By end of 2027, custom and open-architecture AI accelerators will take at least 25% of new data-center AI deployments

Prediction Statement

By December 31, 2027, custom and open-architecture AI accelerators — hyperscaler in-house ASICs, frontier-lab custom inference chips such as OpenAI's Jalapeño, and open-instruction-set parts (for example RISC-V designs) — will account for at least 25% of new data-center AI accelerator deployments by unit volume, up from the mid-single-digit to low-teens share they hold in mid-2026. The merchant GPU incumbent will remain the largest single supplier, but its share of new deployments will fall below 75%.

Reasoning and Analysis

The economic driver is that inference, not training, is now the dominant recurring cost of running AI products, and the supplier margin on every merchant accelerator is a permanent tax on that cost. Above a high volume threshold, designing custom inference silicon converts that margin into captured savings — which is exactly why the largest serving fleets are moving first.

The week of June 24-26, 2026 made the direction concrete: OpenAI and Broadcom unveiled the Jalapeño inference ASIC with deployment targeted for end of 2026; Qualcomm announced its Dragonfly data-center roadmap with Meta as anchor customer, acquired the heterogeneous-execution software company Modular, and entered reported talks to buy RISC-V startup Tenstorrent. Hyperscalers already deploy multiple generations of in-house accelerators. The combination of strong unit economics, multiple credible suppliers, and a software-portability push aimed at the CUDA moat is what pushes a niche into a quarter of new deployments.

Confidence Factors

Raises confidence: inference is now the largest recurring AI cost; multiple hyperscalers already ship in-house silicon at volume; frontier labs have joined the trend with credible design partners; a well-funded attack on the software moat is underway.

Lowers confidence: custom-silicon programs routinely slip; the incumbent's software ecosystem and new-generation cadence are formidable; "new deployment" share is measured imperfectly across vendors that disclose little; a model-architecture shift could favor flexible GPUs over specialized ASICs and slow the transition.

Key Indicators to Watch

• Jalapeño reaching production deployment on schedule by end of 2026.
• Additional frontier labs announcing custom inference silicon programs through 2027.
• Production-grade multi-accelerator support shipping in a portability layer such as Modular's stack.
• Hyperscaler disclosures (or supply-chain estimates) showing in-house accelerators as a rising share of new capacity.
• Any reported Tenstorrent acquisition closing and producing data-center RISC-V parts.

Validation Criteria

Counted correct if credible 2027 industry estimates (analyst supply-chain reports or vendor disclosures) place custom plus open-architecture AI accelerators at 25% or more of new data-center AI accelerator deployments by unit volume, with the merchant GPU incumbent below 75% of new deployments. Counted incorrect if the incumbent holds 75% or more of new deployments through 2027. Where unit-volume data is unavailable, deployed-capacity (token-serving) estimates will substitute, using the same thresholds.