In July 2026, Reuters exclusively reported that Meta Platforms planned to begin manufacturing an artificial-intelligence chip code-named Iris in September, part of a four-generation Meta Training and Inference Accelerators (MTIA) project. The internal memo reviewed by the agency placed the effort inside a larger goal: deploy on the order of 7 gigawatts of computing infrastructure in 2026 and double that figure to 14 GW in 2027.
Iris is not a general-purpose CPU: it is a data-center accelerator aimed at training and inference for the models that power Facebook, Instagram and the rest of Meta’s AI stack. The company tailored the chip with Broadcom’s design help and tapped Taiwan Semiconductor Manufacturing Co (TSMC) for fabrication. The stated aim is to cut massive computing costs and gain relative independence from Nvidia and AMD—without abandoning those purchases.
CNBC and Economic Times repeated the same factual skeleton: Iris was unveiled under its technical name in March alongside other MTIA-roadmap processors; Meta aims to launch a chip about every six months through 2027, a faster cadence than many vendors’ annual cycles. The memo said chip testing took about six weeks with no major issues—a momentum signal after years of uneven internal effort.
By mid-September 2026, follow-on coverage—including independent analysis—described Iris as having “entered production.” Neural Pulse stresses the critical distinction: “in production” is a manufacturing milestone, not a public proof of performance, yield or share of Meta’s total compute. There are no widely published external benchmarks and no official figures for what fraction of Llama/Muse training will run on Iris in 2027.
Power numbers are concrete enough to be falsifiable. One gigawatt can supply on the order of hundreds of thousands of homes; 14 GW implies an energy and industrial footprint comparable to continuous demand of a small country. Per Reuters, the memo said Meta added ~1 GW in the first half of 2026 and projected several more GW to finish near the 7 GW yearly goal.
Custom silicon fits a broader trend: Google with TPU, Amazon with Trainium/Inferentia, Microsoft with Maia, and now Meta accelerating MTIA. The practical difference is that Meta still buys “enormous volumes” of Nvidia/AMD GPUs; Iris is meant to complement specific workloads—recommendation, internal open/closed model training—where a custom ASIC can win on cost per token or per watt.
Supply-chain pressure also bites. The same memo, according to reports, alluded to deals for memory, flash and fiber optics. Building accelerators is useless without HBM, networking and power. That is why the 14 GW target is as much a chips plan as a campuses, power-contracts and permitting plan—terrain where Meta already faces local and climate scrutiny.
For investors and rivals, the mixed message is deliberate: Meta wants better leverage with Nvidia by showing a credible alternative, yet cannot afford a capacity miss if Iris underperforms. “Independence” here means bargaining room, not autarky. Any TSMC delay, stack software bug or memory shortage would reopen third-party GPU dependence.
Neural Pulse summarizes the verifiable core: July Reuters plan (Iris production from September; four-generation MTIA; Broadcom/TSMC; 7→14 GW); September indications that production started; absence of public independent benchmarks. We do not claim Iris already rivals Blackwell/Vera Rubin on raw performance, nor that Meta will stop being a huge Nvidia customer.
The September 15, 2026 editorial angle is infrastructure: while public debate fixates on existential resignations and flagship models, Meta executes the unglamorous part—silicon, megawatts and yields. If that execution sustains a six-month generation cadence, it can reshape the marginal cost of training and serving agents at social-network scale.
What remains to watch in coming quarters is how much of the roadmap becomes energized racks. Until then, Iris is a factory fact in progress and a double-digit-gigawatt energy ambition—both worth tracking, neither yet a complete victory.
For Neural Pulse readers in Spanish and English, the lesson is the same: the AI frontier is measured not only in parameters or coding ELO, but in who controls accelerators, packaging and gigawatts. Meta just moved a piece on that board.
A chip-every-six-months cadence through 2027, if sustained, would force Meta to mature not only silicon but compilers, kernels and tools for partitioning models across GPUs and Iris. Internal ASIC history is full of brilliant chips trapped by mediocre software; the Reuters memo celebrates short testing, not a full software stack proven in massive production.
Environmentally, 14 GW demands a parallel conversation on power grids, cooling water and local permits—topics that already filled recent datacenter coverage. Iris does not absolve Meta of that footprint; if anything, it concentrates it in owned campuses where the ASIC must beat generic GPUs on performance per watt.
