Power delivery is more than campus electricity
Server-level power delivery and compute-energy coordination are different engineering challenges. The October US patent investigation into vertical power delivery is a potential supply-chain risk, not an established interruption of Chinese deployments.
30 SEPTEMBER 2026 · ARCHITECTURE STRENGTHENS
Compute and energy are being co-designed physically, not only coordinated after deployment.
A HyperStrong / 海博思创 project in Baotou has begun construction around grid-side storage coupled with distributed intelligent-compute infrastructure. Xinhua says the design connects energy supply, storage peak-shaving and compute scheduling and uses a dynamic “算随电走、电随算调” mechanism to respond to renewable-power conditions.
Research implication: this extends the thesis from bounded workload shifting toward infrastructure designed from the outset as a joint compute–energy system. It remains construction-stage evidence: operating utilisation, delivered electricity cost, reliability and project returns are not yet established.
25 SEPTEMBER 2026 · THESIS STRENGTHENS
From pilot evidence to bounded operational demonstration.
State Grid Shanghai reports a coordinated two-hour event across 16 data centres, with a maximum 97.8MW load reduction, combining flexible-load adjustment, backup-power switching and cross-region compute transfer. A linked three-day electricity-price-aware inference test across locations including Shanghai, Anhui and Xinjiang handled more than 280,000 requests and more than 33 million tokens; the reported electricity cost per token fell by about 31%.
Research implication: electricity price and grid conditions are now evidenced as inputs to workload placement in a bounded operating test. This strengthens the compute–energy edge of the architecture. It does not establish national repeatability, all-workload portability or net economics after network, software and migration costs.
A separate Zhejiang report says part of a Jiaxing workload was shifted to Qinghai during peak periods, reducing reported token carbon intensity by 62%, with migration completed within three minutes and no service interruption. Treat these as reported operating results, not universal benchmarks.
22 SEPTEMBER 2026 · EARLIER PILOT EVIDENCE
Fujian → Ningxia: the earlier pilot that first made the mechanism concrete
The 16 September exercise, reported by State Grid on 22 September, connects workload relocation with electricity-demand changes and solar availability. This adds an operating example to the compute–energy thesis alongside Alibaba’s capacity ambition.
Research implication: examine who can coordinate workloads and energy economically. The pilot establishes no VNET–CATL participation or funding; scale, service levels and returns remain open.
Read the evidence and implications →22 SEPTEMBER 2026 · THESIS INTERPRETATION
Compute and energy now share centre stage
Alibaba’s ambition to exceed 20 GW of global data-centre capacity by 2032 makes energy a central delivery question alongside chips and compute. Reported announcement · Reuters, 22 September.
Our interpretation: the research must assess compute capability together with access to reliable, economically viable power. A capacity plan only becomes useful infrastructure when sites can be connected, energised, cooled and operated at a return customers will support.
Can it be energised?
Track secured grid connections, connection dates and power availability at each location against the compute deployment timetable.
Can it run economically?
Test delivered electricity costs, cooling overhead and reliability alongside utilisation and revenue. Storage economics and asset ownership need project-level evidence.
Who captures the value?
Identify who funds and owns energy assets, who bears power-price risk and whether the operator retains the resulting savings.
Capacity is not consumption: the target concerns total global capacity, not an additional 20 GW or a China-only build-out. The reporting does not resolve whether the measure refers to IT load or total facility power. Annual electricity use depends on deployment, load and the measurement boundary.
VNET–CATL relevance: this sharpens the question behind the compute–energy relationship: can energy capabilities improve delivery and site economics? It does not establish an Alibaba award for VNET, CATL project funding, or customer commitments from CATL’s ecosystem. Those remain separate evidence tests.
Compute is being specified as a grid-responsive resource.
A new national-network project covers assessment of data-centre adjustable-load potential, strengthening the compute–energy branch without yet proving commercial deployment.
The role in the system
Storage, power electronics, cooling and grid access.
Compute-energy coordination is promising; the commercial outcome remains to be established.
Companies in this layer
CATL · Zhongheng · Huawei Digital Power · Haier
Names identify research roles, not a claim that every company has a partnership with every other company.
What would strengthen the thesis?
Evidence of energisation, delivered power costs and funded customer contracts.
What would weaken it?
Technical progress that fails to improve usable output or economic returns; delays, costs or commercial friction that prevent the value test being met.
Architecture notes
Energy system
Still an open research gate. We are testing whether storage, grid-forming capability, direct green power and flexible workloads make electricity dynamically coupled to compute rather than a passive input.
Update · compute is being specified as a grid-responsive resource
From power supply to compute–electricity coordination
On 20 September 2026 a National Integrated Computing Power Network project was registered specifically to define technical requirements for evaluating the adjustable-load potential of data centres. Related registered projects include a compute–electricity coordination reference architecture and power/compute resource planning. This materially strengthens the proposition that electricity may become dynamically coupled to workload placement rather than remaining a passive input.
The evidence is architectural, not yet commercial. It does not prove that VNET, GDS or CATL currently earn revenue from dynamic workload/grid coordination, nor that hyperscale AI training can be freely shifted in response to grid signals. Those remain key tests.
Follow the evidence
22 SEPTEMBER 2026 · RESEARCH UPDATE
Capacity targets need an electricity bridge
A global capacity target cannot be equated to Chinese grid demand. Location, utilisation and whether capacity is measured at IT load or facility level must be established before estimating energy requirements.