Good news and bad news arrived for Meta on the same day. The good news: its self-developed AI chip, Iris, passed key testing in just six weeks with no major issues. Production starts in September. The bad news: the stock dropped nearly 3% in pre-market trading.
That tells you everything about the AI chip business in 2026. Building the silicon is no longer the hard part. Paying for everything around it is. Iris is a technical win. But it's also a $145 billion reminder of what Meta has committed to this race.
Five Years of Failure, Finally Reversed
Meta has been working on custom AI silicon for over five years. The MTIA program was widely seen as the slowest among the big tech players. The company tried acquisitions. It reportedly offered $800 million for a Korean chip startup. The deal fell through. For years, nothing seemed to click.
Now, finally, something did. Iris — the latest iteration of MTIA — completed validation in six weeks. No major defects found. That's unusually fast for a data-center AI chip. The memo reportedly described deploying new Nvidia GPUs at Meta's scale as "extremely difficult and time-consuming." That frustration is exactly what drove Meta to build its own.
The chip is built for Meta's own workloads. Facebook ranking. Instagram recommendations. Generative AI inference. It's designed with Broadcom. Manufactured by TSMC. It won't replace Nvidia or AMD. The memo is explicit: Iris is a complement, not a substitute. But it's a complement that gives Meta leverage.
The roadmap is aggressive. Meta plans a new chip every six months through 2027 — twice the industry norm. Four chips are in the pipeline: MTIA 300, 400, 450 and 500. The later ones are designed specifically for generative AI inference. The company is no longer waiting for Nvidia's roadmap. It's writing its own.
The $145 Billion Question Wall Street Can't Ignore
Now the bill. Meta plans to deploy 7 gigawatts of compute capacity this year. By 2027, that doubles to 14 gigawatts. Capital expenditure for AI infrastructure this year: up to $145 billion. That's roughly the company's entire net profit for fiscal 2025. Investors are starting to question whether the returns will ever catch up.
The stock drop isn't irrational. It's math. Wall Street sees a company spending more on AI infrastructure than it makes in a year. Meanwhile its core AI product still lags behind OpenAI, Anthropic and Google. The monetization path beyond better ad targeting is unclear. Even Meta might have more compute than it knows what to do with — there's talk of building a cloud business to sell excess capacity. That's not a sign of strength. That's a sign of anxiety.
Market fears have been mounting: token demand could peak before supply needs to, or the full-scale shift to AI agents might still take years, putting companies in an awkward position where they've built out massive infrastructure ahead of actual demand. For Meta, the equation is simple: if AI revenue doesn't grow fast enough, that 14 gigawatts of capacity becomes a liability, not an asset.
Iris is a win. But it's a win inside a much bigger loss. The company spent five years and billions getting to this moment. Now it needs billions more to deploy the chips, build the data centers, power them with dedicated gas plants. And hope the AI revenue eventually catches up. The chip passed testing in six weeks. The real test starts in September.
P.S. If you're a Meta shareholder, you're not worried about Iris. You're worried about whether 14 gigawatts will ever generate 14 gigawatts of profit. That's the question the memo didn't answer.
