---
type: analysis
slug: china-s-power-surplus-cannot-erase-huawei-s-chip-efficiency-gap
title: "China’s power surplus cannot erase Huawei’s chip-efficiency gap"
description: "Huawei can build large campuses where electricity is abundant, but the harder metric is useful compute from each megawatt. The Horinger case suggests energy is an enabler, not the decisive advantage."
published_at: 2026-11-13T11:00:00.000Z
topics: ["trends"]
source_urls: ["https://www.oxfordenergy.org/wpcms/wp-content/uploads/2026/02/Comment-The-China-data-centre-advantage.pdf","https://epoch.ai/data-insights/ai-computing-capacity-additions-doubling","https://epoch.ai/publications/huawei-horinger-china-compute-gap","https://epoch.ai/data/ai-chip-sales?cumulative=false&tab=h100_equivalents&view=graph"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# China’s power surplus cannot erase Huawei’s chip-efficiency gap

Huawei can build large campuses where electricity is abundant, but the harder metric is useful compute from each megawatt. The Horinger case suggests energy is an enabler, not the decisive advantage.

## Summary

China’s energy abundance is helping Huawei build AI data centers, but it does not automatically translate into cheaper or more powerful AI compute. The Horinger case and Oxford Institute for Energy Studies analysis point to the same conclusion: capacity, chips, efficiency and operational constraints determine whether power becomes a competitive edge.

## What happened

The Oxford Institute for Energy Studies argues that China’s power supply is abundant, while its market rules and grid interconnection hurdles can complicate the economics of data centers. Its analysis says new facilities mainly rely on grid power and that national supply is not currently a barrier, even as some eastern regions face restrictions.

Epoch AI’s case study of Huawei’s Horinger campus makes the conversion problem concrete. The facility is large, but Epoch estimates its AI compute at roughly 7% of SpaceX’s Colossus 2. The study attributes the gap partly to Huawei making fewer chips than Nvidia and to lower performance and energy efficiency of domestic chips. More electricity and more buildings help; they do not on their own close the compute gap.

## What the archive adds

New Runtime’s earlier coverage of Epoch AI’s chip-sales estimates puts the buildout in a broader supply curve: quarterly additions of AI-chip compute grew about 2.5× annually from Q1 2024 through Q2 2026, slower than the roughly 6× pace in 2023. That measure counts chips sold, not chips already running in data centers. The Horinger study adds a campus-level check: estimated installed compute can remain far below the global frontier even at a major site.

## Trend signal

AI infrastructure headlines increasingly mix three different quantities: electricity available, chips shipped and compute that is operational and useful. China’s fast construction and energy supply address the first; export controls and domestic chip production shape the second; efficiency, deployment and utilization determine the third. Conflating them makes a large campus look like a competitive result before its delivered compute is measured.

## Why now

Three forces converge: US export controls and Chinese procurement pressure strengthen demand for domestic Ascend capacity (according to the US-China Economic and Security Review Commission); China’s official May 2026 action plan calls for reliable energy for computing infrastructure and coordination between computing and electricity; and Horinger illustrates that power access does not remove chip supply and efficiency constraints (according to Epoch AI).

## Connections to other events

The useful inference is not that China lacks power, but that power is an input whose value depends on the accelerator attached to it. OIES says China’s electricity prices are comparable to US prices and lower than Europe’s, while market inflexibility can blunt potential savings. Epoch’s Horinger analysis says domestic chips deliver less compute per unit of power than leading US chips. Put together, abundant electricity accelerates construction and access, but does not make each megawatt equally productive.

## Why it matters

Builders should evaluate AI infrastructure in delivered compute per megawatt and per deployed chip, not campus size or national generation alone. Buyers comparing capacity should ask whether the quoted figure is planned, commissioned or operational, and which workload and precision underpin the compute equivalence. For Huawei, continued expansion can still create valuable domestic capacity; its strategic payoff depends on increasing chip supply and efficiency while improving software and utilization.

## What happens next

- **Epoch AI will publish an update or new public Horinger data-center entry that reports an operational compute estimate above 81,000 H100-equivalents by 2027-02-11.** (3 months, 55%). Watch for: A new Epoch AI publication or data-center directory entry updates Horinger’s operational compute estimate. Resolves on 2027-02-11, yes if: The Epoch AI data-center directory page lists Horinger at more than 81,000 H100-equivalents on 2027-02-11.
- **Epoch AI will publish a public update to its global AI chip-sales dataset or capacity analysis by 2027-02-11.** (3 months, 70%). Watch for: The AI chip-sales data explorer or capacity insight shows a newer update date or newly added quarterly data. Resolves on 2027-02-11, yes if: The Epoch AI AI chip-sales data explorer page or capacity-analysis page shows a public update dated after its latest October 2026 data and on or before 2027-02-11.

## Limits and uncertainty

The comparison depends on modeled estimates: Epoch uses public information and satellite imagery for Horinger, and H100-equivalents are a peak-operations proxy rather than a full account of workload performance. OIES describes national and regional power conditions, not Huawei’s confidential power contracts or operating costs. The strongest counterargument is that China can build faster, allocate infrastructure strategically and improve domestic chips over time; this conclusion would weaken if those gains produce sustained competitive compute per megawatt.

## Competing readings

- **Energy abundance is chiefly a construction and access advantage; the decisive competitive constraint remains compute delivered by available chips** (likely, about 65%). What would tell it apart: Horinger’s estimated compute remains much smaller than the largest US campus despite substantial power capacity ([epoch.ai](https://epoch.ai/publications/huawei-horinger-china-compute-gap)); OIES finds power abundant nationally but prices and institutional conditions do not guarantee lower data-center costs ([oxfordenergy.org](https://www.oxfordenergy.org/wpcms/wp-content/uploads/2026/02/Comment-The-China-data-centre-advantage.pdf)).
- **China’s buildout is already a durable cost advantage because state coordination and abundant grid power will outweigh chip inefficiency** (unlikely, about 30%). What would tell it apart: OIES says China’s power is abundant, but new facilities mainly rely on grid power and market inflexibility limits cost benefits ([oxfordenergy.org](https://www.oxfordenergy.org/wpcms/wp-content/uploads/2026/02/Comment-The-China-data-centre-advantage.pdf)); Epoch’s Horinger comparison indicates a substantial compute gap against the frontier campus ([epoch.ai](https://epoch.ai/publications/huawei-horinger-china-compute-gap)).
- **The headline figures are too model-dependent to infer any competitive conclusion yet** (unlikely, about 25%). What would tell it apart: Epoch notes chip sales precede deployment and H100e is an imperfect proxy, limiting exact operational comparisons ([epoch.ai](https://epoch.ai/data-insights/ai-computing-capacity-additions-doubling)); Even with those caveats, Horinger’s published estimate is far below the frontier comparison ([epoch.ai](https://epoch.ai/publications/huawei-horinger-china-compute-gap)).

## Base rate

Of 62 earlier research stories about companies, organisations, people or projects in the New Runtime archive over the last 12 months that are at least 3 months old, 42 (68%) were followed within 3 months by a release story on the same company or product line. In the archive, 42 of 62 comparable research stories were followed by a same-entity release story within three months. That is a useful prior for expecting follow-up coverage, not evidence that Huawei will close its compute gap.

## What would change this view

- Epoch AI reports Horinger operational compute per megawatt approaching the level of leading US facilities in a public update (signal): A measured efficiency improvement would show that power abundance is translating into frontier-relevant usable compute.
- Huawei publishes independently matched workload results showing Ascend systems narrowing performance-per-watt gaps against leading accelerators (model-listing signal): The chip-efficiency constraint is shrinking faster than the current campus comparison implies.

## Skill to build now

Practice infrastructure diligence: take one public AI data-center estimate and label each figure as planned, powered, chip-installed or operational; then compare compute per megawatt using the stated chip metric and workload caveats.

## Reader takeaway

When assessing the next AI data-center announcement, separate power capacity from operational accelerator capacity. Ask for commissioned IT load, chip mix, utilization and workload-specific performance per watt before treating megawatts as an AI advantage.

## Open question

Will Huawei’s next generation of Ascend systems and software close enough of the performance-per-watt gap to make abundant electricity a durable cost advantage?

## Sources

- [Huawei, China’s biggest AI chip developer, is drawing on China’s abundant energy to build data centers.](https://www.oxfordenergy.org/wpcms/wp-content/uploads/2026/02/Comment-The-China-data-centre-advantage.pdf)

- [New AI computing capacity is doubling every 9 months.](https://epoch.ai/data-insights/ai-computing-capacity-additions-doubling)

- [epoch.ai](https://epoch.ai/publications/huawei-horinger-china-compute-gap)

- [Much of Broadcom’s increase in shipped compute came from chip volume: Broadcom shipped 65% more chips, or roughly 800,000 in total.](https://epoch.ai/data/ai-chip-sales?cumulative=false&tab=h100_equivalents&view=graph)

- [What Huawei’s Horinger data center reveals about China’s compute gap](https://epoch.ai/publications/huawei-horinger-china-compute-gap?trk=article-ssr-frontend-pulse_little-text-block)

- [2025 Annual Report to Congress](https://www.uscc.gov/sites/default/files/2025-11/2025_Annual_Report_to_Congress.pdf)

- [China to boost mutual empowerment of AI, energy](https://english.www.gov.cn/news/202605/08/content_WS69fde96dc6d00ca5f9a0ad4b.html)

- [Serving Large Language Models on Huawei CloudMatrix384](https://arxiv.org/abs/2506.12708)

- [AI chip sales documentation – Methodology](https://samaritan-research.org/data/ai-chip-sales-documentation/methodology)

- [China to improve cross-regional power allocation](https://english.www.gov.cn/policies/latestreleases/202602/11/content_WS698c83bcc6d00ca5f9a090d4.html)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
