AMD builds its second stack: Meta and Oracle anchor a full alternative to the dominant AI platform

Two anchors in one week

AMD and Meta are deepening a partnership that now spans silicon, systems and software. AMD says AMD Helios rackscale solutions are currently installed and being tested by Meta, and Meta is a lead customer for 6th Gen EPYC CPUs. Meta will also deploy a custom AMD Instinct GPU optimized for its workloads, and the companies are collaborating on future platforms including AMD Instinct MI500 series GPUs and future EPYC processor generations, including Verano. m:…0486

AMD says it will invest $1 billion in Meta's equity offering, a financial tie that aligns the two companies beyond a purchase order. AMD also provided Day-0 support for Muse Spark Open models on AMD Instinct GPUs, and the companies are optimizing recommendation models on AMD Instinct MI350X GPUs. Day-0 support and recommendation optimization are the unglamorous parts of a platform deal, and they are exactly what a buyer needs before committing a training fleet. m:…0486

Oracle, meanwhile, presents AMD Helios as an open, co-designed AI infrastructure platform, integrating 72 AMD Instinct MI455X GPUs with 6th Gen AMD EPYC CPUs and AMD Pensando networking. Oracle says Helios provides up to 31 TB of HBM4 memory per 72-GPU rack and up to 1.7 PB per second of aggregate peak memory bandwidth per rack, with 6th Gen EPYC processors providing up to 256 cores per socket and 16 memory channels with up to 8000 MT per second. m:…1333

The software layer decides it

Hardware announcements are easy; serving software is where alternative stacks usually stall. Oracle says AMD ROCm software supports PyTorch, JAX, vLLM and SGLang, which covers the frameworks and serving engines most teams actually use. AMD Pensando DPUs power Oracle Acceleron SmartNICs, so the networking story is part of the same platform rather than an afterthought. m:…1333

AMD is also publishing the kind of low-level engineering work that builds developer trust. One post details training EAGLE3 speculative decoding draft models on AMD Instinct MI350X GPUs with AMD Quark FP8 quantization, and notes that speculative decoding performance is limited by the quality of the draft model. That is a useful warning for teams chasing serving speed: a faster draft model only helps if the draft is good. m:…4823

A second AMD post examines warp-pipelining for tiled GEMM operations on AMD Instinct, where keeping matrix units busy on compute-bound shapes depends on overlapping data preparation and computation. Resident waves often reach synchronization points in phase, causing both to prepare data or wait simultaneously, and extra buffering does not by itself establish the execution timing needed to hide data preparation. Tiled GEMM operations compute on one pair of operand tiles while preparing the next, so the engineering task is scheduling, not just capacity. m:…9699

Why this matters

The pattern across the week is a stack coming together in public: anchor customers testing the hardware, a financial commitment tying the vendors together, a cloud platform productizing the rack, and engineering posts showing how to squeeze performance from the software. That sequence is what a second stack needs to be credible, and it is a different signal from a single chip announcement. m:…0486 m:…1333 m:…4823

For executives, the decision is not whether AMD hardware exists, but whether the full chain, from Day-0 model support to serving frameworks to kernel-level tuning, holds up on their own workloads. Meta testing Helios and Oracle shipping it on OCI give two places to run that test, and the ROCm posts give developers a preview of how much tuning the stack still expects from them. m:…0486 m:…1333