Meta has started manufacturing its Iris data center AI chip this month, according to an internal memo reviewed by Reuters and cited across multiple technology outlets. The move marks a shift in how the company plans to run its AI workloads over the next two years, and it lands at a moment when GPU supply remains tight across the entire industry. Iris is not a GPU. It is a custom accelerator built under Meta’s Training and Inference Accelerator program, known as MTIA, internally code-named MTIA 400, and it is one of four MTIA generations Meta unveiled in March 2026 to take on a growing share of the ranking, recommendation, and generative AI work that currently leans on Nvidia hardware.
The timing matters. Meta reportedly wants to roughly double its data center AI computing capacity to about 14 gigawatts by 2027, and doing that entirely through GPU purchases would mean competing even harder for the same constrained supply that every other hyperscaler is chasing. Building its own silicon gives Meta a second lever to pull. This article breaks down what Iris actually is, how it fits into the four-generation MTIA roadmap, what the reported numbers say about Meta’s ambitions, and what the move could mean for Nvidia, Microsoft, Google, and the rest of the AI hardware market heading into 2027.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Meta’s Iris Chip Enters Production This Month
According to the internal memo, reported by Reuters in July 2026, Meta’s plan is to begin manufacturing Iris in September 2026, after the chip wrapped six weeks of testing that same July with no major issues reported. The chip is described as a data center AI accelerator rather than a general-purpose CPU or a traditional graphics processor. It is tuned specifically for Meta’s own workloads: the recommendation engines that power the Facebook and Instagram feeds, ad ranking systems, and increasingly, generative AI features baked into Meta’s apps.
That distinction is worth sitting with for a moment. Nvidia GPUs are general-purpose accelerators built to handle a wide range of AI training and inference jobs across thousands of different customers. Iris does not need that flexibility. Meta knows exactly which models it runs and how they behave in production, so it can strip out anything the chip does not need and put the silicon budget into the operations that matter most for its own traffic. That is the same logic Google applied years ago with its Tensor Processing Units, and it is the logic Amazon applies with Trainium.
Reports describe September’s production start as part of a broader transition where Meta shifts a growing share of its AI workloads off third-party accelerators and onto its own MTIA stack across 2026 and 2027. That is a multi-year migration, not a single product launch, and Iris is one piece of it.
Inside the MTIA Program: Four Chips in Two Years
Meta’s own engineering blog frames the effort bluntly, titling one of its roadmap posts “Four MTIA Chips in Two Years: Scaling AI Experiences for Billions,” according to Meta’s official AI blog. Meta unveiled all four chips in that lineup, MTIA 300, MTIA 400 (Iris), MTIA 450, and MTIA 500, in March 2026, and by that same month MTIA 300 was already in production. That post lays out a roadmap where MTIA 500 is scheduled for mass deployment in 2027, alongside earlier generations already running in production clusters. Iris fits into this same family of internally designed silicon, positioned as one of the newer chips in that four-generation plan.
The pace is aggressive by hardware standards. Most chip programs measure their release cadence in years per generation, not months. Meta compressing four distinct accelerator designs into a two-year window signals how much internal pressure there is to reduce dependence on external GPU vendors, and how much capital the company is willing to commit to pull it off.
Some MTIA variants are already deployed in Meta’s data centers today, handling inference workloads for ranking and recommendation systems. Meta is reportedly planning to ship a new AI chip roughly every six months through 2027, and the next step up, MTIA 450, is already specced to double the HBM memory bandwidth of MTIA 400 (Iris), according to Meta. Iris and the newer chips in the pipeline are scheduled to roll out through late 2026 and into 2027, layering on top of what is already running rather than replacing it overnight.
MTIA v1 and v2: The Technical Foundation Behind Iris
Meta has published real technical specifications for its earlier MTIA generations, and those numbers give the clearest picture available of how the chip family actually performs. Describing the first-generation MTIA inference accelerator, Meta’s engineering blog states: “The accelerator is fabricated in TSMC 7nm process and runs at 800 MHz, providing 102.4 TOPS at INT8 precision and 51.2 TFLOPS at FP16 precision,” according to Meta’s official AI blog.
That first chip was built to hit a specific balance of compute, memory bandwidth, and memory capacity for serving ranking and recommendation models rather than chasing the largest possible number on a spec sheet, a design goal Meta explained to Reuters when the chip debuted. The second generation built on that foundation directly. Meta’s blog notes: “Early results show that this next generation silicon has already improved performance by 3x over our first generation chip across four key models we evaluated,” according to Meta’s official AI blog.
A 3x jump between two consecutive generations is a large leap, and it explains why Meta is confident enough to keep investing in the program rather than falling back on GPUs. Meta has not published equivalent process node, clock speed, or throughput figures for Iris itself. Until Meta discloses those details, any specific performance number for Iris beyond what is confirmed for MTIA v1 and v2 should be treated as unconfirmed.
| MTIA Generation | Fabrication Process | Clock Speed | Reported Performance | Status (as of Sept. 2026) |
|---|---|---|---|---|
| MTIA v1 (first gen) | TSMC 7nm | 800 MHz | 102.4 TOPS (INT8), 51.2 TFLOPS (FP16) | Deployed for inference |
| MTIA v2 (next-gen) | Not disclosed by Meta | Not disclosed | ~3x v1 performance across 4 evaluated models | Deployed for inference |
| Iris | Not disclosed by Meta | Not disclosed | Not disclosed | Production started September 2026 |
| MTIA 500 | Not disclosed by Meta | Not disclosed | Not disclosed | Scheduled for mass deployment in 2027 |
The 14-Gigawatt Bet: Why Meta Wants Off Nvidia’s GPUs
The headline number in the memo, reported by Reuters in July 2026, is a target of roughly 14 gigawatts of data center AI computing capacity by 2027, literally double the roughly 7 gigawatts of AI capacity Meta was targeting for 2026. Gigawatts, not chip counts, has become the unit hyperscalers use to talk about AI buildouts, because power availability, not chip supply alone, is now the binding constraint on how fast any company can add compute.
Reports frame Meta’s core objective as cost reduction, aiming to lower GPU spending and soften the effect of an ongoing shortage of AI GPU components that has squeezed the entire industry through 2026. If a hyperscaler can shift even a meaningful slice of its inference workload from a general-purpose GPU to a purpose-built chip it owns end to end, the savings compound across every data center it operates. Ranking and recommendation inference, the workload MTIA chips are explicitly tuned for, runs continuously at enormous scale across Meta’s apps, which makes it the natural first target for in-house silicon rather than training, where the largest and most experimental models still tend to run on Nvidia hardware.
Power is the other half of the equation. Every gigawatt of data center capacity requires securing electricity contracts, permitting, and often new generation capacity years in advance. Meta is not just trying to cut its chip bill. It is trying to make sure that whatever power it does secure gets converted into as much usable AI compute as possible, and a chip designed for its exact workload does that more efficiently than a general-purpose part built to serve every customer’s use case at once.
Broadcom’s Multi-Year Role in Meta’s Custom Silicon
Meta does not design and manufacture MTIA chips alone. Broadcom, which has built a business around helping large technology companies design custom AI accelerators, is reported to be extending its partnership with Meta on MTIA hardware through 2029, according to a Broadcom investor announcement covering the deal. That multi-year framing matters: it signals that both companies expect the custom silicon relationship to keep expanding well past Iris, through several more chip generations.
Broadcom plays a similar role for Google’s TPU program, which makes it one of the few companies with direct visibility into how multiple hyperscalers are approaching custom AI silicon at the same time. That position has turned Broadcom’s networking and custom ASIC business into one of the more closely watched proxies for hyperscaler chip spending, alongside the more familiar Nvidia and AMD data center numbers investors already track.
How Iris Stacks Up Against Google TPU, Microsoft Maia, and Amazon Trainium
Meta is late to custom silicon compared with some of its peers, but it is far from alone. Google has run multiple generations of its Tensor Processing Units for internal search, ads, and cloud AI workloads for close to a decade. Amazon has done the same with its Trainium and Inferentia chips inside AWS. Both programs show that the design-your-own-accelerator strategy can mature into a durable part of a hyperscaler’s infrastructure rather than a side experiment.
The most direct new rival to Iris in the current news cycle is Microsoft’s Maia line. Reports based on The Information describe a Maia 300 accelerator built for internal Azure and Copilot workloads, with an early public unveiling possibly coming as soon as September 2026. Microsoft has not officially confirmed Maia 300 details beyond general statements, but a reported order for more than 300,000 accelerator units of TSMC manufacturing capacity for 2027 delivery has circulated widely, underscoring how much capacity these programs are starting to consume even before public launches.
Set side by side, the picture is a genuine four-way race among Meta, Google, Microsoft, and Amazon to reduce reliance on merchant GPUs for at least part of their AI compute stack, even as all four remain major Nvidia customers for training and for workloads that still need general-purpose flexibility.
| Company | Custom Silicon Program | Primary Reported Use Case | 2026-2027 Status |
|---|---|---|---|
| Meta | MTIA (including Iris) | Ranking, recommendation, generative AI inference | Iris entered production September 2026; MTIA 500 targeted for 2027 |
| Microsoft | Maia (Maia 300 reported) | Internal Azure and Copilot workloads | Unveiling possibly as early as Sept. 2026; officially unconfirmed by Microsoft |
| Tensor Processing Units (TPU) | Search, ads, Google Cloud AI | Multiple generations already in production (established program) | |
| Amazon | Trainium / Inferentia | AWS internal and customer-facing AI workloads | Multiple generations already in production (established program) |
| Nvidia | Data center GPUs (Rubin roadmap) | General-purpose AI training and inference across all customers | Continues to supply hyperscalers including Meta and Microsoft |
Nvidia’s Response: Rubin, Data Center Roadmaps, and Market Position
None of this means Nvidia is being pushed out. Meta, Microsoft, Google, and Amazon all remain among Nvidia’s largest customers, and the reported shift toward custom silicon is aimed at specific, repetitive inference workloads rather than the frontier training runs that still demand the most flexible and powerful GPUs available. Nvidia continues to advance its own data center GPU roadmap, including its next-generation Rubin architecture, and the company’s install base advantage across software, tooling, and developer familiarity remains difficult for any single custom chip program to replicate quickly.
What changes is the mix. If Meta and Microsoft succeed in moving a larger share of inference traffic onto MTIA and Maia hardware, the growth rate of their GPU orders could slow even as their total AI compute keeps climbing, since the custom chips absorb workloads that would otherwise have gone to additional GPU purchases. That is a subtler outcome than Nvidia losing customers outright, but it is the kind of shift that shows up first in growth-rate commentary on earnings calls before it shows up anywhere else.
Market Impact: What a Hyperscaler Chip Shift Means for GPU Demand
The immediate practical effect of Meta’s move is on supply allocation rather than pricing. TSMC capacity is finite, and every wafer devoted to MTIA chips, Maia accelerators, or Google’s TPUs is a wafer not available for Nvidia’s GPU lines or for the dozens of smaller AI chip startups also competing for the same fabs. That allocation fight, not headline GPU prices, is where the near-term pressure will show up first for companies without the scale to negotiate their own long-term capacity deals.
Smaller AI labs and startups that lack the capital or engineering headcount to design their own silicon are the group most exposed to this dynamic. They cannot build an MTIA-style program of their own, so they remain fully dependent on merchant GPU supply precisely as the largest buyers start diverting some of their spending, and some of the available fab capacity, toward custom chips instead. That could tighten availability for everyone else even if Meta’s own GPU orders do not fall sharply in absolute terms.
Historical Context: From First-Gen Inference Chips to Data Center Ambitions
Meta’s first MTIA chip, built on a TSMC 7nm process, was a relatively modest inference accelerator aimed narrowly at recommendation workloads. It did not attempt to compete with Nvidia’s flagship data center GPUs on raw throughput. The second generation closed part of that gap with its reported 3x performance jump. Iris represents the point where the program graduates from a targeted efficiency project into a chip meant to carry a meaningful share of Meta’s total AI compute load, backed by a 14-gigawatt capacity target and a Broadcom manufacturing partnership that now runs through 2029.
That arc mirrors what happened at Google roughly a decade earlier and at Amazon in the years after that: an internal project built to shave costs on one specific workload gradually expands into a full parallel hardware track that sits alongside, rather than replaces, purchased GPUs. Meta is following a path other hyperscalers have already proven works, just doing it faster and later.
Industry Voices on the Custom Silicon Shift
Meta’s own engineering team has been direct about the design philosophy driving MTIA. As the company put it when describing the first-generation chip, its “architecture is fundamentally focused on providing the right balance of compute, memory bandwidth, and memory capacity for serving ranking and recommendation models,” a statement Meta made to Reuters when the chip line debuted. That framing has not changed as the program has scaled up toward Iris and the planned MTIA 500 chip. Meta’s corporate newsroom has continued to frame its infrastructure investments, including custom silicon, as core to keeping pace with AI feature demand across its apps.
On the manufacturing side, Broadcom has described its extended role plainly: the company will “deliver technology supporting Meta Training and Inference Accelerator (MTIA) chips, with plans to extend through 2029,” according to a Broadcom investor release announcing the partnership. Taken together, Meta’s design commentary and Broadcom’s manufacturing commitment point to a program built for the long term rather than a one-off cost-cutting exercise tied to a single product cycle.
Risks and Unanswered Questions in Meta’s Chip Strategy
Custom silicon programs carry real execution risk. Designing a chip is one challenge. Ramping manufacturing yield, validating it against production traffic at global scale, and building the software stack engineers need to actually target the hardware are separate, harder problems that have derailed ambitious chip programs at other companies before. Meta has not disclosed a yield target, a unit cost figure, or a specific performance benchmark for Iris itself, which makes it difficult to verify from the outside whether the chip will deliver the efficiency gains the 14-gigawatt plan assumes.
There is also a training-versus-inference distinction that is easy to lose in the coverage. Every confirmed detail about MTIA chips, including Iris, points to inference workloads: ranking, recommendation, and serving already-trained models to users. None of the reporting suggests Meta is ready to move its largest, most demanding training runs off Nvidia GPUs. That means even a fully successful Iris rollout would only address part of Meta’s total AI compute bill, not all of it.
What to Watch Through 2027
A handful of developments will show whether Meta’s Iris bet is paying off, and they are worth tracking through next year.
- Watch Meta’s capital expenditure commentary on upcoming earnings calls for any explicit reference to Iris production volumes or MTIA’s share of total inference compute.
- Watch whether Microsoft moves forward with a public Maia 300 unveiling in the coming weeks, which would confirm the reported TSMC capacity order and put a second hyperscaler’s custom chip roadmap into the open at the same time as Meta’s.
- Watch TSMC’s capacity allocation disclosures for signs of how much fab space custom hyperscaler chips are consuming relative to Nvidia’s GPU lines.
- Watch whether Meta’s reported 14-gigawatt target for 2027 holds up, gets revised, or slips, since power procurement timelines are often the real constraint behind big compute targets.
- Watch for MTIA 500’s mass deployment, which Meta’s own roadmap points to in 2027, as the next concrete milestone after Iris.
Five predictions follow from where things stand today. First, expect Microsoft to confirm at least partial Maia 300 details before the end of 2026, given how widely the TSMC capacity order has already circulated. Second, expect Meta’s GPU order growth rate, not its absolute GPU spending, to be the metric analysts watch most closely for signs that MTIA is actually displacing Nvidia hardware. Third, expect Broadcom’s custom silicon design revenue to keep climbing as a disclosed line item, given its now-extended role with both Meta and Google. Fourth, expect at least one more hyperscaler-adjacent company to announce or expand its own custom AI chip program in the next twelve months, following the pattern Google, Amazon, and now Meta have each set. Fifth, expect training workloads to remain on merchant GPUs through 2027 even if inference increasingly shifts to custom silicon, since none of the current reporting suggests any hyperscaler is ready to move its largest model training runs off Nvidia hardware yet.
What This Means for the Broader AI Hardware Market
For readers tracking AI hardware more broadly, Iris is a reminder that the GPU shortage story has two sides. On one side, demand for AI compute keeps climbing faster than fabs can add capacity, a dynamic already visible in memory pricing pressure across the PC building market. On the other side, the largest buyers are no longer just competing for a bigger slice of Nvidia’s output. They are increasingly building their own supply, which changes who competes for what. Consumer and enthusiast hardware, from smartphone chips built on leading-edge process nodes to desktop AI accelerator cards aimed at local inference, sits downstream of the same TSMC capacity that Meta, Microsoft, and Nvidia are now all drawing from at once.
That also explains why AMD’s own push into unified-memory AI hardware, seen in products like the Ryzen AI Max PRO 400 platform, and Nvidia’s continued expansion of enterprise server partnerships, visible in deals like the one behind the HPE and Nvidia Vera CPU server launch, are both playing out on the same crowded manufacturing runway that Meta’s Iris chip now competes on for wafer starts and packaging capacity.
Frequently Asked Questions
What is Meta’s Iris AI chip?
Iris is a data center AI accelerator that Meta designed in-house under its MTIA (Meta Training and Inference Accelerator) program. It is built specifically for Meta’s own AI workloads, such as ranking and recommendation systems, rather than as a general-purpose chip sold to other companies.
When does Iris start production?
According to an internal memo reviewed by Reuters, Meta planned to begin manufacturing Iris in September 2026.
Is Iris a replacement for Nvidia GPUs at Meta?
Not entirely. Reports indicate Iris and the broader MTIA lineup are focused on inference workloads like ranking and recommendation. There is no confirmed reporting that Meta plans to move its largest AI model training runs off Nvidia GPUs.
What is MTIA?
MTIA stands for Meta Training and Inference Accelerator, a four-generation internal chip program Meta has been developing to power AI features across Facebook, Instagram, and its other services, with Iris representing one of the newer chips in that lineup.
How much AI compute capacity is Meta targeting?
Reports citing the internal memo describe a plan to roughly double Meta’s data center AI computing capacity to about 14 gigawatts by 2027, using a combination of in-house accelerators and purchased hardware.
Who else is building custom AI chips like Meta’s Iris?
Google has run multiple generations of its Tensor Processing Units for years, Amazon has its Trainium and Inferentia chips for AWS, and Microsoft is reported to be preparing a Maia 300 accelerator for its own Azure and Copilot workloads, with an unveiling possibly coming as soon as September 2026.
Does Broadcom manufacture Iris?
Broadcom is a longtime design and manufacturing partner for Meta’s MTIA program. According to a Broadcom investor announcement, the company’s partnership with Meta on MTIA chips has been extended through 2029.
Will this reduce demand for Nvidia GPUs?
It could slow the growth rate of GPU orders from hyperscalers that succeed in shifting inference workloads to custom chips, but Meta, Microsoft, Google, and Amazon are all expected to remain major Nvidia customers for training workloads and for AI tasks that still require general-purpose flexibility.
