AMD picked an odd week to brag about laptops nobody buys off the shelf. On October 5, 2026, the company published a set of AI benchmarks for its Ryzen AI Max+ Pro 495 chip, code-named Gorgon Halo, two days before Nvidia was widely expected to launch its RTX Spark hardware at a Microsoft event. Buried in the announcement was a bigger claim: AMD says it has now shipped “over half a million agentic PCs” to date, according to Tom’s Hardware. That single sentence, more than the benchmark charts attached to it, is the part worth unpacking.
The timing was not an accident. AMD has spent 2026 building a case that it owns the “agentic PC” category before Nvidia, Intel, or Qualcomm can define it on their own terms. The Gorgon Halo numbers are a skirmish in that fight. The half-million unit figure is the actual headline, because it is the first time AMD has put a hard number on a category it invented largely through marketing. For deeper context on how the two companies have traded benchmark blows this year, shattered.io covered the earlier round of Gorgon Halo vs. RTX Spark claims last week.
AMD’s Half-Million Agentic PC Claim, Explained
AMD’s headline figure is simple to state and hard to verify: more than 500,000 units it classifies as agentic PCs have shipped to date. The company did not break the figure down by quarter, by OEM, or by chip generation in the materials reviewed by outlets covering the announcement. There is no public methodology explaining what separates an “agentic PC” from any other laptop running a Ryzen AI chip with an NPU on board.
That ambiguity matters because “agentic PC” is not an industry standard the way “Wi-Fi 6E” or “USB4” is. It is AMD’s own label, first pushed through its Advancing AI event in mid-2026, where the company framed its full product stack around what it called the agentic AI era, and reinforced through a string of blog posts and partner announcements since, including a late-September tie-up with Perplexity highlighted in AMD’s own agentic PC press kit. Half a million units sounds sizable until you compare it against the PC market as a whole, where laptop shipments run in the tens of millions per quarter. AMD framed the number as proof of early traction. Skeptics will read it as a modest base being inflated by a flattering label.
What Actually Counts as an “Agentic PC”?
Ask five people in the industry to define an agentic PC and you will likely get five different answers. AMD’s own materials describe the category loosely, as machines built around NPUs and unified memory pools large enough to run local AI agents that can plan and execute multi-step tasks without constant cloud calls. The company’s press kit lists partner integrations and developer platforms rather than a hard technical bar a chip must clear to earn the label.
That looseness is deliberate. A strict, published spec would let Intel or Qualcomm claim compliance for their own chips just as easily. By keeping the term brand-specific, AMD gets to control who counts. The risk is that reporters and buyers eventually tune out the number the way “AI PC” itself became a marketing throwaway term in 2024 and 2025, used by every vendor regardless of what the hardware could actually do locally.
Inside the Gorgon Halo Benchmarks
Setting the shipment claim aside, the benchmark portion of the announcement is more concrete, even if it comes from AMD’s own lab rather than an independent reviewer. AMD tested its top-spec Ryzen AI Max+ Pro 495 configuration, carrying 192GB of unified memory, against a system built on Intel’s Core Ultra X9 388H with 64GB. Both machines ran ComfyUI, a popular generative-AI workflow tool, across several open models.
The ComfyUI Test Suite
AMD reported a performance swing ranging from 1.1x to 32.2x in its favor across the tested workloads, a spread wide enough that the top end is almost certainly an outlier driven by a specific model hitting Intel’s memory ceiling rather than a representative average. On GLM 5.3 Flash, a 320-billion-parameter model that activates roughly 18 billion parameters per token, AMD’s system produced about 20 tokens per second. According to Club386’s reporting on the same disclosure, the Ryzen system also hit 42 tokens per second on Qwen 3.8 Flash, a smaller 125-billion-parameter model with about 5 billion active parameters.
GLM 5.3 Flash and the Memory Ceiling Problem
The GLM 5.3 Flash result is the clearest illustration of why AMD picked this comparison. A 320-billion-parameter model simply does not fit in 64GB of system memory at any reasonable quantization level, which means Intel’s machine was likely forced to either fail the test or fall back to disk swapping that tanks throughput. AMD is not just claiming a faster chip here. It is claiming a chip that can run workloads the competing configuration cannot run at all, which is a different and more durable kind of advantage than raw tokens-per-second bragging rights.
The Memory Capacity Argument
Memory, not core count, is the argument AMD keeps returning to. The 192GB pool on the top Gorgon Halo configuration is enough, according to AMD’s own marketing language, to run models with up to 300 billion parameters locally. That claim traces back to AMD’s broader Ryzen AI Max PRO messaging rather than a Gorgon Halo-specific technical disclosure, so it should be read as a product-line claim rather than a lab-verified result for this exact chip.
Compare that to Nvidia’s own DGX Spark, which ships in a 128GB configuration as the standard retail version, with a separate 64GB variant reserved for OEM partners, according to Nvidia’s published DGX Spark specifications. Nvidia’s own documentation caps a single 128GB unit at roughly 200 billion parameters, scaling up to 700 billion only when four units are clustered together. On paper, AMD’s single-chip 192GB configuration undercuts Nvidia’s single-unit ceiling by a wide margin, assuming AMD’s 300-billion-parameter claim holds up under independent testing.
Gorgon Halo vs. Intel Core Ultra X9 388H: The Numbers
Here is how AMD’s own test configuration compares, drawn from the figures reported by Tom’s Hardware and Club386 following the October 5 disclosure.
| Spec or Result | Ryzen AI Max+ Pro 495 (Gorgon Halo) | Intel Core Ultra X9 388H |
|---|---|---|
| Unified memory tested | 192GB | 64GB |
| Benchmark suite | ComfyUI generative-AI workloads | ComfyUI generative-AI workloads |
| GLM 5.3 Flash (320B total, 18B active) | ~20 tokens/sec | Not disclosed by AMD |
| Qwen 3.8 Flash (125B total, 5B active) | 42 tokens/sec | Not disclosed by AMD |
| AMD-reported performance range | 1.1x to 32.2x faster | Baseline |
| Announcement date | October 5, 2026 | n/a |
Two things stand out. AMD never disclosed Intel’s raw tokens-per-second figures, only the resulting multiplier, which is standard practice for vendor-run comparisons but still makes the numbers harder to audit independently. And the memory gap, 192GB against 64GB, is three times larger on the AMD side, which does most of the explanatory work behind the wider end of that 1.1x-to-32.2x range.
The RTX Spark Launch AMD Didn’t Test Against
The most interesting line in the entire disclosure is what AMD left out. The benchmarks never include RTX Spark hardware, because Nvidia’s product was not yet publicly available for testing when AMD ran its numbers. AMD built an entire competitive narrative around a chip it has not actually benchmarked against. That is a defensible move given the circumstances, but it also means every comparison in this story is really AMD versus Intel, dressed up as AMD getting ahead of Nvidia.
RTX Spark itself was expected to launch around a Microsoft event on October 7, 2026, the same day this article publishes. Shattered.io has tracked the lead-up closely, including Nvidia’s RTX Spark N1X chip shipping with 6,144 CUDA cores and RTX Spark systems rolling out across six PC makers. Until independent reviewers get RTX Spark hardware in hand and run the same ComfyUI workloads AMD used, every percentage point in this story should be read as AMD’s framing of a fight Nvidia has not entered yet.
How the Broader Unified-Memory Chip Field Lines Up
Gorgon Halo does not exist in isolation. It is one entry in a small but growing field of chips built around large, unified memory pools specifically to run local AI models without a discrete GPU card. Nvidia’s DGX Spark predates it in that niche, and AMD’s own prior-generation Ryzen AI Max+ 395, known as Strix Halo, set the template AMD is now iterating on.
| Chip / Platform | Maker | Max unified memory | Status as of October 7, 2026 |
|---|---|---|---|
| Ryzen AI Max+ Pro 495 (“Gorgon Halo”) | AMD | 192GB | Shipping in OEM systems |
| DGX Spark | Nvidia | 128GB (64GB OEM-only variant) | Shipping |
| RTX Spark N1X | Nvidia | Not yet independently tested | Launch expected October 7, 2026 |
| Ryzen AI Max+ 395 (“Strix Halo”, prior gen) | AMD | 128GB | Shipping since 2025 |
| Core Ultra X9 388H | Intel | 64GB (as tested by AMD) | Shipping |
The pattern across this table is consistent. Every vendor chasing local AI workloads is converging on the same idea: pair a modest GPU with a huge, fast pool of memory the CPU and GPU can both address, rather than relying on a discrete card with its own separate VRAM. AMD just happens to have pushed that pool furthest, for now.
Who Can Actually Buy a Gorgon Halo Machine
None of this matters much if the hardware stays out of reach. Early Gorgon Halo systems are not cheap. Club386 reported pricing on two of the first machines built around the Ryzen AI Max+ Pro 495: a Minisforum MS-S1 Max-P495 priced at $7,399, and an HP ZBook Ultra G3a listed at $7,449. Both sit firmly in workstation territory, not the price range of a typical consumer laptop buyer comparing AI PCs on a store shelf.
That price point reframes the half-million-unit shipment claim. If the flagship 192GB configuration costs north of $7,000, the bulk of those half-million agentic PCs almost certainly carry lower-memory Ryzen AI chips sold at mainstream laptop prices, not the Gorgon Halo flagship AMD just benchmarked. The marketing umbrella is wide. The chip generating headlines this week sits at its most expensive, least representative edge.
The Agentic PC Race: AMD, Intel, Qualcomm, and Nvidia
AMD is not fighting this battle alone against Intel. Qualcomm has pushed its own Arm-based Snapdragon X2 chips as efficiency leaders, and shattered.io’s own testing found a Snapdragon X2 system beating an Intel Core Ultra configuration by 31% while costing $750 less. Microsoft, for its part, has leaned toward Qualcomm and Nvidia silicon for its flagship Surface line, a shift documented in coverage of Surface going five years without an AMD-powered model.
That makes AMD’s agentic PC push partly a bid to win back design wins it has not held in Microsoft’s own hardware for half a decade. Winning the OEM conversation with Minisforum and HP is one thing. Winning it with Microsoft, Dell, and Lenovo at consumer price points is a different and much larger prize, and AMD’s current benchmarks do not yet speak to that fight directly.
A Short History of the AI PC Marketing Push
None of this is new behavior from the chip industry. Intel spent 2023 and 2024 pushing the “AI PC” label onto any laptop with an NPU capable of a few dozen TOPS, regardless of whether a single consumer application actually used it. Qualcomm followed with its own Copilot+ PC marketing tied to Windows on Arm. AMD watched both campaigns play out, then arrived in 2026 with “agentic PC,” a term built around the idea that local AI should not just classify a photo but actually complete multi-step tasks on a user’s behalf.
The pattern each time is the same: a vendor defines a category on its own terms, ships a wave of hardware under that label, and only later do independent benchmarks and real-world software catch up enough to say whether the label meant anything. AMD’s half-million-unit claim is an early checkpoint in that cycle, not a conclusion.
Market Impact: What This Means for Buyers and OEMs
For OEMs like Minisforum and HP, the calculus is straightforward. A chip that can legitimately run 300-billion-parameter models locally opens a workstation niche that previously required a discrete GPU with 48GB or more of VRAM, hardware that costs far more than $7,399 on its own. For buyers evaluating options across Nvidia’s own split DGX Spark pricing tiers, Gorgon Halo adds a genuine third option rather than just another marketing variant of the same silicon.
For everyday laptop shoppers, the direct impact is close to zero this week. The flagship Gorgon Halo configuration is a workstation part wearing a consumer-facing category name. The real effect will show up a few quarters out, if and when 32GB to 64GB Gorgon Halo variants reach mainstream laptops at prices under $1,500, which is where AMD’s half-million-unit shipment number actually lives today.
Why AMD Moved Before RTX Spark Had Even Launched
Releasing benchmarks against a product that is not yet out, rather than waiting for Nvidia to ship and then responding, is a specific kind of narrative play. AMD gets to set the first impression in tech press coverage, control which comparison chip gets used, and frame the memory-capacity argument before reviewers have RTX Spark units to test independently. It is a defensive move dressed up as an offensive one.
It also carries risk. If RTX Spark ships with its own large unified-memory configuration and beats AMD’s numbers once independent labs get hands-on, AMD’s preemptive framing will read as premature rather than prescient. The company is betting that its 192GB ceiling and the 300-billion-parameter claim tied to it will hold up even after Nvidia’s hardware reaches reviewers.
Predictions: Where This Goes From Here
- Independent reviewers will get RTX Spark hardware within weeks of its launch and run comparable ComfyUI workloads, which will either validate or puncture AMD’s preemptive framing.
- AMD will keep citing the half-million agentic PC figure in earnings commentary without publishing a clear per-quarter breakdown, treating the category label as a durable marketing asset rather than a one-time stat.
- Lower-memory Gorgon Halo variants, likely in the 32GB to 64GB range, will reach mainstream laptop price points within two to three quarters, doing more to grow the half-million-unit base than the $7,000-plus flagship configuration ever will.
- Intel and Qualcomm will respond with their own unified-memory messaging rather than matching AMD’s raw capacity, since neither currently ships a 192GB consumer-adjacent configuration.
- Microsoft’s choice of silicon for its next Surface refresh will matter more to AMD’s agentic PC ambitions than any single benchmark release, given AMD’s five-year absence from that lineup.
The Bottom Line
Strip away the category label and two separate stories are tangled together here. One is a genuinely interesting hardware result: a 192GB unified-memory chip that can run models a 64GB competitor physically cannot load, at a price that undercuts comparable GPU-based workstations. The other is a marketing claim, half a million units shipped under a term AMD itself defined, that nobody outside AMD can independently verify yet. Both deserve attention. Neither should be taken as the final word before RTX Spark actually ships and gets tested under the same lights.
Frequently Asked Questions
What is AMD’s Gorgon Halo chip?
Gorgon Halo is the code name reported for AMD’s Ryzen AI Max+ Pro 495 processor, a chip built around a large pool of unified memory, up to 192GB in its flagship configuration, aimed at running generative-AI workloads locally without a discrete GPU.
Did AMD actually test its chip against RTX Spark?
No. AMD’s October 5, 2026 benchmarks compared the Ryzen AI Max+ Pro 495 against Intel’s Core Ultra X9 388H, not against Nvidia’s RTX Spark hardware, which was not yet publicly available for independent testing at the time.
What does “over half a million agentic PCs” actually mean?
It is AMD’s own shipment claim for machines it classifies under its self-defined “agentic PC” category. AMD has not published the methodology behind the classification or a breakdown by chip generation or OEM, so the figure should be treated as a company-reported total rather than an independently audited number.
How much memory does the Ryzen AI Max+ Pro 495 support?
The configuration AMD benchmarked carried 192GB of unified memory. AMD’s broader marketing for the Ryzen AI Max PRO line claims support for models up to roughly 300 billion parameters at 4-bit quantization, though that figure traces to general product-line materials rather than a Gorgon Halo-specific independent test.
How does this compare to Nvidia’s DGX Spark?
Nvidia’s DGX Spark ships in a 128GB retail configuration, with a 64GB variant reserved for OEM partners, and is rated by Nvidia to handle models up to roughly 200 billion parameters per unit. AMD’s 192GB Gorgon Halo configuration claims a higher single-unit ceiling, though the two chips have not yet been benchmarked head-to-head by an independent reviewer using the same test suite.
When is RTX Spark launching?
RTX Spark was widely expected to launch around a Microsoft event on October 7, 2026, the same day AMD’s benchmark story broke into wider coverage. Final specs and independent reviews were not yet available at the time of this report.
Can regular consumers buy a Gorgon Halo laptop today?
The first systems built around the flagship 192GB Ryzen AI Max+ Pro 495, including a Minisforum mini PC and an HP ZBook Ultra laptop, are priced above $7,000, putting them in workstation territory rather than mainstream consumer reach. Lower-memory Gorgon Halo variants at more typical laptop prices had not been detailed as of this report.
Why did AMD release these benchmarks now?
The timing, two days ahead of Nvidia’s expected RTX Spark launch, let AMD set the first narrative in tech press coverage and frame the memory-capacity argument before reviewers could test Nvidia’s hardware directly. It is a common tactic in chip marketing: get ahead of a competitor’s launch cycle rather than wait to respond to it.
