Meta starts building its own AI chip in September. Follow the parts and the independence gets small.

Sometime in September, if an internal memo reported by Reuters this month is accurate, a line somewhere in TSMC's network will begin turning out wafers for a processor that Meta designed for itself. The chip is codenamed Iris. It is not the first accelerator the company has built — that programme is years old and has a public name of its own — but it arrives with more weight than its predecessors, because it arrives at the moment when every large buyer of compute is trying to answer the same question: how do you stop paying Nvidia's margin on the most expensive thing you buy? The headlines wrote the answer for Meta before the first wafer starts. Meta is escaping Nvidia. It is a satisfying sentence. It survives about ten minutes of following the parts.

Start with what is actually established, because on this story the gap between what a company has confirmed and what has been reported about it is unusually wide, and it matters. The September production date, the testing timeline, the supplier list — all of it traces to a single internal memo reviewed by Reuters and published on 9 July. Meta declined to comment on it. One memo, one outlet. Worth saying plainly, because a week later the same claim came back around as a fresh headline announcing that Iris had passed its testing, which was not new reporting but the July memo restated. It is worth adding that TechCrunch, working from the same document, rendered the testing claim more cautiously still — that at least one chip cleared its testing phase in about six weeks, without naming Iris at all. So even the subject of the six-week figure is softer than the coverage implies. None of which makes the memo wrong. It makes it one memo, and the correct posture toward a manufacturing date that has not happened yet is patience.

What Meta has actually said out loud

The company's own record is narrower and, in its way, more revealing. In March, Meta published the shape of its silicon roadmap: four generations of its Meta Training and Inference Accelerator line, with MTIA 300 already in production and aimed at training the ranking and recommendation models that decide what appears in a feed, and the 400, 450 and 500 following into 2027, optimised first for inference — for running models rather than making them. The cadence is the number that should catch your eye. Meta says it intends a new generation every six months or less, against an industry norm of one to two years. That is not a claim about transistors. It is a claim about organisational metabolism, and it is much harder to fake than a benchmark.

Note what is not in any of it: the word Iris. Meta's public posts do not use the codename, and no source I can verify maps Iris onto a specific MTIA number. Plenty of coverage has quietly assigned it one. I am not going to, because on the factory floor the difference between a part you can name and a part you have inferred is the difference between reporting and decoration.

The second thing on the record is the piece that actually explains the chip, and it was announced in April with both companies' names on it. Meta and Broadcom are co-developing multiple generations of MTIA silicon, built on Broadcom's XPU platform, in an arrangement that covers not only chip design but advanced packaging and networking — the two things that turn a die into a machine. The first phase alone is described as exceeding a gigawatt of deployed capacity, the opening instalment of a multi-gigawatt rollout. There is a governance detail buried in that announcement that tells you more about its scale than the engineering does: Hock Tan, Broadcom's chief executive, moved off Meta's board into an advisory role because the partnership had grown too large to sit comfortably alongside a directorship. Companies do not rearrange their boards over a component order.

The design is Meta's. Almost nothing else is.

This is where the escape narrative starts coming apart, and it comes apart in the ordinary way — not through any single dramatic dependency, but through the accumulation of unglamorous ones. The chip is designed in-house, on a platform licensed from Broadcom. It is fabricated by TSMC, on the same leading-edge processes and through the same advanced packaging step that every other AI accelerator on earth is queued behind. According to the memo, its memory comes under extended contract from Samsung and its flash from Sandisk. And its optical interconnect — the fibre that carries data between racks, without which a gigawatt of accelerators is a gigawatt of expensive space heaters — comes from Sumitomo Electric, a company almost nobody outside the supply chain could name and which is, at this moment, closer to the critical path of frontier AI than most of the labs you read about.

Follow that one link further, as you always should. Meta is not the only company doing this. Google's TPU line is a Broadcom co-design and has been for years. OpenAI unveiled a custom inference part with Broadcom last month. The pattern is not that hyperscalers are building their own chips; it is that hyperscalers are building their own chips through the same design house, on the same platform, at the same fab, using the same packaging step, waiting on the same two memory makers. Broadcom's most recent quarter reported artificial-intelligence semiconductor revenue of $10.8 billion, up roughly 143 percent year on year, inside total revenue of about $22.2 billion; the company has guided the AI line to around $16 billion in the current quarter. On its earnings call in June, management described AI bookings of more than $30 billion against that $10.8 billion shipped — call commentary rather than a filed figure, so hold it loosely, but the direction is not ambiguous. The order book is running well ahead of what the supply chain has been able to deliver.

Everyone's escape from Nvidia is being designed by the same company, fabricated at the same fab, and fed by the same two memory makers.

The target is quoted in gigawatts, which tells you what is actually scarce

The most instructive figure in the Reuters memo is not the September date. It is the pair of capacity numbers: roughly seven gigawatts of compute online across 2026, growing toward fourteen by the end of 2027. Sit with the unit. Meta is not describing its ambition in chips, or in FLOPS, or in racks. It is describing it in electrical power, because power is what it has learned to worry about — the same lesson every operator of this scale has now absorbed, and the reason state regulators have started writing data-centre rules that read like utility law.

Hold the two disclosed numbers against each other and the strategic reality resolves. The Broadcom partnership's first phase is described as exceeding one gigawatt. The 2027 target, as reported, is fourteen. Even on generous assumptions about how fast the later MTIA generations ramp, Meta's own silicon will be carrying a minority of Meta's own compute for years. The rest gets bought, and there is only a short list of people to buy it from. Iris is not a replacement for that purchase order. It is a lever against it — real, financially significant, and a great deal smaller than the framing suggests.

The money says the same thing in plainer language. In April, Meta raised its capital-expenditure guidance for this year to a range of $125 to $145 billion, up from $115 to $135 billion, against $72.2 billion actually spent in 2025. The company's stated reason for the increase is the sentence worth reading twice: higher component prices, and additional data-centre costs to support future-year capacity. A company designing its own accelerator raised its spending guidance by ten billion dollars partly because the parts got more expensive. That is a supply-chain disclosure hiding in an earnings release, and it is the most honest thing anyone has published about the economics of this build-out all year.

The constraint nobody has moved

I want to be careful here, because this is the part of the story where confident claims outrun their sourcing badly. A great deal of what circulates about advanced packaging — that CoWoS is sold out through the year, that packaging rather than lithography is now the binding constraint on AI accelerators — traces back to sites that cite each other and no one else. I could not stand any of it up. What I can report is a market-research figure from TrendForce in June, itself relayed from institutional-investor estimates: TSMC's advanced-packaging supply-demand gap is expected to narrow from around twenty percent to around ten percent by the end of this year, with monthly CoWoS capacity reaching roughly 120,000 to 140,000 wafers at TSMC and another 50,000 to 60,000 across outsourced assembly partners. Second-hand, and worth treating as an estimate rather than a measurement.

But read it for what it does say. The gap is narrowing. It is not closed. After two years of the most aggressive capacity expansion in the history of the packaging business, demand is still running ahead of the ability to assemble finished parts — and every custom accelerator programme announced in that period, Meta's included, adds to the demand side of exactly that equation. This is the quiet arithmetic that the independence story never does. A dozen companies deciding to build their own silicon does not reduce the load on the chokepoint. It increases the number of parties queuing at it, all of them now holding designs they are contractually and reputationally committed to shipping.

What Iris actually buys

Strip the framing away and what remains is still substantial, just different in kind. Iris buys margin — every accelerator Meta builds itself is bought closer to the cost of manufacture rather than at a system vendor's price. It buys roadmap control, the ability to shape silicon around the specific workload of ranking a feed rather than around whatever the merchant market considers general-purpose. It buys negotiating leverage, which is often worth more than the chips: a buyer with a credible in-house alternative is a different buyer across the table. And if the six-month cadence holds, it buys something rarer, which is the capacity to iterate hardware on roughly the same clock as the models it runs. None of that is independence. All of it is worth doing.

Two other moves this month suggest the ambition runs past Meta's own workloads. Bloomberg reported on 1 July that the company is standing up a business — organised internally as Meta Compute — to sell surplus AI capacity to outside customers, the way a cloud provider does. And in the middle of the month the Wall Street Journal reported that Dave Brown, the Amazon Web Services executive who ran compute and machine-learning services after nearly nineteen years at the company, is leaving for Meta's infrastructure organisation. Read those together and the picture sharpens. Hiring the person who ran the world's largest compute rental business is not a data-centre operations hire. It is a product hire. You do not need that person to run your own servers. You need them to sell someone else's time on them.

So here is the shape of it, as best the record supports. In September, or near it, a fab will start producing a chip that Meta drew, on a platform Broadcom licensed, from a process TSMC owns, wired with optics from a Japanese firm most readers have never heard of, fed by memory from a Korean one, destined for buildings whose real constraint is the local grid. Meta will own the part of that chain with its name on the die and none of the parts that decide whether it ships on time. That is not a criticism of the strategy; it is the strategy, correctly understood. The largest technology companies on earth have concluded that vertical integration in silicon is worth tens of billions of dollars, and every one of them has arrived, by an independent path, at the same three or four upstream firms. Concentration of this kind never announces itself. It accretes, one rational decision at a time, until the whole industry's plan for escaping a single supplier depends on a handful of others — and independence, at this scale, means nothing more than choosing which queue you stand in.