Credibility & Infrastructure
We publish what the hardware said, not what the narrative needs.
SigPulse exists because China’s AI and hardware reality is reported mostly from press releases, secondary summaries, and synthetic guesses. We invert that: claims are validated on physical machines we own, in the places they operate, and every article carries the config that lets you re-run the test.
Who we are
SigPulse is run by a small research desk working across the China–global boundary: bare-metal GPU labs in Shenzhen, cloud nodes in the US, and field work inside China’s payments and connectivity stack. Our background is cross-border technology — taking frontier Chinese models and hardware claims, and testing them the way a Western infrastructure team would: versioned, reproducible, on the record.
We are independent. No vendor pays for coverage. When a number in an article is ours, the article says which machine produced it, on which date, with which flags — and when a number is someone else’s, we cite it and say why we trust it or don’t.
Hardware & testbed disclosure
Full inventory behind the “Proof of Execution” tags. If an article cites hardware not on this list, treat that as an error and tell us.
| Node | Role | Specification | Status |
|---|---|---|---|
| AWS US-East node | Editorial, routing & telemetry edge — clean non-CN egress for citation-grade sources | 2 vCPU · 4GB RAM · Ubuntu 22.04 | Always-on |
| Shenzhen Lab · Station A | Primary model-inference testbed — throughput, VRAM & power measurements | RTX 4090 24GB (450W envelope) · 16-core host, 96GB DDR5 · Ubuntu 24.04 · PyTorch 2.5 | Active |
| Shenzhen Lab · Station B | Quantization testbed — long-run stability, thermal soak, multi-runtime A/B | RTX A4000 16GB · Ubuntu 24.04 · llama.cpp / Ollama / vLLM matrix | Active |
| Field kit (Shenzhen / Shanghai / Beijing) | Payments, connectivity & KYC field protocols — merchant-terminal samples, SIM/KYC runs | Passport-based accounts · corporate & personal eSIM · 2026 wave | Continuous |
Methodology, in one paragraph: each benchmark is a median of ≥3 runs on a cold cache; power figures are board-power readings (not TDP); prompt/response lengths and runtime versions are pinned in the article’s replication appendix. Raw markdown of every article is exposed at /posts/<slug>.md for diffing and citation.
Advisory & B2B inquiries
We advise executives and research teams on China AI capability claims, local compute economics, and hardware procurement diligence — and run commissioned reproductions of vendor benchmarks on our testbeds.
Write to intel@sigpulse.com — signal over noise, links to numbers. We answer every substantive brief within two business days.