I reverse-engineered the TiinyAI Pocket Lab from marketing photos

I wanted the TiinyAI Pocket Lab to be real.

I just finished a five-month painful journey with NVIDIA’s DGX Spark, a $3,999 “personal AI supercomputer” that turned out to be bandwidth-kneecapped to protect NVIDIA’s datacenter business. I wrote about it here. Long story short, NVIDIA hid the one number that mattered, 273 GB/s memory bandwidth, behind impressive-sounding specs like “128GB unified memory!” and “1 PFLOP FP4!” I trusted them, I bought it, I got played and I sold it at a loss.

That was a tough experience that definitely taught me not to trust marketing blindly. After being lied to once, I learnt to look for what’s not on the spec sheet. The numbers they lead with are the ones that might not matter that much, kinda like when a magician is doing misdirection. The numbers they don’t tell you are the ones that matter the most.

So after selling the DGX at a loss I was back on the horse searching for a solution that would allow me to run inference locally. Just like everyone else I wanted the biggest memory pool (ideally) for the lowest cash. That is when I found a startup called TiinyAI claiming they’d built a pocket-sized device that runs a 120-billion-parameter model at 20 tokens per second for $1,299. My first thought? Too good to be true. I opened a notepad, then I opened their website and looked at their marketing photos, their spec sheet and got to work. How does that go again? Fool me once, shame on you. Fool me twice shame on me? I was not going to be fooled twice. The first thing I thought was: 120B parameters on LPDDR5X? C’mon now. The DGX Spark has 128GB of unified LPDDR5X at 273 GB/s and I just spent five months learning why that wasn’t enough to run 70B models at interactive speeds. You’re claiming 120B on the same memory technology. What is this Voodoo magic you have…

The next few days became a scavenger hunt trying to piece together how does this device work, what is all this tech they claim to be so ground breaking and how can they defy the physics that NVIDIA were not able to fool.

What I found was three layers of technical misdirection stacked on top of forked academic research, wrapped in a corporate structure designed for opacity, and sold through a Kickstarter campaign with a $10,000 funding goal that has now raised $1.7 million from 1,266 backers.

Let’s get started.

Part 1: The Promise

TiinyAI’s pitch is simple, the Pocket Lab is “the first pocket-size AI supercomputer.” It plugs into your laptop via USB-C, has 80GB of LPDDR5X RAM, an ARM SoC with a 30 TOPS co-located NPU, a 160 TOPS dedicated external NPU (190 TOPS total), and can run large language models with up to “120 billion” parameters locally. No cloud, no GPU, no subscription fees. Zero token costs, always on, private and offline.

The Kickstarter launched March 11, 2026. It hit $1 million from 728 backers in its first five hours. As of today, it sits at $1,737,722 from 1,266 backers, 17,377% of its $10,000 goal. The super early-bird price is $1,399. If you placed a $9.90 deposit on their website beforehand, you get it for $1,299. Estimated delivery: August 2026.

They have a Guinness World Record awarded for the “smallest mini PC (100B LLM Locally).” They were featured at CES 2026. Their press release, distributed through Hong Kong, quotes Samar Bhoj, “GTM Director,” talking about “edge-cloud synergy” and introducing “AgentBox” as a new hardware category.

The press coverage has been uniformly positive. Mashable, TechRadar, WCCFTech, Benzinga, Geo News. All echoed the “120B” and “20 tok/s” claims like parrots without doing any due diligence that would take 5 minutes and raise more questions than answers. None spent two minutes on a LinkedIn search that would have revealed the “hottest AI hardware startup in US” has four employees: a hidden VP, a GTM director with no career history, a VC analyst in Hong Kong, and a Hong Kong intern.

TiinyAI’s website and Kickstarter page now display a “Trusted by Top Media” banner listing these publications by name. Write a press release, pay for PR Newswire distribution through Hong Kong, watch outlets reprint it without fact-checking, then reference the coverage as third-party validation. This is how credibility laundering works.

I did the math.

Part 2: The Hardware They Won’t Name

What’s missing when you read TiinyAI’s marketing materials? They don’t tell you what SoC is inside the device.

Every legitimate hardware company tells you what’s under the hood. Apple tells you it’s M-series silicon. NVIDIA tells you it’s GB10. Qualcomm tells you it’s Snapdragon. Even budget SBC makers like Orange Pi and Radxa prominently name their processors. TiinyAI tells you the Pocket Lab has “an Armv9.2 CPU” and “a dedicated NPU with up to 160 TOPS.”

But they published photos. And I want to be very clear about sourcing here: every piece of hardware identification in this post comes from TiinyAI’s own publicly shared marketing materials and social media posts. Nobody leaked schematics or broke an NDA or sent me confidential docs. I looked at what they voluntarily showed the world, and I read it with 20+ years of engineering experience.

The SoC: Process of Elimination

ARMv9.2 is bleeding-edge. Very few SoCs in production implement this ISA. I went through every candidate I could find:

Qualcomm Snapdragon X Elite: 12-core ARMv9.2. Matches the core count. But it’s Windows-exclusive silicon with heavy Qualcomm branding. If you had a Snapdragon inside, you’d market it. Eliminated.

MediaTek Dimensity 9300: ARMv9.2, but only 8 cores. Wrong count, mobile SoC, can’t support 32GB LPDDR5X. Eliminated.

Apple M-series: Custom ARM cores, not licensable, Apple-only. Eliminated.

Ampere Altra: ARMv9.0, 128-core server chip. Wrong core count, wrong power envelope, wrong ISA revision. Eliminated.

That leaves exactly one chip: the CIX P1 (CD8180) from CIX Technology, a Shenzhen-based fabless semiconductor company. It’s the only publicly announced 12-core ARMv9.2 SoC in production.

My confidence level is very high, there literally isn’t another chip matching these specs.

This is not exotic hardware. You can buy this exact SoC in a Radxa Orion O6 mini-ITX board for $200-300. Minisforum sells the MS-R1 mini PC with this chip and up to 64GB of LPDDR5X for about $500-700. Initially I thought that perhaps they used off-the-shelf components but the SBC boards that integrate this chip are too big to fit in the enclosure.

I want to be clear, the CIX P1 is a perfectly fine ARM chip for general computing and light AI workloads. It is not, by any stretch, a “supercomputer.”