Meta's 'Iris' AI Chip Enters Production in September, Targeting 14 Gigawatts of Compute by 2027
Meta Platforms is preparing to manufacture its first large-scale custom AI accelerator chip, code-named Iris, with production set to begin in September. According to an internal memo reviewed by Reuters, the chip cleared six weeks of testing without a major issue — an unusually swift development cycle for silicon of this ambition. The accelerator is the newest and most ambitious member of Meta's Meta Training and Inference Accelerator programme, known as MTIA, a line that has shipped inference chips since 2023 but never at this scale. Designed in partnership with Broadcom and manufactured by Taiwan Semiconductor Manufacturing Company, Iris is intended to power the recommendation engines and generative AI features that underpin Facebook and Instagram for billions of users daily.
The strategic calculus behind Iris is straightforward: every inference query Meta serves on its own silicon is one that does not pay Nvidia's margin. Meta has projected total capital expenditure of between $125 billion and $145 billion for 2026, the overwhelming majority of it directed at AI infrastructure. According to the Reuters memo, the company intends to deploy seven gigawatts of compute in 2026 and double that figure to 14 gigawatts by 2027. The internal memo frames the new chip as a supplement to, rather than a replacement for, the Nvidia and AMD processors Meta already spends billions on each year — but industry analysts view the distinction as one that will erode with each successive MTIA generation. Broadcom, notably, is also the design partner behind Google's newest tensor processing units and OpenAI's first custom chip, cementing its role as the silent architect of hyperscaler independence from Nvidia.
Wall Street's reaction to the Iris disclosure was ambivalent. Shares in Meta fell briefly after the memo circulated before recovering later in the same trading session, rising approximately 4.6 percent by late afternoon on July 9, according to reporting from Let's Data Science. The recovery was partly attributed to a separate announcement of developer access to a new AI coding model. Investor patience with Meta's AI spending is not unlimited; in April, a surge in first-quarter capital expenditure prompted a JPMorgan downgrade, and the stock shed roughly ten percent of its value over two sessions on concerns that returns were not keeping pace with outlays. Iris does not resolve that tension — it extends the timeline over which Meta hopes to demonstrate that owning its silicon stack will generate durable cost advantages.
The Iris production timeline fits into a broader partnership with Broadcom formalised earlier this year, covering multiple MTIA generations through 2029, as Yahoo Finance reported. Meta's move mirrors an industry-wide pivot among hyperscalers. Google has its TPUs, Amazon has Trainium and Inferentia, and Microsoft is developing its own Maia AI accelerator. What distinguishes the current moment is the sheer scale of the bets: meta alone projects up to $145 billion in infrastructure spending this year, and the combined capex of the five largest US technology platforms is expected to approach or exceed $500 billion in 2026, according to Goldman Sachs estimates. Each dollar spent on proprietary silicon is a dollar that does not flow to Nvidia, raising the stakes for a company whose market capitalisation has been built on AI infrastructure dominance.
The September production milestone will be closely watched by investors and competitors alike. If Iris ships on schedule and performs reliably in Meta's vast data centres, it will validate a custom-silicon strategy that took years and multiple failed iterations to reach this point, and it will give Meta a credible roadmap toward hardware autonomy. If it stumbles, it will revive longstanding questions about whether even the largest technology platforms can match the engineering velocity of a dedicated chipmaker like Nvidia. For the broader semiconductor industry, the answer matters enormously: a successful Iris would signal that the era of hyperscaler dependence on merchant silicon is giving way to a more fragmented, vertically integrated hardware landscape that rewrites competitive dynamics from Taiwan's fabs to Wall Street's analyst models.