Meta-Broadcom MTIA Deal: 1GW of 2nm Custom AI Silicon

TL;DR

On April 14, 2026, Broadcom and Meta announced an extended strategic partnership through 2029 under which Meta will deploy over 1 gigawatt of custom MTIA (Meta Training and Inference Accelerator) silicon in an initial phase, with a "multi-gigawatt rollout" to follow¹². Broadcom described the deal as delivering "the industry's first 2nm AI compute accelerator" built on its XPU platform¹. On the same day, Broadcom CEO Hock Tan — who joined Meta's board in February 2024³ — said he would not stand for reelection and will transition to an advisor role on Meta's custom silicon roadmap, a governance change reflecting the scale of the new commercial relationship⁴.

The deal extends Broadcom's dominance in hyperscaler custom accelerators. Broadcom reported $8.4 billion of AI revenue in its Q1 FY2026 (+106% year over year) and guided to $10.7 billion for Q2 FY2026 (+140% YoY)⁵. Tan has told investors Broadcom now has "line of sight" to AI chip revenue "significantly in excess of" $100 billion in fiscal 2027⁶.

What You'll Learn

• What Meta and Broadcom actually announced on April 14, 2026, and what "1 GW of MTIA" means in practice
• How the four MTIA generations (300, 400, 450, 500) scale, and which one hits 2nm first
• Why Hock Tan stepped off Meta's board on the same day the deal was announced
• How the Broadcom deal fits alongside Meta's separate 6 GW AMD agreement and its ongoing Nvidia spend
• What Broadcom's Q1 AI revenue and $100B-per-year forecast tell us about the custom-silicon wave

What Was Actually Announced

The April 14 announcement is not Meta's first custom-chip disclosure — it is the commercial scaffolding underneath a roadmap Meta already unveiled on March 11, 2026, when it detailed four successive MTIA generations dubbed MTIA 300, MTIA 400, MTIA 450, and MTIA 500, all slated to ship by the end of 2027 on a roughly six-month cadence⁷⁸.

What is new on April 14 is the depth of the Broadcom commitment:

One gigawatt, in the data-center sense, is a power envelope rather than a chip count, but it is enormous. For comparison, Anthropic's recently expanded Google TPU deal totals roughly 3.5 GW of training and serving capacity coming online in 2027⁹. A single 1 GW facility — the category Meta's Prometheus supercluster in New Albany, Ohio is expected to occupy when it comes online in 2026 — would plausibly consume on the order of hundreds of thousands of accelerators depending on chip, rack density, and overhead¹⁰. The point of the April 14 announcement is that Meta now has a contracted silicon pipeline — not just a single chip — to fill capacity of that size with its own design rather than a third-party GPU.

The MTIA Roadmap, Chip by Chip

Meta's March disclosure gave enough specification to compare the four generations directly⁷⁸:

Two numbers from Meta's own roadmap slide are worth pulling out: from MTIA 300 to MTIA 500, HBM bandwidth rises 4.5x and compute FLOPS rises 25x (comparing MTIA 300's MX8 throughput to MTIA 500's MX4 throughput, i.e. across lower-precision formats)⁷. That is the generational climb Broadcom is now under contract to co-deliver, from chiplet integration to advanced packaging to Ethernet-based scale-out networking¹.

Hock Tan Leaves Meta's Board the Same Day

A deal of this magnitude between a customer and a supplier is hard to govern when the supplier's CEO also sits on the customer's board. Hock Tan, who has been Broadcom's president and CEO since March 2006¹², joined Meta's board on February 14, 2024³. On April 14, 2026, he informed Meta he would not stand for reelection, moving instead to an advisor role focused specifically on Meta's custom silicon roadmap⁴.

Neither company framed the move as compelled by a specific conflict — the Broadcom release called it a transition "given the scale of this expanded partnership"¹ — but the optics required it. With a multi-year, multi-generation chip commitment running through 2029, a Broadcom CEO on Meta's board would be voting on AI infrastructure decisions that could directly affect Broadcom's largest customer relationship.

Where the MTIA Deal Fits in Meta's Broader Spend

This is one of three major silicon relationships Meta is simultaneously running, and the April 14 deal explicitly does not replace Nvidia. The map looks roughly like this:

On top of that, Meta has guided 2026 capital expenditure to a range of $115–$135 billion (up from $72.2 billion in 2025) as it pushes data-center capacity from today's footprint toward the 5 GW Hyperion cluster in Louisiana and the 1 GW Prometheus supercluster in Ohio¹⁰¹⁵. The pattern is deliberately multi-vendor: Meta gets pricing and supply leverage by running custom Broadcom silicon, custom AMD GPUs, and Nvidia GPUs in parallel rather than betting the farm on any one of them.

What This Says About Broadcom's AI Business

For Broadcom, the April 14 announcement is one more brick in an AI revenue wall that has been growing faster than almost anything else in the semiconductor sector:

• Q1 FY2026: $8.4 billion in AI revenue, +106% year over year, ~43% of total quarterly revenue of $19.31 billion⁵.
• Q2 FY2026 guide: $10.7 billion in AI revenue, +140% year over year⁵.
• Multi-year outlook: Tan told investors on the March 4 earnings call that Broadcom has "line of sight" to AI chip revenue "significantly in excess of" $100 billion in fiscal 2027⁶.

The customer list making those numbers possible is narrow but marquee. Broadcom is the design and networking partner for Google's Ironwood TPU v7, with coverage reportedly extending through 2031¹⁶; it co-develops MTIA with Meta; and it serves additional hyperscaler customers whose custom silicon it builds on its XPU platform. The April 14 Meta deal is notable less because it is a new customer and more because it quantifies the pipeline: one customer, multiple gigawatts, through 2029, at a process node Broadcom is first to commercialize for AI accelerators.

What It Means for Developers and Builders

If you work on AI applications rather than on data-center procurement, the direct impact is indirect but real:

1. More inference capacity, more often priced competitively. Every gigawatt of custom silicon Meta operates is a gigawatt Meta does not need to buy on Nvidia's price curve. To the extent that pressure flows into what Meta charges for its own products (Llama hosting, ads inference, AI features in WhatsApp and Instagram), it eases the cost floor for builders who consume those services².
2. An open model pipeline that outlasts a single GPU generation. Meta has been the most aggressive open-weights publisher at the frontier; multi-gigawatt MTIA capacity is the training and serving substrate that lets it keep doing that without being capacity-gated by any one vendor¹⁷.
3. The "custom chip" becomes the default, not the experiment. Between Google's TPUs, Amazon's Trainium line, Microsoft's Maia, and now Meta's MTIA on Broadcom 2nm, every major hyperscaler now has a multi-year, multi-generation roadmap for silicon it designs itself. For enterprise buyers, that changes the competitive landscape more than any one model release.

The Bottom Line

The April 14 Meta-Broadcom announcement is a pricing and supply-chain story dressed as a press release. By locking in >1 GW of custom MTIA silicon now and multiple gigawatts through 2029 — on a process node Broadcom says it is first to ship for AI — Meta secures a path to train and serve its next generation of models without being price-constrained by any single GPU supplier. For Broadcom, it further cements a business line that grew 106% in a single quarter and whose CEO now talks about $100 billion a year as a line of sight, not a stretch goal. For the rest of the industry, it confirms a pattern every hyperscaler is now running in parallel: buy Nvidia, buy AMD, and build your own.

References