All sold out. Yes, I know that resellers have some stock. I don’t think even you believe that $3000 price increases can be explained by more expansive memory.
According to the teardown, it is 8x H58G78CK8BX185 chips. According to Google. current retail price is $364.39 USD x 8 = $2915. The large volume prices are typically much better than retail. If you look at LPDDR5X pricing you can see DRAM price increase but it is not that big : LPDDR5X 128Gb | MemoryMarket - memory prices, market analysis, industry news, research report
We are talking about premium $6000+ device. More expensive than a typical Windows x64 workstation, which runs WSL quite fine btw. I am not a big fan of Mac either but why would somebody buy RTX Spark? Play games? Obviously, because of all that compute in GB10 that can be used in the compute intensive creative workflows. Clearly, the original idea was to create a a compute supercharged MacBook Pro competitor. The amount of spit and polish needed to make it work is big though.
I was 20-years windows-faithful until I started building python apps in WSL2 and could not stop swearing at process handling, semaphoring, dispatching on processes inside WSL2 that are intercepted by Windows kernels. It’s a nighmare on its own. MacOS is as close to Linux as it can be - everything just works. Plus all these broken patches, BSOD’s etc. Corporate up their tits with Microslop so if you have to use it - its inevitable, but if you have a choice - MacOS is insanely better. And Linux too of course.
I have already ported 80% of my windows-dotnet codebase into either dotnet.core or python and cannot be happier. Only small front-loading stuff that need to work with windows-only 3rd party frameworks stay on windows.
I cannot understand how you can like Windows if you have a choice not to use it. Its pure BDSM. But hey, some people are into it
In my daily experience with mixed development teams, developing software intended for Linux servers using WSL2 allows for a workflow that is far closer to - and more compatible with - the production environment than working on a Mac.
The integration of CUDA and NVidia’s Container Toolkit also works relatively seamlessly in WSL2, meaning that development for such a stack stays much closer to the final product.
Additionally, the ability to use snapshot-like features when you want to experiment with minimal risk is another convenient advantage.
However, these observations are based on the x86/AMD64-based WSL2 integration; it remains to be seen how things will play out with RTX Spark on Windows ARM.
I never claimed to be a fan of Windows, either. Quite the opposite: ever since Windows 11, I’ve tried to avoid it wherever possible. That said, I find the restrictions on freedom inherent in the Apple ecosystem even more off-putting.
I get it, I have similar restrictions for frameworks I have. But I work and develop on mac and just ship windows-dependent code to windows box sitting under my desk or remote server. This way it avoid feeling filthy everyday with ungodly disgust of Microslop anathema hahaha
Since the $4,999 64GB announcement, more vendors have run out and several remaining128GB FE units jumped to ~$10,000 per unit.
Early $3,999 Spark adopters expanding their bundle were already questioning last months $5,999 rates for unchanged First-Gen units.
For ~$9,998 per128GB boxes retailers must be targeting the next wave of adopters who would be willing to invest 2.5x more for a unit launched 12 months ago with unresolved system-failing thermal throttling priced exponentially high due to what, resource scarcity.
It would be interesting to see if the 160GB AI Cube comes through with better rates and better speed connectivity. If they do, they will sweep the early adopter market.
The AI Cube prototype looks cube shaped. What pictures are you looking at?
To me, the first-gen DGX FE Spark does not yet qualify for a 2.5X scarcity investment without proven exponential ROI.
Xiaomi has a long-standing public pledge to cap its hardware profit margins at 5%. That’s why I’m interested in tracking the AI Cube closely for capability, connectivity and price.
True! The AI Cube pictures I saw, it looked like a low rectangular box. The reported 80 GB of unified memory is a non-starter, unless the price is below $2K. Not sure how good the 2x10Gbps ports will work for clustering.
High-bandwidth AI acceleration chip with 6nm3D wafer-level stacked advanced package, with 28,672 valid data lines, and support 1.22TB/s ultra-high near-store computing bandwidth.
20 core high-performance CPU
16 core high computing power NPU
Up to160GB unified memory
Rumored speculative price range is minimum $5,000 while some believe with the higher 160GB RAM and 1.22 TB/s memory bandwidth it may land in the $10,000 range.
I wonder if what you’re remembering is another AI mini computer and not the AI Cube
That looks promising but will it be compatible with one of the established GPU frameworks and runtimes? It does me little good if only their own (MiMo) models run on it in a proprietary way.
While we have no way to be certain until the prototype becomes something we can actually test, I would assume the AI Cube wouldn’t be restricted to MiMo model optimizations only as that would be very restrictive. Testing on their own proprietary, frontier-class flagship makes sense because it’s their own model.
I’d look at MiMo optimized and compare capabilities (potentially) to 42B close match open-weights equivalents like DeepSeek V3/V4, Qwen 2.5 72B/ 3.8 Max, and GLM 5.1/5.3
With the AI Cube vs. DGX Spark we’d be looking at different hardware and different software with a hardware infrastructure switch or multi-node integration. The DGX Sparks have a fully mature CUDA tool chain at our disposal with native training, quantizing and tuning. I haven’t seen anyone talk about a public software stack for the AI Cube.
AI Cube is pulling in XRING O3 technology and if the memory bandwidth delivers anywhere near 1.22 TB/s with that AI Accelerator O100 it could be game changing. The claimed RING to BROADCAST smart AI switch is also a smart intuitive upgrade if executed as intended. To make any model utilize that multi-die layout it might need to pass through a Xiaomi proprietary compiler perhaps … and that may be OK or create the friction you’re concerned about.
Price, actual release date, true runtime and what its real-world performance will remain speculative until/ unless it launches.
If global RAM price doesn’t skyrocket further Xiaomi’s public 5% net hardware profit margin cap position could place entry units with smaller disk size at $2,500 - $3,500. It would be curious to see if with up to 160GB unified RAM, 1.22 TB/s Memory Bandwidth and Ring to Broadcast training component remain at the ~$5K mark or if they also join the 10K club.
& with Thermal Throttling quickly becoming a bottleneck (…our cooling cage thread is telling…) the marketed AI Cube’s Massive chassis equipped with 33,874 CNC cooling holes adds another appealing perk.
So until 2027 and when we can see actual product real-world numbers the AI Cube is a nice to watch prototype and an excellent multi-node lab component to have at the right price, network connectivity and software agnostic capability.
It is mind boggling. Xiaomi has a Flash model which is 310B parameters. If they release an AI box, then maybe their device should run their own model well, at least a 4bit quant with a reasonable context at 8bits. Don’t get me wrong, 160GB is much better then 128GB. Still, both 80GB and 160GB are very odd sizes. At least Qwen 3.8 Flash Next might run really well on 160GB device, not sure what 80GB is for. Qwen 3.8 27B + some image/video generation with comfyui? They have access to the cheap CXMT RAM, why not at least 256GB? Maybe they will do some sleek thunderbolt clustering. That will be the only saving grace.
I think the 80G figure was on that Reddit community thread where folks were speculating Xiaomi could launch a lower-tier binned version, or if only a portion of the memory architecture will be efficiently usable for real-time, heavy LLM token generation.
On the spec sheets I’ve only seen up to 160G circulating quoted from the 3-Chip Xuanjie O3, O100 and D100 architecture.
Xuanjie O3 — their flagship SoC
Built on a 3nm process with 24 billion transistors, the O3 uses a ten-core all-big-core CPU (six ultra-large plus four large cores) at up to 4.35 GHz, a 16-core G2-Ultra NX GPU, and 16 MB of SLC cache. Claimed as the first mobile processor to support LPDDR6, reaching 113.8 GB/s.
The O3 NPU delivers 200 TOPS of tensor compute and 3.13 TFLOPS of vector compute.
Xuanjie O100 — dedicated fast near-memory compute bandwidth AI accelerator at 1.22 TB/s
The first 6nm AI accelerator using wafer-on-wafer 3D stacking, with hybrid bonding at a 1.4 µm pitch, yielding 28,672 effective data lines and a 14-core high-bandwidth NPU delivering on-device inference of up to 330 tokens per second.
Xuanjie D100 — a 3nm automotive-grade AI chip with a 20-core CPU and 16-core NPU.
This is the capacity chip: it supports up to 160 GB of unified memory and, per Xiaomi, local deployment of models up to 200B parameters.
Xiaomi hasn’t shared how memory is partitioned across the three chips, what quantization the 120B model runs at, or how the 330 TPS figure was measured.
The 160GB may be a chip capacity ceiling vs. a question of RAM access.
With Thunderbolt clustering 160GB RAM and 1.22 TB/s that could be a fascinating game changer question is on availability the O3 should be out this year, the O100 and D100 are slated for commercial use next year before the AI Cube can come together for further prototyping and GA.
I found the 80GB tied to the August announcements on Weibo and IT Home
Even if they launch a 80GB … if at a reasonable price with Thunderbolt Clustering, I’d be paying close attention to what workloads it can support at single/ clustered
I’ve stopped comparing them across model → GPU → memory → kernel
And looking at a potential shift in fast intuitive distributed machine workload task → workload phase → compute substrate → memory locality → data-movement topology
80GB (160 GB max), access to 1.22 TB/s bandwidth, intelligent Ring / Wafer distribution, with potential clustering, makes it worthwhile to watch whether its architectural design unlocks new potentials.
At this rate of exponential growth, by the end of this year a Spark is going to cost somewhere between $20k and $160k.
And by the end of next year it’s going to cost [checks math] approximately somewhere between $1.28M and $2.68T (trillion – that’s if the price doubles twice a month, hopefully it won’t come to that). Or maybe it’ll cost the same but it’ll only have [checks math] 512 bytes of VRAM 🤔️
