Meta’s Iris Chip Feeds a 14 Gigawatt Compute Bet

Meta planned to start making its Iris AI chip in September to help lift computing power to 14 gigawatts in 2027, an internal memo showed. Testing took six weeks and found no major issues. Broadcom is helping design the part. Taiwan Semiconductor Manufacturing Co. is set to fabricate it.

Iris does not take Nvidia or AMD boards out of Meta’s halls. It is a cheaper way to fill a doubling of compute that still needs merchant GPUs, memory contracts, and new power plants in Louisiana.

The Chip Code-Named Iris Cleared Testing in Six Weeks

The July memo, later corrected on a capacity figure, said the data-center part code-named Iris sits inside Meta’s four-generation Training and Inference Accelerator project, or MTIA. Meta declined to comment on the memo. Broadcom and TSMC were not quoted on the production date.

The same memo said the in-house effort had struggled for more than five years. A clean six-week bug pass is the first public sign that a new die was ready to go to the fab. Meta had already shown Iris under its technical name in March, beside three other processors, and said it wants a new chip about every six months through 2027. Most vendors still ship on a yearly cycle or slower.

Mike Gualtieri, a vice president and principal analyst at Forrester, put the motive in blunt terms. “You can’t become an AI titan if you are dependent on another company for chips,” he said. The memo itself was more practical. Adopting the latest GPUs at Meta’s size “has been a heavy lift, and it has cost us time.”

That line is the tell. Iris is a scheduling tool as much as a science project. If a new Nvidia generation is slow to land in Meta’s racks, a homegrown inference part that drops into the same chassis can keep Facebook ranking, Instagram recommendations, and Llama-class serving moving while the GPU queue clears.

Seven Gigawatts This Year, 14 in 2027

The memo’s capacity plan is larger than the chip. Meta planned to deploy seven gigawatts of computing infrastructure in 2026 and to double that to 14 gigawatts in 2027. One gigawatt is enough to power about 800,000 homes.

To get to seven gigawatts, the company added 1 gigawatt in the first half of 2026 and forecast another 2.5 gigawatts by year-end. An earlier copy of the same story had said 5.5 gigawatts for that second-half add; the published memo text was corrected to 2.5. The seven-gigawatt figure is a fleet total, not a single-year install of seven new gigawatts from a standing start.

THE 2026 COMPUTE BILL

• 2026 capex range: Meta told investors it expects capital spending, including principal payments on finance leases, of $130 billion to $145 billion.
• Last year’s spend: 2025 capex was $72.2 billion, so the top of the new range is about double.
• June quarter hit: Second-quarter capex was $31.1 billion against free cash flow of $784 million.
• The top line still grew: Revenue was $60.8 billion, up 28%, with net income of $15.8 billion, or $6.18 a share.

Those June-quarter numbers landed on July 29. Total expenses were $42 billion, up 55%, and included $2.4 billion tied to legal proceedings and $1.2 billion of severance from a May headcount cut. Full-year expenses were guided to $165 billion to $169 billion. Chief financial officer Susan Li said Meta is “demand constrained” and still has “numerous ROI-positive places” to put compute if it had more of it.

Chief executive Mark Zuckerberg told the same call that a large share of the new compute will go to training, the core apps, and personal agents, and that Meta also wants a large business serving outside customers. The company is already taking offers to sell spare capacity at a premium to what it paid. Iris is one way to make those watts cheaper to own. It is not a substitute for the power feed or the building.

Broadcom Builds the Silicon Meta Calls Homegrown

On April 14, Meta and Broadcom announced an expanded deal to co-develop multiple MTIA generations through 2029. The first phase is a commitment that exceeds 1 gigawatt, described as the opening of a multi-gigawatt rollout. Broadcom works across chip design, advanced packaging, and Ethernet networking on its XPU platform. Because of the size of the contract, Broadcom chief executive Hock Tan left Meta’s board and moved to an advisor seat on the silicon roadmap.

Meta is partnering with Broadcom across chip design, packaging, and networking to build out the massive computing foundation we need to deliver personal superintelligence to billions of people. As we roll out more than 1GW of our custom silicon to start and then multiple gigawatts over time, this partnership will give us greater performance and efficiency for everything we’re building.

Mark Zuckerberg, founder and CEO, Meta newsroom, April 14, 2026

Tan called the first MTIA drop “just the beginning of a sustained, multi-generation roadmap.” On a later Broadcom earnings call he said the Meta work remained on track, with three MTIA generations due by the end of 2027 and line of sight to 3 gigawatts through 2028. That is still a slice of a 14-gigawatt fleet. The rest of the hall still has to be filled with GPUs and older MTIA parts.

The “in-house” label is doing a lot of work. Meta sets the workload, the compiler path, and the rack. Broadcom supplies the custom-accelerator platform. TSMC cuts the wafers. Samsung is on the memo for memory, Sandisk for flash storage, and Sumitomo Electric for fiber-optic gear, each on multi-year supply deals. Independence from Nvidia, in this setup, is a shift of vendor risk, not a retreat from the supply chain.

That is the pattern across the large cloud builders. Google’s TPUs, OpenAI’s first custom part, and Meta’s MTIA all run through Broadcom’s design shop and TSMC’s advanced lines. The company that collects a fee every time a hyperscaler tries to pay Nvidia less is Broadcom. The company that still has to find a leading-edge slot for the die is TSMC.

Why Nvidia GPUs Still Fill Most of the Hall

Meta’s own engineering post is clear about mix. The company says it will keep a diverse silicon portfolio and use the best parts it can buy as well as the parts it designs. MTIA is for cost-effective serving of Meta’s own models. It is not a product for sale, and it is not the default training engine for the largest runs.

The March 11 hardware blog says Meta has already deployed hundreds of thousands of MTIA chips, onboarded internal production models, and tested the line with Llama. Those early chips, now called MTIA 100 and MTIA 200, started as ranking and recommendation parts. The new stack, four successive generations of MTIA chips, stretches from recommendation training into generative-AI inference.

Meet the actual device, not the press name. An MTIA 300 compute chiplet is a grid of processing elements. Each element has two RISC-V vector cores, a dot-product engine for matrix math, a special-function unit for activations, a reduction engine, and a DMA block for local memory. Later parts glue two compute chiplets together, then a 2-by-2 of smaller chiplets, with network dies and HBM stacks around them. A rack of 72 MTIA 400 devices shares one scale-up domain on a switched backplane, with air-assisted liquid cooling so the boxes can land in older halls.

Software is the other half of the bet. MTIA is built to run PyTorch, vLLM, and Triton, and to sit in Open Compute racks. Meta says production models can compile with torch.compile and torch.export and land on GPUs and MTIA without a rewrite. That is how a custom part can take inference load off a GPU without forcing every research team onto a private stack.

The implication is unromantic. Training clusters will still want dense GPU islands for the jobs the custom die was not built to win. Inference and ranking, which run all day on Facebook and Instagram, are where a six-month MTIA loop can cut the bill. Bank of America analyst Justin Post has argued that Iris was unlikely to produce large capacity-cost savings in 2026 because production was only slated to start in September. The 2027 math is the one that matters, and even that math assumes the dies yield, the racks accept them, and the power is there.

Louisiana Power Plants Have to Match the Chips

A 14-gigawatt computing plan is a utility plan with a chip on top. Meta’s Richland Parish campus, which the company also calls Hyperion, is the clearest example. Rachel Peterson, Meta’s vice president of data centers, said the site has the potential to scale up to 5 gigawatts and that Meta has worked with Entergy so other consumers are not paying Meta’s costs.

Entergy Louisiana’s March 27 release lays out what that looks like on the grid. Meta is to pay for seven new combined-cycle gas plants totaling more than 5,200 megawatts, with room later for carbon capture and hydrogen co-firing. The same package includes about 240 miles of new 500-kilovolt lines, batteries at three sites, nuclear uprates, help funding up to 2,500 megawatts of new solar, and a memorandum to study future nuclear use.

WHAT ENTERGY SAYS META IS FUNDING

• Gas generation: Seven combined-cycle units totaling more than 5,200 megawatts, paid for by Meta rather than spread across the residential base.
• Wires: About 240 miles of new 500-kilovolt transmission linking south Louisiana to north Louisiana and Arkansas.
• Storage and carbon-free add-ons: Batteries at three locations, nuclear uprates, and up to 2,500 megawatts of additional solar, plus a nuclear development memo.
• Bill credits claimed: About $2 billion of customer savings over 20 years on top of $650 million from an earlier pact, or $2.65 billion combined, plus $120 million for Entergy’s Power to Care program and $140 million for efficiency aid.

Phillip May, Entergy Louisiana’s president and chief executive, said the structure is meant to keep rates in check while the plants get built. Local groups have spent 2026 fighting over how much of Meta’s load data the public may see. As of late September, two hearings still sat between the company and the generation it wants, and more process was on the October calendar. A chip that clears the lab in six weeks can still wait on a gas turbine and a 500-kilovolt line.

The same constraint shows up in the bill of materials. The memo’s Samsung, Sandisk, and Sumitomo contracts exist because HBM, NAND, and optical gear have been as tight as GPU allocations. A custom accelerator that needs more HBM bandwidth each generation, which Meta says it does, is still in that queue. Iris does not create extra memory. It competes for it.

Meta Wants a New Accelerator Every Six Months

The March blog is the cleanest map of what Iris is feeding. Meta published research papers on the first two chips at ISCA in 2023 and 2025, then sped up. From MTIA 300 to MTIA 500, it says HBM bandwidth rises 4.5 times and compute FLOPS rise 25 times. The later parts reuse the same chassis, rack, and network, so a new die can drop into a footprint that is already in the hall.

MTIA LINE META DESCRIBED IN MARCH

Iris is the code name the July memo used for a chip in that line as it headed to manufacturing. Meta did not map the nickname onto a single row in the public blog. What the company did say is that MTIA 450 and 500 are inference-first parts that can still pick up ranking and some training, while mainstream GPUs remain the tools built for the heaviest pre-training jobs.

HOW THE SILICON PLAN TIGHTENED

1. March 11, 2026: Meta publishes the 300 through 500 roadmap and says it can ship a new chip roughly every six months.
2. April 14, 2026: Meta and Broadcom extend the partnership through 2029 and commit to more than 1 gigawatt of custom silicon as the first phase.
3. July 9, 2026: An internal memo says Iris cleared six weeks of testing with no major issues and would enter manufacturing in September, on the way to 14 gigawatts in 2027.
4. July 29, 2026: Meta narrows 2026 capex to $130 billion to $145 billion after a quarter in which $31.1 billion of spending left $784 million of free cash flow.

The six-month loop is the part of the plan that actually fights Nvidia’s pricing. If Meta can refresh an inference die twice a year, it can chase new low-precision formats and HBM generations without waiting for a GPU architecture cycle. The catch that the memo already named is time. Getting those dies through TSMC, into cooled racks, and onto a grid that is still adding gas plants is slower than a six-week test.

By early October, Meta had not put out a public confirmation that Iris had moved from the July plan into volume at the foundry. The capacity targets, the Broadcom gigawatt, and the Louisiana plants were still the load-bearing facts. Fourteen gigawatts in 2027 only holds if the wafers, the memory, and the turbines show up with the chip.