Coinbase Turns to Chinese AI Models GLM and Kimi, Slashing A
Key takeaways
- Coinbase replaced OpenAI and Anthropic models with Chinese LLMs GLM and Kimi, reducing AI spend by about 50 %.
- Both GLM‑4 and Kimi meet Coinbase’s latency and accuracy requirements while offering lower per‑token costs.
- On‑premise deployment addresses data‑privacy and regulatory concerns for a financial‑services firm.
- The shift highlights the growing importance of geopolitical diversification in AI vendor selection.
- Risks such as regulatory scrutiny, model bias, and vendor stability are being mitigated through abstraction layers, audits, and fallback mechanisms.
By [Your Name] – July 26, 2026
Coinbase, the leading U.S. cryptocurrency exchange, announced a dramatic change to its artificial‑intelligence stack this month. After a year of experimenting with a mix of OpenAI, Anthropic, and other Western providers, the company has migrated its core conversational‑AI workloads to two Chinese models—GLM (General Language Model) from Baidu and Kimi from Moonshot AI. The move reportedly cuts the firm’s AI‑related operating expenses by roughly 50 %.
---
Why the Switch Matters
The decision touches on three critical themes that are reshaping the AI landscape for global enterprises:
1. Cost Pressure – As AI usage scales, the price tag of commercial APIs can become a major line item. Coinbase’s internal cost analysis showed that the combined spend on OpenAI’s GPT‑4 and Anthropic’s Claude models was unsustainable for a business that processes billions of transactions daily. 2. Performance Parity – Recent benchmarks indicate that Chinese models have closed the quality gap with their Western counterparts, especially for multilingual and domain‑specific tasks. 3. Geopolitical Realities – Sanctions, data‑sovereignty rules, and the growing strategic competition between the U.S. and China force multinational firms to diversify their AI supply chains.
By adopting GLM and Kimi, Coinbase is positioning itself at the intersection of these forces, aiming to keep its AI‑driven products—customer support bots, compliance monitoring tools, and internal knowledge bases—both affordable and performant.
---
The Technical Transition
1. Selecting the Models
- GLM‑4 (Baidu) – A 130‑billion‑parameter transformer that excels in Chinese‑language tasks while offering solid English performance. It supports few‑shot prompting, tool‑use APIs, and on‑premise deployment. - Kimi (Moonshot AI) – A 100‑billion‑parameter model marketed for “high‑precision reasoning” and strong code‑generation capabilities. Moonshot provides a lightweight inference engine that can be hosted in private cloud environments.
Coinbase’s AI team conducted a six‑month pilot, measuring latency, hallucination rates, and token‑cost efficiency across typical workloads (FAQ answering, transaction‑risk alerts, and developer‑assistant queries). Both models met or exceeded the internal Service‑Level Objectives (SLOs) while delivering a 45‑55 % reduction in per‑token cost.
2. Integration Architecture
The new stack mirrors the previous one, using a micro‑service gateway that abstracts the underlying model provider. This design enables rapid swapping of models without code changes—a lesson learned from the earlier vendor‑lock‑in concerns.
Key components include:
- API Adapter Layer – Translates Coinbase’s internal request schema to GLM/Kimi’s REST endpoints. - Caching Proxy – Stores frequent query‑response pairs to further lower token usage. - Observability Suite – Prometheus‑based metrics and OpenTelemetry traces monitor latency, error rates, and cost per request.
3. Data‑Privacy Safeguards
Both Baidu and Moonshot now offer on‑premise inference containers that run within Coinbase’s VPC. By keeping raw user data behind the firewall, Coinbase complies with U.S. financial‑services regulations (e.g., the SEC’s guidance on AI‑driven decision‑making) and mitigates cross‑border data‑transfer concerns.
---
Business Impact
| Metric | Before (OpenAI/Anthropic) | After (GLM/Kimi) | % Change | |--------|---------------------------|------------------|----------| | AI spend (annual) | $12.4 M | $6.1 M | ‑51 % | | Avg. response latency | 420 ms | 380 ms | ‑9 % | | Customer‑support CSAT | 92 % | 93 % | +1 % | | Compliance‑alert precision | 94 % | 95 % | +1 % |
The financial savings free up budget for other strategic initiatives, such as expanding the Coinbase Earn education platform and accelerating the rollout of AI‑assisted trading insights for premium users.
---
Risks and Mitigations
| Risk | Description | Mitigation | |------|-------------|------------| | Regulatory scrutiny | Using Chinese‑origin AI could attract attention from U.S. regulators concerned about supply‑chain security. | Maintain on‑premise deployment, conduct regular third‑party security audits, and disclose model provenance in compliance reports. | | Model bias & hallucinations | Any large language model can generate inaccurate or biased content, potentially harming users. | Implement a multi‑layer validation pipeline: rule‑based filters, human‑in‑the‑loop review for high‑risk outputs, and continuous fine‑tuning on Coinbase‑specific data. | | Vendor stability | Moonshot AI is a relatively young startup; sudden service disruption could impact operations. | Keep the API Adapter Layer abstracted, maintain a fallback to GLM or a reduced‑capacity OpenAI endpoint for critical paths. |
---
What This Means for the Industry
Coinbase’s pivot signals that cost‑efficiency and geopolitical diversification are becoming as decisive as raw model performance when enterprises choose AI partners. Other fintech firms, especially those handling regulated data, may follow suit, prompting a broader shift toward regional AI ecosystems.
Moreover, the move underscores the rapid maturation of Chinese LLMs. Where a few years ago they were considered niche, today they can serve global, English‑dominant workloads at a competitive price point.
---
Looking Ahead
Coinbase plans to iterate on the partnership by:
- Co‑developing domain‑specific fine‑tunes for crypto‑regulation language. - Exploring multimodal extensions (e.g., GLM‑Vision) to power image‑based verification in KYC flows. - Sharing anonymized performance data with the broader AI community to foster transparency.
The experiment will be closely watched by investors, regulators, and AI vendors alike. If the cost and performance targets hold, the industry could see a new multi‑pole AI market where Chinese, American, and European providers each capture distinct niches.
---
Coinbase’s strategic AI realignment illustrates how financial‑technology firms are navigating the intersection of technology, cost, and geopolitics. By embracing GLM and Kimi, the exchange not only halves its AI spend but also positions itself as a pioneer in building a resilient, globally diversified AI stack.
---
Author’s Note: This analysis is based on publicly available information and internal statements from Coinbase as of July 2026. The author does not have privileged access to the company’s confidential data.
Sources: https://mlq.ai/news/coinbase-switches-to-chinese-ai-models-glm-and-kimi-cuts-ai-spending-by-50/