Meta has decided it is done being at the mercy of Nvidia’s pricing, and it is building its own way out. According to an internal memo reviewed by Reuters, the company will begin manufacturing its first in-house data-center AI chip, codenamed Iris, in September. It is the flagship of Meta’s MTIA programme (Meta Training and Inference Accelerators), the first of four planned chip generations, and it cleared its bug-testing phase in about six weeks without any nasty surprises.
Broadcom is the design partner, TSMC is handling fabrication, and the strategic logic is the same one driving Google, Amazon, OpenAI and Anthropic down the identical road: if you are spending an obscene amount of money renting someone else’s silicon, building your own starts to look cheap. And Meta’s spend is genuinely obscene. The company expects to lay out as much as $145 billion on AI infrastructure this year alone, a slice of Big Tech’s collective $700 billion-plus bonfire.
The number that should make you sit up
Here is the figure worth staring at: the memo lays out a plan to go from seven gigawatts of computing capacity coming online in 2026 to fourteen gigawatts by 2027. Fourteen gigawatts. For one company’s AI. That is roughly the output of fourteen large nuclear reactors, dedicated to serving you a marginally better Instagram feed and a chatbot that still cannot reliably count the number of R’s in “strawberry.”
A quick word about sand
Now for the part that never makes it into these triumphant chip memos. Those chips are made of silicon, and silicon comes from sand. Specifically high-purity quartz sand, the good stuff, which is nothing like the sand on your local beach and is already in worryingly short supply. Sand is, astonishingly, the second most-extracted resource on the planet after water, and we are tearing through it faster than nature can make it, ripping up riverbeds and dredging coastlines to feed the construction and electronics industries.
So let me get this straight. The richest companies in human history, having already decided they need fourteen gigawatts of power and $145 billion of kit, now also want to help drain the rivers dry and vacuum up the f*cking beaches to get the raw sand for their chips and the concrete for the data centres to house them? Get f*cked, honestly. It is the most on-brand thing imaginable: an industry that promises to solve every problem while steadily consuming the physical world in order to do it.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
The honest read
None of this makes Iris a bad move for Meta. Building your own chips to escape Nvidia’s margins is, coldly, good business, and it is exactly what a rational trillion-dollar company should do. The point is the externalities. Every one of these “we’re making our own silicon” announcements is another few gigawatts of demand, another vast draw on power, water and raw material, presented as a pure win because the only ledger anyone reports on is the company’s own. The sand, the rivers, the strained local grid and the towns next to the data centres are all off the balance sheet. For now.