A newly funded project claims a tenfold performance boost for domestic GPU clusters. The numbers don't add up. The product is a black box. And the Token Factory component—whatever it is—smells like a compliance trap waiting to spring.
Last week, QianVision Technology announced the launch of its "wylon" Super Node System, a turnkey AI infrastructure platform built around a collection of domestic GPUs from six Chinese vendors—Cambrian, Birun, Muxi, Xiwang, Haiguang, and Moore Threads. The press release, distributed through blockchain-oriented media channels, boasts of “over ten times comprehensive performance improvement” achieved through a proprietary operating system called HitenOS. It also mentions a mysterious “Token Factory” that is already testing in East China.
As a crypto security auditor who has spent the last decade dissecting smart contracts and DeFi infrastructure, I have learned one hard rule: Silence in the logs speaks louder than the code. And this announcement is filled with silence.
Let’s start with the numbers. The system is built from 72 GPUs per rack, four racks total—288 GPUs. That is a modest cluster by today’s standards. A typical large language model training run at a major lab uses thousands of H100s or GB200s. Even a mid-tier Chinese AI company like Baidu or Alibaba operates clusters in the thousands. So QianVision is not aiming for the frontier of foundation models. They are targeting mid-scale training, fine-tuning, and inference—the bread-and-butter of enterprise AI, especially in government and state-owned enterprises that are under pressure to decouple from NVIDIA due to US export controls.
But here is where the audit trail goes cold. The press release provides zero specific benchmarks. No training throughput in tokens per second. No inference latency at batch size 1. No MFU (Model FLOPS Utilization) numbers. For a 10x claim to be credible, you need a baseline. The unnamed baseline is probably “running the same workload on raw, unoptimized cards with no software stack.” That is not an honest comparison. In my experience auditing performance claims in DeFi—where projects love to tout “2x gas savings” that turn out to be measured against a straw-man contract—every exploit is a confession written in gas fees. Here, the exploit is a missing benchmark.
The core of the system is HitenOS, described as a “large model operating system” that optimizes GPU memory, scheduling, and inter-node communication across heterogeneous domestic accelerators. This is essentially an AI compute middleware—think of it as a custom orchestration layer that tries to compensate for the weak software ecosystems of each individual GPU vendor. That is a legitimate engineering challenge. Domestic GPUs typically lag far behind NVIDIA in software maturity: their drivers are buggy, their collective communication libraries are slow, and their memory management is primitive. A good middleware can squeeze real gains—maybe 2x in throughput, 3x in memory efficiency if the baseline is truly terrible. But 10x? That requires a miracle or a lie.
Precision kills the illusion of complexity. HitenOS claims to deliver “fast interconnects” and “hundreds of terabytes of dedicated cache.” But what is the actual interconnect topology? Are the 72 GPUs per rack linked via a proprietary NVLink-equivalent, or are they stitched together with PCIe 5.0 switches? What is the cross-rack bandwidth—100GbE RoCE? InfiniBand? These details determine whether the system can scale. A 4-rack cluster is one thing. If they try to push to 100 racks, the communication overhead will crush performance unless the network is engineered correctly. The press release is silent on this.
Now, the elephant in the room: Token Factory. The name alone triggers every red flag in a crypto auditor’s mind. It could mean a distributed training framework (like an AI version of Filecoin or Akash) that tokenizes compute resources. Or it could mean a DePIN (Decentralized Physical Infrastructure Network) project that issues a native token to incentivize GPU contributors. Either way, if Token Factory involves a token offering, the regulatory risk in China is existential. The People’s Bank of China has repeatedly banned crypto trading and ICOs. The State Council has classified virtual currency mining as a prohibited industry. A project that advertises a “Token Factory” in a blockchain news outlet is signaling that it intends to issue tokens, probably to raise capital from retail investors.
I have seen this pattern before. In 2021, I audited a DeFi protocol that claimed to be a “decentralized computing network.” The whitepaper was full of mathematical formulas and references to distributed systems. The code was a mess of copy-pasted Uniswap contracts with a governance token bolted on. The team disappeared six months after the token sale, leaving investors holding worthless bags. The lesson: Trust is the vulnerability they never patched. QianVision’s silence on the legal structure, the tokenomics, and the compliance status of Token Factory is a vulnerability that no amount of GPU optimization can fix.
Let’s also examine the competitive landscape. The domestic AI infrastructure market is dominated by Huawei Ascend, which offers a vertically integrated stack: own chips (Ascend 910B/910C), own AI framework (MindSpore), own servers, and own operating system. Alibaba Cloud has its own chips (Hanguang 800) and a mature PaaS platform. QianVision is an integrator—it has no chip, no framework, and no cloud platform. Its only differentiation is hardware neutrality: it claims to support six GPU vendors, giving customers an escape from lock-in. But that flexibility comes at a cost. Each GPU vendor has its own driver, SDK, and quirks. HitenOS must maintain compatibility across all six, which is a massive engineering burden. If one vendor updates its firmware, the whole stack may break.
Moreover, the scale of operations is tiny. A 288-GPU cluster is a drop in the ocean. Even if HitenOS delivers a perfect 10x improvement (which I doubt), the absolute performance is still far below what a single H100 node can achieve. The cost per token generated will be higher, not lower. The only customers who will buy this system are those who have no choice—government agencies and state-owned enterprises that are mandated to use domestic hardware. That is a finite market, and it is already being served by Huawei and a handful of other state-backed vendors.
Every exploit is a confession written in gas fees. In this case, the exploit is the 10x claim, and the gas fee is the trust of investors who believe in the narrative without seeing the receipts.
Now, the contrarian angle. What if QianVision is actually onto something? The domestic GPU ecosystem is fragmented. Companies like Moore Threads and Cambrian are trying to compete with NVIDIA but their software stacks are immature. A well-designed middleware that abstracts away the hardware differences and provides a unified programming model could be genuinely valuable. Huawei’s CANN is closed and vendor-specific. An open-source alternative would be welcomed by the community. If HitenOS is released as open source (the press release doesn’t say), it could become the de facto standard for domestic GPU programming, much like CUDA is for NVIDIA. That would give QianVision influence far beyond its hardware sales.
Furthermore, the Token Factory could be a legitimate compute marketplace that allows idle GPUs to be rented out by third parties, similar to Together AI or RunPod but using domestic hardware. If it’s a permissionless network that lets anyone contribute GPU cycles and get paid in tokens, it could unlock a new source of compute for Chinese AI startups that cannot afford cloud or dedicated clusters. The token would be a utility for measuring and exchanging compute, not a speculative asset. That is a defensible use case.
But I have seen too many projects abuse the term “utility token” to cover up a Ponzi. The burden of proof is on QianVision. They need to publish the Token Factory whitepaper, the tokenomics model, and a legal opinion from a reputable Chinese law firm confirming compliance. They need to release a full benchmark suite, including MLPerf results, with detailed hardware and software configuration. They need to open source HitenOS—or at least publish a developer API and documentation—to prove it is not vaporware.
Until then, my recommendation as a security auditor is simple: Verify everything. Trust nothing. The blockchain industry has a long history of promising revolutionary infrastructure only to deliver a token sale and a ghost town. QianVision’s wylon system may be different. But the absence of evidence is evidence of absence. The silence in the logs is deafening.
The takeaway is a rhetorical question: In a bull market where every project claims to be the next NVIDIA, who is auditing the auditors?