Open Source AI Is A Price War

All this fuss about Open source AI is a price war.

Sorry open-source lads and lassies. I know there are philosophical bits. Safety bits. Licence bits. Community bits. All true, all amazing stuff.

But the reason this suddenly matters to normal people is money.

Anthropic and OpenAI need the world to believe that the best AI has to come through expensive, “safe”, American API pipes. Controlled by, well, them.

Meanwhile, Chinese labs are shipping opensource models that are more than good enough for a lot of real work and a hell of a lot cheaper.

The very existence of these opensource alternative undercuts the whole US/closed source mission. And threatens to bring the entire economy of the West crashing down.

So…this is not just about software development principals!

WTF is opensource?

What actually is opensource? The annoying thing is that everyone uses "open source AI" to mean different things.

Sometimes they mean free. Sometimes downloadable. Sometimes local. Sometimes the training data is visible. Sometimes the licence lets you use it commercially. Sometimes they mean "I saw DeepSeek on Twitter and now I am pretending to understand geopolitics."

This is further confused by the fact that most opensource LLMs aren’t actually opensource. They are open-weights. Grr…

The useful split is this:

• open source means you can see the code, the weights, the training pipeline and sometimes even the underlying data.

• open weights means you can download the trained model files

Most of the models people call open source are really open weights. You can get the trained numbers. You can run them somewhere. You might be able to fine-tune them. But you usually cannot rebuild the full thing from scratch because you do not have the training data, process, and boring details behind the model.

Here’s a useful analogy from Dentro.

Closed source models like ChatGPT and Claude are like a taxi. You can tell the driver where to go, but you do not own the vehicle. You definitely can’t take it home and start modifying it - the cabbie would get very annoyed I imagine. And, like black cabs, they are expensive!

Open weights are more like owning a car. You have the thing. You can drive it where you want. You can modify bits of it. You are responsible for the faff that comes with ownership. Most of the time this works out cheaper than taking a taxi but on the whole it’s a little more of a faff.

True open source is the kit car version. You get the parts, the plans and enough information to rebuild the thing yourself. Very few models are truly open-source, indeed none of the well-known production grade ones.

Also whether a model is free to use and whether it’s local are entirely separate issues from the model being open weights.

GLM 5.2 for example is technically open-weights. You can go right now and download the whole thing, all 1.5TB of it. And then run it on your local computer for “free” (minus the cost of electricity of course).

BUT…your local computer is going to have to be a data centre… you aren’t running this on any commercially available computer.

Most local models are opensource. But not all opensource models can be run locally! So don’t conflate the two.

Where does all this live?

Let’s actually have a look at some open-source (open-weights, see even I’m using the term wrong..) models. How do you find them? Where can you download them? What can you do with them?

The first site to look at is Artificial Analysis. Their open-source model page compares open-weight models by intelligence, openness, size and more.

These are pretty much ALL Chinese models and the first non Chinese model is Nemotron (Nvidia).

Let’s have a look at the intelligence against cost-per-task. Basically your bang for your buck:

The smartest models are (duh) Anthropic’s Fable, Opus, Sonnet and GPT5.5. They are also the most expensive.

GLM5.2(max) is sitting at a very similar intelligence level (same as Sonnet 5 which came out this week) but at a fraction of the cost.

The x-axes here is log scale by the way so the costs are getting much higher in that riht hand side - look how it jumps from $1 to $2 to $3 in smaller steps.

Opensource models are coming in either i) slightly worse but much cheaper or sometimes ii) equivalent quality but much cheaper. That is where the price war becomes obvious.

You do not always need the smartest model. You need the cheapest model that clears the bar for the job. Customer service reply? Internal summary? Classification? Simple data clean-up? A frontier model is often total overkill.

We have a pigeon that keeps flying into the garden and making a mess right now. The sensible method to get rid of it is to wave and clap until it buggers off. The Claude Fable method would be to hit it with an Intercontinental Ballistic Missile.

Often we don’t NEED so much power. And when that power comes at 50-100x the cost it’s silly to use it.

The second site to know about is Hugging Face. This is where the models live. Here for example is GLM-5.2:

The tags at the top tell us what it can do (text gen), its languages, its license etc.

On the right we get the model size - here it is 753 billion parameters (big!).

Hugging Face has almost 3 million other models for you to browse, learn about and download.

But remember that most won’t run on your local computer. Opensource is NOT the same as local! This GLM-5.2 model is the opposite of a casual laptop download - it’ll need a small data centre to run it!

Still incredibly important though. You can inspect it. You can see the files. You can see how this stuff is actually distributed. It’s all right there for you to explore.

China is playing a blinder

This is where the geopolitics come in.

The US labs have a very expensive story to sell. More chips. More data centres. More money. More closed models. More funding rounds. More valuation nonsense.

Scaling has worked so far. Just throw MORE at the problem and the AIs get better. It’s actually pretty amazing. And noone is quite sure why it works…

OpenAI and Anthropic are trying to justify obscene numbers and the market is nodding along because AI is the big growth story. About 2% of US GDP is being invested into AI this year and somewhere north of 50-60% of all VC money is going into AI startups.

It’s a massively intense concentration of investment. In one basket. A basket that is currently not laying eggs? I’ve lost the metaphor…you know what I mean.

A lot is now riding on 2 companies. Anthropic and OpenAI. And their upcoming IPOs (going public on the stock market). Both of their IPOs are based on valuations just shy of $1T with revenues being a tiny multiple of that - much lower than the revenue multiples we expect.

It’s one hell of a big bet. And making such a big bet requires investor confidence. If they get wind of this not working out, of their money not coming back to them, of potentially losing everything then they’ll panic and withdraw.

The IPOs of Anthropic and OpenAI are big litmus tests here. Do the markets have the balls to support the dream? Big, scary risk.

…Enter China.

China is bringing opensource models to market. Not with one model. With a flood of them. DeepSeek, Qwen, Kimi, GLM, MiniMax, Hunyuan and the lot. Some are brilliant. Some are a bit crap, much like US models! But the direction is clear: cheaper, good-enough intelligence keeps arriving.

And if the best closed model is only needed for the hardest 5-10% of work, then a lot of the Western AI economy starts looking very different.

Why would I pay Anthropic or OpenAI a sizeable chunk of my revenue every single month for tokens when I can deploy a Chinese model locally within my organisation, fine tune it and run it basically for the cost of electricity thereafter? Shit, I even get better security because my information is not being sent up and down to Anthropic / OpenAI. It sits locally in a secure server in my office.

The very existence of these cheap, efficient, Chinese local models is a threat to the narrative the closed-source American labs are pushing. It undercuts their whole business model.

Even if very few companies actually deploy local Chinese based models the fact that they even exist as an alternative sheds doubt on OpenAI and Anthropic.

This is why the safety argument from closed labs always feels a bit fishy to me. To be clear: the safety argument is real. You cannot release powerful weights into the world and then recall them. Once a model is out, it is out - Pandora’s box is open. Governments should care about that. Cyber, fraud, bio, state actors…lots of scary stuff.

BUT...

The business incentive is also real.

Closed labs fear open models because they are dangerous. Sure. They also fear them because they are becoming good enough. Good enough kills margins. Good enough gives enterprise buyers wiggle room. Good enough means developers can switch vendors. Good enough means "pay us whatever we ask" stops working.

That is a wedge. If a huge chunk of US market confidence is tied up in AI capex, and Chinese open-weight models keep making intelligence cheaper, this starts to look much bigger than a nerd fight.

If the Chinese opensource models can destablise the Anthropic and OpenAI IPOs and cause sufficient investor doubt then the whole AI industry could come tumbling down, bringing with it the American and subsequently Western economy.

This is not a fight over closed-source vs. open-sourced software development.

This is a fight for who owns the top spot in the global economy: the US or China.

Pick a stack, not a religion

So what do you actually DO?

Do not become an open-source purist. Also do not become a closed-model fanboi. Both are annoying. And you know it.

Instead play it smart.

Use frontier models for genuinely hard work. Strategy. Proper reasoning. Important writing. Research. Stuff where being 5% better actually matters. Use the VERY best you can afford here and keep updating as new models and tools release.

Use hosted models for cheap volume. Either ChatGPT/Claude subscriptions OR open-source subscriptions for more usage. Or both! This is where routine automations, internal workflows live. Anywhere "pretty good and much cheaper" beats "best in class and wildly expensive". This is the workhorse.

Then use local models for privacy, fallback, learning and simple repetitive work. LM Studio is the friendly route. Ollama is more developer-ish but useful. Gemma is worth watching. Chinese models are worth testing if your data/privacy situation allows it. These models will live on your device and deal with the day to day and anything they can’t deal with gets bounced up the chain.

This setup gives you and your business the most flexibility to deal with whatever is coming down the tracks.

To the Task,

Kyle