The important development is not that Meta may have one more AI chip. It is that large AI operators are building portfolios of purpose-built silicon. That shifts opportunity—and risk—into the less visible parts of the supply chain: foundry capacity, advanced packaging, high-bandwidth memory, testing, cooling and the local engineering needed to keep complex systems running.
Reports in 2026 described Meta’s next custom AI accelerator under the codename “Iris”, with Broadcom as a design partner and TSMC as the intended manufacturer. Meta has not used the Iris name in its public announcements, so it should be treated as a reported codename rather than a confirmed product specification.
What Meta has confirmed is substantial. The company says it has deployed hundreds of thousands of its Meta Training and Inference Accelerator (MTIA) chips for inference workloads, is developing and deploying four new MTIA generations within two years, and has expanded its partnership with Broadcom to co-develop multiple future generations. This is a useful case study for European and Taiwanese suppliers: custom AI silicon is no longer an experiment at the edge of the market. It is becoming part of the infrastructure plans of the largest technology companies.
First, separate the confirmed story from the reported one
The public facts are clear enough to explain the strategic direction without overstating what is known about a single chip.
Meta describes MTIA as a family of custom chips designed for its own AI workloads. Its current deployment supports inference for ranking, recommendations and advertising. Meta says MTIA 300 is already in production for ranking and recommendation training; later generations—MTIA 400, 450 and 500—are intended to support a wider set of workloads, with generative-AI inference the near-term priority.
In April 2026, Meta and Broadcom announced an expanded partnership to co-develop several generations of next-generation MTIA chips. Meta also continues to source hardware from a range of partners, including Nvidia, AMD, AWS and Arm.
The reported Iris production plan fits this wider strategy. It does not mean Meta is replacing all external accelerators with one in-house chip. It means Meta is trying to match more of its high-volume, predictable workloads to hardware designed around its own models, software and data-centre systems.
For suppliers, that distinction matters. A shift to custom silicon can reduce demand for some general-purpose capacity per workload, while increasing demand for specialised design services, manufacturing, packaging, memory, validation and systems integration.
Why hyperscalers want custom chips alongside GPUs
General-purpose GPUs remain essential for many AI workloads, especially frontier-model training and rapidly changing research. They benefit from mature software ecosystems and can be redeployed across many tasks.
But a hyperscaler also runs enormous volumes of repeatable work. Recommendation systems, advertising ranking and many inference tasks have known models, known data flows and predictable operating constraints. A custom accelerator can be designed around those constraints.
The potential advantages are practical:
- better performance per watt for a defined workload,
- lower cost per inference once deployment reaches sufficient scale,
- tighter integration with networking, memory and server design,
- less reliance on a single external hardware roadmap,
- faster iteration when hardware and internal software teams work together.
This is why the industry should not frame custom silicon as a simple “GPU killer” narrative. Google has long used TPUs; Amazon has developed Trainium and Inferentia; Microsoft, Meta and other large operators have pursued their own designs. The likely outcome is a mixed infrastructure: flexible GPUs for some workloads, custom accelerators for others, and CPUs, networking and storage designed around the whole system.
A chip design is only the beginning
The headline tends to focus on who designed the processor. In practice, an AI accelerator becomes a deployable product only after it passes through a tightly connected chain of capabilities.
Foundry capacity
If a reported Meta programme is manufactured by TSMC, it reinforces a broader reality: custom-chip customers still depend on a small group of manufacturers able to produce leading-edge logic at scale. Developing an ASIC may diversify a company away from Nvidia’s product roadmap; it does not remove dependence on advanced foundry capacity.
This is why production timing should be read cautiously. A design may be ready, but volume availability still depends on wafer allocation, yield, process maturity and the priorities of a crowded customer base.
Advanced packaging and high-bandwidth memory
Modern AI accelerators are not only logic dies. Their performance depends on how closely compute, memory and interconnects are brought together. High-bandwidth memory (HBM), substrates, interposers, bonding, assembly and test can be as consequential as transistor density.
The bottleneck can therefore move. More custom accelerator designs do not automatically make AI capacity easier to obtain if advanced packaging or HBM supply remains constrained. For suppliers, that creates opportunities in materials, equipment, inspection, metrology, yield management and reliability services.
Power and thermal management
Higher compute density creates a thermal and electrical problem at the same time. The relevant market includes thermal-interface materials, cold plates, liquid-cooling components, pumps, manifolds, sensors, power delivery, monitoring and qualification under load.
This is an important opening for industrial companies that do not describe themselves as “AI chip companies”. Precision machining, fluid systems, specialty materials, industrial automation and test engineering all become more valuable when the performance target is a stable, serviceable rack rather than an isolated processor.
Testing, validation and operations
Every new accelerator also requires validation: package test, board test, system bring-up, firmware, software optimisation, reliability work and field diagnostics. A hyperscaler can build much of this capability internally, but it still relies on a wide ecosystem of equipment and specialist suppliers.
The commercial opportunity is not limited to a factory gate. It can include spare parts, application engineering, failure analysis, documentation, local support and lifecycle management.
What European suppliers should take from the Meta case
European suppliers should not wait for a direct purchase order from Meta, Broadcom or TSMC before assessing the trend. The more realistic question is where their capability fits within the expanding custom-silicon supply chain.
Potentially relevant supplier categories include:
- semiconductor materials, chemicals and specialty gases,
- substrates, interconnects and packaging materials,
- inspection, metrology and test equipment,
- reliability, failure-analysis and qualification services,
- thermal-management and liquid-cooling systems,
- power electronics, connectors, cables and sensors,
- cleanroom, contamination-control and industrial-automation services,
- logistics for sensitive, high-value or regulated components,
- field service, spare-parts management and application engineering.
The opportunity is strongest where a company has a specific technical advantage. “We serve AI” is not a market position. “We can qualify and service this specialised subsystem for European customers within an agreed response time” is closer to one.
What this can mean for Taiwanese suppliers entering Central Europe
Taiwanese suppliers already understand the technical intensity of the semiconductor ecosystem. Their challenge in Europe is usually not proving that they can manufacture. It is adapting the operating model to customers who may expect local service, European documentation, shorter response times, audited quality systems and predictable logistics.
Czechia can be relevant when the required European function is clear. Brno can support R&D, chip design, application engineering and university cooperation. Prague can support commercial coordination, institutions and international management. Northern Czechia can be useful for industrial logistics and access to the semiconductor cluster around Dresden.
None of these locations is automatically right for an AI-supply-chain project. The decision should start with the operating need:
| If the European operation must do this | First capability to verify |
|---|---|
| Provide field service | locally available engineers, spare-parts stock and response-time commitments |
| Support advanced materials or chemicals | permitted facilities, safe handling, transport and waste-management processes |
| Build application engineering | specialist recruitment, laboratory access and customer proximity |
| Supply cooling or precision components | qualification requirements, production capacity and cross-border logistics |
| Develop partnerships | named target customers, local decision-makers and a credible commercial owner |
This avoids a frequent mistake: opening a representative office because Europe is strategically attractive, before defining what the office must deliver. A local entity without technical support, qualification evidence or a customer plan may create meetings, but not a durable supply-chain position.
A practical checklist before investing
For a supplier looking to benefit from the custom-AI-silicon cycle, the first work should be specific and evidence-led:
- Identify the exact product or service that becomes more valuable as AI systems scale.
- Map the customer tier: chip designer, foundry, packaging house, server maker, integrator or data-centre operator.
- Check qualification, traceability, cybersecurity and export-control requirements before quoting business.
- Test whether a European customer needs local inventory, engineers or service-level commitments.
- Compare Czechia with neighbouring locations based on the function, not an incentive headline.
- Validate infrastructure: electricity, water, cleanroom standards, chemical handling, transport and expansion capacity.
- Start with a measurable commercial or technical pilot, then expand once the demand and operating assumptions are proven.
The main conclusion
Meta’s MTIA roadmap—and the reported Iris programme within it—matters because it illustrates a structural change in AI infrastructure. Large technology companies are not choosing between custom chips and GPUs. They are assembling portfolios of silicon tailored to different workloads.
That broadens the market beyond the most visible chip designers. It increases the importance of the companies that can manufacture, package, cool, test, transport, qualify and support AI hardware at scale.
For European and Taiwanese suppliers, the useful response is not to make broad claims about the AI boom. It is to identify one operational problem created by denser, more specialised AI systems, prove the capability to solve it, and build the European service and qualification model customers require.
How Kodo can help
Kodo helps international technology and industrial companies prepare the commercial and operational side of entering Czechia. We support early market validation, partner mapping, local positioning, English and Czech communication, web content and coordination of the first implementation steps.
For legal, tax, regulatory, certification and investment-incentive matters, companies should work with qualified specialists and the responsible authorities.
Related reading
- From Taiwan to Czechia: What Semiconductor Companies Should Check Before Choosing a Location
- Onsemi’s Delayed Rožnov Expansion: What It Signals for Semiconductor Investors in Czechia
- Onsemi’s Rožnov Investment Faces a Deeper Reset: What Foreign Suppliers Should Watch Now
- How Foreign-Owned Manufacturers in Czechia Can Fund ASME and Quality Certification
Sources
- Meta: Expanding Meta’s Custom Silicon to Power Our AI Workloads
- Meta: Partnership with Broadcom to Co-Develop Custom AI Silicon
- Meta: What Is Compute Power? Meta’s AI Infrastructure Explained
- Radio Taiwan International: TSMC to Manufacture Meta’s New AI Chip
- Meta: Introducing Our Next Generation Infrastructure for AI
This article provides general market and supply-chain information, not legal, tax, regulatory or investment advice. Companies should verify requirements for their specific products, customers and operating model.
