Nick AI

tensorcash:native might be one of the most misleading "decentralized AI" narratives I've looked into. I got this shilled over and over again, so I did a deep dive, especially after pearl-2:native, when everyone started looking for alternatives bc facts > everything else. 42% of circulating supply came from the first 5,004 blocks. Two addresses mine 77% of the network. Those same two dominate model-admission votes. Verification can fall back to a centralized server. And the supposed useful work? There's no clear evidence of meaningful paying demand. I went through the chain data and the code. Here's what I found, based on a snapshot from October 27. 1. Two addresses control almost the entire network. Over the last 7 days, across 1,081 blocks, two addresses mined 40.9% and 36.2% of all blocks. That's 77% between two addresses. Add a third, and you're at 88%. Every public pool combined barely makes a dent. Sure, anyone can technically connect a miner. But permissionless mining doesn't mean much if the vast majority of blocks are already being produced by a handful of operators. 2. Those same miners effectively decide what gets mined. The process for admitting the new 27B model required 50 support commits within 100 blocks. 44 of those 50 came from the same two dominant miners. And it doesn't stop there. The reference audit service requires human approval before a model passes. So the supposed miner vote comes down to two dominant operators and a centralized gatekeeper. The 27B model they voted in also receives a 20x target adjustment and requires roughly 54GB of model weights. That's well beyond what a single 24GB GPU can handle without quantization or other memory-saving techniques. The people deciding which model gets admitted are also the people dominating the mining process. Make of that what you will. 3. Verification has a centralization problem, too. In v1.2.4, the code falls back to verify.tensorcash.org when no primary verifier is configured. That server also determines which detector policy applies. Meaning nodes using this fallback aren't operating with fully independent verification. A central service becomes part of the process that determines which chain state they accept. That's a pretty important detail for a project selling decentralization. 4. Now let's talk about the "no premine" narrative. Blocks 1 through 5,004 paid out approximately 1.86M TSC. That's roughly 42% of the entire circulating supply today. The developers have stated in an interview that they mined early alongside early adopters. Four addresses associated with those early blocks also co-signed an official project statement. Meanwhile, 5,003 of those 5,004 early addresses have never spent a single coin. To be clear, I can't prove who owns each address, and I can't prove those coins will ever be sold. But 1.86M TSC sitting in early, largely untouched addresses is a massive potential supply overhang for a token trading around $48K in daily volume. Calling something a fair launch doesn't make the distribution fair. 5. The "20K funded addresses" metric is misleading. The first 5,004 blocks paid out to 5,004 fresh addresses. One address that did spend its coins took a 100 TSC chunk and split it into 1,500 tiny outputs. That's how you end up with impressive-looking address counts. Wallets aren't holders. Outputs aren't users. And neither tells you how much independent demand actually exists. 6. Where is the actual demand for the work? This is probably the biggest issue. When real traffic is missing, the miner generates synthetic prompts to keep the system running. So a high number of "tokens computed" doesn't necessarily mean customers are paying for inference. It can simply mean the GPU is processing work generated by the system itself. That's not revenue. That's activity. There's also a potential problem with confidentiality: if a winning proof reveals enough information to reconstruct the prompt and answer, that raises serious questions about whether sensitive inference workloads can safely use the network. The entire thesis depends on useful work. Where are the paying customers? 7. The market structure makes the supply situation even more interesting. At the October 27 snapshot: Price: approximately $5.11 Market cap: approximately $22.6M FDV: approximately $108M Daily trading volume: approximately $48.5K Daily issuance value: approximately $45K New issuance alone is worth almost the entire reported daily trading volume. That's before accounting for existing holders who may eventually want to sell. The three dominant miners' payouts also flow to three receiving addresses that have spent together across multiple transactions. That could indicate shared custody or a common exchange desk, although the transactions alone don't establish ownership. In a market this thin, these details matter. So what's the actual picture? I can't prove the founders control today's dominant miners. I can't prove the early holders are insiders, and the untouched early coins haven't necessarily been sold. Those distinctions matter. But the observable facts still raise serious questions. Two operators produce the overwhelming majority of blocks. Those same operators dominate model-admission voting. Verification can rely on a central server. Early issuance accounts for roughly 42% of circulating supply. And the project's headline metric of computational activity doesn't establish real customer demand. At some point, you have to separate the marketing from the mechanics. Based on the data I reviewed, TensorCash looks a lot less like a decentralized, fair-launch useful-work network and a lot more like a highly concentrated mining operation with a token attached. And if you're buying tensorcash:native based on the promise of decentralized compute, you should understand exactly what you're betting on. Overall, TensorCash isn't the only decentralized-compute play, and plenty of smaller-cap projects already handle these problems better. Crynux is one example. Rewards are issued only for confirmed jobs, so synthetic mining earns nothing. That doesn't prove paid demand, but it ties rewards to actual work rather than inflated activity. Applications pick the model they need via Hugging Face IDs, instead of miners deciding what runs. TensorCash can add models too, the difference is who chooses. Quantization lets larger models run on modest hardware. One example is Qwen3.8-27B at 4-bit, which needs a 24GB minimum and is live on five nodes. Community checkpoints run on the network as well. There are OpenAI-compatible APIs with streaming, tool calling and vision, so developers can build on it today. Founder Aaron Yu spent ten years at Google on Gemini and TensorFlow inference, though the project isn't Google-backed. None of this is solved. Verification, security and real demand still have to show up. The point is simple: if a network's value rests on mining activity rather than verifiable work, compare it with what its cheaper competitors are already shipping. I'm much more interested in networks where useful work is the foundation of the incentive model, rather than something the marketing has to convince you is happening. Do your own research. Check the chain data, read the code, and verify the claims yourself before buying into either narrative.