# Huawei Horinger vs Colossus 2: Epoch AI Puts China's Compute Gap at 7%

> Source: <https://explainx.ai/blog/huawei-horinger-7-percent-colossus-2-epoch-ai-compute-gap-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-15 00:00:00+00:00

**Key takeaway:** Epoch AI estimates Huawei's Horinger campus delivers about 81,000 H100-equivalents from roughly 105,000 Ascend 910C and 8,000 Ascend 910B chips at ~170 MW, versus ~1,112,000 H100e at Colossus 2's ~946 MW, because U.S. data centers yield 2.5x more AI compute per unit of power at current chip mixes.

**Huawei's Horinger AI campus in Inner Mongolia runs on domestic Ascend chips, yet Epoch AI estimates it holds only about 7% of the AI compute at SpaceX's Colossus 2 in Tennessee — roughly 81,000 versus 1.1 million H100-equivalents in late 2026.** The gap is not mainly electricity. [Epoch's November 6, 2026 report](https://epoch.ai/publications/huawei-horinger-china-compute-gap?ref=explainx) argues China can site hundreds of megawatts, while **chip volume and per-watt efficiency** still favor U.S. Blackwell stacks and export-controlled Nvidia supply.

For builders comparing **where frontier training runs**, **which cloud regions matter**, and **how open-weight Chinese labs stretch domestic silicon**, Horinger is the clearest public sizing exercise we have this week. It sits beside our [Colossus 2 capacity thread](https://explainx.ai/blog/spacex-colossus-2-1-2-million-blackwell-gpus-gun-range-spam-claimed-vs-verified-2026), [Anthropic's Colossus partnership](https://explainx.ai/blog/anthropic-spacex-colossus-1-supercomputer-partnership-2026), and [Southeast Asia routing fights](https://explainx.ai/blog/singapore-nvidia-ai-chip-routing-megaspeed-ft-export-networks-2026) over who actually gets H100-class throughput.

## TL;DR

| Metric | Huawei Horinger (Epoch point estimate) | Colossus 2 (Epoch directory) | 
|---|---|---|
| AI compute | **~81k H100e** (90% CI 36k–141k) | **~1,112k H100e** | 
| IT power | **~170 MW** current | **~946 MW** current | 
| Chips cited | **~105k Ascend 910C** ,**8k 910B** | **Blackwell B200/B300** class | 
| 2028 projection | **~141k H100e** | **≥1.8M H100e** | 
| Share vs Colossus 2 | **~7% today** ; similar or smaller share in 2028 | Baseline | 

## What Epoch AI measured at Horinger

[Epoch AI](https://epoch.ai/publications/huawei-horinger-china-compute-gap?ref=explainx) treats Horinger as a case study in China's **East Data West Computing** push. The campus sits in Horinger New Area, Inner Mongolia, split into **West, South, and North** zones marketed as Huawei Cloud availability regions. Epoch bought **high-resolution satellite imagery** and counted **rooftop cooling towers and chillers** to infer IT power. Four dense AI-style buildings in South and West zones sum to about **286 MW** of IT capacity in the cooling model, aligned with substation and generator cross-checks near **304 MW** and **230 MW**.

## How the three zones differ

Satellite map of Huawei Horinger data center zones in Inner Mongolia for China AI infrastructure analysis
*Image: [Epoch AI](https://epoch.ai/publications/huawei-horinger-china-compute-gap?ref=explainx)*

Epoch notes the **North zone** shows far fewer cooling towers than South and West, so those buildings likely host **lower-density** workloads. **South and West** match published renderings of **hypernode** halls aimed at large-model training. By August 2026 imagery, **three of four** AI-focused buildings looked operational; one West building remained idle with a **December 2026** start assumption. Vapor plumes and moving fan "ghosts" in satellite frames support active cooling on running halls.

Occupancy matters. Epoch assumes **76% rack fill**, matching a reported rate for the wider Horinger area in September 2026, then maps **Ascend sales with a four-month deployment lag**. That yields **~105,000 Ascend 910C** and **8,000 Ascend 910B** chips on site, or **81,000 H100e** central estimate.

## Ascend 910C vs Nvidia: why power is not enough

Epoch cites **CloudMatrix-Infer** evidence for **~1.504×10¹⁵ INT8 OP/s** per 910C package (two dies), about **0.76 H100e** versus an H100 peak near **1.979×10¹⁵ OP/s**. Blackwell B200/B300 classes sit near **2.5 H100e** with **~2.5× better energy efficiency** in Epoch's framing.

The report's punch line for infrastructure planners: **Chinese megawatts do not equal U.S. megawatts in useful FLOPs**. Epoch writes U.S. AI data centers can support **2.5× more AI compute per unit of power** than China's best domestic halls at current chip mixes. That connects to **[DeepSeek Ascend pretraining rumors](https://explainx.ai/blog/deepseek-chat-demo-static-browser-client-github-2026)** and **[entity-list procurement workarounds](https://explainx.ai/blog/northwestern-polytechnical-nvidia-h200-entity-list-procurement-2026)** — labs chase both **domestic Ascend** and **diverted Nvidia**.

## The Colossus 2 comparison

Bar chart comparing Huawei Horinger AI compute at 81,000 H100-equivalents versus SpaceX Colossus 2 at about 1.1 million
*Image: [Epoch AI](https://epoch.ai/publications/huawei-horinger-china-compute-gap?ref=explainx)*

Epoch's [Colossus 2 directory page](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=explainx) lists roughly **1.112 million H100e** today, tied to **SpaceXAI** ownership with customers including **Anthropic** and **Cursor**, plus **946 MW** IT power and **~$35.8B** capital cost in their model. Horinger at **81k H100e** is still **large by global standards** — Epoch says only **VNET Bayin Ulanqab** and **Huawei Wuhu** likely match or slightly exceed it among Chinese sites they depth-reviewed.

Caveats Epoch flags:

- Campuses often split into **multiple clusters** that cannot merge into one training job without special networking.
- **10%** of Horinger (~8,100 H100e) could still train a**DeepSeek-V4-Pro scale** model in**one month** in their scenario.
- An **optimistic unified run** might need**~one year** for**GPT-6 Astra scale** — a thought experiment, not a product roadmap.

## Build-out through 2028

Construction imagery shows another South-zone building on track for **February 2028**. A public line mentions **100,000 accelerator cards** and **Ascend 950** silicon; Epoch assumes **~76,000 950DT** chips in that hall with **135 MW** IT power. Grid plans suggest substation expansion could eventually support **~3.5×** current campus power.

Even then, Epoch projects Horinger at **~141k H100e** (90% CI 65k–227k) by early 2028 while Colossus 2 could pass **1.8 million H100e**. **Huawei Wuhu** may grow faster, reaching **~151k H100e** by December 2027 — so China's frontier is **multi-campus**, not a single Inner Mongolia site.

## Who might use the capacity

Epoch lists plausible tenants without confirming contracts:

- **Huawei Cloud** itself, including driver-assistant workloads (**~60k H100e** cited elsewhere for April 2026).
- **Fintech and data aggregation** lines from a local announcement on a West-zone building.
- **Zhipu (Z.AI)** and**iFlytek** as companies tied to large domestic stacks and Horinger New Area offices — placement unverified.

That diversity matters for builders: **not every Ascend rack trains a chat model**. Much capacity likely serves **inference, cloud SKUs, and vertical AI**, similar to how U.S. hyperscalers mix training halls with general regions.

## What this means if you build on AI compute

**If you procure GPUs or rent clusters**, treat Epoch's H100e numbers as **capacity planning anchors**, not lease pricing. China's **power advantage** shows up in land and cooling; **frontier scale** still tracks **Nvidia export dynamics** and **Ascend ramp curves**.

**If you benchmark open models**, a capable **GLM or DeepSeek** release on Ascend does not disprove the gap — it shows **algorithmic efficiency** under tighter FLOPs. Compare against [Artificial Analysis-style cost posts](https://explainx.ai/blog/grok-4-7-republicans-gpt-6-1-sol-democrats-convovo-midterm-vote-study-2026) only when the same chip class is named.

**If you follow policy**, Horinger supports **[Pax Silica allied supply-chain lists](https://explainx.ai/blog/pax-silica-joint-ai-supply-chain-list-san-francisco-declaration-november-2026)** and **[Commerce remote-access rules](https://explainx.ai/blog/commerce-final-rule-chinese-ai-firms-remote-nvidia-chips-the-information-2026)** narratives: Washington worries about **rented compute**, Beijing invests in **domestic halls**, and both sides argue with **megawatt headlines** while **FLOPs per watt** decide practical frontier access.

## Sister campuses and CloudMatrix384 rollouts

Epoch lists **eight reported CloudMatrix384 supernode sites** from **April–December 2025**, mixing **Huawei-owned** halls and **China Telecom, China Mobile, China Unicom, and Shenzhen Longgang** facilities. **Horinger** is one node in that mesh; **Wuhu** may outgrow it by **December 2027** in Epoch's projections. Treat **national Chinese frontier capacity** as the **sum of many campuses**, not Horinger alone — the same way U.S. readers should add **[Starbase notices](https://explainx.ai/blog/spacex-starbase-super-compute-facility-24-billion-notice-hinojosa-2026)** to Colossus 2 rather than conflating them.

The report also estimates **907,000 Ascend 910C** chips sold by **end-March 2026**, with Horinger consuming **~11%** of that stock by **August** deployment assumptions. That ratio matters for **supply-chain analysts**: a flagship hall can be **large** without exhausting domestic inventory, leaving chips for **telecom edge sites** and **supernodes** listed in Epoch's table.

## Satellite methodology in plain language

Epoch's workflow for readers who will cite the report in briefings:

1. **Buy commercial satellite passes** over Inner Mongolia and sister sites.
2. **Outline cooling hardware** on rooftops — towers, chillers, ghosted fan motion — to infer**heat rejection** and**IT megawatts** .
3. **Cross-check** substation transformer counts and**backup diesel banks** against cooling totals.
4. **Apply rack occupancy** and**chip-deployment lag** from sales databases.
5. **Convert installed Ascend counts to H100e** using documented INT8 peaks, with confidence intervals.

The org withholds some vendor sources to protect continued access — contact **[info@epoch.ai](mailto:info@epoch.ai)** per the article footer if you need the withheld annex for research. explainx.ai did not receive a private annex; our numbers come from the public page and [Huawei Horinger directory entry](https://epoch.ai/data/ai-data-centers/directory/huawei-horinger?ref=explainx).

## Efficiency gap table for procurement teams

| Factor | U.S. frontier (Colossus-class) | Huawei Horinger (Epoch estimate) | 
|---|---|---|
| Dominant silicon | Nvidia Blackwell B200/B300 | Ascend 910C / 910B mix | 
| H100e per MW (qualitative) | Higher FLOPs per watt | ~40% of leading U.S. halls per Epoch | 
| PUE context (2025 averages) | U.S. ~1.49 in Uptime survey | China ~1.40 — power not the bottleneck | 
| Networking caveat | Multi-building NVLink domains | Clusters may not merge for one run | 

If your compliance team blocks **China-region training**, these numbers still inform **red-team benchmarking**: adversaries with **~81k H100e** can host **credible mid-scale experiments** even when they cannot match **million-GPU** campuses.

## How Epoch AI fits other trackers

Epoch is a nonprofit research org publishing **[AI data center directories](https://epoch.ai/data/ai-data-centers?ref=explainx)** and compute-gap analyses. Searchers often ask **what Epoch AI does** — here, it fuses **satellite evidence**, **chip sales estimates**, and **normalized H100e** math so policymakers and ML engineers share one vocabulary. Our [prior Epoch trajectory coverage](https://explainx.ai/blog/epoch-trajectory-gemini-4-argon-live-website-impersonation-report-2026) shows the same team sizing **agent risk**; this report sizes **physical AI capacity**.
