Monthly Macro: China's Open-Weight Challenge and the Future of AI Economics

00:05, Chris

Welcome back to another episode of Monthly Macro. Today is Tuesday, August 4th, 2026. This month, we're taking a slightly different angle and digging into one of the bigger debates happening across the AI landscape, which is the rise of low cost, open weight Chinese models and what that could mean for model providers, hyperscalers, software companies, and the broader enterprise adoption curve.

The conversation has moved quickly. Not long ago, open weight models were either dismissed as too rudimentary or viewed as a niche alternative to proprietary frontier models.

But with models that are approaching frontier level performance at a fraction of the cost, investors are starting to ask whether the price of intelligence is falling faster than expected, and what that means for the economics of AI.

So, joining Richard de Chazal to help him unpack all this is Arjun Bhatia. He's William Blair's co-group head of technology, media and communications research and somebody we've had on this podcast a few times.

Arjun recently published a report looking at how low cost, open weight Chinese models could reshape competition at the model layer, drive greater inference demand, and create both risks and opportunities across the AI ecosystem.

So, Richard, I'll turn it over to you to kick it off, and I'll come back at the end. Thanks very much.

01:51, Richard

Chris and Arjun, thanks for coming on to the podcast to discuss this. I think, unquestionably, this really has to be the most important topic for financial markets and the economy at the moment. And, you know, as Chris was saying, it is really incredible how quickly these things are moving.

I mean, I think about two years ago when we were sort of first introduced to these open weight models, we had that big threat of deep seek that came in, and I think that was, sort of, brushed aside because they were kind of using frontier model data.

And again, most of it seemed a little bit more basic compared to what the US was churning out. So that, sort of, died down.

But I think over the last few months, that threat has really jumped up once again, and we're starting to see some incredible models coming out of China. And we're starting to see some more significant stock market volatility taking place on the back of that.

So, I thought maybe it would be good to sort of start with the basics here. What’s the difference between these kinds of open weight models and the models that are being produced in the US, these, sort of, closed weight models? And, maybe, why does that matter?

03:13, Arjun

Yeah, absolutely.

Well, first of all, thanks for having me on. I couldn't agree more. I think in terms of just how fast things are moving, and it almost seems that the pace of change is accelerating. You know, this open weight topic obviously is very important, but in general, it just seems that things are moving very, very quickly.

But yeah, the open weight debate… I think it's super interesting, super relevant, has implications across the entire AI landscape, from what it means for compute and hyperscalers to what it means for the frontier model providers, which everyone, you know, from the investor audience to even, kind of, like people are paying a lot more attention to industry experts, etc..

And basically, what's happening is, you know, these open weight models, and most of them are Chinese, they are, you know, think of it as similar to open source.

Basically, what they've done is they've exposed the weights and parameters of the model, and they make those public, and they post them on an open model repository, like Hugging Face for anybody to go and download.

And, you know, think of the model weights and the parameters as basically the algorithm or the intelligence that will make up the model.

So, this is really the secret sauce. You're not getting, you know, a few things. You're usually not getting the data the models are trained on. You're not getting, you know, the source code, but you're getting the weights, which are defining how the model thinks, and how the model connects dots, and generates output.

You know, the reason this is, sort of, having such an impact is these open weight models on various third-party performance benchmarks are now almost matching the frontier models that are mostly US based companies OpenAI, Anthropic, Google. And these three are closed proprietary models, meaning the frontier model providers don't make the weights or the parameters of the model public.

You have to buy the model and usually buy tokens through the model provider or some sort of authorized partner. Whereas, for the open weight models, you can, as I said, you know, you can download it, you can go to a dedicated inference provider that's hosting it, and you get access to the weights, and you can customize it how you like for your specific use case.

And the economics are also different, right? Because the open source or the open weight model providers don't really make any money. Actually, they make no money from selling the model because, they're not selling the model.

They’re posting it on Hugging Face for the world to download. They try to make money from selling inference, but they're not the sole inference provider because, you know, a US company could easily download the model hosted on AWS or their hyperscaler of choice. They can host it on their own data centers, in which case, you know, there's no, sort of, economics that will flow back to the Chinese open source model providers.

So that's, sort of, at a high level the difference and what's happening in the ecosystem. But I think in a lot of cases, it's just the importance of this is that there's intensifying competition in the space.

And it just happens to be coming from these open weight Chinese models that are now matching performance from the frontier model providers that, you know, had a big lead.

06:53, Richard

One of the things you said, actually in your report that you and the tech team put out on the implications of low cost models was that you said, actually the debate, maybe, the open versus closed is maybe the wrong debate. Can you can you unpack that?

07:14, Arjun

Yeah, it is definitely a debate. But, you know, I think just making it about open and closed, you know, we might be missing the point a little bit. What this is really about, I think, is that, as I said, competition is intensifying. And, you know, if we, kind of, rewind back two years, and, you know, since AI started two years ago, the market was just hyper-focused on the Frontier Labs.

And I would argue, even, you know, up until June of this year, the market was focused on the Frontier Labs, OpenAI, Anthropic, and Google.

But now you have this group of model providers—again, open-weight model providers out of China. Whether it's Moonshot, Kimi—and that's Moonshot, Kimi is the one that really made a splash about a month ago.

But you also have Alibaba, you have MiniMax, you have Z-AI. There's a host of them.

But the big thing is, they're matching performance of the frontier models. It doesn't matter if they're open or closed. It's just that you have, now, someone that is as performant as OpenAI and Anthropic, or almost as performant, right? Which we didn't think would, you know, would happen.

We thought these three companies in the US just had such a big lead. They had access to the best hardware from Nvidia and others, and that they would just continue pushing the frontier. But now, you know, you have competitors that are catching up. They have not pushed the frontier yet, but they're certainly catching up. And, you know, they're getting sort of 80%, 90% of the functionality, maybe at a tenth of the price in some cases.

So, the price question is also important. And I think a lot of the times it gets looked at through this open-versus-closed lens. But it doesn't necessarily have to be an open-versus-closed debate, right? Like, you can have US closed-model companies that can produce models that are lower inference costs, right?

So, the US model companies can bring down their token costs to be more competitive with these Chinese open-weight model companies. Or China can make their models closed, and they could serve them at a low token cost to be more price-competitive with the US model company. So, I think the idea here is the cost of intelligence is coming down.

And what does this mean? You know, what does this mean, a little bit, for the US closed-model companies is, you know, can you push the frontier enough to maintain premium pricing power? Or is the gap between the frontier and close-to-the-frontier just permanently sort of shrunk, that, you know, the business model may not allow you to capture, you know, outsized economics, as you would have if you had, you know, a lead that was, you know, years or miles long?

So, yeah, I think you can look at this through open versus closed. But I really think the point is, there are now a group of other companies, model companies, that have almost matched frontier model performance.

10:31, Richard

And it seems to me that, you know, from companies that are facing suddenly huge compute costs, like I saw, you know, someone on LinkedIn saying, "Does anyone know how to reduce Claude costs from $100,000 to $4,000? If you do, please DM me."

You know, I think the cost thing is certainly a factor. And, you know, at the end of the day, it seems like some of these open-weight models, even though they're maybe not fully as good as the frontier models, you know, they're good enough.

You know, at the end of the day, we don't all have to drive Ferraris to the grocery store when maybe, you know, a Ford Pinto will do. What do you think about that?

11:18, Arjun

Yeah. I mean, I think good enough, at a fraction of the cost, it's a pretty big change in the landscape. And, you know, we've seen all these articles and all this media reporting about companies and their employees token-maxing and running through their AI budgets for the year in a matter of, you know, a couple of weeks.

That's not sustainable, right? So now, if you're posed with, "Oh, we have this good-enough model that, you know, again, you're not getting 100% of the functionality, but you're getting it at a tenth of the cost."

Yeah. You're going to see companies, I think, flock to that. I think we're already seeing that, right? Like, you know, you see consumers will do that.

I think businesses across the Valley, like startups across Silicon Valley, are already using, you know, these low-cost providers, even though they're not the best. You know, like I said, they're 90% of the way there.

If you can manage your costs, you know, it makes the ROI hurdle so much easier. And now, you, even if you're, you know, reselling intelligence, and I kind of use that in a, in a loose way, but, you know, if you bundled it into another service, and, you know, intelligence is a part of your value stack, now you, as the service provider, can get a bigger margin.

And so there's so many benefits, I think, of, you know, lowering of the cost of, you know, of intelligence. And I think, naturally, you'll see businesses flock to it, from startups to even more established companies, right? That have announced, or either considering, or are already using lower-cost models.

So, I think, in this case, right, there's many use cases, I'll say, where good enough is fine, right? Like, in customer service, your AI doesn't need to, you know, know the history of the universe. It needs to know your product portfolio and your customer service policies. And that's good enough.

13:18, Richard

Yeah, which leads us to the next question then.

You know, if I'm an investor, and if some of this is becoming commoditized, like electricity or maybe cloud computing, presumably there are some winners there, ie the adopters.

You know, how should we think about what happens to the enablers? Are they now just putting a ton of weight on Jevons Paradox, that sort of, it all gets made up by volume?

13:48, Arjun

Yeah, I do think there is, sort of, truth to that analogy, or both of those, maybe, of cloud computing, you know, electricity, others, where if you lower the cost, there should be more consumption, as I was saying earlier, right?

I think the businesses, enterprises will look at this through an ROI lens, you know, if we're adopting a new technology, you know, what's the ROI hurdle we need to hit to make sure it makes sense from a business perspective to use it?

And if the cost is lower, you know, there's going to be more adoption, because there's more use cases now that fit this ROI hurdle that enterprises look at. So, I do think it's sort of unambiguously good for the adopters, and for the users of AI, and for the enablers, for the model companies.

Yeah, this is, you know, I think their growth will come through volume.

Maybe not price, going forward. And so, yeah, you're sort of relying on Jevons Paradox, of just saying, "Hey, the usage growth will outpace the downtick in price over time, so I think these can still be good businesses," right?

And I'll use Cloud as an example. So, you know, AWS, and this is true for the other hyperscalers as well, you know, since AWS started, they've had 130 price cuts for their compute services since 2006, when they launched.

And today it's a $200 billion run-rate business. The price cuts did not slow down adoption. And you can look at that in aggregate for Azure and GCP, also, it's the same thing. It only increased it, because businesses could then say, "Okay, hey, I can move this to the cloud, I can move that to the cloud."

And I think it's going to be the same thing in AI, where you can say, "I can apply AI to this use case and that use case," and now, you know, this n of maybe one or two ends up being an end of a n of hundred.

And, you know, I think the exciting thing about AI in general is that it's a platform technology. So, the use cases are only, you know, as limited as, sort of, the imagination of the users that sort of build on top of it.

So, I think there's so many use cases right now that we can't even imagine that will, you know, emerge over the next several years and decades, that should drive volume higher, which ends up actually being a pretty good outcome for the enablers, who should benefit from scale over time.

16:23, Richard

So, thinking about the hyperscalers, and they're sort of building out the infrastructure, do you think they need to be a little bit more cautious there?

That, you know, that doesn't seem to be what we're hearing so far coming out of earnings season. They still seem to be sort of full thrusters on. No?

16:39, Arjun

Yeah. Yeah.

The costs, CapEx costs, are going up. No one seems to want to, you know, pull their foot off the gas. And I think it's still full steam ahead in, you know, building out the infrastructure.

I do think part of it is just that we're so early in this still. And the hyperscalers are, you know, they're seeing their demand pipeline, and they're saying, "Oh, it doesn't make sense for us to stop building."

But I think, even in the long run, right, like I think it sort of makes sense. So there's two reasons, I guess I'll give, for why I think they're continuing to spend, with, you know, the likelihood that '27 CapEx is going to be higher for the hyperscalers than '26.

One is that the US model companies, right, even though they're seeing this competition intensify, Anthropic and OpenAI seem to have, seem to be giving no indication that they're going to stop pushing the frontier, which means you need more training capacity, right? And so that's one part of the CapEx puzzle.

And then the other part is, you know, once everyone starts using this, if the cost of inference goes down and gets lower, as we've been talking about, then you're going to have an explosion of usage of AI, which also requires compute to deliver it.

And, you know, I think the hyperscalers, in general, will be big beneficiaries of delivering AI inference into the enterprise. They have a lot of enterprise data sitting in their clouds already. And when you're running AI workloads, you want your enterprise data, and you want your AI workloads to be as close as possible. It reduces latency, it helps improve the output quite nicely, and it helps manage costs if they're both in the cloud.

So, I think for those two reasons, you'll see hyperscalers continue to press spending.

The other part that's interesting, though, is, you know, I think almost having more of these open-weight model companies, and having more model companies in general, I think is good for the hyperscalers, because that means there's more models that need to get trained, right?

Obviously, the Chinese models aren't going to use AWS infrastructure to train, but if there's US open-model companies, or there's more US model companies, whether they're closed or open, they're going to require training, which will require hyperscale compute, and then inference, whether it's US or Chinese models. You know, the entity that decides where inference is being run is the customer.

So, even if it's a US company using a Chinese model, that's most likely going to get run on AWS or Google Cloud or Azure. So, the hyperscalers benefit from that also.

And the hyperscalers benefit from their customer base becoming less concentrated. So it's not just dependent on OpenAI and Anthropic. You know, you have OpenAI, Anthropic, plus, you know, maybe thirty other model companies that are your customers.

So, I think there's, you know, good visibility for the hyperscalers to, in a good case, to be made for them to keep building out capacity.

19:51, Richard

So you would be sort of more on the Zuckerberg side of this, i.e., "You know, let's not close the doors and not allow these companies to compete in the US. Let's see how we can fix our own bottlenecks next, which are preventing us from competing more aggressively on that side."

I did see an article in the Wall Street Journal over the weekend that, you know, some American open-weight companies were trying to raise money, and they were having a lot of difficulty with sort of VC funds that are closing the door in their face, because they have so much vested interest in these sort of existing frontier models working, and, I guess, not cannibalizing their own investments there.

So, you know, it's good to hear that you think sort of a more competitive landscape is probably the right way to go. Correct?

20:45, Arjun

Yeah, I, yeah, I do, I do think so, and I don't think, you know, banning the Chinese competition is, you know, a good path for us to go down.

I think it probably just ends up hurting the competitiveness of US businesses, right? Because if you look elsewhere across the world, they're all going to have access to these lower-cost AI.

And then it's US businesses that get hurt, because they're having to pay for more expensive AI. And I think even for Anthropic and OpenAI, right, like they should want to compete. They should, you know, they should want to figure out, "Okay, how can we make our training runs more efficient, so that we can lower the cost of inference, you know, to compete with these newcomers in China?"

It's a tricky question, right? I don't want to say it's that simple, because we don't know how China is doing this. I think that's still a big question that's to be answered, right? There's a lot out there about, are they distilling, you know, the US models, meaning, you know, they just, like, blast Anthropic with a million questions, and use that data then to train their own models.

And so it's a shortcut, and, you know, IP theft, in a way, depending on who you listen to. And so that's one thing, right?

But they also have a very different sort of setup, right? They have a different energy setup, land laws can be, you know, a little bit more prioritized for data centers if they need to, data copyright laws, right?

Like, these are all things US model companies have to deal with that Chinese companies don't. And so, yeah, maybe, is the answer here, do we just, you know, try to attack the bottlenecks, as you said? But you also have to sort of balance it with, you know, we have sort of this growing nimbyism with data centers. We have a lot of, I think, anxiety about where AI is going.

And so, I mean, those are all, I think, things that regulators, policymakers will have to balance with the public to, you know, if they want to speed up the pace of AI deployment.

And, you know, I would also say, like, you know, I do think that VC question is quite interesting, of, you know, they don't want to sort of cannibalize their investment in Anthropic and OpenAI.

But I think, like, you know, we're moving in one direction, which is, you know, enterprises want more model choice. And it's almost like, you know, it's going to happen one way or another. So, you might as well, you know, start to look at some alternatives in the US also. And, again, like I said, I don't think it's bad. I don't think it's bad for OpenAI or Anthropic even. I think we're in a growing market, and I think there can be multiple winners.

23:37, Richard

I agree with that.

Maybe the time we have left, maybe a sort of a two-part question. So, you know, if we're having this conversation three years from now, which seems a very long way away, given how fast things are moving, what do you think is going to be the single most important metric that investors should be watching to see who's going to win this race?

And, you know, a somewhat related question, what's the one thing about AI that investors today are most likely getting wrong?

24:13, Arjun

Yeah. Good questions.

So, on the first part, in terms of metrics, you know, I think right now we are in this sort of build-out phase of AI infrastructure. So the metrics that are most closely watched are the input costs, sort of like CapEx, and, you know, how many gigawatts of data center capacity do you have, or how many data centers do you have.

But I think that will change, right? I think that will become less important over time. It won't go away, right, that's still relevant.

But I think, like, inference token growth might be the best, because that is an indicator of, "Is your AI actually getting used in real life?"

And if inference token volumes are growing, that means, you know, your end users are getting return from it, and they're increasing deployments, and you're seeing, you know, more people sort of jump into the pool and use AI more and more, which means, you know, you're growing and you're having an impact.

So, I think that's a metric that will be important. And, you know, it's sort of analogous to even what happened with the prior tech wave of cloud, right? It was sort of defined by the apps, right, which are, you know, what's sort of in the hands of the end users, and that's where the economic impact is.

So that's your first question. And then, on the second part, what are investors getting wrong about AI today? I think, just like other tech cycles, you know, I just think it's not going to be linear. You know, we already hear about cybersecurity incidents with the models.

You know, I think there could be sort of just hiccups along the process. There will be bumps. There will be setbacks. In some cases, it might even be that, you know, users or enterprises aren't ready to adopt the technology. So it's not that the tech needs to catch up, it's that, you know, the end users need to catch up to how quickly the tech is moving.

And that might be sort of a bottleneck in the adoption process, in some cases, right? The ecosystem needs to get developed.

So, I think those are all sort of, you know, drivers of just a non-linear journey. But I think if you zoom out, I think it will be moving up into the right, adoption will be increasing. But, you know, I think there certainly will be…

26:43, Richard

Some bumps along the way?

26:43, Arjun

Yeah. Yeah.

26:46, Richard

Well, I think that's probably all the time we have, Arjun. So, I think we'll leave it there. But thank you very much, once again extremely interesting and helpful.

26:55, Arjun

Awesome. Thank you.

26:57, Chris

Big thanks again, Richard, Arjun, for walking us through what is clearly becoming one of the more important debates in AI. Thanks, everybody, for listening. We'll be back next month with another episode of Monthly Macro.