Introduction

On May 11, 2025, Intel unveiled “Foundry AI Services,” a new business unit that blends advanced packaging, IP licensing, and contract manufacturing for AI‑focused chips. The strategy is a direct counter‑move against NVIDIA’s vertically integrated stack, inviting startups and hyperscalers to co‑design purpose‑built accelerators on Intel process nodes.

“Foundry AI Services is about unlocking co‑innovation—whether you’re building a transformer chip, a sensor‑fusion module, or something we haven’t thought of yet,” said Stuart Pann, Intel EVP of Foundry Services.¹

The program combines Intel’s Open Compute Interface (OCI), Foveros Direct 3‑D stacking, and optional silicon‑photonics tiles, allowing customers to stitch together CPUs, NPUs, HBM4, and custom logic on a single substrate.

Why it matters now

• AI hardware is fragmenting beyond GPUs; custom accelerators handle sparsity, low‑precision, and modality‑specific workloads.
• Startups need advanced packaging and supply‑chain guarantees without surrendering control to a platform vendor.
• Intel must pivot beyond x86 to stay relevant as packaging bandwidth outshines raw transistor counts.

Call‑out: Packaging is the new performance frontier

Intel’s demo system showed three chiplets—CPU, photonics I/O, and NPU—bonded with Foveros Direct, delivering 3.2 TB/s aggregate bandwidth while cutting interconnect power by 47 %.

Business implications

AI hardware startups can skip years of CAPEX by tapping Intel’s toolkit instead of raising nine‑figure funds for custom fabs. Hyperscalers gain leverage against single‑vendor lock‑in, while regulators welcome more diversified supply chains.

Looking ahead

Expect Intel to name its first three Foundry AI partners at Hot Chips 2025 in August. Gartner projects that by 2030, 55 % of AI silicon will be custom accelerators rather than GPUs—making open foundry ecosystems indispensable.

The upshot

 Disruption in AI is shifting from software to silicon supply chains. By opening its toolkit, Intel aims to be the enabling layer for specialized AI hardware—and to keep the ecosystem from collapsing into a single‑vendor silo.

––––––––––––––––––––––––––––

¹ Stuart Pann, press statement, May 11, 2025.