China’s Green Token Factory: Where the Unit of Account Is a Token, Not a Tokenised Asset
Seven Chinese parties — including Alibaba Cloud, Wangsu, a State Power Investment subsidiary and Guotai Haitong’s research arm — published a technical whitepaper on 20 October for turning green power into AI tokens, with a five-layer stack and a new efficiency metric, TPW (tokens per watt). There is no blockchain in it, and that is the point worth understanding. China’s route to monetising renewable energy is not securitising a power plant’s cash flow; it is converting the watt into a metered, dispatchable service unit and selling that. It is a different asset class, with a different legal requirement — contract law and metering standards rather than securities law.
What was published, and what the word means
At the 2026 Physical AI and World Model Conference in Nanjing on 20 October, seven organisations released the 2026 Technical Whitepaper on the Green Token (Token) Factory: Cathay Smart Computing (Zhejiang), Power Investment Integrated Energy (Jiangsu), Shantie Digital Technology (Shanghai), Alibaba Cloud, Wangsu Science and Technology, the Guotai Haitong Securities research institute, and DeepEngine (Changzhou). The document sets out a five-layer vertical architecture — energy infrastructure, computing infrastructure, token production, token operations, token circulation — explicitly framed as a value chain running from a watt of green electricity to a standardised, metered, traceable token.
The first thing a reader from this industry needs to fix is vocabulary. Here, "token" means the billing unit of an AI model — a word-piece — not a blockchain asset. "Token circulation" means commercial delivery with metering, provenance and carbon labelling, not secondary-market trading. There is no ledger, no smart contract and no transfer agent anywhere in the architecture. Reading it as an RWA document will produce confident and entirely wrong conclusions.
The second thing worth noting is the metric. Traditional data centres measure electricity with PUE and compute with FLOPS; neither answers the question the whitepaper poses, which is how much intelligence each watt ultimately becomes. So it introduces TPW — tokens per watt — and makes it, alongside the green-power consumption share and carbon intensity, one of three defining criteria. The text is careful to add that these criteria apply "rather than any specific physical form of power supply", which quietly loosens the assumption that green direct connection must mean a dedicated line from a nearby solar farm.
That last clause is the most commercially consequential sentence in the document. It converts a debate about grid topology into a debate about measured outcomes, which is the only version of the debate that can be put into a procurement contract.
The numbers that make the argument
The demand figures are the ones to hold onto. According to the National Data Administration, China’s average daily token calls passed 140 trillion in March 2026; by August, CCTV reported the figure above 500 trillion. Inference now accounts for more than 70% of compute demand, and inference spending overtook training during 2026. Electricity is more than 30% of the production cost of an inference token — the single largest variable cost. The China Academy of Information and Communications Technology puts the growth rate of data-centre electricity demand at 3.6 times that of total social electricity consumption.
Those four numbers explain why this stopped being an engineering question. If power is a third of unit cost, then the price of a kilowatt-hour is not an input to the business — it is the business. Wholesale prices in northwestern provinces run at roughly RMB 0.10 to 0.20 per kWh; the same electricity sold to a computing centre fetches RMB 0.31 to 0.33. That spread, not the token, is the actual commercial motive for coupling compute to power, and it was stated plainly on the same day by Tian Dayong, president and COO of Bicheng Energy, speaking at a sustainability conference in Shanghai.
The policy timeline is equally load-bearing. "Compute-power coupling" was written into the Government Work Report for the first time in March 2026, and in May four ministries jointly issued an action plan on two-way empowerment between AI and energy, making the green-power consumption ratio a reference indicator, encouraging direct green connection at computing facilities and pushing for local consumption of new energy. A whitepaper published five months after a four-ministry action plan is not an academic exercise; it is an attempt to standardise before procurement rules harden.
It is worth being precise about what these figures do and do not establish. They establish that the load is large, growing fast and price-sensitive to electricity. They do not establish that any individual token factory earns a return, because return depends on the price of a token, and the whitepaper is silent on token pricing.
Two city-level templates, and what they measured
Changzhou is the reference case: a 4 MW green token factory signed in July 2026 into the Changzhou high-tech zone, co-built by seven parties including the municipal data bureau. Its distinguishing features are distributed solar plus fishery-solar hybrid multi-source green direct connection, solid-state transformers, phase-change liquid cooling and a physical-AI model for scheduling. Reported outcomes: green-power share steady above 50%, compute utilisation lifted from the industry-typical 30–40% to above 65%, and annual output projected at no less than 60 trillion tokens.
Haining is the second: a 7 MW facility co-built by Cathay Smart Computing, Wangsu, Shantie Digital and DeepEngine, positioned as a Yangtze River Delta demonstration, with an expected full-year green-power share above 80%. The engineering specification is unusually concrete for a document of this kind — 0.5–1 MW N-type TOPCON arrays, lithium storage sized for 0.5–2 hours of backup, an 800 V HVDC bus, solid-state transformers, per-rack metering via intelligent PDUs, full-chain efficiency of at least 96%, dual 10 kV feeds plus N+X diesel generation.
Two caveats belong next to those numbers. First, the Changzhou figures are projections attached to a signed project, not audited operating results — the 60 trillion tokens is an annual output estimate, and the utilisation improvement is a design target benchmarked against an industry range. Second, and more important, utilisation and green share are efficiency metrics. They say a factory produces more tokens per unit of constrained input; they say nothing about whether anyone buys those tokens at a price that covers the capital charge.
The document is candid about this gap in a way that most infrastructure whitepapers are not. It lists the unresolved problems directly: the intermittency of wind and solar output, barriers to cross-regional green-power trading, the absence of cross-industry standards, and the security risk of scheduling across the IT and OT domains. A vendor consortium naming its own unsolved problems is a better signal of seriousness than a projection of total addressable market.
Why this matters to an RWA reader
Read the whitepaper as an asset document and something useful appears. This site has spent months documenting a recurring failure: tokenised energy projects that put a cash-flow claim onchain before anyone had verified the cash flow, and then discover that verification, not the chain, is the binding constraint. The Chinese document attacks the same problem from the opposite end. It builds the metering, the provenance, the carbon label and the delivery standard first, and defers the question of whether the resulting service unit can be financed.
Tian Dayong’s remarks the same day pushed that deferral further than the whitepaper does. Compute centres, he argued, are moving from centralised to distributed — "much like renewables a dozen years ago" — with fragmented, movable compute cabinets absorbing surplus distributed green power. On the asset side he described two directions: compute companies building their own generation because self-supplied green power has a higher return than the compute centre itself, and power companies moving from selling electricity to selling tokens, where the value is expected to be eight to ten times the wholesale price. He then used the language of structured finance almost exactly: compute and power may eventually be accounted as "one book", with the entity below it fragmenting into compute assets, power assets, hardware assets and software assets, for equity products and fund products.
"Asset fragmentation" is securitisation vocabulary. What is missing is the three things that turn fragmentation into a security: a legal form for the token service unit (a service contract? a revenue right? a usufruct?), a register and custody arrangement (who books it, who holds it, what happens on default), and performance guarantees (how an SLA breach is compensated). The whitepaper’s fifth layer gives metering, provenance, carbon labelling and circulation. It does not give legal form.
That gap defines the difference between the two routes now running in parallel. The OpenWorld/Meridian structure documented here last week treats physical assets — Tesla Powerwalls, solar panels, GPU nodes — as collateral and issues onchain financing against them; it needs securities law. The Chinese route converts green power into a service unit and sells the service, carrying the green attribute as a price premium; it needs contract law and metering standards. Both produce investable cash flow. Only one, so far, produces a tradeable instrument.
What to watch from here
First, whether TPW reaches a procurement document. A whitepaper metric becomes economically real the moment a cloud provider or model vendor writes it into a tender — "green-power share above X%, carbon intensity below Y grams per kWh, TPW above Z". Until then it is a design aspiration. Watch the tender documents of state-owned computing buyers, not the conference slides.
Second, whether Changzhou publishes a full operating year. The 60 trillion token figure and the 50%-plus green share are projections. A first audited year — actual output, actual green share, actual utilisation, actual electricity cost per million tokens — would be the first hard datapoint this model has produced anywhere. If those numbers diverge materially from the projections, the template is weaker than it looks.
Third, how the Haining 80% is verified. Direct green connection, green certificates and a power purchase agreement are three different claims with three different carbon attributes, and the whitepaper is clear that it is the criteria rather than the physical form that matter. If "green token" is to command a premium from an AI buyer with its own Scope 2 reporting, the provenance method has to be specified, not asserted.
Fourth, and the one that would actually connect this to RWA: the first financing product secured on token service units. Not a tokenised asset — a lease, a bank facility, an ABS, anything whose collateral is a contracted stream of metered AI tokens with a stated green share. That is the moment the two literatures meet. Until then, this is an unusually competent engineering document about how to build the factory, and a financing story still waiting for its first term sheet.
Related analysis
China Puts a Number on Its Carbon Credit Market — and Four Words on Cross-Border Trading
At a State Council Information Office briefing on 21 October, Vice Minister of Ecology and Environment Li Gao put the voluntary carbon market on the record: 44 projects registered with over 25 million tonnes of reductions, more than 24 million tonnes traded, turnover approaching RMB 2 billion, 18 methodologies published with 3 more imminent and 11 out for comment, and full coverage of key sectors targeted for 2027. He also gave the single most useful number this site has found all year: a 50 MW concentrated solar plant in Qinghai whose crediting-period revenue from CCER sales comes to nearly RMB 100 million. On cross-border trading the language is precise and deliberately modest — study, formulate, rely on the domestic registry, propose an alliance.
The Second Half of Tokenisation: Hong Kong Opens 24/7 Secondary Trading in Tokenised Public Funds
On 26 October 2026 the SFC announced its first authorisation for tokenised funds offered to the public to trade on the secondary market of a licensed virtual asset trading platform. Two tokenised money market funds from China Asset Management (Hong Kong) — the HKD and USD digital money market funds, pairs CAMFF-HKD and CUMFU-USD — began round-the-clock trading through PantherTrade, the Futu-owned licensed platform, with retail investors eligible. The tokens are not new instruments and cannot leave the platform; ownership stays in an off-chain register and transfers are recorded on the venue's ledger. Which makes this precisely the right thing to look at: for the first time in Hong Kong's tokenisation build-out, a tokenised instrument can trade away from its net asset value. That is the difference between a product and a market.
Three Days After "Successfully Issued", Another Agreement: Dragon Rise Signs a Non-Binding RWA Framework — With an Unnamed Partner
On 23 October, Dragon Rise Group Holdings (6829.HK) — the Hong Kong engineering-turned-new-energy company whose battery-swap taxi tokenised note this site examined on 20 October — announced that its wholly-owned subsidiary Dragon Rise Investment has signed a non-binding strategic cooperation framework agreement with an unnamed "new energy enterprise" to pursue RWA financing, diversified funding and asset management. Three days earlier, the company had declared the first tranche of its tokenised note "successfully issued". The three hardest follow-up questions from that article — was the HK$30 million first tranche fully subscribed, does the second tranche keep the parent guarantee, and what are the KPIs behind the 6 per cent performance interest — remain unanswered. What arrived instead was another agreement: vaguer, unpriced, and with a counterparty that does not even have a name.