📑 Table of contents

Huawei circumvents the embargo: an Ascend 950DT cluster exports AI tokens, not chips

Skynet Watch 🟢 Beginner ⏱️ 13 min read 📅 2026-11-19

Huawei circumvents the embargo: an Ascend 950DT cluster exports AI tokens, not chips

🔎 Compute crosses the border, not silicon

On November 13, 2026, a datacenter lit up in Shantou, Guangdong. On the surface, just another project: 1 billion yuan (about 150 million dollars), racks of servers, fiber. In reality, it is the first Chinese cluster designed from the ground up to sell AI inference abroad — without exporting a single chip.

The mechanism, dubbed "token export," is disarmingly simple. Requests from foreign clients are processed locally, on Huawei Ascend 950DT chips. Only the results — the tokens — cross the border, via the submarine cable that lands right next to the datacenter.

Why now? Because demand is shifting. Daily token usage on the site went from 100 million to 100 billion in less than five months, according to figures reported by GD Voices. And declared demand from foreign companies registered in Shantou exceeds 3 trillion tokens per day. At that scale, compute becomes a strategic export — exactly what US controls were meant to prevent.


The essentials

  • The first Huawei Ascend 950DT cluster dedicated to AI token exports has been operational since November 13, 2026 in Shantou (Guangdong), as DIGITIMES and Digital Today confirm.
  • The setup brings together Huawei (chips), China Telecom (network and data center, installed on a submarine cable landing station) and Guangdong Tianyima (investment and operations).
  • Phase 1: 1 billion yuan (~$150M) for a target of 2,000 petaflops; the park ultimately aims for 400 MW of IT capacity, for a total investment of 120 billion yuan.
  • The business model: inference runs in China, and only the tokens leave the country, heading to Southeast Asia. No goods subject to licensing cross the border.
  • Growth: from 100 million to 100 billion tokens/day in under 5 months; recorded foreign demand exceeds 3 trillion tokens/day.

This news is geopolitical first and foremost. And yet it directly concerns developers: tokens have become a billed, localized commodity — and, now, a geopoliticized one. Two levers for keeping control of your own stack:

Tool Main use Price (November 2026) Ideal for
Hostinger Host your AI applications and endpoints on a VPS, without depending on a dominant cloud from ~€3/month (check on hostinger.com) Developers, small teams
Self-hosting open models (Kimi K2.6, GLM-5) Sovereign inference, controlled cost per token Variable (server + model license) Teams that care about data location

Of course, a VPS won't replace a 1,000-card SuperPoD. But the principle remains the same as in Shantou: whoever controls the compute controls the tokens.


What exactly is the Shantou cluster?

The first Chinese datacenter explicitly designed to sell AI inference abroad without exporting hardware. Commissioned on November 13, 2026, it is backed by the landing station of a China Telecom submarine cable — a detail that is not just a detail, as we will see later.

The project rests on a triangle with clearly defined roles. Huawei supplies the Ascend 950DT chips. China Telecom provides the network and the datacenter. Guangdong Tianyima invests and operates. According to the South China Morning Post, the phase 1 investment reaches 1 billion yuan, or about 150 million dollars.

Phase 1 figures

Item Value Source
Commissioning November 13, 2026 Digital Today, Work Optional
Phase 1 investment 1 billion yuan (~$150M) SCMP, Work Optional
Target capacity 2000 petaflops Work Optional
Phase 1 footprint ~413 mu (~27.5 hectares) GD Voices
Planned IT capacity (campus) 400 MW GD Voices
Total campus investment 120 billion yuan GD Voices

Two figures deserve a closer look. First, the 400 MW: that's the electricity consumption of a mid-sized city, dedicated to computing. Then, the 120 billion yuan — nearly 17 billion dollars: the ambition goes far beyond the Shantou prototype.

A SuperPoD, not just a datacenter

The other technical detail comes from GD Voices: the Ascend 950 SuperPoD brings together more than 1000 compute cards organized into a single logical computer. We're not talking about independent servers taking turns, but a single machine at the scale of an entire building.

It's the only way to compensate for the silicon gap with Nvidia: where an American chip retains the per-unit advantage, Huawei bets on numbers and interconnection.


How does "token export" work?

Inference runs in China, and only the generated tokens leave the country. A request's journey takes four steps:

  1. The client sends its request from Southeast Asia.
  2. It travels via the submarine cable to Shantou.
  3. The Ascend 950DTs run the inference on Chinese soil.
  4. The generated tokens return through the same cable.

The location is no accident. The datacenter is built on China Telecom's submarine cable landing station, which reduces latency to Southeast Asia to the physical minimum. Huawei and China Telecom literally built the datacenter where the data already enters China.

Hence the model's name: "token export". By analogy with manufacturing exports, China imports demand (the requests) and exports value added (the responses). Except the factory is a datacenter and the raw material is electricity.

Why it disarms US export controls

Because US export rules target goods: chips, equipment, software. A computing service running on Chinese soil is not a good. There is nothing to license, nothing to block at customs, nothing to trace at a reseller.

Chip exports Token exports
What crosses the border Hardware (accelerators, servers) Textual output (tokens)
US leverage Licenses, embargoes, blacklisted entities No direct leverage
Strategic asset The chip The datacenter, the energy, the cable
Risk for the exporter Unauthorized re-export Nearly zero: the compute stays home
Example H200 licenses to China Shantou Ascend 950DT cluster

The last row of the table sums it all up: in this model, the strategic asset never leaves the country. It's the exact inverse of embargo logic.


100 Billion Tokens Per Day: The Curve That Changes Everything

Yes, the growth is spectacular: daily usage has jumped from 100 million to 100 billion tokens in less than five months, a factor of 1,000. Furthermore, the recorded demand from foreign companies listed in Shantou exceeds 3 trillion tokens per day (3 × 10¹²).

Caution first: "recorded demand" is not billed volume. Provincial authorities have an interest in displaying ambitious figures for a flagship project. But even with a significant discount, the trend is real—and it is explainable.

This is because consumption per request is exploding mechanically. Context windows are lengthening (up to 12 million tokens at certain players), agents are chaining tool calls, and every point of quality comes at the cost of reasoning tokens. An agent that thinks consumes ten times more than a chatbot that simply answers.

At this scale, billing becomes an engineering subject in its own right: our guide Tokens, Context, Costs: Understanding LLM Billing details why the token has become the unit of account for AI. Shantou is applying the same logic on a national scale: selling tokens down to the kilowatt-hour.


A circumvention of US sanctions?

Legally, no — and that's precisely the problem for Washington. US export controls govern the sale of goods: chips, equipment, software. An inference service delivered from China exports nothing that can be controlled.

The context makes the story all the more piquant. While Washington was granting sales licenses for the H200 to around ten Chinese companies — deliveries subsequently blocked on Beijing's side (our analysis) — China took the opposite path: stop depending on American chips, and sell compute.

The reasoning is airtight. As long as the embargo targets hardware, every chip denied accelerates the building of an alternative. And once the alternative exists, it gets monetized — including with the very customers the American chips were supposed to "protect" from China.

Washington has few direct levers at its disposal. What remain are the indirect levers: diplomatic pressure on buyers in Southeast Asia, secondary sanctions, control of cable infrastructure. None of them is fast, and all are costly in political capital.

My take: this is a quiet but major turning point. The American embargo was designed for a world where AI = chips. Shantou demonstrates a world where AI = service. And a service doesn't go through customs.


Ascend 950DT: what we know — and what we don't

Huawei has mastered scaling: more than 1000 cards in a single logical computer, with a target of 2000 petaflops for phase 1. On the other hand, the capacity actually delivered is not disclosed, notes Work Optional — and that opacity is itself information.

Let's sort out the confirmed from the speculative. Confirmed: the SuperPoD architecture (more than 1000 cards), the commissioning, the three-party setup, the export orientation. Not confirmed: the per-unit performance of the 950DT against a current Nvidia accelerator, the models being run, the actual utilization rate of the cluster.

My bet: the chip remains, on a per-unit basis, behind the best of American silicon. But that no longer matters much. Shantou's competitive advantage is not the chip — it's geography (the cable), energy (400 MW planned), and price. Compute has become AI's scarcest resource, to the point that Google is rationing Gemini for Meta. In this world, whoever owns megawatts and fiber sells exactly what they own.


What this changes for developers and businesses

In the short term, little changes in Europe; in the medium term, a new player on the global inference market — and added pressure on prices.

For Southeast Asia, the offering is real as of today: low latency, relative sovereignty (the compute stays "in Asia," even if it's in China), and probably aggressive pricing. For Western companies, it's more complicated: having customer data processed in China raises compliance questions that few legal departments will sign off on.

On the model side, the Chinese software layer is no longer the weak link. DeepSeek V4 Pro (Max) scores 88 on the latest overall leaderboard, ahead of Claude Opus 4.6 (87); Kimi K2.6 lands 88.1 on agentic tasks. There's no indication that these models run on the Shantou cluster — but they prove that demand can be absorbed by a 100% Chinese stack, from chips to models.

The lesson to take away, whatever your geography: compute localization is becoming an architectural criterion, on par with latency or cost. The choice of a model will soon no longer be merely technical. It will be geopolitical.


The Gray Areas to Watch

Three blind spots to keep in mind before declaring a revolution.

One, delivered capacity. The 2,000 petaflops are a target; the actual figure has not been published. A showcase cluster can be sized for demonstration more than for actual load.

Two, registered demand. The 3 trillion daily tokens represent stated intentions from companies registered in Shantou — not signed contracts, let alone revenue.

Three, customer geopolitics. A Southeast Asian CIO who routes their data to China takes a real political risk: regional tensions, secondary sanctions, cable cuts. The export token sells sovereignty — Beijing's, not the customer's.


❌ Common Mistakes

Mistake 1: believing that Huawei "exports chips"

That's the whole subtlety of the story: no chip leaves China. The product being sold is the token, a service. If you read the news as a hardware sale, you miss the mechanism — and its significance.

Mistake 2: confusing announced capacity with delivered capacity

The 2,000 petaflops are a phase 1 target; the capacity actually in production is not disclosed. A rule of thumb for reading any Chinese compute news: distinguish the announcement from the megawatt actually plugged in.

Mistake 3: mistaking "registered demand" for revenue

The 3 trillion tokens/day is demand declared by companies registered in the zone. It's an indicator of interest, not a billed volume. Compare like-for-like figures before concluding that the market is exploding.

Mistake 4: believing the story only concerns geopoliticians

It concerns anyone who pays for tokens. The rise of a non-American inference offering will change prices, latency, and architecture choices — including for a freelancer hosting their own agent.


❓ Frequently Asked Questions

What is "token export"?

A business model where AI requests from foreign clients are processed in China, and only the results—the tokens—cross the border. The hardware remains on Chinese soil, which mechanically neutralizes US export controls designed for physical goods.

What chips equip the Shantou cluster?

Huawei Ascend 950DT, organized into SuperPoDs of over 1000 cards forming a single logical computer. China Telecom provides the network and the datacenter, installed on an underwater cable landing station; Guangdong Tianyima invests in and operates the site.

How many tokens does the cluster process per day?

Daily usage has increased from 100 million to 100 billion tokens in less than five months, according to GD Voices. The recorded demand from foreign companies listed in Shantou exceeds 3 trillion tokens per day—a figure of intent, not billing.

Yes, under current rules: US controls focus on the export of goods (chips, equipment), not services executed in China. Washington could react through pressure on Asian buyers or secondary sanctions, but no direct leverage exists.

Who are the target clients?

Southeast Asia, prioritized, served by the subsea cables landing in Shantou. Low latency and likely competitive prices target regional companies seeking low-cost inference without dependence on US clouds.

When was the cluster commissioned?

November 13, 2026, according to Digital Today and Work Optional. DIGITIMES confirmed the opening of the first Ascend 950DT cluster dedicated to token export in its November 19, 2026 edition—a one-week delay, consistent with a showcase project.


✅ Conclusion

By installing its datacenter where the cable touches land, Huawei turned a constraint — the chip embargo — into a product: Chinese tokens for export, 100 billion per day. The next time someone talks to you about the chip war, watch the cable war instead.