AMD says it has built the first x86 chip that can run a 300-billion-parameter language model without ever touching the cloud. The claim landed on May 20, 2026, when AMD unveiled the Ryzen AI Max PRO 400 Series, a pair of mobile and workstation processors that pack up to 192GB of memory into a single chip package. That is enough headroom, AMD says, to load models several times larger than what today’s high-end gaming GPUs can handle on their own.
The announcement matters because it reframes what “local AI” means on a PC. Instead of renting GPU time in the cloud or settling for a scaled-down 7B or 13B model squeezed onto a laptop GPU, AMD is betting that unified memory, shared between CPU and GPU cores on the same chip, is the fastest way to put frontier-scale models directly into the hands of developers, researchers, and enterprise IT departments. ASUS, HP, and Lenovo are expected to ship systems built on the new silicon in Q3 2026, according to AMD’s announcement.
What AMD Actually Announced on May 20
AMD’s announcement introduced the Ryzen AI Max PRO 400 Series and described it as the world’s first x86 client processors capable of running 300 billion parameter models locally. That framing is deliberate. AMD has spent the past two years building out its Ryzen AI Max lineup around the idea that memory capacity, not raw GPU horsepower, is the bottleneck holding local AI back. A gaming GPU with 16GB or 24GB of VRAM can run small and mid-size models comfortably, but anything approaching frontier scale needs memory in the hundreds of gigabytes, which is exactly what data center GPUs provide and consumer graphics cards do not.
The PRO 400 Series targets that gap directly. Rather than pairing a CPU with a discrete GPU that has its own dedicated VRAM pool, AMD’s design shares a single large memory pool between the CPU cores, the integrated GPU, and the neural processing unit on the same package. That is the same basic approach AMD used in its earlier Ryzen AI Max chips (known by the code name Strix Halo), but the PRO 400 Series pushes the maximum memory ceiling higher and adds the “PRO” manageability and security features that AMD bundles for commercial and enterprise buyers.
Meet the Two Chips: Ryzen AI Max PRO 485 and Max+ PRO 495
AMD’s official product pages list two model names in the new series: the AMD Ryzen AI Max PRO 485 and the AMD Ryzen AI Max+ PRO 495. Both product pages carry the same May 20, 2026 publish date as the wider announcement, and both list an identical maximum memory figure: 192GB. The “+” in the Max+ PRO 495 name typically signals a higher-end configuration within AMD’s Ryzen AI Max branding, though AMD’s public materials do not spell out every architectural difference between the two SKUs.
The Memory Split: 192GB Total, Up to 160GB as GPU VRAM
The detail that has generated the most attention is how AMD splits that 192GB pool. According to AMD’s announcement, the platform supports up to 192GB of total memory, with as much as 160GB of that able to be allocated specifically as GPU VRAM. That is a striking number: 160GB of addressable graphics memory is more than six times what a flagship RTX 5090 gaming GPU carries, and it puts the PRO 400 Series in the same conversation as some data center accelerators, at least on paper.
The remaining memory stays available to the CPU and system processes, which matters for anyone planning to run a large model alongside a normal desktop workload rather than on a stripped-down inference server. AMD has not published exact figures for memory bandwidth on the PRO 400 Series, so any claim about tokens-per-second throughput on a specific model should be treated as unconfirmed until AMD or independent reviewers publish benchmark numbers.
The 300 Billion Parameter Claim, Explained
AMD’s headline claim, that the Ryzen AI Max PRO 400 Series can run 300 billion parameter models locally, is a marketing and launch claim rather than a fully documented benchmark result. AMD’s own product-page facts confirm the memory ceiling (192GB total, up to 160GB as VRAM) but do not spell out which specific model was tested, what quantization level was used, or what test conditions produced that number. In practice, running a 300B-parameter model on 160GB of VRAM requires aggressive quantization, likely 4-bit or lower precision, since a 300B model at full 16-bit precision would need roughly 600GB of memory just to hold the weights.
None of that makes the claim false. Quantized inference is standard practice across the industry, from cloud providers to hobbyists running open-weight models on consumer hardware. But it does mean buyers evaluating the PRO 400 Series for a specific workload should look for independent, reproducible benchmarks once review units ship, rather than assuming “runs 300B models” means running them at full precision and full speed. Tom’s Hardware, covering the same announcement, framed the story around a different but related constraint: memory bandwidth, not just capacity, determines how usable a large model actually is once it is loaded, and that bandwidth question arrives at an inconvenient moment, with DRAM prices sitting at a 15-year high.
How Unified Memory Changes the Local AI Inference Math
To understand why AMD’s approach is notable, it helps to compare it with the two other ways people currently run large language models: cloud inference and discrete-GPU inference. Cloud inference (calling an API run on data center hardware) offers effectively unlimited memory and the fastest models, but every request leaves the user’s machine, which raises cost, latency, and data-privacy concerns for regulated industries. Discrete-GPU inference (running a model on a gaming or workstation GPU) keeps everything local but is capped by whatever VRAM that card carries, typically 16GB to 32GB on today’s high-end consumer cards.
Unified memory APUs like the Ryzen AI Max PRO 400 Series occupy a third lane. Because the CPU and GPU draw from the same physical memory pool, the system does not need a discrete graphics card with its own dedicated VRAM at all. It also means the practical memory ceiling is set by how much system RAM the machine ships with, rather than by whatever a graphics card manufacturer decided to solder onto a board. That is precisely the argument AMD used to justify the “world’s first” framing: no other x86 client chip lets a single processor package address 192GB of shared memory for AI workloads.
The tradeoff is bandwidth. Unified memory pools built around LPDDR-class RAM are typically slower than the dedicated high-bandwidth memory (HBM) found in data center GPUs or the GDDR memory on gaming cards. That is the tension WccfTech highlighted in its coverage of the launch, reporting the platform under AMD’s internal “Gorgon Halo” development code name and framing the 192GB unified pool as the series’ single biggest architectural upgrade over the prior generation.
AMD vs Nvidia: Two Different Bets on Local AI Hardware
Nvidia’s answer to the same problem is the DGX Spark, a compact desktop AI computer built around Nvidia’s own Grace Blackwell architecture. DGX Spark ships with 128GB of unified LPDDR5x memory and a 273GB/s memory bus, according to Nvidia’s own hardware documentation. It launched at a Founders Edition price of $3,999 in October 2025, before Nvidia raised the MSRP to $4,699 in February 2026, citing memory supply constraints, the same DRAM crunch Tom’s Hardware pointed to in its coverage of AMD’s announcement.
The comparison is not perfectly apples-to-apples. DGX Spark is a dedicated appliance built and sold by Nvidia as a single product, complete with Nvidia’s own software stack for AI development. The Ryzen AI Max PRO 400 Series is a processor that OEMs like ASUS, HP, and Lenovo will build into their own laptops and workstations, meaning final system prices, memory configurations, and software bundles will vary by manufacturer. But on raw memory capacity, AMD’s 192GB (with up to 160GB as VRAM) outnumbers Nvidia’s 128GB by a wide margin, even if Nvidia’s dedicated memory bus may still out-pace AMD’s shared pool on bandwidth once independent tests are available.
AMD vs Qualcomm and Apple in the Broader AI PC Race
AMD is not the only chipmaker chasing the “run a big model on your own machine” pitch. Qualcomm’s Snapdragon X2 Elite platform, aimed at Windows on Arm laptops, is documented by Qualcomm as supporting a maximum memory capacity of “128+ GB,” a figure Qualcomm itself frames as an open-ended minimum rather than a hard ceiling. Apple, meanwhile, has pushed unified memory further than anyone on the consumer side: its M5 Ultra Mac Studio, which Apple launched earlier in 2026, tops out at 512GB of unified memory paired with an 80-core GPU, well beyond what any x86 chip currently offers.
That puts AMD’s “world’s first x86” qualifier in context. Apple’s Mac Studio has offered larger unified memory pools for longer, but Apple’s chips run on its own ARM-based architecture and macOS software stack, which locks out a large share of enterprise Windows environments and the x86-native software those environments depend on. AMD’s pitch is narrower but arguably more practical for IT departments already standardized on Windows and x86: the PRO 400 Series lets them add large-scale local AI capability without replacing their operating system or their existing software licenses.
Spec Comparison: Ryzen AI Max PRO 400 vs the Field
The table below lines up AMD’s new chips against the closest competing local-AI hardware platforms on the market, using each vendor’s own published specifications.
| Platform | Architecture | Max Unified Memory | Max GPU-Allocated Memory | Starting Price |
|---|---|---|---|---|
| AMD Ryzen AI Max PRO 485 | x86 (unified CPU/GPU/NPU) | 192GB | Up to 160GB | Not yet disclosed |
| AMD Ryzen AI Max+ PRO 495 | x86 (unified CPU/GPU/NPU) | 192GB | Up to 160GB | Not yet disclosed |
| Nvidia DGX Spark | Arm (Grace Blackwell) | 128GB | Shared pool, 273GB/s bus | $4,699 (Founders Edition, 2026) |
| Qualcomm Snapdragon X2 Elite | Arm (Windows on Arm) | 128+ GB | Shared pool | Varies by OEM |
| Apple Mac Studio (M5 Ultra) | Arm (Apple Silicon) | 512GB | Shared pool | Varies by configuration |
Two things stand out. First, AMD is the only x86 platform in this group, which is the entire basis of its “world’s first” claim. Second, every one of these platforms relies on the same underlying idea, shared memory between CPU and GPU, rather than a discrete GPU with its own separate VRAM. That convergence across four major chipmakers in the space of about eighteen months suggests unified memory has become the industry’s default answer to local AI, not just AMD’s.
OEM Rollout: ASUS, HP, and Lenovo Target Q3 2026
AMD’s announcement named three OEM partners expected to ship systems built on the PRO 400 Series: ASUS, HP, and Lenovo, with availability targeted for Q3 2026. That is a tight window from the May 20 announcement date, giving each manufacturer roughly one fiscal quarter to finalize designs, validate thermal and power delivery for a chip carrying up to 192GB of memory, and get units into retail and enterprise procurement channels.
AMD has not disclosed pricing for systems built on either chip, and no specific product names, model numbers, or configurations from ASUS, HP, or Lenovo have been confirmed publicly at the time of this article. Given that 192GB of high-capacity memory alone represents a substantial bill-of-materials cost, especially with DRAM prices elevated industry-wide, expect PRO 400 Series systems to launch as premium workstation or mobile workstation products rather than mainstream consumer laptops.
Enterprise and Developer Impact
For enterprise IT buyers, the pitch is straightforward: run a large language model entirely inside the company’s own hardware, with no data ever leaving the device. That matters most for regulated industries, healthcare, finance, legal, and government contractors, where sending proprietary or sensitive data to a third-party cloud API is either against policy or against the law. A workstation that can hold a 300B-parameter-class model locally removes that constraint entirely, provided the quantized model’s accuracy holds up for the task at hand.
For individual developers, the calculus is different. A PRO 400 Series machine is likely to cost more than a typical developer laptop, but far less than the cumulative API bill of running large-model inference in the cloud over the life of a multi-year project. Whether that math works out depends heavily on the still-unconfirmed price, and on real-world tokens-per-second figures that no independent reviewer has published yet. Developers weighing a purchase decision today are, in effect, betting on AMD’s claim ahead of any hands-on verification.
Historical Context: From Strix Halo to the PRO 400 Series
AMD’s unified-memory APU strategy did not begin with the PRO 400 Series. The company’s earlier Ryzen AI Max chips, built on the Strix Halo design, established the basic template: combine CPU cores, GPU cores, and a neural processing unit on one package, and let them share a single large memory pool rather than splitting memory between a CPU’s system RAM and a GPU’s dedicated VRAM. That approach let AMD compete on memory capacity even when its integrated graphics could not match a discrete GPU’s raw compute throughput.
The PRO 400 Series represents the next step in that strategy: take the same unified-memory concept, push the ceiling from whatever the previous generation supported up to 192GB, and add the PRO-tier manageability, security, and long-term support commitments that commercial and enterprise buyers expect. It is also AMD’s clearest statement yet that it sees local AI capability, not just raw CPU or GPU benchmark scores, as a core differentiator in the commercial PC market going forward.
Market Impact: What This Signals for the AI PC Segment
The AI PC category has struggled to articulate a clear value proposition beyond marketing checkboxes and NPU TOPS ratings that most consumer software does not yet use. AMD’s PRO 400 Series is a bet that “run a genuinely large model, entirely offline” is a concrete enough capability to justify a premium price, in a way that abstract NPU performance claims have not been. If AMD, Nvidia, Qualcomm, and Apple are all converging on unified memory as the mechanism to deliver that capability, memory capacity, rather than clock speed or core count, is likely to become the headline spec OEMs market around for the next generation of AI-capable PCs.
There is also a supply-chain angle worth watching. Tom’s Hardware’s framing of the announcement, tying it directly to DRAM prices at a 15-year high, points to a real tension: every chipmaker chasing bigger unified memory pools is competing for the same constrained memory supply. If that shortage persists through the Q3 2026 launch window AMD has targeted, it could push OEM system prices for PRO 400 Series machines well above what the underlying chip announcement alone would suggest.
What’s Confirmed and What’s Still Marketing
It is worth separating what AMD has confirmed from what remains an open question. Confirmed, directly from AMD’s own announcement and product pages: the Ryzen AI Max PRO 485 and Ryzen AI Max+ PRO 495 model names, a maximum memory ceiling of 192GB for both chips, up to 160GB of that allocatable as GPU VRAM, and a target OEM rollout through ASUS, HP, and Lenovo in Q3 2026. Not yet confirmed: pricing for either the chips or the systems built on them, the exact model, quantization level, and test conditions used to validate the 300-billion-parameter claim, and independent, third-party benchmark results for real-world inference speed.
That gap between confirmed specs and unverified performance claims is normal at this stage of a product launch, but it is also exactly where buyers should focus their attention before committing budget. A memory ceiling is a hard, verifiable number. A “runs 300B models” headline is a claim that only becomes fully meaningful once someone outside AMD has run the test.
Predictions: Where On-Device LLM Hardware Goes Next
- Independent benchmarks will surface within weeks of the Q3 2026 OEM launch. Reviewers will almost certainly test tokens-per-second on popular open-weight models once ASUS, HP, or Lenovo systems reach labs, which will either validate or puncture the 300B-parameter marketing claim.
- Memory capacity becomes the new headline spec war. Expect AMD, Nvidia, Qualcomm, and Apple to keep pushing unified memory ceilings higher in each successive generation, the same way clock speed and core count drove chip marketing in past decades.
- DRAM pricing pressure will show up in final system prices. With Tom’s Hardware already flagging 15-year-high DRAM costs tied to this announcement, PRO 400 Series laptops and workstations are likely to launch at a premium relative to non-AI-focused systems in the same category.
- Enterprise procurement, not consumer retail, will be the first real market. Given the likely price point and the PRO-tier manageability features AMD bundles with this series, expect early adoption to concentrate in regulated industries and enterprise IT rather than general consumer laptops.
- Quantization tooling becomes a competitive differentiator. Since running a 300B-parameter model on 160GB of VRAM requires aggressive quantization, expect AMD to invest heavily in software tooling (through ROCm or partner frameworks) to make that quantization process easier for developers, since the hardware capability alone will not matter if the software experience is difficult.
The table below summarizes the confirmed specs at a glance.
| Detail | Ryzen AI Max PRO 485 | Ryzen AI Max+ PRO 495 |
|---|---|---|
| Announcement date | May 20, 2026 | May 20, 2026 |
| Max total memory | 192GB | 192GB |
| Max memory allocatable as GPU VRAM | Up to 160GB | Up to 160GB |
| Target OEMs | ASUS, HP, Lenovo | ASUS, HP, Lenovo |
| Expected availability | Q3 2026 | Q3 2026 |
| Confirmed price | Not disclosed | Not disclosed |
The Bigger Picture for Local AI
Whatever the final benchmark numbers show, AMD’s PRO 400 Series announcement marks a clear shift in how chipmakers are pitching AI capability to buyers. For the past two years, “AI PC” marketing centered on NPU TOPS figures that most users never directly experienced, since few mainstream applications were built to take advantage of on-device neural processing at all. A processor that can plausibly hold a 300-billion-parameter-class model in memory is a different kind of claim entirely: it is concrete, testable, and directly tied to a capability, running a large model without a cloud connection, that developers and enterprises can immediately understand and evaluate.
The next few months will tell whether AMD’s hardware lives up to the claim. AMD has done its part by shipping product pages with a confirmed 192GB memory ceiling. The rest, real-world speed, actual OEM system pricing, and independent verification of the 300B-parameter claim, is now in the hands of ASUS, HP, Lenovo, and the reviewers who will get their hands on the first units later this year.
Frequently Asked Questions
What is the AMD Ryzen AI Max PRO 400 Series?
It is a new line of AMD processors, announced May 20, 2026, built around unified memory shared between the CPU, GPU, and NPU on a single chip package. AMD markets the series, which includes the Ryzen AI Max PRO 485 and Ryzen AI Max+ PRO 495, as the first x86 client processors able to run 300-billion-parameter AI models locally.
How much memory does the Ryzen AI Max PRO 400 Series support?
Both the Ryzen AI Max PRO 485 and Ryzen AI Max+ PRO 495 support up to 192GB of total memory, according to AMD’s official product pages. Up to 160GB of that pool can be allocated specifically as GPU VRAM.
Can these chips really run a 300 billion parameter model?
AMD says so in its announcement, but the exact model, quantization level, and test conditions behind that claim have not been publicly detailed. Running a model that size within 160GB of VRAM would require significant quantization. Independent benchmarks have not yet been published, so treat the claim as an AMD marketing statement pending third-party verification.
When will laptops and workstations with this chip be available?
AMD’s announcement pointed to Q3 2026 for systems from ASUS, HP, and Lenovo. No specific product names or launch dates for individual OEM systems had been confirmed publicly at the time of this article.
How does this compare to Nvidia’s DGX Spark?
Nvidia’s DGX Spark ships with 128GB of unified memory and a 273GB/s memory bus, and launched at $3,999 before rising to a $4,699 Founders Edition MSRP in 2026. AMD’s PRO 400 Series offers a larger 192GB memory ceiling, with up to 160GB allocatable as VRAM, but is a processor for OEM systems rather than a standalone appliance like DGX Spark, and AMD has not disclosed pricing.
Is the Ryzen AI Max PRO 400 Series meant for consumers or businesses?
The “PRO” branding and the OEM partners named (ASUS, HP, Lenovo) point toward commercial and enterprise workstation buyers rather than mainstream consumer laptops. Given the likely cost of a system with up to 192GB of memory, early adoption is expected to concentrate among enterprise IT and professional developers rather than general consumers.
Does this replace the need for cloud-based AI models?
Not entirely. It gives users the option to run large models locally without sending data to a cloud API, which matters for privacy and regulated industries. But cloud providers still offer larger, unquantized models and faster inference speeds than any single local machine can currently match.
What is AMD’s code name for this chip family?
Industry reporting, including coverage from WccfTech, has referred to the underlying design by the internal development code name “Gorgon Halo,” continuing AMD’s practice of using Halo-family code names for its unified-memory APU designs.
