Intel, Arm, and Nvidia propose a new open, license-free specification using 8-bit floating point, or FP8, precision as a common interchange format for AI
Context & Ripple Effects
This proposal extends a familiar playbook: Intel has repeatedly co-authored open standards with its own competitors when the prize is an interchange layer rather than the silicon itself. The consortium behind CXL, the interconnect standard Intel debuted with Google and Microsoft followed the same logic, as did Microsoft and Facebook's Open Neural Network Exchange format for moving models between frameworks.
FP8 is the precision-level version of that move: by making 8-bit floating point a license-free common format, Intel, Arm, and Nvidia are agreeing on the plumbing so competition can stay focused on chips and software. The Linux Foundation's later push for composable, multi-provider AI systems via the Open Platform for Enterprise AI shows where this interoperability instinct has been heading.
First-order effects
- AI developers gain a single license-free target for 8-bit precision exchange, removing the need to maintain separate FP8 implementations per vendor's hardware.
- Intel, Arm, and Nvidia each trade a slice of format control for broader ecosystem adoption of their accelerators, since a shared spec lowers the cost of running mixed-vendor fleets.
Second-order effects
- Other accelerator vendors and framework maintainers face pressure to adopt the spec or risk incompatibility with the three companies that dominate training and inference silicon.
- Lower-precision interchange makes heterogeneous deployments — pairing CPUs with AI accelerators, as in the OpenAI talks with Arm about a CPU alongside Nvidia and AMD AI chips — cheaper to assemble, strengthening demand for exactly the mixed compute these three sell.
Third-order effects
- If the pattern from ONNX through CXL to FP8 holds, the industry's pre-competitive layers keep thickening, pushing differentiation further up the stack toward software, systems integration, and services.
- License-free interchange standards reduce the lock-in value of proprietary formats, shifting leverage toward whoever controls the fastest implementation of the shared spec rather than the spec itself.
The trend: Rival chipmakers are increasingly cooperating on open, license-free interchange standards for AI while reserving their competition for the silicon and software above them.