Custom AI ASICs refer to application-specific integrated circuits purpose-built for AI workloads rather than general computation. In 2026, they are reshaping the AI server market at a pace that would have seemed implausible two years ago. Projections suggest custom AI ASICs will account for 27.8% of AI server shipments this year — the highest share since 2023 — representing 44.6% year-over-year growth. That’s not a footnote in a quarterly report. That’s a structural shift.
Key Takeaways
- Custom AI ASICs are projected to reach 27.8% of AI server shipments in 2026, the highest share since 2023.
- Meta has announced four MTIA chip generations — 300, 400, 450, and 500 — with deployments spanning 2026 and 2027.
- Broadcom confirmed a $10 billion custom AI processor order, widely believed to be from OpenAI, with delivery targeted for Q3 2026.
- The industry is splitting workloads: custom silicon handles high-volume inference while GPUs retain dominance in training.
- Google, AWS, and Microsoft have each made comparable custom-silicon commitments on their own timelines.
Why custom AI ASICs are surging in 2026
The core logic is simple: inference is a volume game. Once a model is trained, running it billions of times daily is a predictable, repeatable task — exactly the kind of workload where purpose-built silicon crushes general-purpose GPUs on efficiency. Custom AI ASICs don’t need to handle every conceivable computation. They need to handle one job extremely well, at massive scale, for as cheaply as possible. That’s why the economics are finally clicking into place for hyperscalers.
The workload segmentation story is now well established across the industry. Training remains GPU-heavy — Nvidia isn’t going anywhere — but inference is increasingly the territory of custom silicon. OpenAI, for example, has secured access to roughly $100 billion worth of Nvidia GPU hardware across multiple generations for training purposes. Yet simultaneously, Broadcom confirmed that an undisclosed client plans to procure $10 billion worth of custom AI processors, with delivery targeted for Q3 2026. The industry broadly believes that client is OpenAI, though this has never been formally confirmed. If accurate, that single deal could represent roughly one to two million XPUs depending on unit pricing — a staggering volume of purpose-built compute.
Meta’s MTIA roadmap: four chips, two years
Meta’s approach to custom AI ASICs is the most publicly detailed of any hyperscaler right now. On March 11, Meta announced four successive generations of its Meta Training and Inference Accelerator — the MTIA 300, 400, 450, and 500 — all scheduled for deployment over the next two years. The chips are optimised primarily for AI inference, though training is a secondary use case.
The MTIA 300 is already in production, handling ranking and recommendations training. The MTIA 400 has completed lab testing and features a 72-accelerator scale-up domain; it’s on the path to data center deployment. The MTIA 450 and 500 are scheduled for mass deployment in early 2027 and later in 2027, respectively, designed to cover AI inference production through that year. Meta’s roadmap also points toward a version built on TSMC’s N3 fabrication process with HBM3E memory, signalling that the company is investing in leading-edge packaging alongside chip design.
What makes Meta’s announcement significant isn’t any single chip — it’s the cadence. Four generations in two years, publicly committed, with staggered deployment timelines. That’s the kind of roadmap discipline that historically only Nvidia has demonstrated in AI silicon. Meta is signalling it intends to own its inference stack long-term, not just experiment with it.
Broadcom’s role and the Nvidia comparison
Broadcom sits at an interesting intersection in the custom AI ASIC market: it’s not a hyperscaler building chips for its own data centers, but a chip design partner enabling hyperscalers to build their own. The $10 billion processor agreement — whether for OpenAI or another party — illustrates the scale at which this model now operates. Broadcom secured the agreement in October, with deployments potentially beginning as early as this year.
Compare that to Nvidia’s position. Nvidia’s GPUs remain the default choice for AI training, and OpenAI’s $100 billion GPU commitment underscores that training workloads aren’t migrating to custom silicon anytime soon. But inference is a different story. Custom AI ASICs built through Broadcom partnerships, or developed entirely in-house by Google and Meta, are increasingly handling the high-volume, cost-sensitive inference layer that sits between a trained model and the end user. Nvidia dominates training; custom silicon is eating inference. Those aren’t the same market.
Google, AWS, and Microsoft: the broader hyperscaler push
Meta and the Broadcom/OpenAI story get the headlines, but Google, AWS, and Microsoft have each made equivalent commitments to custom AI silicon on their own timelines. Google’s TPU program is the longest-running hyperscaler ASIC effort in the industry — it predates the current AI boom by nearly a decade. AWS has its Trainium and Inferentia lines. Microsoft has been investing in custom silicon for its Azure infrastructure. The pattern is consistent: every major cloud provider has concluded that relying entirely on third-party GPU vendors is a strategic risk they’re not willing to carry indefinitely.
The 2026–2027 window looks like the period when these parallel investments converge into meaningful deployment scale. Custom AI ASICs reaching 27.8% of AI server shipments isn’t a ceiling — it’s likely a floor, given the commitments already in place across the industry.
Is the GPU era for AI inference ending?
Not ending — segmenting. GPUs remain essential for training large models, where flexibility and raw compute matter more than efficiency. But for inference, the economics increasingly favour custom silicon. Purpose-built ASICs can deliver the same output per query at lower power and cost, which matters enormously when you’re running billions of inferences daily. The question isn’t whether custom AI ASICs will take share from GPUs in inference — that’s already happening. The question is how fast.
What does the Broadcom and OpenAI ASIC deal actually mean?
Broadcom confirmed a $10 billion custom AI processor order with an undisclosed client, with delivery targeted for Q3 2026. The industry widely believes the buyer is OpenAI, though neither party has formally confirmed this. If the deal implies one to two million XPUs, it would represent one of the largest single custom silicon procurements in the industry’s history.
When will Meta’s MTIA chips reach full deployment?
Meta’s MTIA 300 is already in production. The MTIA 400 is completing its path to data center deployment after lab testing. The MTIA 450 and MTIA 500 are both scheduled for mass deployment in 2027, with the 450 arriving in early 2027 and the 500 later that year.
The custom AI ASIC story in 2026 isn’t about one company or one chip — it’s about a coordinated, industry-wide bet that purpose-built silicon is the only economically rational way to run AI inference at hyperscale. Meta’s four-generation roadmap, Broadcom’s massive custom chip programs, and Google’s long-running TPU effort are all expressions of the same conviction. Nvidia will keep the training market. But inference? That’s being rebuilt from silicon up, and 2026 is the year the scale finally becomes undeniable.
Edited by the All Things Geek team.
Source: Tom's Hardware
