Throughout 2026, Meta was the forgotten child of artificial intelligence. Everyone was talking about Nvidia, Alphabet, and OpenAI. Its stock kept moving up and down without really going anywhere.
And suddenly, within a few weeks, two things changed. It closed its biggest legal battle. And it announced that it would put its own artificial intelligence chips into its data centers.
THE TWELVE CHIPS THAT CAME FROM TAIWAN
Meta is currently testing the third generation of its own processor. It's called MTIA 450, code-named Arke, and it will enter data centers during the first half of next year.
And it doesn't stop there. The next generation, MTIA 500, also known as Astrid, will finish its design in about a month and will go into machines by the end of 2027.
Now, here's the most interesting part of the story. On September 1, the first twelve chips from TSMC arrived at Meta. And their performance came within 2% to 3% of what the simulations had predicted.
In other words, there were no unpleasant surprises in the design.
In fact, from the very first day, the team managed to run not only Meta's models on them, but also models from DeepSeek and Alibaba.
"And who actually makes all of this?" you might be wondering. Broadcom is the design partner. TSMC handles manufacturing. And the person running the project is Yee Jiun Song, the company's vice president of engineering.
WHY IS IT DOING THIS?
This is where things get even more interesting. Because the reason Meta is doing this isn't prestige, but cost.
"Each generation takes on slightly more technological risk and gives us better performance," says Song. Better performance per watt of energy. And per dollar spent.
And to understand the scale, the company has committed to more than one full gigawatt of chips within twelve months. After that, it expects to accelerate, assuming, of course, that the artificial intelligence market doesn't collapse.
There's also one decision that clearly shows the logic behind this strategy. Meta had been designing a chip called Olympus, which would handle both model training and inference. And it canceled it.
"Why?" you might ask. Because it would have cost around 30% more. And as Song himself puts it, when you're building gigawatts upon gigawatts of capacity, a cost like that suddenly becomes "completely unacceptable."
So Meta has chosen a side. It's going exclusively after inference. What does that mean? In very simple terms, it's the moment when the trained model gives you an answer. The everyday work, not the studying.
"These are the workhorse chips," says Song. And Meta Superintelligence Labs is helping with this as well, telling the team in advance what the company's next models are likely to demand.
AND THE STOCK?
Here we have a real rollercoaster. The stock has risen 22% from its August low and is heading toward its best month since May 2025.
Of course, there's a big "but." During 2026, Meta has had four rallies of around 20%. And all of them faded, with each subsequent peak coming in lower than the previous one. That's why the stock is essentially flat since the beginning of the year, while the Nasdaq 100 is up 15%.
So what has changed this time?
First, the $18 billion settlement related to lawsuits surrounding social media. And pay attention to this. Based on its own calculations, losing the case could have cost the company as much as $1.4 trillion. Yes, you read that correctly. Since the agreement was announced on August 26, the stock has fallen in only 4 of the 12 trading sessions.
Second, Muse, the new AI assistant for everyday tasks. JPMorgan upgraded the stock to overweight last week specifically because of this.
And now, its own AI chip.
Morgan Stanley makes another comparison worth noting. It says the situation resembles Alphabet in late 2025, when it received a favorable decision in its antitrust case, then released a series of AI products, and has since gained more than 50%.
And the valuation? Around 18 times estimated earnings, below the ten-year average of 20 and more than 30% below its 2025 peak.
Of course, it's not all positive. Infrastructure investments are doubling this year to $140 billion and are heading toward $200 billion in 2027. And free cash flow is turning negative. Minus $6.3 billion this year, with a forecast of minus $30 billion in 2027.
BAD NEWS FOR NVIDIA?
You might say now... isn't this bad news for Nvidia?
No. And here's why.
Meta's chips are designed for general-purpose inference. They aren't targeting the very high-speed inference market, where models need to respond almost instantly. And model training is completely excluded, since Olympus was canceled.
And pay attention to how Song himself puts it. The goal is to build chips that are "competitive with the ones our suppliers build for us."
In other words, Meta isn't cutting out Nvidia. It's taking control of a piece of the workload that it runs every day and understands better than anyone else. And at the same time, let's not forget, its total capex is increasing, not decreasing.
