RTX Spark laptop: ASUS, MSI, Dell, pricing and date

#What is actually announced for the RTX Spark

The NVIDIA RTX Spark is not a new graphics card that you screw into a tower. It is a SoC—a single chip combining the processor, GPU, and memory—derived directly from the GB10 that already powers the DGX Spark, NVIDIA's compact AI development workstation. The Computex 2026 announcement boils down to one sentence: this same chip is coming to consumer laptops and compact machines in fall 2026.

The factor that changes everything for local AI is memory: 128 GB of unified memory shared between the CPU and GPU. Whereas a conventional graphics card partitions its VRAM (24 GB on a RTX 4090, for example), the RTX Spark makes 128 GB available to the model. That's what lets it load models that would never fit on a consumer GPU.

What to remember right now: the chip’s specs are known because they are the specs of the GB10, which is already commercially available. What remains unclear are the exact machine models, their configurations, and especially their prices. NVIDIA set the direction; manufacturers have not yet opened preorders.

#The GB10 chip, explained

The GB10 is a Grace Blackwell architecture: an ARM “Grace” CPU and a “Blackwell” GPU on the same package, connected to a shared memory pool. This “unified memory” design follows the same philosophy as Apple Silicon chips (M4 Pro/Max)—and that is precisely what makes it interesting for running large LLMs without a dedicated €2,000 graphics card.

Unified memory

128 GB shared between the CPU and GPU. This is the key figure: it determines the maximum model size the machine can load.

Architecture

Grace Blackwell (GB10)—Grace ARM CPU + Blackwell GPU on the same SoC, with CUDA support on the GPU.

Target

LLM inference and local AI workloads, not high-end gaming. It is not a replacement for RTX 5090 for 3D rendering.

Format

Built into compact laptops and desktops at the factory. You can't buy it separately to add to an existing PC.

Ecosystem

GPU-side drivers and tools NVIDIA (CUDA), making it compatible with most standard inference engines.

#Manufacturers and the fall 2026 timeline

NVIDIA will not sell the RTX Spark directly in this consumer form: PC manufacturers will integrate it into their machines, as they do with mobile RTX GPUs. The partners named at Computex 2026 are Dell, Asus, MSI, HP, and Lenovo—the five biggest names in laptops.

Dell

Partner announcement. Specific models and product lines not disclosed at this time.

Asus

Partner announced. Laptop and/or compact desktop positioning expected.

MSI

Announced partner. Historically active in creator and AI systems.

HP

Announced as a partner. Likely a mobile workstation variant.

Lenovo

Announced partner. Natural candidates for pro and workstation product lines.

In practice, each brand will probably offer the RTX Spark in several configurations (battery life, cooling, display, storage), but the chip and its 128 GB of unified memory will remain the common foundation. This guide targets that foundation, not one commercial model or another that does not yet exist.

#Price: clearly distinguish estimates from official figures

This is the point that needs the clearest explanation, because it is widely repeated. The $1,799 figure associated with the RTX Spark is an analyst estimate, not a price announced by NVIDIA or any manufacturer. No official price exists as of the date of this guide.

What’s official

The existence of the RTX Spark, the GB10 chip, 128 GB of unified memory, the five partner manufacturers, and the “fall 2026” window.

What it isn't

The price (€1,799 is an estimate), exact configurations, preorder dates, and prices in euros.

Worth watching

Official product announcements from Dell, Asus, MSI, HP, and Lenovo as fall approaches—the only reliable sources for the actual price.

#Which LLMs can run on 128 GB of unified memory

This is where the RTX Spark becomes interesting. With 128 GB of unified memory, it targets a class of models beyond the reach of a typical consumer GPU. To give you a sense of the scale, here are the approximate memory requirements with Q4_K_M quantization, the recommended format for a good quality-to-size tradeoff.

7B–14B models (Q4)

≈ 5 to 9 GB. They run effortlessly, with plenty of headroom for a long context and several models loaded in parallel.

32B models (Q4)

≈ 19 GB. Comfortable, with room for a large context cache.

70B models (Q4)

≈ 40 GB. This is where the RTX Spark makes sense: a 70B model fits comfortably, whereas you would need two RTX 4090 on a typical machine.

Large MoE (100B+ parameters)

Possible depending on the quantization. The 128 GB opens the door to models that 24 GB of VRAM cannot load at all.

On the software side, there's nothing specific to set up: the chip exposes a CUDA-compatible GPU, so the usual stack works. A Ollama daemon on http://localhost:11434 with Open WebUI or LM Studio on top remains the simplest starting point. Q4_K_M quantization is the default setting, with Q5_K_M or Q8_0 when available memory allows you to target higher quality.

# Un daemon Ollama expose l'API sur le port 11434
ollama serve

# Charger un gros MoE — un usage typique de ces 128 Go unifiés
ollama pull qwen3.6:35b
ollama run qwen3.6:35b

# Vérifier la répartition mémoire GPU/CPU
ollama ps

#RTX Spark vs. conventional GPU: two different approaches

Comparing the RTX Spark with a RTX 4090 or 5090 only makes sense if you understand that they are playing different games. One focuses on memory capacity, the other on raw speed.

RTX 4090 / 5090 (24–32 GB VRAM)

Very high bandwidth, fast generation. Ideal for up to 32B models in Q4; beyond that, you need multiple cards or offloading.

RTX Spark (128 GB unified)

Massive memory capacity in a compact, mobile form factor. It can handle 70B models and large MoE models, but at a more moderate speed.

Mac M4 Pro/Max (24–48 GB unified)

Same unified philosophy as Apple, but the RTX Spark takes memory much further while staying in the CUDA ecosystem.

In short: if you need to run models up to 32B quickly, a dedicated GPU is often still the best choice. If you want to load very large models on a portable machine without building a dual-GPU workstation, that’s exactly the niche the RTX Spark is targeting.

#Should you wait until fall 2026?

1. 01
   Clarify your memory needs

   If your target models fit in 24 GB (up to ~32B in Q4), a current GPU will serve you today, faster and without waiting. The RTX Spark is only worthwhile if you regularly target 70B models or large MoE models.
2. 02
   Don't budget around $1,799

   Until an official price is published, treat this figure as a hypothesis. Wait for manufacturers' product sheets to learn the actual price, especially in euros.
3. 03
   Monitor vendor announcements

   Dell, Asus, MSI, HP, and Lenovo will publish configurations and pre-order dates as fall approaches. They are the only reliable sources for pricing and availability.
4. 04
   Check the bandwidth before buying

   When independent tests come out, look at the measured tokens/second on a 70B, not just the 128 GB shown. Capacity says nothing about actual speed.

#Frequently asked questions

Is the NVIDIA RTX Spark a graphics card?

No. It is an SoC (GB10 Grace Blackwell chip) integrated at the factory into laptops and desktops. You cannot buy it separately to add to an existing tower.

What is the price of the RTX Spark?

There is no official price. The $1,799 figure circulating is an analyst estimate, not a price announced by NVIDIA or a manufacturer. The actual French price will depend on the configuration, VAT, and margins.

When will it be released?

NVIDIA is targeting fall 2026, but no firm pre-order date has been announced. Treat this timeline as an indicative window.

Which LLM models can it run?

With 128 GB of unified memory, it can load 70B models in Q4 (≈ 40 GB) and large MoE models, beyond the reach of a consumer GPU with 24 GB. However, speed remains limited by memory bandwidth.

Is it better than a RTX 5090?

It depends on your needs. The 5090 is faster with models that fit in its 32 GB; the RTX Spark loads much larger models but more slowly. Two different approaches.

Which manufacturers will offer it?

Dell, Asus, MSI, HP, and Lenovo were cited as partners at Computex 2026. The exact models and configurations have not yet been detailed.

Is there a Dell, Asus, HP, or Lenovo with RTX Spark?

The five brands (Dell, Asus, MSI, HP, Lenovo) were announced as partners at Computex 2026, but none has yet published a product sheet or pre-order date for a specific RTX Spark model.

#Go further

The RTX Spark fits into a hardware landscape that these guides already detail:

NVIDIA DGX Spark

The mini PC that already includes the GB10 chip and its 128 GB of unified memory, with concrete performance benchmarks against a RTX 5090 and a Mac Studio.

Choose your GPU for local AI

To choose between unified memory and dedicated VRAM: RTX 4070, 4090, Mac M-Max—the buying guide that focuses on the right criteria.

Choose your quantization (Q4, Q5, Q8, FP16)

Essential for understanding why a 70B fits in 40 GB in Q4, and what each quantization step costs in quality.