On Thursday July 9, 2026, Reuters reporters Katie Paul, Max A. Cherney and Stephen Nellis published an exclusive — sourced to an internal Meta memo — saying that Meta will begin manufacturing "Iris," its first in-house data-centre AI chip, in September. The chip cleared six weeks of validation with no significant issues. Meta is targeting 7 gigawatts of compute deployed by end of 2026 and 14 gigawatts by 2027. Broadcom is the design partner, TSMC does the fabrication, and by the close of the New York session Nvidia had fallen 1.2 per cent to $201.63 while Broadcom had risen 2.6 per cent to $398.63. Meta itself closed roughly flat at $602.86, then jumped 4.6 per cent in late trading once the memo hit the tape.

The leak does two things at once. It tells Meta employees the six-month chip cadence Iris is being built at is now the plan through 2027 — a rate at which Meta will ship a new AI processor every half-year, while the industry norm is closer to a year and Nvidia refreshes its data-centre line roughly every eighteen months. And it tells the market — via an internal memo that leaked on a Thursday afternoon, which is not an accident — that a company spending as much as $145 billion on AI infrastructure this year intends to spend a growing share of it on itself.

What "no major issues" means

Six-week validation is fast. Meta's memo says testing found nothing worth escalating. That reads like a marketing note until you look at the roadmap it belongs to: MTIA 300, 400, 450 and 500, one every six months, all designed in-house at Meta, all riding Broadcom design services and TSMC leading-edge nodes. MTIA 300 is already in production for the ranking and recommendation systems that drive Facebook and Instagram. 450 and 500 are being tailored for generative image and video inference — the workloads Meta needs cheap-per-token for Reels, Threads, and whatever the Ray-Ban glasses turn into next year.

The Loop covered OpenAI's Broadcom-designed Jalapeño a few weeks back: nine months from spec to first silicon, 26 gigawatts of committed inference capacity, same TSMC line. Meta's timeline is if anything more aggressive, and it maps onto the same partner. If you were writing Broadcom's investor pitch this week you would put it in one sentence: the three largest custom-silicon AI programmes on Earth are ours. That is roughly what the stock did on Thursday.

The arithmetic of $145 billion

Meta's 2026 capex guidance is $125–145 billion. Meta's total revenue for the full year 2023 — three years ago — was roughly $135 billion. The company is spending, on AI infrastructure alone in a single year, more than it earned in the year Threads launched. As Forrester analyst Mike Gualtieri put it to Reuters, "You can't become an AI titan if you are dependent on another company for chips." That is a sentence written to reassure Meta shareholders, and it is a sentence written for Jensen Huang. Both readings are correct.

The gigawatt maths does the same work. Meta added 1 gigawatt of compute in the first half of 2026 and plans to add another 2.5 gigawatts by year-end to hit the 7 GW target. Doubling that to 14 GW in 2027 requires either a lot more Nvidia H200 orders — which Meta has been placing, and continues to place — or a rapid substitution into Iris-class silicon at a per-watt cost Meta controls. The memo is the announcement that the second half of that sentence is now the plan.

What Meta's supply chain is telling you

The named-supplier list Meta locked in this year reads like a hardware map:

• Design: Broadcom, contracted through 2029, covering all four MTIA generations
• Fabrication: TSMC, leading-edge node
• Memory: Samsung Electronics
• Flash storage: Sandisk
• Fibre-optic interconnect: Sumitomo Electric

Notice what is missing. No Nvidia, no AMD, no Intel. Meta is not ripping Nvidia out — the H200 fleet will keep growing — but it is building the parallel stack such that new marginal capacity can flow into Iris rather than into a fifth-year H-series purchase order. This is the play Google ran with the TPU over a decade, and the play Amazon has been running with Trainium and Inferentia. The difference is speed: Google took ten years to get to serious internal share; Meta is trying the same substitution in eighteen months and change.

The market reaction

Nvidia at $201.63, down 1.2 per cent on the day. Broadcom at $398.63, up 2.6 per cent. Meta flat during the session, then up 4.6 per cent in late trading once the memo hit the tape. There is nothing subtle about the pricing. When the largest AI capex buyer on Earth announces a credible manufacturing timeline for its own inference silicon — and does it via a memo leak on a weekday afternoon — the market marks the incumbent down and the design partner up, in that order.

Nvidia has been the entire AI trade for two years. A 1.2 per cent single-session move is small in absolute terms and large in what it signals about how much slack was priced into Nvidia's expected 2027 share of Meta's spend.

What this means

Three claims, ordered by confidence.

1. The custom-silicon transition is now a schedule, not a strategy. Meta has committed dates: September manufacturing, MTIA 300 already shipping ranking workloads, 450 and 500 in the pipeline on a six-month rhythm. OpenAI has committed dates on Jalapeño. Google's 8th-generation TPU is in the field. Anthropic is in talks with Samsung on 2nm inference silicon. The four biggest inference buyers in the world are on the same clock, and Broadcom's fingerprints are on three of the four programmes.
2. Nvidia's data-centre share peaks somewhere in the next four quarters. Not because the H200 stops selling — it will not — but because the marginal gigawatt of AI capacity starts flowing into custom parts before the H200 line has finished shipping to Meta. The revenue mix inside Nvidia's data-centre segment starts to look qualitatively different by mid-2027. The stock market will figure this out earlier than the earnings prints do.
3. Broadcom is the shadow winner of the 2026 capex cycle. It has become the design house of record for hyperscaler AI silicon. It is the only large-cap semiconductor company that gets paid whichever labelled chip the customer buys — because it engineered a chunk of both.

What to watch

• Whether Meta's Q3 2026 earnings call discloses Iris deployment numbers or keeps the memo's targets internal. The gap between the leaked memo and the eventual disclosure is where the story gets sharper.
• Whether Nvidia's next data-centre segment print shows a slowdown in the Meta line specifically. Nvidia does not disclose per-customer revenue; analysts will try to back it out from Meta's capex composition.
• Whether MTIA 450 — Meta's tailored generative-image inference chip — ships anywhere close to its notional early-2027 slot. That is the chip that decides whether Meta can run Reels and generative-ad workloads at a per-request cost that undercuts Nvidia-only competitors.
• Whether Broadcom signs a fourth hyperscaler custom-silicon programme. It already has three publicly known (Meta, OpenAI, Google). AWS uses Marvell for Trainium. If Broadcom takes even part of a next-generation AWS design, it becomes the effective supplier-of-record for the entire hyperscaler custom-silicon industry.

Six weeks. Zero significant bugs. Six-month cadence. $145 billion of capex behind it. Meta looked at the two-year run in which Nvidia has been the entire trade and decided the honest answer is we would like to be an AI titan, and titans do not rent their arithmetic. The memo tells the employees. The stock ticks tell the story. Iris ships in September.