So, take a look at what’s happening with Meta.
We have a company that is set to spend up to $145 billion on AI infrastructure this year. Yet its stock is down about 10% year to date, because investors look at those numbers and ask, “Where is all that money actually going?”
And what does Zuckerberg do?
He comes out and says he’s going to give away his models... for free.
In an Instagram video yesterday, Mark Zuckerberg announced two things.
The first is Muse Glimmer.
It’s a 30 billion parameter open-weight model released under the Apache 2.0 license. That means anyone can download it, run it, modify it, and even use it commercially without paying a cent.
“What exactly are weights?” you might ask.
Put simply, they are the calculations and rules that determine how the model works and behaves. They are essentially the heart of the system.
The second announcement is that Meta will soon open up Muse Spark 1.2, the company’s latest foundation model.
And pay attention to the pace here. Muse Spark 1.1 was released just last month.
But the most interesting thing about Glimmer is where it runs.
You don’t need a data center. It runs locally on a Mac or PC with a single consumer-grade graphics card. With 24GB of VRAM, it can operate without sacrificing reliability.
And it’s not just a chatbot.
Alexandr Wang, Meta’s Chief AI Officer and founder of Meta Superintelligence Labs, explained that it functions as a full AI agent. It can plan, call tools, verify its own outputs, and recover when something goes wrong.
The model weights have already been uploaded to Hugging Face, along with documentation for anyone who wants to build their own agent.
Oh, and one more thing.
Within the next few months, a new model internally codenamed “Watermelon” is expected to arrive, reportedly more powerful than Muse Spark.
Will it be open or closed? Zuckerberg didn’t say.
Meta’s stock rose as much as 2.1% before ultimately closing up 0.26%.
To understand why Meta is doing this, we need to look at who dominates the open-source AI landscape today.
The answer is simple:
The Chinese.
Companies like Alibaba, DeepSeek, and Moonshot AI have been aggressively releasing open-weight models that, in some areas, compete directly with the best American offerings.
Meanwhile, OpenAI and Anthropic have chosen the opposite approach: closed models, subscriptions, and APIs.
So what’s left in the middle?
A massive gap.
As Neil Shah of Counterpoint Research put it on CNBC:
“If Western tech giants build only walled gardens, developers and enterprises will naturally gravitate toward Chinese open-weight models.”
He added that there is enormous demand for non-Chinese open models.
That is exactly where Meta wants to position itself.
Alongside the announcement, Zuckerberg also published a 6,500-word essay, roughly 14 pages long.
In it, he does two things.
First, he calls on Washington to change its approach.
He argues that foreign AI labs currently enjoy advantages because American companies face additional restrictions on training data. He also says that blocking foreign open-source models is not the answer. Instead, the goal should be to make American models the best in the world.
Second, he takes aim at rivals.
Without naming them directly, he criticizes the rhetoric coming from parts of the AI industry.
He writes that much of the discussion around AI is filled with excessive doom and gloom and questions why someone who truly believes AI will eliminate most jobs would be racing to build that future.
He concludes that the idea that AI is so dangerous that it requires extreme concentration of power is itself a troubling concept.
His vision?
Rather than concentrating superintelligence, it should be distributed widely.
In Zuckerberg’s view, everyone should eventually have a highly capable personal AI agent that understands who they are, what they want, and what matters to them.
At this point you might be wondering:
“If Meta is giving away the technology for free, how does it make money?”
That’s where the strategy gets interesting.
First, standards.
When millions of developers build tools and applications on Meta’s architecture, Meta effectively sets the technical standards for the industry.
The ecosystem starts revolving around its technology.
Second, cost.
AI agents require continuous processing to execute multi-step tasks.
If that processing happens locally on your device, GPU costs shift away from Meta’s data centers and onto the user’s hardware.
That means Meta can deploy advanced AI across WhatsApp and Instagram at a much lower operating cost per user, improving margins.
Third, competitors.
OpenAI, Anthropic, and Google depend heavily on subscription-based cloud APIs.
By releasing a powerful 30B model for free, Meta turns the core technology into a commodity.
That puts pressure on competitors’ margins while Meta continues earning its money elsewhere: advertising.
Fourth, capital spending.
Every release like this serves as proof to Wall Street that Meta’s massive AI investments are producing tangible results.
And the technology doesn’t sit on a shelf.
It flows directly into advertising targeting systems, Reels recommendation engines, and products like Ray-Ban Meta Smart Glasses.
Fifth, hardware.
Glimmer is already optimized for AMD, Nvidia, Intel, Apple Silicon, and ARM platforms.
Hardware manufacturers now have an incentive to promote Meta’s models as a way to sell the next generation of “Agentic PCs.”
In other words, Meta gets distribution and marketing for free.
Put all of this together and a larger picture emerges.
The market is beginning to see Meta not merely as a social media company, but as foundational infrastructure for the AI ecosystem.
Historically, companies viewed as infrastructure providers tend to command higher valuation multiples than those viewed as simple application businesses.
