tiny corp is shipping tinybox red v2 at $12,000 with four 9070 XT GPUs and 64 GB of GPU memory, alongside higher-end Blackwell systems. Buyers are weighing the bundled tinygrad stack against DIY rigs, model-fit limits, and cloud economics.
Posted by albelfio
The tinybox is a high-performance AI compute system sold by tiny corp for deep learning training and inference. Available in variants: red v2 (4x 9070XT GPUs, 778 TFLOPS FP16, 64GB GPU RAM, $12,000), green v2 blackwell (4x RTX PRO 6000 Blackwell GPUs, 3086 TFLOPS FP16, 384GB GPU RAM, $30,000), and upcoming exabox (~1 EXAFLOP). Features include 32-core AMD CPUs, up to 192GB system RAM, Ubuntu 24.04 OS, rack-mountable designs, now shipping worldwide.
The shipping announcement is straightforward. According to the tinybox page, red v2 is the entry configuration: four 9070 XT GPUs, 64 GB of aggregate GPU memory, 778 TFLOPS FP16, up to 192 GB of system RAM, and a rack-mountable Ubuntu 24.04 box for $12,000. The same page also positions the line upward, with a $30,000 green v2 Blackwell configuration and an “upcoming exabox” described as roughly 1 exaflop.
What makes this more than another small GPU server launch is the software angle. The thread summary says engineers are treating Tinybox as a test of whether bundled hardware plus tinygrad integration is worth paying for versus assembling comparable parts yourself. That framing got sharper in the newer fresh discussion, which says the main appeal may be a vertically integrated, “just works” setup that avoids CUDA and PyTorch toolkit friction.
Posted by albelfio
For AI engineers, the thread is mainly a reality check on local compute economics: whether Tinybox’s bundled hardware and tinygrad integration justify the premium over DIY rigs, colo, or cloud, and whether the advertised configurations can plausibly serve the model sizes implied.
The pushback is equally concrete. In the community highlights, one commenter said “there is apparently no reason whatsoever to buy this hardware” given the pricing, while another argued there is “no way” the red v2 is doing meaningful work with a 120B model. The same discussion also questioned the exabox claim that it can “function as a single GPU,” which leaves the most ambitious part of the roadmap looking more like an architectural question than a settled product detail.
Posted by albelfio
Thread discussion highlights: - gymbeaux on pricing and support: $12,000 gets you 1Gb/s networking and vanilla Ubuntu 24.04... margins are around 50%... there is apparently no reason whatsoever to buy this hardware, even if you plan on using tinygrad exclusively - bastawhiz on hardware sizing and model feasibility: There's no way the red v2 is doing anything with a 120b parameter model... I'm also confused why this is 12U... For $65k, I'd expect a much better CPU and 256gb of RAM. - andai on exabox architecture: Can someone explain the exabox? They say it "functions as a single GPU". Is there anything like that currently existing?
Posted by albelfio
Today’s new signal is less about raw specs and more about the software-hardware product story: one commenter framed the main value as tinygrad’s vertical integration, arguing the appeal is a “just works” environment that avoids CUDA/PyTorch driver and toolkit friction. That adds a clearer rationale for the product beyond benchmark comparisons. There’s also a fresh curiosity around the exabox and how it is supposed to “function as a single GPU,” suggesting interest in whether the roadmap is a real architectural step or just branding for a clustered system.
