is what I would advise executives, please don’t miss out on this opportunity.

ADI IGNATIUS: Yeah, so this is fascinating. I mean, this is the market to watch, and I think you’ve put your finger on the fact that there really has been this sense of distinctive and parallel development, and now we’ll kind of watch how it evolves and it seems really smart to say companies need to look at all of this that’s out there and figure out the best solutions for themselves. So why don’t we end what’s something listeners could do right now to kind of educate themselves better about these possible opportunities?

AMIT JOSHI: Figure out your favorite or go online, go to Google, go to your favorite search engine, find out a few sources of information, a newsletter from China that you would want to subscribe to, get more information about what’s happening in their ecosystem. Then go ahead and play with a couple of these things. I’m a huge believer in the fact that these tools are not something that you can learn from watching a YouTube video. This is so that you really need to get your hands dirty with, so create a sandbox for yourself. Don’t put anything private or confidential in there for obvious reasons, but then fool around with it. Ask it nasty questions, ask it politically incorrect questions and ask it normal questions and see what kind of output it gets and compare.

ADI IGNATIUS: All right, Amit, thank you for the article that you co-wrote for HBR and thank you for being on IdeaCast today.

AMIT JOSHI: It was my pleasure. Thank you so much for having me.

ADI IGNATIUS: That’s IMD professor Amit Joshi, who co-authored the HBR article, How Savvy Companies are Using Chinese AI.

Next week, Alison will speak with futurist Nick Foster on how to reframe your future planning.

If you found this episode helpful, share it with a colleague and be sure to subscribe and rate idea cast in Apple Podcasts, Spotify, or wherever you listen. If you want to help leaders move the world forward, please consider subscribing to Harvard Business Review. You’ll get access to the HBR mobile app, the weekly exclusive Insider newsletter, and unlimited access to HBR online. Just head to hbr.org/subscribe.

And thanks to our team, senior producer Mary Dooe, audio product manager Ian Fox and senior production specialist Rob Eckhardt. And thanks to you for listening to the HBR IdeaCast. We’ll be back with a new episode on Tuesday. I’m Adi Ignatius.

Adjuvanted imasapepimut and etimupepimut (Cylembio) in combination with pembrolizumab (Keytruda) did not achieve statistical significance with the primary end point of progression-free survival (PFS) vs pembrolizumab alone in patients with unresectable or metastatic melanoma in topline results from the phase 3 IOB-013/KN-D18 trial (NCT05155254), according to a press release from the developer, IO Biotech.¹

It was reported that though the treatment did not demonstrate statistical significance, it did yield a clinically meaningful improvement in PFS. The early and sustained separation of the PFS curves showed an improvement with the combination (HR, 0.77; 95% CI, 0.58-1.00; P = .056); the noted threshold for significance was P <.045.

In an intent-to-treat analysis, the combination treatment elicited a median PFS of 19.4 months compared with 11.0 months with pembrolizumab alone. A trend towards improvement regarding overall survival (OS) was observed (HR, 0.79; 95% CI, 0.51-1.10); the data were not fully mature, though the developer expects OS to mature over the next 6 to 9 months.

The improvement to PFS was noted across all subgroups. A profound effect was observed in patients with PD-L1 negative tumors, with a median PFS of 16.6 months with the combination and 3.0 months with the monotherapy (HR, 0.54; 95% CI, 0.35-0.85; P = .006). A post-hoc analysis showed that those without prior anti-PD-1 treatment demonstrated a median PFS of 24.8 months vs 11.0 months, respectively (HR, 0.74; 95% CI, 0.56-0.98; P = .37).

It was reported that more detailed results will be shared at an upcoming medical conference. The developer plans to meet with the FDA in the Fall of 2025 to discuss the data and the next steps for submitting a biologics license application for the treatment in advanced melanoma.

“In this study, patients treated with [adjuvanted imasapepimut and etimupepimut] in combination with pembrolizumab have achieved the longest median PFS ever observed in a phase 3 clinical study in advanced melanoma, and in the PD-L1 negative population, patients achieved a remarkable 16.6 months of median PFS, compared with 3.0 months in patients treated with pembrolizumab alone,” stated Omid Hamid, MD, director of Clinical Research and Immunotherapy at The Angeles Clinic and Research Institute, A Cedars Sinai Affiliate, in the press release.¹ “The significant benefit seen across patients with poor prognostic factors, including patients who are PD-L1 negative, cannot be overlooked. Given the notable safety profile and the strong clinical effect observed with [adjuvanted imasapepimut and etimupepimut], as well as the unmet need in [patients with] advanced melanoma. [Adjuvanted imasapepimut and etimupepimut], if approved, has the potential to become a new standard of care for patients with advanced melanoma.”

IOB-013/KN-D18 is a randomized, open-label trial that enrolled a total of 407 patients who were randomly assigned to receive either adjuvanted imasapepimut and etimupepimut with pembrolizumab (n = 203) or pembrolizumab alone (n = 204). Treatment consisted of subcutaneous adjuvanted imasapepimut and etimupepimut at 85 µg every 3 weeks for up to 35 cycles, with additional doses given on day 8 of cycles 1 and 2; and intravenous pembrolizumab at 200 mg every 3 weeks for up to 35 cycles. In the monotherapy group, the pembrolizumab dosage was the same.²

Eligible patients were diagnosed with histologically or cytologically confirmed stage III or IV melanoma and were treatment-naïve, having not received previous systemic anticancer therapy for unresectable or metastatic melanoma. Patients also had at least 1 measurable lesion per RECIST v1.1 and provision of archival or newly acquired biopsy tissue not previously irradiated.

Trial exclusion criteria include known or suspected central nervous system metastases, receipt of prior radiotherapy within 2 weeks of start of trial treatment, and BRAFV600-positive disease that experienced rapid progression and/or receipt of standard first-line therapy with BRAF or MEK inhibitors.

The primary end point of the trial was PFS. Secondary end points included overall response rate, OS, durable objective response rate, complete response rate, duration of response, time to complete response, disease control rate, and incidence of adverse events.

No new safety signals were observed with the therapy, and the combination was well tolerated. The most reported adverse effects were infusion site reactions, which were resolved on treatment; 56% of patients who received the combination reported an event.

“Since reporting the positive outcome of our Phase 1/2 study (MM1636; [NCT03047928]) in a similar patient population, we have been eagerly awaiting these results supporting the activity of [adjuvanted imasapepimut and etimupepimut] combined with an anti-PD-1 in patients with advanced melanoma,” said Inge Marie Svane, MD, PhD, professor and director of the National Center for Cancer Immune Therapy at the Copenhagen University Hospital, Herlev, and principal investigator in the phase 3 trial, in the release.¹ “These data provide evidence that a therapeutic cancer vaccine can improve PFS in patients with metastatic disease.”

References

1. IO Biotech announces clinical improvement in progression free survival demonstrated in pivotal phase 3 Trial of Cylembio® plus KEYTRUDA® (Pembrolizumab) for the treatment of first-line advanced melanoma, but statistical significance narrowly missed. News release. August 11, 2025. Accessed August 11, 2025. https://tinyurl.com/y7k29pub
2. IO102-IO103 in combination with pembrolizumab versus pembrolizumab alone in advanced melanoma (IOB-013 /​ KN-D18). ClinicalTrials.gov. Updated January 9, 2024. Accessed August 11, 2025. https://tinyurl.com/umnaztt8