The battle for AI dominance: open-weight versus closed-weight models

You can also listen to this podcast on iono.fm here.

SIMON BROWN: I’m chatting now with Dr Michael Power, an independent global strategist and economist. Michael, appreciate the early morning time. The battle for AI dominance is very much shaping up to be one between the sort of closed-weight and open-weight models – broadly US versus China.

DR MICHAEL POWER: You’re absolutely right. The battle royal between open-weight and closed-weight is developing. And while it’s not precise, it’s mostly US versus China. There are open-weight US models, including one from Nvidia called Nemotron, but generally speaking it’s a China-versus-US battle.

It’s playing out in a way which is not very welcome to the US guys like OpenAI and Anthropic.

Read:
Navigating the US-China AI technology split
Nvidia warns Trump curbs on China AI chips to cost $5.5bn

SIMON BROWN: And part of that is just good old-fashioned price, or perhaps power or compute, the open models. And I’m thinking, I use some. I use DeepSeek from time to time. It is massively cheaper than what I would pay Anthropic or OpenAI, and it’s frankly not a bad model at all.

DR MICHAEL POWER: The old story of delivering 95% of the quality for 5% of the price is a very hard one to beat. And that is what the Chinese models are essentially banking on – and you’re absolutely right.

The Americans as a result are now in an ill-disguised price war, particularly between OpenAI and Anthropic, as they try and first beat each other and then keep the Chinese at bay.

Read:
China’s cheap AI is shifting investors’ focus to internet stocks
Anthropic’s Amodei rejects open model ban but calls for testing
China state media says support for Open AI models has limits
Amazon will offer OpenAI models to customers for first time
Apple to open AI models to developers
China floods the world with AI models after DeepSeek success

SIMON BROWN: The Chinese then comes. I hear what you’re trying to do, absolutely – but I’m thinking around a business, particularly if you want to run on premises, you’re worried about sending your proprietary data off.

We’re going to see a lot more of that move towards open-weight because I can download it, I can run it on my own service. At some point, I will be able to run it on what will admittedly be a chunkier laptop, but it’ll run on a laptop.

DR MICHAEL POWER: Indeed, and one small detail, but rather an important one – you can run it for free. You don’t need to pay for your tokens. But you’re absolutely right, and the game in town now is to reduce the size of these models.

There’s a technical term called quantise, which means basically to shrink the size of these models so they can fit on your laptop. It doesn’t need to be that chunky; and very soon indeed – there are a couple of rare examples – you can fit it on your phone.

Read: Alibaba releases AI model with performance rivalling Anthropic

SIMON BROWN: Are we going to see – not Anthropic, not OpenAI – they’re not going to move towards open-weight. Meta came out initially with an open-weight, Llama. You mentioned [what] Nvidia’s got. Are we perhaps going to see some of the closed-weight models sort of tinker with and try their hand at open-weight if we continue to see this trend?

Open-weight – to me, this is the Linux of compute. You know, when Linux came out 30-odd years ago [it was] deeply derided. It runs most of the servers these days.

DR MICHAEL POWER: Indeed it does. It was deeply derided by Microsoft, and today it runs all their servers. So it’s absolutely the same trend. And, of course, it’s open-weight, and that really is what makes it; it just basically becomes a language, and language is not something you have to pay to use.

The thing about it is that the Americans are essentially blinking. You’re right, Meta has gone back to having an open-weight model. Google already has one called Gemma. They haven’t emphasised it enough, but it’s an extremely good model and not to be underrated.

Read: Biden to further limit Nvidia AI chip exports in final push

Elon Musk and his Grok – or whatever he wants to call it – is now flirting with the idea of going open-weight. He was always, deep down, an open-weight guy.

And so that just leaves, as you said before, OpenAI. Again, OpenAI was founded as an open-weight model, hence its name. OpenAI and Anthropic [are] as the last two holdouts.

SIMON BROWN: And then a last question for the investors out there listening. This is maybe less about Nvidia chips, less about the power and the data centres. Maybe it’s about Tencent or one of these Chinese apps, to your point, who are utilising the open-weight models [and] are going to benefit from it and [be] able to price it.

DR MICHAEL POWER: They will, too. Yes, but I also think that there are opportunities from some of the American models to realise that actually the action is in the adjacent space now, not in the model itself.

And those who are smart enough to move first and quickly into those adjacent spaces can also make money. So there will be opportunities to make money across the board, but they’re not going to be the mega-zillions that many people thought was going to be the case.

Read:
China’s cheap AI is shifting investors’ focus to internet stocks
Nvidia mega-deal ushers MediaTek into top AI chipmaker club
Tencent-backed chipmaker Enflame to debut after $911m IPO
Tencent’s revenue miss heightens pressure for faster AI payoff

SIMON BROWN: We’ll leave it there. Dr Michael Power, independent global strategist and economist, appreciate the early morning time.

#battle #dominance #openweight #closedweight #models