Meta Platforms and Broadcom have extended their custom AI chip partnership through 2029, committing to an initial deployment exceeding 1 gigawatt of Meta Training and Inference Accelerator (MTIA) silicon manufactured on a 2-nanometer process. The agreement, announced on April 15, 2026, anchors Meta’s $135 billion 2026 capital expenditure plan and positions Broadcom to capture up to $100 billion in custom AI chip revenue by fiscal 2027. It also reshapes the balance of power in a $650 billion AI infrastructure buildout that has, until now, been dominated by Nvidia’s merchant GPU business.
The MetaâBroadcom deal is the clearest signal yet that hyperscalers are no longer willing to outsource the most strategic layer of their AI stack. With Broadcom shares climbing back to the $400 level on the news and Meta earmarking roughly 6 gigawatts of AMD GPU capacity and millions of Nvidia chips alongside the MTIA ramp, the announcement codifies a three-vendor strategy that will define Meta’s AI infrastructure through the end of the decade. This analysis unpacks what the contract includes, what it means for Broadcom’s AI revenue trajectory, how it compares with Google’s TPU and Amazon’s Trainium programs, and why the MTIA 2nm roadmap matters for the next wave of frontier models.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Inside the MetaâBroadcom 2029 Extension: Scope, Scale and Silicon
The agreement disclosed on April 15, 2026 is not a one-off order. It is a multi-generation partnership that locks Broadcom in as Meta’s primary co-development partner for MTIA accelerators through 2029, with an initial commitment exceeding 1 gigawatt of compute capacity. Broadcom described the deal as “the first phase of a sustained, multi-gigawatt rollout,” language that implies follow-on tranches well beyond the initial gigawatt. To put the scale in perspective, 1 gigawatt of deployed compute is enough electrical load to power roughly 750,000 U.S. homes, making this one of the largest single-customer silicon commitments in the history of semiconductors.
The technical centerpiece is the next-generation MTIA chip, which Broadcom and Meta say will be the first custom AI silicon fabricated on a 2nm process node. Meta has already shipped four generations of MTIA accelerators, starting with the MTIA 300 that powers ranking and recommendation systems across Facebook, Instagram and WhatsApp. The current production MTIA design is targeted at specific training workloads, while the new 2nm chips are engineered to handle the full spectrum of Meta’s AI workloads with a strong emphasis on inference, where unit economics matter most at Meta’s scale. Meta and Broadcom have telegraphed three additional chip generations by 2027, implying a roughly annual cadence of new silicon drops rather than the 18â24 month cycles typical of previous custom ASIC programs.
Governance around the deal is also evolving. Broadcom CEO Hock Tan will step down from Meta’s board and transition into an advisory role specifically focused on custom chip strategy. That structure removes any lingering fiduciary conflicts as Meta becomes, in effect, a meaningful portion of Broadcom’s long-term order book, while keeping Tan directly engaged on roadmap decisions. It is a template other hyperscalers are likely to copy as custom silicon partnerships deepen.
The 1 Gigawatt Number: What It Actually Buys Meta
The “1 gigawatt” headline is the unit of analysis that matters, because capex dollars are increasingly translated into power capacity rather than chip count. At roughly 1,000 watts per advanced accelerator at the system level, 1 gigawatt of deployed MTIA compute translates into the equivalent of roughly 1 million accelerator packages, once networking, CPUs and cooling overhead are stripped out. In practice, Meta will not hit a literal million-chip figure because a large fraction of that gigawatt is consumed by liquid cooling, optical interconnect and redundant power, but the order of magnitude underscores why a single agreement can reshape Broadcom’s revenue mix.
For Meta, the gigawatt allocation sits inside a staggering 2026 capex plan of up to $135 billion, nearly triple the company’s 2024 capex run rate and a step function above the $60â65 billion range it was guiding to just 18 months ago. Meta has publicly committed to roughly 6 gigawatts of AMD MI-series GPUs on top of millions of Nvidia accelerators, plus the MTIA buildout, spread across Louisiana, Ohio and its new campuses. The MTIA 2nm fleet is slated to handle the workloads where Meta has the deepest software control: Reels ranking, ads auction models, Llama 4 and Llama 5 inference, and on-device personalization signals feeding into Ray-Ban Meta and its smart glasses roadmap.
Crucially, Meta is not replacing Nvidia. It is partitioning workloads. Frontier training for the largest dense and mixture-of-experts models will continue to run on Nvidia Blackwell and Rubin clusters, where CUDA, NCCL and the mature software stack still offer an unmatched combination of throughput and developer productivity. AMD’s Instinct footprint handles a growing slice of training and mid-tier inference where Meta wants price-performance use. MTIA takes the workloads with the most predictable shapes – embeddings, ranking, recommendation and certain inference paths – where a purpose-built ASIC can beat a general-purpose GPU on cost per token by a wide margin.
Broadcom’s $100 Billion AI Revenue Target: How Meta Gets It There
Broadcom has guided to delivering approximately $100 billion worth of custom AI chips by fiscal 2027. To understand why the Meta agreement is the cornerstone of that target, it helps to look at Broadcom’s current AI revenue mix. In fiscal 2025, Broadcom generated roughly $20 billion of AI-related revenue out of a total company revenue run rate below $64 billion – meaning AI has already grown from a sliver of the business in 2023 to roughly a third today. Outside of Meta, Broadcom counts Google (TPU), ByteDance and now Anthropic as anchor custom-silicon customers, and the Anthropic order book alone is pegged at approximately $21 billion this year.
The Meta extension effectively locks in a multi-year revenue curve. A 1 gigawatt MTIA deployment, using reasonable assumptions on chip, package, substrate and networking content per watt, implies a chip ASIC bill-of-materials in the $10â15 billion range over the first phase, with additional revenue from switching silicon (Tomahawk, Jericho) and co-packaged optics as the fleet scales. Each successive chip generation lifts ASPs and adds more Broadcom content, particularly if Meta adopts Broadcom’s 51.2 Tbps Ethernet fabric to stitch MTIA pods together instead of Nvidia’s NVLink-based reference design.
Broadcom’s stock market reaction captures the strategic weight of the news. Shares have climbed back to roughly $400 in the days around the announcement, a level that reflects both the Meta headline and a wave of related custom-silicon expansions with Google and Anthropic. On a trailing 12-month basis, AI revenue is now the single biggest swing variable in Broadcom’s valuation, eclipsing the slower-growing VMware and legacy networking businesses that made up the bulk of its identity a few years ago.
MetaâBroadcom Deal at a Glance: Key Numbers
| Metric | Figure | Context |
|---|---|---|
| Partnership end date | 2029 | Multi-generation MTIA co-design extension |
| Initial compute commitment | 1+ gigawatt | Equivalent to powering 750,000 U.S. homes |
| Process node | 2 nanometer | First 2nm custom AI silicon in industry |
| MTIA generations shipped | 4 | Starting with MTIA 300 for ranking/recsys |
| Additional generations planned by 2027 | 3 | Roughly annual refresh cadence |
| Meta 2026 capex | Up to $135 billion | Vs. ~$45B in 2024 |
| Meta AMD commitment | ~6 gigawatts | Parallel GPU deployment |
| Broadcom AI revenue FY2025 | ~$20 billion | Out of <$64B total revenue |
| Broadcom FY2027 AI target | ~$100 billion | Custom silicon + networking |
| Broadcom Anthropic order book (2026) | ~$21 billion | Separate custom chip pipeline |
| Broadcom share price around news | ~$400 | Back to prior high-water mark |
Why 2nm Matters: The Process Node Arms Race
TSMC’s 2nm (N2) node is the industry’s first use of gate-all-around (GAA) nanosheet transistors at volume, replacing the FinFET architecture that has defined leading-edge logic since 16nm. Compared with 3nm, TSMC has publicly targeted roughly 10â15% performance gains at iso-power or up to 25â30% power reduction at iso-performance, plus a material density uplift. For AI accelerators – which are bottlenecked on memory bandwidth, power and thermals – every percentage point of efficiency compounds across a gigawatt deployment.
By anchoring its next MTIA generation on 2nm before Nvidia’s Rubin successor reaches the node in volume, Meta and Broadcom are effectively buying a one-generation leapfrog in transistor-level efficiency. In an inference-heavy workload profile, where models like Llama 4 and the ranking stacks behind Reels run continuously across billions of user sessions, total cost of ownership is dominated by power and real estate rather than chip price. Every 20% reduction in watts per token delivered can translate into hundreds of millions of dollars in annual opex savings at Meta’s scale.
Broadcom brings the physical design, SerDes, HBM controllers, chiplet packaging and co-packaged optics IP. Meta brings the architecture – systolic arrays tuned to its recommendation and transformer workloads, the PyTorch-based MTIA compiler stack, and the model-co-design feedback loop where architecture changes in MTIA drive software choices in Llama and vice versa. That tight co-design is the core moat that Nvidia cannot replicate on generic silicon.
Custom Silicon Showdown: MTIA vs TPU vs Trainium vs Maia
Meta is not alone in this move. Every hyperscaler now has a custom AI accelerator program in production, each with its own strategic rationale, partner ecosystem and exposure to Broadcom. The comparison matters because it reveals whether Meta’s gigawatt-scale MTIA bet is a defensive catch-up or an offensive leap.
| Hyperscaler | Custom Chip | Design Partner | Process | Primary Workload | Public Deployment Signal |
|---|---|---|---|---|---|
| Meta | MTIA (Gen 5 on 2nm) | Broadcom | TSMC 2nm | Ranking, recsys, Llama inference | 1+ GW initial phase through 2029 |
| TPU v7 / Ironwood-class | Broadcom | TSMC 3nm | Gemini training + inference | Multi-GW fleets across GCP regions | |
| Amazon (AWS) | Trainium2 / Trainium3 | Marvell / Alchip | TSMC 3nm | Anthropic Claude training, AWS Bedrock | Project Rainier multi-GW cluster |
| Microsoft | Maia 200 | GUC / internal | TSMC 3nm | Azure OpenAI inference | Azure-only; smaller scale than peers |
| OpenAI | Titan (forthcoming) | Broadcom + Samsung HBM4 | TBD (advanced node) | Frontier training + inference | Multi-year ramp starting 2026â2027 |
The pattern is unmistakable: Broadcom now touches three of the five frontier custom-silicon programs (Meta, Google, OpenAI), Marvell anchors the fourth (Amazon Trainium), and Microsoft is the one holdout still leaning heavily on internal ASIC design and external packaging partners for Maia. That concentration of design IP inside Broadcom is the structural reason Wall Street is willing to price in a $100 billion-plus AI revenue outlook.
Wall Street Reaction and Analyst Commentary
The analyst community has broadly interpreted the extension as a net positive for Broadcom, a neutral-to-positive for Meta, and a modest incremental negative for Nvidia at the margin. Vivek Arya of Bank of America has repeatedly flagged Broadcom’s custom silicon book as the single most under-appreciated driver of 2026â2027 semiconductor earnings. Harlan Sur of JPMorgan has argued that the Broadcom AI narrative is transitioning from “hope” to “confirmed visibility,” with Meta’s multi-year commitment serving as the anchor.
On the Meta side, Mark Shmulik of Bernstein has framed the MTIA ramp as essential margin defense: without custom silicon, Meta’s reported AI compute unit costs would be materially higher, compressing operating margins in a year when capex is already absorbing an unprecedented share of free cash flow. Meta CFO Susan Li has publicly committed to “driving efficiency across the full stack” and described MTIA as a “critical lever” for long-run unit economics – consistent with the company’s guidance that blended cost per inference should decline through 2027 even as the workload footprint expands.
The Nvidia read-through is more nuanced. Stacy Rasgon of Bernstein has argued that custom silicon does not cap Nvidia’s total addressable market so much as it caps the rate at which Nvidia can absorb 100% of hyperscaler AI wallet share. In other words, Nvidia’s 2027 revenue can still grow sharply in absolute terms while its share of aggregate hyperscaler AI spend slides from roughly 80% today toward something closer to 60%. Broadcom and AMD divide up most of the ceded share, with Marvell, Alchip and a long tail of specialty players taking the rest.
Meta’s Capex Math: $135 Billion and the Power Problem
Meta’s 2026 capex guidance of up to $135 billion is the number that makes the MTIA deal financially legible. Roughly 70% of that capex is now AI infrastructure: servers, networking, data center construction and the substation-class power delivery required to land gigawatt-scale campuses. The MTIA 2nm tranche and the parallel 6 gigawatts of AMD GPUs are the largest known line items, with Nvidia commitments accounting for the bulk of the remainder.
Power is now the binding constraint. Meta’s Louisiana and Ohio campuses have run into the same grid interconnection timelines that have forced half of U.S. AI data centers to face delay or cancellation in the Department of Energy’s recent capacity review. Meta has responded with direct deals for gas turbines, long-term power purchase agreements with nuclear operators, and on-site battery storage. A gigawatt-class MTIA fleet will draw roughly the same electrical load as a mid-sized U.S. city, and every watt of efficiency improvement delivered by 2nm is watts Meta does not have to source, stand up and pay for.
That is also why Broadcom’s co-packaged optics roadmap is strategically relevant. Traditional pluggable optics burn 10â20% of a rack’s power envelope just moving data between trays. If Broadcom can co-package 1.6T optics alongside MTIA pods by 2027, the fleet-wide power savings scale linearly with the deployment, offering tens of megawatts of reclaimed capacity that can instead be spent on more compute.
Historical Context: From 2023 MTIA v1 to 2029 Multi-Gigawatt
Meta first unveiled the MTIA program publicly in May 2023, positioning it as an inference-only accelerator for its deep learning recommendation models (DLRM). That first-generation chip was a far cry from today’s roadmap: it was fabricated on TSMC’s 7nm node, pulled roughly 25 watts per chip, and was deployed in tightly scoped production roles where the DLRM workload shape was well understood. MTIA v1 made it possible to offload a meaningful share of recsys inference from Nvidia A100/H100 GPUs, but it did not displace them as the center of gravity of Meta’s AI stack.
The 2024 MTIA v2 generation – sometimes branded externally as MTIA 300 for ranking and recommendation systems – lifted performance by roughly 3x per watt versus v1 and brought Meta’s first credible story around serving ads models at scale on custom silicon. MTIA v3, which entered production in 2025, widened the workload envelope to include some training tasks for targeted model components. The new 2nm generation announced this week represents a fourth step: a general-purpose AI accelerator that can carry both heavy inference and significant chunks of training, explicitly co-designed with Broadcom’s networking and packaging stack.
In short, the MTIA story has evolved from a narrow inference ASIC in 2023 to a multi-gigawatt, multi-generation platform that underwrites a $135 billion capex year in 2026. That trajectory is the closest analog yet to what Google did with the TPU between 2015 and 2024, compressed into a shorter timeline and amplified by an AI workload footprint that Meta did not even have when MTIA v1 taped out.
Nvidia’s Read-Through: Still Dominant, But No Longer Monopolistic
Nvidia remains the dominant AI compute provider, and nothing in the MetaâBroadcom agreement changes that for 2026. Meta’s publicly disclosed plans call for “millions” of Nvidia chips across its fleet, with Blackwell and the newly announced Vera Rubin platform carrying the bulk of frontier training. For the most performance-sensitive workloads, where every percentage point of throughput matters more than cost per token, Nvidia’s CUDA and NVLink advantage is still structural.
The longer-term picture, however, is clearly shifting toward a multi-vendor architecture. Nvidia’s own management has been blunt about this on recent earnings calls, arguing that the AI market is growing fast enough that even with rising custom-silicon competition, Nvidia’s absolute dollar revenue can continue to compound at high rates. Jensen Huang has repeatedly made the case that “every AI workload is a Nvidia workload eventually,” pointing to the company’s expanding software footprint as the real moat. But the Meta extension, combined with the Google TPU rollout, the Amazon Trainium ramp and the OpenAI Titan program, collectively defines a ceiling on how much of aggregate hyperscaler AI wallet Nvidia can realistically command in 2027 and beyond.
There is also a competitive implication for AMD. Meta’s 6 GW AMD commitment remains a crucial counterweight and a reminder that custom silicon is not a winner-take-all substitute for a strong merchant GPU. AMD’s Instinct roadmap, anchored by the MI400 series, will occupy the “second source” training and general-purpose inference role that neither Nvidia nor MTIA naturally serves. The three-vendor architecture – Nvidia for frontier training, AMD for flexible general-purpose compute, Broadcom-built MTIA for inference-heavy and recsys workloads – is now Meta’s explicit blueprint.
Market Impact: Hyperscaler AI Spend Hits a Structural Inflection
The MetaâBroadcom agreement arrives at a moment when total hyperscaler AI capex is tracking toward roughly $650 billion in cumulative spend for 2025â2026, with continued growth expected through 2027. Custom silicon is quickly moving from a single-digit share of that spend toward something in the 20â30% range as MTIA, TPU v7, Trainium3 and Maia 200 deployments all ramp in parallel. For Broadcom, Marvell, Alchip and the broader ecosystem of IP and packaging partners, this is the structural inflection that converts AI capex into sustained, visible, multi-year revenue.
Three downstream markets are especially exposed. First, advanced packaging – CoWoS, SoIC, InFO and their successors at TSMC, Samsung and Intel Foundry – is now the binding supply constraint for both Nvidia and custom accelerators. Meta’s 2nm MTIA will consume a meaningful slice of TSMC’s advanced packaging capacity starting in 2026. Second, high-bandwidth memory. HBM3E and HBM4 demand is being pulled forward aggressively as every new AI accelerator ships with more stacks and wider interfaces; SK Hynix, Samsung and Micron are the direct beneficiaries. Third, power infrastructure. Every incremental gigawatt of AI compute drags roughly $1â2 billion of substation, transformer, switchgear and on-site generation spend with it.
The investor takeaway is that AI is no longer a chip story. It is an end-to-end infrastructure story, and the MetaâBroadcom deal is the clearest articulation yet that the highest-margin layer is shifting from general-purpose GPUs toward vertically integrated custom silicon plus the networking, optics and packaging that surround it.
Five Predictions for the MTIAâBroadcom Partnership Through 2029
1. A second gigawatt phase is announced before the end of 2026. Given Broadcom’s “first phase of a sustained, multi-gigawatt rollout” language, a follow-on tranche pegging out through 2027â2028 is almost certainly already in planning. Expect a 2â3 gigawatt follow-on commitment by late 2026, aligning with the ramp of MTIA’s 2nm generation.
2. MTIA takes on frontier training for the Llama 5 generation. To date, MTIA has been pitched as inference-plus-selective-training. By 2027, we expect Meta to run a meaningful share of its frontier Llama 5 pretraining on MTIA pods, using Broadcom’s scale-up and scale-out Ethernet fabric instead of NVLink. That will be the symbolic moment when a hyperscaler credibly runs a frontier model on custom silicon end-to-end.
3. Broadcom crosses $50 billion in AI revenue in fiscal 2027. With Meta, Google, Anthropic and OpenAI (Titan) all in the pipeline, and an approximate AI run rate of $20 billion in FY2025, Broadcom’s AI revenue compounding at 50â60% annually puts the business above $50 billion within two years – a faster ramp than Nvidia’s data center segment achieved in its initial AI phase.
4. Nvidia’s share of hyperscaler AI spend falls from ~80% to ~60% by 2028. Nvidia will continue to grow in absolute dollars, but the combined MTIA, TPU, Trainium and Maia footprints will erode its share of hyperscaler AI wallet. The rebalance is gradual but directionally certain, and it is already visible in 2026 capex disclosures from Meta, Alphabet, Amazon and Microsoft.
5. At least one more hyperscaler signs a multi-gigawatt Broadcom deal by 2027. With Microsoft currently reliant on GUC and internal design for Maia, a Broadcom-backed custom silicon program for Azure is a logical next move. Given Broadcom’s share of Google, Meta and OpenAI design IP, the commercial logic for Microsoft to bring Broadcom in on a bigger share of Maia’s successor is increasingly hard to ignore.
Risk Factors: What Could Derail the MTIA Ramp
The biggest risk is not demand – it is supply. TSMC’s 2nm capacity is spoken for years in advance, and Meta is competing for wafer allocation with Apple, Nvidia, AMD, Google, Qualcomm and other 2nm customers. Any slippage in TSMC’s Arizona or Hsinchu 2nm ramp, any yield issues in the early quarters of production, or any macro shock that forces TSMC to reprioritize capex will hit the MTIA timeline before it hits Nvidia or Apple. Meta and Broadcom have reportedly locked in early wafer allocations, but the dependency chain is real.
The second risk is power. U.S. data center buildout has already hit grid interconnection bottlenecks severe enough that roughly half of proposed AI data center capacity has been delayed or canceled. Even with Meta’s direct investments in power infrastructure, landing a gigawatt of MTIA and parallel AMD/Nvidia compute on schedule will require regulatory approvals, transmission line upgrades and, in some cases, bespoke generation deals that can slip by 12â24 months.
The third risk is workload evolution. If frontier models shift in architectural shape – for example, toward sparse mixtures of experts with sharply different memory bandwidth profiles, or toward world-model architectures of the kind Yann LeCun has been advocating – the MTIA design bets locked in for 2026â2027 could prove less optimal for the workloads that actually dominate in 2028â2029. Broadcom and Meta have hedged against this by planning three additional chip generations by 2027, but the risk of architectural obsolescence is non-trivial in a market this volatile.
What It Means for Developers and Enterprise Buyers
For developers and enterprise buyers, the MetaâBroadcom partnership is mostly a background story – MTIA does not ship as a merchant product, and Meta’s consumer apps will not expose the underlying silicon. But there are two second-order implications worth tracking. First, MTIA enables Meta to offer more generous free or low-cost AI-powered product features (Meta AI, Llama-based assistants, recsys quality) because its unit economics improve. That raises the bar for competitors like ByteDance, Snap and X that cannot match Meta’s custom silicon investments.
Second, the Llama open-weights family is going to benefit from workload-aware co-design. Llama 4 and future releases are already being tuned with MTIA characteristics in mind; that tuning tends to produce models that also run efficiently on merchant GPUs, which is good news for every downstream developer running Llama on AMD, Nvidia, Groq, SambaNova or their laptop. Broadly, hyperscaler custom silicon plus open-weights models is becoming a distinct alternative to the proprietary frontier stack from OpenAI and Anthropic.
For enterprise cloud buyers, the shift to custom silicon matters most through pricing. As Meta, Google and Amazon absorb more of their own compute costs into custom ASICs, the headline prices for their AI APIs should decouple from Nvidia GPU pricing over time. Expect per-token pricing for leading open and proprietary models to decline materially through 2027, driven as much by ASIC unit economics as by model distillation and efficiency gains.
Frequently Asked Questions
What is the MetaâBroadcom AI chip deal?
It is an extension of Meta’s custom AI chip partnership with Broadcom through 2029, announced on April 15, 2026, starting with more than 1 gigawatt of Meta Training and Inference Accelerator (MTIA) compute capacity fabricated on TSMC’s 2nm process. The agreement covers multiple chip generations and positions Broadcom as Meta’s primary custom silicon partner.
How much is Meta spending on AI in 2026?
Meta has guided to up to $135 billion in capital expenditure for 2026, with the majority earmarked for AI infrastructure: servers, networking, data center construction and power delivery. The MTIA gigawatt tranche, around 6 gigawatts of AMD GPUs and millions of Nvidia chips are the largest known line items.
What is MTIA?
MTIA stands for Meta Training and Inference Accelerator. It is a family of custom AI chips Meta has designed in partnership with Broadcom. The first generation shipped in 2023 on a 7nm process for recommendation inference. The latest generation, announced with the April 15, 2026 extension, is a 2nm design capable of both inference and training.
Does the MetaâBroadcom deal hurt Nvidia?
It does not hurt Nvidia in 2026; Meta has publicly committed to buying millions of Nvidia chips for frontier training. Over the longer term, MTIA – combined with TPU, Trainium and Maia – will cap the share of hyperscaler AI wallet Nvidia can capture, even as Nvidia’s absolute revenue continues to grow.
How does MTIA compare with Google’s TPU?
Both MTIA and TPU are hyperscaler custom AI accelerators co-designed with Broadcom, but Google’s TPU program is older (starting with TPU v1 in 2015) and has shipped many more generations. MTIA is newer but scaling rapidly, with a 2nm node, multi-gigawatt deployment and a roadmap of three additional chip generations by 2027.
Why is the 2nm process node important?
TSMC’s 2nm is the first volume node to use gate-all-around nanosheet transistors, replacing FinFETs. Compared with 3nm, 2nm offers approximately 10â15% better performance at the same power or up to 25â30% power reduction at the same performance. For a gigawatt-scale AI deployment, those efficiency gains translate into hundreds of millions of dollars in annual opex savings.
How will the deal affect Broadcom’s revenue?
Broadcom generated roughly $20 billion in AI revenue in fiscal 2025 and has guided to approximately $100 billion in custom AI chip revenue by fiscal 2027. Meta’s multi-gigawatt MTIA commitment, alongside similar pipelines with Google, Anthropic and OpenAI, is a core pillar of that target.
What role will Hock Tan play going forward?
Broadcom CEO Hock Tan will step down from Meta’s board and move into an advisory role focused on custom chip strategy. That structure avoids any conflicts as Meta becomes an even larger Broadcom customer, while keeping Tan closely engaged with Meta on roadmap decisions.
When will the new MTIA chips be deployed?
The initial gigawatt-plus deployment phase begins in the 2026â2027 window, with additional chip generations planned at roughly an annual cadence through 2027 and beyond. The extended partnership covers commitments through the end of 2029.
What are the biggest risks to the MTIAâBroadcom plan?
The three main risks are TSMC 2nm wafer availability and yield, grid-scale power interconnection delays for Meta’s new data center campuses, and the possibility that AI workload architectures shift faster than the MTIA roadmap can adapt. All three are being actively managed but cannot be fully eliminated.
External references: The Next Web, Meta Newsroom, CNBC Broadcom coverage, Tom’s Hardware AI coverage, Reuters Technology.
