At a roundtable on the sidelines of Computex in Taipei, a reporter asked AMD's Rahul Tikoo a simple question: did AMD welcome Nvidia as a new competitor in the local AI PC market? Tikoo stood up, picked a laptop off the table, and asked HP's Jim Nottingham when it launched. "CES 2025," came the reply. Tikoo held up another device. "February or March 2025." He sat back down. "We have 35 products with Strix Halo in market," he said. "Welcome, Nvidia, to the modern compute journey." The point was made without a slide deck.
The product Tikoo was gesturing at, and that Nvidia has now arrived to compete with, is a class of laptop and small-form-factor PC built around a chip that tightly couples a powerful CPU, a capable GPU, and up to 128 gigabytes of shared memory in a single package. AMD calls its version Strix Halo, carrying the Ryzen AI Max branding. Apple pioneered this architecture with the M-series chips starting in 2020. The key insight is that when CPU and GPU share a memory pool rather than shuffling data between separate DRAM banks, you can load much larger AI models locally without them being split or truncated. A 70-billion-parameter language model, which would be impossible to run locally on a conventional laptop, becomes practical on a machine with 96 gigabytes of usable VRAM.
Nvidia's entry into this space, the RTX Spark, pairs a 20-core ARM-based processor co-developed with MediaTek with a Blackwell GPU carrying 6,144 CUDA cores and up to 128 gigabytes of unified LPDDR5X memory. On paper, it is a credible competitor. The announced AI compute figure of 1 petaflop at FP4 precision matches or exceeds AMD's current numbers, and Nvidia's memory bandwidth specification of 300 gigabytes per second beats AMD's 256 gigabyte ceiling. OEM partners including Dell, HP, Lenovo, Asus, and Microsoft Surface are committed. Jensen Huang called it "the reinvention of the PC."
Two things temper the announcement. The first is timing: RTX Spark devices will not reach consumers until autumn 2026, and Nvidia has not announced prices. AMD has been selling comparable hardware for roughly eighteen months, at prices that now extend from $3,300 for a compact desktop to $3,999 for its Ryzen AI Halo developer machine, which comes with 128 gigabytes of memory, dual boot Windows and Linux support, and a pre-validated software stack. The second is that Nvidia's own paper specification numbers have historically landed slightly below their announced figures after independent measurement, a fact that reporters at the Computex briefings noted directly.
The more interesting question, the one that Tikoo himself identified as AMD's remaining challenge, is software. For years, the shorthand answer to "should I buy AMD for local AI?" was "no, buy Nvidia, because CUDA." CUDA is Nvidia's proprietary software layer for GPU computing, and it has twenty years of library development, framework support, and developer familiarity behind it. AMD's equivalent, ROCm, was for a long time a frustrating project that required manual compilation, accepted that some things would never work, and generally felt like an afterthought. That characterisation is no longer accurate, and the gap is closing faster than most observers expected.
PyTorch, the framework that underpins a large proportion of modern AI work, now treats AMD GPUs as a first-class target. Successive releases have added AMD-specific optimisations for attention mechanisms, memory management, and matrix operations. Ollama, LM Studio, and ComfyUI, the tools that most people actually use to run local models, all work on AMD hardware. Quantisation libraries that were once CUDA-only have added AMD support. ROCm is not CUDA: training workloads, new model releases, and performance-critical workflows still arrive on CUDA first, and AMD's software for specialised tasks like agentic sandboxing still needs work. But for an enthusiast running a 70B model on a Strix Halo machine, the things that used to justify picking Nvidia over AMD have largely disappeared.
AMD's broader validation point is not lost. Nvidia entering the unified-memory AI PC category is an endorsement of the architecture. Both companies now converging on chips that couple CPU, GPU, and a large shared memory pool signals that the category is real, not a niche. AMD's next step, the Ryzen AI MAX 400 series codenamed Gorgon Halo, will push unified memory to 192 gigabytes and support models with up to 300 billion parameters, due in the third quarter. When Nvidia's devices arrive in autumn, they will be competing with a second-generation AMD product at roughly equivalent specifications. The software gap, which was AMD's most significant vulnerability, is no longer the obvious reason to wait for Nvidia. That, more than the specific benchmark numbers, is why Tikoo could afford to smile when he welcomed the new competition.