From "Piling Up Chips" to "Building Systems" – The Compute Race Enters a New Phase

In the second week of August 2026, the AI compute sector did not see blockbuster financing news or major disclosures of capital expenditure from tech giants. Instead, two more symbolic events took place: one was the official launch of the world's largest AI compute supercell in Ulanqab, Inner Mongolia; the other was the commercial availability of the first domestic cloud‑based supernode instance capable of running models with over 2 trillion parameters. Though seemingly independent, both point to the same underlying trend: the AI compute race is shifting from "building more powerful individual chips" to "efficiently orchestrating a greater number of chips" – a system‑level engineering challenge. As large language models scale from hundreds of billions to trillions or even tens of trillions of parameters, linear gains in single‑chip performance can no longer keep pace with the exponential growth in computational demand. The industry is forced to re‑examine a fundamental question: is the essence of compute the peak number on a chip spec sheet, or the "effective compute" that can actually be utilised by models?

I. The World’s Largest "Compute Aircraft Carrier" Sets Sail: A Density Revolution

On 6 August, Envision Group officially put into operation the world's largest AI compute supercell – the "Envision Ulanqab Xinghe Base" – in Ulanqab. Covering 120,000 square metres (roughly the size of 20 standard football pitches) and with a planned total capacity exceeding 2 gigawatts, the facility is designed to house up to 1 million AI accelerators working in concert, delivering a theoretical compute output of 1 million PFLOPS. This scale is sufficient to support the parallel training of dozens of trillion‑parameter models, with no global precedent in terms of sheer size.

What truly shook the industry was not the scale, but the density. According to disclosed information, the facility’s compute output per unit area is up to ten times that of traditional data centres. This breakthrough did not come from simple hardware stacking, but from a complete re‑engineering of building structure, cooling design, power distribution, and network topology. The Xinghe supercell adopts a purpose‑built architectural layout, vertically partitioning compute layers, network layers, power distribution, and cooling zones. Each "supernode" unit can house thousands of accelerators, interconnected via a high‑speed non‑blocking network, with intra‑node communication latency controlled at the microsecond level. Traditional data centres often squeeze servers into existing building frameworks, whereas the Xinghe base was designed from the ground up for AI computing – cooling pipes no longer detour but run directly to every rack; fibre optic cabling uses a star topology to minimise signal transmission paths.

Even more noteworthy is its power architecture. Conventional data centres rely on public power grids, with grid connection approvals and transmission line construction taking 3 to 7 years – and a gigawatt‑scale data centre consumes as much electricity annually as a medium‑sized city’s residential use. The Xinghe project, however, bypasses this by directly connecting to self‑built wind and solar farms, coupled with gigawatt‑hour‑scale energy storage, enabling "islanded operation" – even if the regional grid fluctuates, the facility can maintain stable operation independently. This "source‑grid‑load‑storage" integrated model not only circumvents lengthy grid infrastructure cycles, but also pushes the proportion of green electricity used to near 100%. Zheng Zihao, General Manager of Envision AIDC, remarked: "China's chips are taking a path of cluster‑based breakout. The Envision supercell can provide the strongest support for this 'group army'." In his view, while catching up in individual chip process nodes is important, without a matching environment for power, cooling, and interconnection, even the most powerful chips cannot deliver their full combat capability.

II. Cloud‑Based Supernode Goes Live: From "Single‑Card Compute" to "Cluster Efficiency"

Almost simultaneously, on 12 August, Alibaba Cloud announced the general availability of its Lingjun Zhenwu M890 supernode instance, with the first batch launched in the Ulanqab region. Enterprise customers no longer need to build their own computer rooms – they can now provision a 64‑card high‑speed interconnected compute unit on the cloud, capable of handling inference for MoE models with up to 10 trillion parameters. This marks the first domestic supernode‑form compute that has successfully run models exceeding 2 trillion parameters, with Kimi K3 and Qwen3.8 Max already serving customers through this instance. For small and medium‑sized AI startups, this means they no longer have to spend months building their own clusters; within minutes on the cloud, they can access the kind of high‑efficiency compute pool previously available only to tech giants.

The technical details are worth delving into. The instance fully supports FP8 and FP4 low‑precision computing – these precision formats can reduce data transfer and computation volume to one‑quarter or even one‑eighth of traditional FP16, with almost no loss in model accuracy, thereby significantly boosting effective compute output per unit of power. It is equipped with a self‑developed ICN Switch 1.0 chip, expanding the Scale‑up interconnect scale from 16 cards to 64 cards per single cluster, with inter‑card communication bandwidth reaching 800 GB/s. This leap in interconnect capability is critical: in MoE architectures, each token generation requires frequent exchange of gradients and activations among multiple expert modules, and communication overhead often accounts for more than 30% of total training time. Alibaba Cloud has substantially reduced this overhead through its self‑developed switch chips and optimised network protocols. In training scenarios for autonomous driving and embodied intelligence, performance is tripled compared to the previous generation Zhenwu 810E. Underpinned by the Lingjun intelligent compute platform and the HPN 8.0 training‑inference integrated network, a single cluster can orchestrate up to 130,000 heterogeneous compute cards, with scalability towards one million cards – meaning the platform has already reserved ample headroom for even larger models over the next two to three years.

The industry has dubbed 2026 the "first year of domestic supernode deployment." During WAIC, over 20 companies unveiled supernode‑related products, and the China Academy of Information and Communications Technology has issued standard definitions for supernodes, covering hardware architecture, interconnect bandwidth, management and scheduling dimensions. Kaiyuan Securities pointed out that 2026 could be the initial year for mass adoption of domestic supernodes, making them the single largest incremental opportunity in the domestic AI supply chain. Whether supernodes or supercells, the core logic is the same: aggregating scattered compute "atoms" into efficient "molecules", so that the overall capability of a cluster far exceeds the sum of its parts.

III. The Shift in Compute Competition Logic

Taken together, these two events send a clear signal: the logic of AI compute competition is undergoing a fundamental shift.

In recent years, the industry focus was on "building more powerful chips" – higher compute, larger memory, more advanced process nodes. Nvidia, AMD, and various domestic chipmakers have all been racing on this track. But as model parameters climb from hundreds of billions to trillions and beyond, improvements in single‑card performance alone cannot meet the demand. In MoE architectures, each token generation requires frequent parameter synchronisation across different GPUs, and the time GPUs spend waiting for communication is growing ever longer. Studies show that in a 10,000‑card cluster, compute loss due to communication and load imbalance can be as high as 40% – meaning that nearly half of the theoretical peak compute is idling. The deciding factor in the compute race is shifting from "peak compute per single chip" to "effective compute of the cluster".

This is precisely the industrial backdrop against which both "supernodes" and "supercells" are emerging. Whether it is the million‑card physical supercell or the 64‑card cloud supernode instance, both are fundamentally addressing the same challenge: how to make more chips work together more efficiently. Envision tackles density and energy bottlenecks through building‑level vertical zoning and islanded green power; Alibaba Cloud tackles communication and efficiency ceilings through self‑developed switching chips and low‑precision computing. The two technological paths converge on the same goal – pushing the "system efficiency" of compute clusters closer to theoretical limits. This system‑level optimisation offers far greater long‑term compounding benefits than simply advancing chip processes, because once a mature cluster architecture is established, each new generation of chips can be seamlessly integrated, continuously unleashing performance gains.

It is worth reflecting that this shift is reshaping the value distribution across the entire AI industrial chain. Previously, chip manufacturers captured the lion’s share of profits; in the future, companies mastering system‑level technologies – cluster scheduling, network interconnection, cooling management, etc. – will gain greater pricing power. This explains why an increasing number of internet giants and energy conglomerates are crossing over into compute infrastructure – they are not building chips, but building "compute factories."

IV. The Rise of Western China as a Compute Hub

Notably, the two landmark events share a common location – Ulanqab. Both the Envision Xinghe base and the Alibaba Cloud supernode debut are sited in this Inner Mongolian city.

This choice is no accident. Ulanqab is one of the eight national "East‑Data‑West‑Compute" hub nodes, with a unique advantage in network latency – the direct fibre‑optic link to Beijing is only 5 milliseconds, sufficient for most real‑time inference needs. It also boasts abundant wind and solar resources, with green electricity accounting for 67% of the regional power mix. More importantly, the electricity price is significantly lower – around 0.32‑0.35 RMB per kWh, compared to 0.6‑0.9 RMB in East and South China – which can mean a difference of 5 billion RMB annually in electricity costs for a gigawatt‑scale data centre. Compute‑intensive training infrastructure is migrating westward, with data centre projects from Huawei, Apple, Alibaba, Kuaishou, Tencent, and other leading players already established in the region. In just the first half of this year, Ulanqab added over 50,000 PFLOPS of intelligent compute, leading the growth rate among the top ten national clusters.

This migration is not purely cost‑driven; it also reflects the deeper strategic layout of national compute planning. The eastern coastal regions face growing constraints in land, power, and water, making large‑scale supercomputing clusters increasingly difficult to sustain. In contrast, the vast land and abundant renewable energy in the west provide a natural stage for million‑card‑scale supercells. At the same time, through the "East‑Data‑West‑Compute" initiative, western regions are rapidly developing their digital industries, generating local employment and tax revenues, and forming a virtuous cycle of "compute for industry." It is foreseeable that over the next three to five years, western compute hubs such as Ulanqab, Zhongwei, and Qingyang will shoulder more than 60% of the nation's AI training workloads, while eastern data centres will increasingly focus on low‑latency inference and edge computing scenarios.

Conclusion

From Envision's "compute aircraft carrier" to Alibaba Cloud's "cloud supernode," the events of the second week of August 2026 jointly herald a new phase in the AI compute race: the era of single‑point breakthroughs is fading, and the era of system integration has arrived. When million‑card clusters can operate efficiently within a single building, and when ten‑trillion‑parameter models can be summoned on demand from the cloud, compute is no longer just a matter of "having it or not" – it becomes a system engineering challenge of "how to organise, how to schedule, and how to afford it." The winners of this race may not be those with the highest chip compute, but those who best understand how to "orchestrate chips together."

Looking ahead, as supernode technology matures and supercells are replicated at scale, the way we acquire AI compute will increasingly resemble today's electricity – plug in and use, without needing to know about the power plants and transmission networks behind it. In this transformation toward "compute as a service," China is leveraging the institutional advantages of its "East‑Data‑West‑Compute" program and its industrial strengths in new energy to chart a compute development path that differs markedly from other regions. Whether this path ultimately prevails depends on our ability to fill every gap in the system engineering chain – from chip interconnection to cooling, from power scheduling to network protocols – because every detail is a key chess piece determining the final conversion rate of "effective compute." The two events of this week are solid footprints on that long journey.