AMD executives did not flinch when NVIDIA unveiled the RTX Spark at Computex 2026. In fact, they did something far more telling: they welcomed it, dismissed it as a validation of their own strategy, and declared that anyone not buying a Strix Halo notebook today is simply making a mistake. The two companies are now on a direct collision course in the emerging category of local AI workstations powered by unified memory, and the battle lines drawn in Taipei reveal how each side intends to win.
AMD Executives on RTX Spark: “You’re Just Wrong” Without a Strix Halo
Rahul Tikoo, Senior Vice President and General Manager of AMD’s Client Business, described NVIDIA’s entrance into the large-memory local AI segment as a welcome development. “We have been the sole player in this field for nearly two years,” Tikoo said. “As large local memory becomes extremely important for agent-type AI workloads, we welcome NVIDIA joining the space.” That measured tone gave way to something bolder when Andrej Zdravkovic, AMD’s Chief Software Officer, was asked directly about the choice developers face. “If at this point you don’t buy a Strix Halo notebook,” Zdravkovic stated, “you are simply wrong.”
Tikoo also addressed the specification comparisons head-on. Regarding the 128 GB unified memory in the RTX Spark N1X, he noted that Strix Halo already delivers that capability. When pressed on NVIDIA’s 20-core Arm Grace CPU versus AMD’s 16-core Zen 5 configuration, Tikoo countered by pointing to the 16-core, 32-thread design of the Strix Halo and added that the upcoming Gorgon Halo, scheduled for the third quarter of 2026, will push even further ahead.
What the Specification Sheets Actually Show
The raw numbers paint a picture of two products that are closer than either company might want to admit, yet the differences in architecture and strategy are profound.
The RTX Spark N1X combines a 20-core Arm Grace CPU with a Blackwell-generation GPU equivalent to an RTX 5070-class part featuring 6,144 CUDA cores. It supports up to 128 GB of LPDDR5X unified memory and achieves roughly 600 GB/s of bandwidth between CPU and GPU via NVLink-C2C. Built on TSMC 3 nm, NVIDIA claims FP4 AI compute performance of 1 petaflop. Shipments are expected in fall 2026.
AMD’s Gorgon Halo, part of the Ryzen AI Max 400 series, integrates up to 16 Zen 5 cores, 40 compute units of RDNA 3.5 graphics, and an XDNA 2 NPU rated at 55 TOPS. The critical differentiator is memory: Gorgon Halo supports up to 192 GB of LPDDR5X, with 160 GB allocable as VRAM. AMD positions this as the first x86 client processor capable of running an LLM with over 300 billion parameters entirely locally. The flagship Ryzen AI Max+ PRO 495 runs at a boost clock of 5.2 GHz on the CPU side and 3.0 GHz on the GPU side, representing a modest 100 MHz uplift over the existing Strix Halo part.
Where the spec sheets go silent, the real asymmetry appears. NVIDIA’s NVLink-C2C interconnect gives the RTX Spark a dedicated high-bandwidth path between CPU and GPU that AMD’s shared-memory architecture lacks. AMD counters with substantially higher total memory capacity and the full backward compatibility of x86, which remains the dominant ISA for PC software.
Is the RTX Spark a threat to AMD’s Strix Halo and Gorgon Halo? The answer depends on whether memory capacity or software ecosystem matters more for the target customer.
The CUDA Moat Is Not Dead Yet
What NVIDIA brings to this market is not just hardware. The RTX Spark carries the full weight of the CUDA ecosystem, a software stack built over more than 15 years. When a developer runs CUDA, TensorRT, DLSS, Reflex, G-SYNC, and RTX ray tracing on a data center GPU, the exact same code runs on the RTX Spark without modification. This was the reason Qualcomm’s Snapdragon X Elite, despite competent hardware, never gained traction for local AI workloads. Software compatibility, not hardware capability, determined the outcome.
Zdravkovic acknowledged the CUDA moat but argued it has eroded significantly. “If you had asked me the same question three years ago, I would have said that is a big problem,” he said. “But today it is not that big anymore.” He emphasized that AMD’s ROCm software platform has become easier for developers to use and that migration is now straightforward. “The only remaining issue is when an application uses specific commands that only NVIDIA has, or vice versa.”
That confidence does not fully align with the experience of the developer community. ROCm has made genuine progress in supporting PyTorch, JAX, and TensorFlow, but for tasks beyond inference — particularly training and optimization libraries — CUDA remains the default. Delayed support for FlashAttention has been a recurring complaint across developer forums for years. Even RDNA 4-class GPUs do not have full official ROCm support, and the ROCm stack for Strix Halo is currently limited to Windows. The gap between “migration is simple” and the real-world thickness of CUDA’s software library is not trivial.
Memory Capacity as a Double-Edged Sword
192 GB is an impressive number. It exceeds the RTX Spark N1X’s 128 GB by 50 percent, and on paper it gives AMD a clear advantage for developers working with very large models. But delivering that memory at a price consumers can stomach is becoming harder by the quarter.
TrendForce projects that DRAM contract prices in the second quarter of 2026 will rise 58 to 63 percent quarter-over-quarter, following a 90 to 95 percent increase in the first quarter. Cumulative since early 2025, DRAM prices have effectively more than doubled. The structural cause is clear: production capacity is being diverted to HBM for AI accelerators, squeezing supply of general-purpose DRAM. No short-term relief is in sight.
The pricing implications are already visible. According to Morgan Stanley Research, RTX Spark N1X notebooks are expected to start at roughly $2,899, with the lower-end N1 model starting around $1,799. AMD’s own Ryzen AI Halo desktop workstation, based on Strix Halo, launched with preorders at $3,999. Corsair’s AI Workstation 300, equipped with 128 GB, was raised to $3,399 after a $1,100 price increase. What a 192 GB Gorgon Halo machine will cost, given current DRAM market dynamics, is not difficult to imagine.
Why NVIDIA’s Entry May Help AMD
Tikoo’s most strategically revealing comment was his rationale for welcoming competition. “NVIDIA entering this space legitimizes the category,” he said. “With both NVIDIA and AMD in this segment, the entire AI ecosystem on Windows — not just the cloud — moves forward.”
This is likely genuine, not posturing. Since Strix Halo launched in early 2025, compatible notebooks have been limited to a handful of models — the ASUS ROG Flow Z13, the HP ZBook Ultra 14 — and 128 GB configurations are nearly impossible to find in many markets, including Japan. It has been a niche within a niche. By putting Jensen Huang’s keynote presence and NVIDIA’s marketing machine behind the concept of a large-memory local AI machine, the category itself gains mainstream recognition. AMD, which could not create that market alone, may benefit from NVIDIA creating it for them.
The Real Fight Begins This Fall
The RTX Spark ships in “fall” 2026. Gorgon Halo ships in “Q3” 2026. AMD may reach the market slightly earlier, but the two product lines will effectively arrive in the same window.
On hardware alone, the comparison breaks down into clear axes: AMD wins on memory capacity and x86 compatibility; NVIDIA wins on GPU compute performance and software ecosystem maturity. Neither side has a knockout advantage. The deciding factor will be what developers actually choose to run on these machines. CUDA’s library depth, tool quality, and tutorial ecosystem are not captured on any specification sheet.
NVIDIA has already disclosed its roadmap beyond RTX Spark, including the Vera Rubin generation (2027-2028) and Rosa Feynman (2029-2030). At least four more years of generational investment are locked in. AMD has its own follow-up, Medusa Halo based on Zen 6 and RDNA 5, planned for 2027 or later. But how AMD closes the software stack gap — the one structural weakness that specification improvements alone cannot fix — remains the unanswered question.
Zdravkovic’s declaration that skipping Strix Halo is a mistake reflects the confidence of a company that has been alone in this space for two years. That confidence will face its first real test when NVIDIA’s RTX Spark ships this fall.
