Article — "China Builds a Parallel AI Compute Stack on Huawei Silicon as Cash Burns": Four Huawei Ascend 384 SuperNode servers are being bolted together in the Wuxi high-tech zone — 384 cards each, four racks of 384 GPUs wired into a single super-cluster, 1,536 domestic accelerators treated not as a research bench but as a production line.
The people building it do not call it a data center. They call it a "Token factory," and they say this first phase ranks among the top such market-facing factories in China.
That word choice is the whole argument. Once compute becomes a "factory" and its output becomes a "manufactured good" flowing through what Wuxi's officials call a "Token价值流通体系" — a Token value-circulation system — the catastrophic cash burn underneath stops being a bonfire and becomes capital expenditure. The question this piece asks is whether that reframing is accounting or theology.
Once compute becomes a factory and its output a manufactured good, the catastrophic cash burn stops being a bonfire and becomes capital expenditure. Is that accounting or theology?
Last week the open question on this thread was whether Huawei could actually ship domestic silicon at volume — whether Atlas-class SuperNodes were real hardware or a roadmap. This week the question moves downstream: someone is buying them, wiring them, and betting a business on running them. The delta is a Token-factory buildout in Wuxi and a 40,000-server telecom tender that went overwhelmingly to ARM.
Start with the physical reality, because it is no longer aspirational.
The Wuxi factory's operator, Hongxin Electronics, booked more than ¥7 billion in 2025 revenue, of which roughly ¥2.8 billion — nearly forty percent — came from compute-related business. That is a components maker reinventing itself as a compute landlord in a single year.
The second data point is heavier. On June 26, China Telecom awarded its 2026–2027 high-performance server procurement: 40,000 servers, roughly ¥11.5 billion. Split into 28,000 ARM-A servers and 12,000 C86 servers, with Sichuan Huakun Zhenyu — reportedly a Huawei Kunpeng integrator — taking the top slot at ¥8.149 billion.
An ARM/Kunpeng-dominated architecture bet, made by a state telecom carrier, at eleven figures. This is what decoupling looks like once it stops being a slogan: a purchase order.
And the other two carriers are moving the same way — China Mobile has tendered for nearly 62,900 PC servers this year. The demand for domestic compute is not coming from a lab. It is coming from state-owned infrastructure buyers with budgets.
Now the supply-chain evidence, which is where the ecosystem shows it can feed itself.
The profit is rippling up and down the stack. Cambricon, the domestic AI-chip designer, posted 2025 earnings growth of 555.24%. That is not a rounding artifact; that is a chip company that has found a market.
The more telling number is one layer down, at the integrators — the companies that bolt the cards into racks. Unisplendour (紫光股份) grew Q1 net profit 126.06% year on year. Inspur's revenue hit a record ¥164.8 billion, up 43%, with contract liabilities and inventory both ballooning — the signature of a full order book.
When profit migrates from the scarce component down into the assembly layer, that is the signature of an ecosystem that has stopped depending on a single hero product and started reproducing itself. The whole chain — chips, optical interconnect, PCBs, integration — is now making money at once.
So the industrial base exists. That is settled. The fault line is financial, and it runs straight through the biggest buyer of all.
Alibaba's fiscal Q4: revenue ¥243.38 billion, up a modest 3%. AI-related product revenue ¥8.97 billion, its eleventh straight quarter of triple-digit growth, an annualized run rate above ¥35.8 billion.
And free cash flow of negative ¥17.3 billion for the single quarter. Adjusted EBITDA down 84% year on year, against a capital-expenditure plan already above ¥380 billion that CEO Wu Yongming hinted would grow larger still.
Revenue accelerating; cash incinerating faster. That is the picture, and Wu did not hide it. He reframed it.
"AI is more like manufacturing," he told analysts. "To earn more revenue you must build two core factories — an AI training factory and an AI inference factory."
This is the same metaphor now painted on the wall in Wuxi, spoken from the top of China's largest cloud. Wu's version is not Jensen Huang's shovel-seller pitch; it is a justification for why Alibaba Cloud's margins are thin and it is spending anyway.
The logic is seductive and internally coherent. In manufacturing, heavy up-front capex maps to future output and future revenue with reasonable predictability — so it is "building capacity," not "burning cash for scale," the internet playbook that famously never penciled out.
Wu even hands you the option value. The same server deployed today costs more than double what it cost two years ago — so the data center Alibaba builds first carries a built-in "price-appreciation option," and every latecomer must chase it with more expensive steel.
Whoever's factory finishes first locks in the revenue. That is the pitch. It is a good pitch.
But notice what the entire structure rests on: that a Token is a genuine manufactured good with a predictable output-to-revenue conversion. Steel mills know what a ton of steel sells for. Does anyone know what a trillion tokens reliably earns?
Alibaba's own tell is in the numbers it chooses to celebrate. Model-and-application ARR crossed ¥8 billion, targeting ¥30 billion by year-end — a claimed 10x-plus jump since last November, with token consumption growing sharply quarter over quarter.
If that curve holds, ¥380 billion is capacity, not combustion. If it bends, ¥380 billion is a glut with a very expensive cost basis.
That is the unresolved core. The unit economics here are being asserted, not demonstrated. Wu's three "hidden margin levers" — higher-margin model-API services, rising tokens-per-card from inference optimization, and cheaper in-house T-Head silicon — are all forward promises, forecast to lift cloud gross margin "significantly" over the next one to two years. None of them is a printed result yet.
And the doubling server cost cuts both ways. It rewards whoever built early; it punishes anyone whose demand ramp arrives after the capacity does, because they will be servicing depreciation on the most expensive vintage of hardware ever deployed while the tokens trickle in.
For the domestic stack specifically, this is sharper than it is for a US hyperscaler, because much of the demand is being manufactured by the state. The China Telecom tender, the carrier server buys, the Wuxi factory launched hand-in-hand with the municipal data bureau — this is allocation, not a market clearing on price.
State allocation can guarantee that the servers get bought. It cannot guarantee that the tokens they produce get sold at a margin. Those are two different questions, and the industrial base only answers the first.
So the parallel stack is now real enough to be self-reproducing — the chips profit, the integrators profit, the carriers buy, the factories rise. That much is no longer in doubt.
What is in doubt is whether it is a business or a bet dressed as a business. The "AI-is-manufacturing" frame converts a negative-¥17.3-billion quarter into a virtue — but only if token demand compounds faster than the doubling cost of each new server.
Absent that, the frame is theology, and the state is subsidizing inventory. Whoever's cash outlasts the demand ramp wins. The rest are capacity glut with a Huawei logo, waiting for a buyer who was assumed rather than found.