---
title: "At Web Summit in Lisbon, European Companies Show They Now Run Chinese Open-Source Models in Production"
date: 2026-11-18
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-websummit-china-opensource
language: en
---

# At Web Summit in Lisbon, European Companies Show They Now Run Chinese Open-Source Models in Production

> British fintech Revolut swapped OpenAI and Anthropic coding assistants for Zhipu's open GLM models across roughly 2,000 engineers, while Orange, Renault, and Siemens publicly use DeepSeek and Qwen — self-hosting and cost, not leaderboard glory, drive the shift.

The most consequential AI story out of Lisbon this month was not a keynote demo. It was a procurement decision. At the 2026 Web Summit (网络峰会), held November 9–12 in the Portuguese capital, co-founder Paddy Cosgrave told the opening session that British fintech Revolut has replaced its OpenAI and Anthropic coding assistants with the open-source GLM model family from China's Zhipu AI (智谱). Roughly 2,000 engineers now use GLM as their default coding model, and the company reports average coding time down about 60% with AI-related call costs down nearly 90%.

## A coding model becomes infrastructure

The Revolut case matters because of where the model sits: not in a side experiment, but on the default path of a large engineering organization in a regulated industry. According to Xinhua's reporting from the summit, Revolut's AI infrastructure lead described GLM as the company's default programming model across its engineering staff. Those efficiency and cost figures are the company's own, but the structural fact — a European fintech routing daily engineering work through Chinese open weights — is verifiable and new.

The economics explain the logic better than any benchmark chart. Closed frontier APIs bill per token at prices set by the vendor. An open-weight model can be downloaded, fine-tuned, run on the buyer's own servers, and version-locked. For a company processing sensitive financial data, that last part is not a discount — it is a requirement. The model stops being a rented capability and becomes part of the infrastructure, with all the stickiness that implies.

## The wider European pattern

Revolut is not an outlier. Agence France-Presse reporting ahead of the summit noted that French telecom operator Orange, carmaker Renault, and German industrial group Siemens have all publicly said they use Chinese models such as DeepSeek (深度求索) and Alibaba's Qwen (通义千问) — positioned as complements to American and European models rather than wholesale replacements. AFP's framing of the appeal: most Chinese models can be deployed directly on European companies' own servers, keeping data in-house and costs predictable.

The summit itself amplified the theme. Of the roughly 70,000 executives, founders, investors, and policymakers attending, organizers said Chinese participation hit a record this year, and several European attendees reportedly put the China pavilion on their must-visit route. Speakers framed open-source models as a technical common language that lets countries at different development stages build on the same substrate.

## Two very different beneficiaries

The deployment story splits into two groups with different motives.

**Developing countries buying AI infrastructure.** ModelBest (面壁智能) CEO Li Dahai (李大海) told Xinhua his company participated in building Laos' national large model, handling training work alongside domestic Chinese partners. For governments without trillion-parameter compute budgets or elite AI talent pools, open weights mean they can stand up a national model grounded in their own language, data, and regulations without training from scratch. The model is a starting point, not a product they must build themselves.

**Developed countries diversifying their stack.** For European incumbents, the calculus is different: procurement leverage and data control. Industry observers have noted that open-weight models now match closed systems on many routine tasks at a fraction of the price. Adding a DeepSeek or Qwen deployment alongside US models turns a single-vendor relationship into a portfolio — useful both for negotiation and for resilience if API terms change.

Hua Zhiwei of the Center for China and Globalization argued that Chinese open models give European companies a new option for cutting digital-transformation costs, and Rohit Sharma, a partner at US venture firm True, said open technology would keep developing across model development and inference infrastructure as AI matures.

## Why this summit felt like a turning point

Last year's conversation about Chinese models overseas centered on rankings and lab evaluations. This year's evidence is production deployments with named companies, headcounts, and internal metrics. The shift from "how good is the model" to "who runs it in production and why" tracks how mature technologies actually get adopted: not when they top a chart, but when a CFO signs off on the cost model and a CISO signs off on the data boundaries.

There is also a second-order effect worth watching. Every high-profile migration like Revolut's lowers the perceived risk for the next buyer. Conversely, it raises pressure on closed-API vendors to justify their pricing against a downloadable alternative that keeps improving.

## Honest limitations

The headline numbers — 60% faster coding, nearly 90% lower call costs — come from Revolut's own statements relayed through Xinhua, not from an independent audit, and "coding time" metrics are notoriously sensitive to how tasks and baselines are defined. The Orange, Renault, and Siemens examples are public statements about using Chinese models as supplements; none of the companies disclosed scale, workload share, or results. The Laos national model involvement is described by ModelBest's own CEO, again via Xinhua. Finally, none of this tells us about model quality trends going forward — a deployment chosen for cost and control in 2026 can be swapped out if closed systems change their terms, which is precisely the flexibility these buyers are preserving.

## What readers can do now

- **Run the pilot on your own hardware.** GLM, DeepSeek, and Qwen weights are downloadable; a one-week trial on your own servers with your real coding tasks tells you more than any summit anecdote.

- **Build a two-vendor stack deliberately.** Pair a closed frontier model for hard reasoning with an open-weight model for high-volume routine work — that is the pattern the named European companies describe.

- **Check the license before embedding.** Open weights are not all alike; confirm commercial-use terms and any usage thresholds for the specific model version you deploy.

- **Track the cost line, not the hype line.** Log your actual per-task token spend before and after any migration; the Revolut figure is a company report, and your workload may differ.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-websummit-china-opensource
Free to quote with attribution and a link to the original.
