Meta Iris : the custom AI chip that joins the hyperscalers club — and Broadcom becomes the second most important name in AI silicon

Meta Iris: the custom AI chip that completes the hyperscalers club — and Broadcom becomes the second most important name in AI silicon

🔎 The club is complete, and the real winner doesn't wear a hyperscaler logo

September 2026 marks a silent but massive turning point for the AI industry. Meta is going to launch the production of its custom chip Iris, designed with Broadcom and manufactured by TSMC. An internal memo seen by Reuters confirms it: manufacturing starts in two months.

What had seemed inevitable for three years is becoming reality. Every major hyperscaler now has its own AI silicon. Google has its TPUs (7th generation), Amazon has Trainium, Microsoft has Maia, OpenAI has Jalapeño, and Meta joins the movement with Iris.

But the most interesting story isn't Iris itself. It's the fact that Broadcom is simultaneously designing for Meta, OpenAI, and Google. This discreet partner is accumulating an unprecedented position of power in the AI silicon value chain.

The key points

• Meta Iris enters production in September 2026, designed with Broadcom, manufactured by TSMC, for high-volume recommendation and inference workloads.
• The hyperscaler club with custom silicon is complete: Google TPU, Amazon Trainium, Microsoft Maia, OpenAI Jalapeño, Meta Iris — all following the same model.
• Broadcom is the hidden designer behind several of these programs (Iris, Jalapeño, TPU), with $8.4 billion in AI revenues in Q1 FY2026 (+106% in one year).
• Meta is doubling its compute capacity to 14 GW by 2027, with a CA$13 billion, 1 GW data center in Alberta, its 33rd global site.
• Meta is considering selling its surplus compute to external customers, which would put it in direct competition with AWS, Azure, and GCP.

Recommended tools

The universal pattern: custom inference, Nvidia training

An identical strategy across all hyperscalers

The reasoning is the same at Meta, Google, Amazon, Microsoft, and OpenAI. Inference workloads are predictable, repetitive, and now represent the bulk of the operational cost of an AI model deployed at scale.

An ASIC chip optimized for a specific type of inference costs significantly less to manufacture and run than a versatile Nvidia H100 or B200 GPU. The return on investment is direct and measurable.

Frontier training, on the other hand, remains on Nvidia GPUs. Models like GPT-5.5 (agentic score of 98.2) or Gemini 3.1 Pro (92 overall) require the flexibility and software ecosystem that only CUDA offers today. No one takes the risk of training a frontier model on custom silicon.

The deterrent negotiating lever

Owning your own chip is not just a matter of cost. It is a massive negotiating lever against Nvidia. Every order for H200 or Rubin GPUs is now the subject of a discussion where the hyperscaler can say: "If your price isn't competitive, we will accelerate the deployment of our internal silicon."

The threat is credible because the chips actually exist. Iris is not a research project — it is a product entering industrial production.

Meta Iris: what we know technically

Designed with Broadcom, manufactured by TSMC

Iris's design follows the same pattern as Jalapeño at OpenAI: Broadcom provides ASIC design expertise, TSMC handles manufacturing. This hyperscaler-Broadcom-TSMC trio has become the industry's standard template.

According to Tom's Hardware's May 2026 report, TSMC now manufactures chips for the five major hyperscalers and for Broadcom, the latter being the "dominant designer" of custom AI ASIC chips.

Targeted workloads, not a general-purpose GPU

Iris is not designed to do everything. Meta primarily intends it for two types of workloads: recommendation and ranking systems (the core of its advertising business) and a portion of the inference for its LLM models.

This is a rational choice. Meta's recommendation systems generate trillions of predictions per day. Each prediction is a relatively simple calculation but at a monstrous volume. An ASIC tailored for this profile crushes a general-purpose GPU in energy efficiency and cost per request.

Complement, not replace

Meta is explicit: Iris complements Nvidia and AMD GPUs, it does not replace them. Meta's compute architecture will remain hybrid. GPUs will continue to handle training and complex inference, while Iris will absorb the predictable volume.

According to HPCwire, Meta plans to introduce a new AI chip approximately every six months until 2027. This pace is radically faster than the traditional cycle of the semiconductor industry.

The Scale of Meta's Compute Expansion

From 7 GW to 14 GW: The Doubling

The internal memo revealed by Reuters and confirmed by TechDogs sets a clear goal: Meta is initially deploying 7 gigawatts of compute power, then doubling to 14 GW by 2027. To put this into perspective, a typical data center consumes between 50 and 200 megawatts. Meta is speaking in gigawatts.

This scale is unprecedented in the history of computing. Even at the peak of the cloud expansion in the 2010s, deployments were measured in tens of megawatts.

Alberta: The 1 GW, CA$13 Billion Data Center

On July 8, 2026, Meta officially announced the construction of its first Canadian data center in Sturgeon County, Alberta. According to Meta's official press release, it is an AI-optimized facility with a capacity of 1 GW, representing an investment of over CA$13 billion.

CNBC reports a cost of approximately US$9 billion, with a construction timeline of 2 to 3 years. This is Meta's 33rd data center in its global fleet.

The choice of Alberta is not coincidental. The province offers affordable energy, a cold climate that reduces cooling costs, and access to the Canadian energy mix, which includes a significant share of renewables. Energy Digital specifies that the site will be liquid-cooled, now the standard for high-density AI facilities.

Why This Raw Scale

The answer lies in the models. Each generation of LLM is significantly more expensive to serve than the previous one. Benchmark scores are climbing (GPT-5.5 reaches 98.2 in agentic, Claude Opus 4.7 Adaptive reaches 94.3), but this quality comes at a compute cost.

Meta serves models to billions of users across Facebook, Instagram, WhatsApp, and Threads. Every interaction with an integrated AI assistant, every AI-enhanced content suggestion, every image generation — all of this consumes compute. Demand is structurally higher than the current supply.

Broadcom : the second most important player in AI silicon

Numbers that speak volumes

Broadcom generated $8.4 billion in AI semiconductor revenue in the first quarter of fiscal year 2026, up 106% year-over-year. According to Hashrate Index, the company guided to $10.7 billion for the second quarter.

CEO Hock Tan stated during an analyst conference that AI chip revenue would exceed $100 billion by 2027, citing "direct visibility into six hyperscaler customers." That is a figure sufficient to understand the magnitude of the phenomenon.

The designer behind the curtain

Broadcom does not manufacture chips. It designs them. And it does so for several clients who compete with each other. According to Tom's Hardware and Introl, Broadcom is involved in the design of Meta Iris, OpenAI Jalapeño, and Google TPUs.

This is a remarkable strategic position. Broadcom accumulates expertise from each program, identifies design patterns that work, and reapplies them. In practice, each new program benefits from the lessons learned on previous ones.

Jon Peddie describes this strategy as a "$100 billion bet on custom silicon," and notes that Broadcom's position is unique: no other player has this cross-sectional view of hyperscalers' AI silicon programs.

The concentration risk

This centralization has a downside. If Broadcom encounters a problem (design delay, IP leak, talent constraint), all programs are impacted simultaneously. It is a single point of failure masked by the apparent diversity of the end brands.

The global custom ASIC market: $660-690 billion in capex in 2026

Hyperscalers are (partially) leaving Nvidia

Oplexa's report on the custom ASIC market in 2026 is unequivocal: Google, Meta, and Amazon are all building their own chips. Hyperscaler capital expenditures (capex) will reach $660 to $690 billion in 2026, with 75% directed toward AI infrastructure.

All this capex is not going to Nvidia. A growing share is funding the design and manufacturing of internal chips. Nvidia remains dominant for frontier training and flexible deployments, but high-volume serving is gradually migrating to custom ASICs.

The economic calculation that justifies everything

Let's take a concrete example. If Meta serves 100 billion recommendation requests per day, and each request costs $0.00001 on a GPU but $0.000005 on a custom ASIC, the savings amount to $500,000 per day — or more than $180 million per year.

It is for workloads of this volume that ASICs make sense. Below a certain threshold, the chip design cost (hundreds of millions in R&D) is never amortized. Above it, it is an economy-generating machine.

Meta, with its scale of billions of users, is precisely in the zone where the ASIC is profitable. Shifting 20 to 30% of serving costs to Iris potentially represents billions in annual savings.

The most aggressive signal: Meta will sell external compute

From consumer to provider

This is the most strategic piece of information in Meta's internal memo. The company is considering selling its surplus AI compute capacity to external customers. According to sources consulted by Build Fast With AI and Prompt Injection in their July 2026 roundups, this direction is being taken seriously by analysts.

If Meta executes this plan, it enters into direct competition with AWS, Azure, and Google Cloud. This is a radical shift in posture. Until now, Meta was the world's largest consumer of AI compute without reselling it. Becoming a provider disrupts market dynamics.

Why Meta can afford it

With 14 GW of planned capacity, Meta will inevitably have a surplus. Internal demand, however massive it may be, will not fill 100% of this capacity 100% of the time. Data centers operate with a typical utilization factor of 60-80%.

Reselling the surplus improves the return on investment of the CA$13 billion spent in Alberta and the tens of billions invested elsewhere. It is pure financial optimization.

The reaction of cloud hyperscalers

AWS, Azure, and GCP will not remain passive. Their economic model relies on the margin between the cost of compute and the selling price. If Meta enters the market with a structurally lower cost (no cloud margin to take, infrastructure optimized primarily for its own use), the pressure on prices will be real.

It is a scenario reminiscent of how Amazon entered the cloud market in 2006: by monetizing a surplus capacity built for its internal needs. History could well repeat itself.

The software ecosystem: the Achilles' heel of all custom chips

CUDA remains the de facto standard

If custom silicon is so economically advantageous, why hasn't everyone switched? The answer is CUDA, Nvidia's software ecosystem. Twenty years of development have created an advantage that no one has caught up with yet.

Models like DeepSeek V4 Pro (Max), which scores 88 on a general benchmark, or Kimi K2.6 (84 in general, 88.1 in agentic in self-host), are optimized for CUDA first. Porting them to custom silicon requires significant engineering effort.

What this means for Iris

Iris will likely be limited to a restricted set of workloads where the software is controlled by Meta. Recommendation systems are a perfect candidate: Meta writes its own software stack end-to-end. For general LLM inference, the reliance on CUDA-compatible frameworks makes the transition more complex.

This is precisely why the "custom inference, Nvidia training" pattern is so widespread. Training requires maximum software flexibility. Inference of specific models on mastered workloads makes it possible to bypass the CUDA dependency.

Comparison: Hyperscaler Custom Silicon in July 2026

The pattern is striking. TSMC manufactures everything. Broadcom designs a growing share. And every chip targets inference first, not training.

What Compute Expansion Means for Developers

More Compute = More Possibilities

For developers working with models like GPT-5.4 (89 overall, 87.6 agentic), Claude Sonnet 4.6 (83 overall, 81.4 agentic) or Gemini 3.1 Pro (92 overall, 87.3 agentic), Meta's compute expansion translates concretely into downward pressure on API costs.

As the supply of compute increases, the price per million tokens tends to drop. This is already visible over the last 18 months: the cost of inference has plummeted by over 90% for equivalent models.

Beware of Silicon Vendor Lock-in

If Meta actually sells external compute based on Iris, developers who adopt it will have to optimize their models for this specific chip. It's a silicon lock-in disguised as low prices. The day Meta changes its infrastructure priorities, workloads optimized for Iris will need to be recompiled or migrated.

It's the same lesson as with the cloud: the low entry price rarely compensates for the exit cost.

The impact on the AI model race

More compute for training too

If Iris frees up Nvidia GPUs currently used in inference, these GPUs can be redeployed for training larger models. This is an indirect but powerful leverage effect.

Meta has also demonstrated its ambition regarding models with recent initiatives. Even if Meta's first closed model marked a strategic turning point, compute capacity determines the ceiling of what is possible in terms of model size and quality.

The scaling race continues

Benchmarks show it: each additional tier of performance requires significantly more compute. The jump from Claude Opus 4.6 (87 overall) to Claude Opus 4.7 Adaptive (90 overall, 94.3 in agentic) did not come free in terms of compute resources.

With 14 GW of capacity, Meta is giving itself the means to stay in the frontier training race while optimizing its operational costs via Iris. It is a two-pronged strategy that only a handful of players in the world can finance.

❌ Common mistakes

Mistake 1: Confusing complement and replacement

Iris does not replace Nvidia GPUs. It complements them on specific workloads. Thinking that Meta will move away from Nvidia is a misread. Meta remains one of the world's largest GPU customers. Iris optimizes the margin on a specific segment of compute, nothing more.

Mistake 2: Underestimating Broadcom's role

The common mistake is to see Iris as a "Meta" chip. In reality, the design is largely driven by Broadcom, which applies expertise accumulated across several programs. Ignoring Broadcom in the equation means ignoring the true center of gravity of custom silicon.

Mistake 3: Projecting compute sales as immediate

Meta is "considering" selling external compute. This is not a product announcement. The operational, regulatory, and strategic hurdles are considerable. Interpreting this signal as an imminent commercial offering is premature.

Mistake 4: Believing that custom ASICs are accessible to everyone

The ASIC model only works at very large scale. A company with fewer than a few billion requests per day will never amortize the design cost of a custom chip. The hyperscaler club is closed economically, not technically.

❓ Frequently Asked Questions

Will Meta Iris replace Nvidia GPUs at Meta?

No. Iris is designed for recommendation workloads and a portion of high-volume inference. Frontier training and complex inference will remain on Nvidia and AMD GPUs. This is a complementarity strategy, not a substitution strategy.

Why is Broadcom so important in this story?

Broadcom simultaneously designs chips for Meta (Iris), OpenAI (Jalapeño), and Google (TPU). With $8.4 billion in AI revenue in Q1 FY2026 and a projection reaching $100 billion by 2027, it is the second most powerful player in AI silicon after Nvidia.

Will the Alberta data center be enough for Meta's needs?

The Alberta site (1 GW) is part of a broader plan aiming for 14 GW of total capacity by 2027. It is Meta's 33rd data center. It contributes to the goal of doubling capacity, but is just one piece of the puzzle.

Is Meta really going to sell compute to external customers?

This is an intention revealed by an internal memo, not an announced product. If executed, this would put Meta in competition with AWS, Azure, and GCP. But the timelines and uncertainty are significant. Keep a close eye on this in 2027.

What is the concrete advantage of an ASIC chip over a GPU for inference?

An ASIC is optimized for a specific type of computation. It eliminates the unnecessary circuits of a general-purpose GPU, reducing the manufacturing cost and energy consumption per request. At the scale of billions of daily predictions, the savings amount to hundreds of millions of dollars per year.

✅ Conclusion

Meta Iris does more than just add to the club of hyperscalers with custom silicon — it confirms that the center of gravity for AI silicon is shifting toward ASICs designed by Broadcom and manufactured by TSMC. With 14 GW of planned capacity, a CA$13 billion data center in Alberta, and the ambition to sell external compute, Meta is no longer just consuming compute: it is preparing its transformation into AI infrastructure. The real signal to take away is not Iris itself, it's that Broadcom is becoming the silent designer behind a growing share of the world's compute.