From the evening of March 29 to March 31, 2026, China‑based large AI model service provider DeepSeek experienced varying degrees of service anomalies for three consecutive days, drawing wide attention. The major outage from the evening of March 29 into the morning of March 30 lasted nearly 12 hours (official records indicate about 7 hours 13 minutes), marking the longest single service interruption since DeepSeek's founding. Users reported they could not log in to the web or app, conversations were interrupted, pages froze, and repeated "server busy" prompts appeared; related topics repeatedly trended. At 18:05 on March 31, DeepSeek's official status page confirmed all anomalies were resolved and web, app and API services had fully returned to normal.
According to DeepSeek's official service status page, the incident timeline is as follows:
- March 29, 21:35: Initial web/app performance anomalies detected.
- 23:23: First fix implemented, some users briefly restored access, but fluctuations recurred.
- March 30, 00:20: Second performance anomaly investigation; 01:24: remediation measures implemented.
- Around 10:33 on March 30: The first major outage was marked "resolved", but user tests showed actual full recovery close to 10:00 AM, totaling nearly 12 hours of interruption.
- March 31, 17:02: Web/API performance anomalies recurred and were swiftly resolved at 18:05.
This incident is not isolated. Public records show that in the past 15 months DeepSeek has experienced at least seven notable service interruptions, many previously linked to sudden user surges, server resource strain, or malicious attacks. In January 2025, it suffered multiple short‑term outages due to large‑scale malicious attacks and temporarily restricted registrations from non‑+86 phone numbers.
Possible causes: user surge, compute bottlenecks and overlapping model iterations
DeepSeek did not disclose a specific technical root cause, labeling the issue only as a "performance anomaly." Industry experts and media analyses point to multiple compounding factors behind the prolonged outage:
- High concurrent user volume and insufficient compute scaling: DeepSeek's daily active users reportedly rose from 120 million to 200 million, while compute expansion was limited (in some periods increasing by only 8.3%). Although MoE (Mixture‑of‑Experts) architectures are efficient, they can consume massive memory and compute during peak inference, risking performance collapse.
- Graduation season/weekend traffic peaks: At the end of March students preparing for exams and extended work hours drove higher dependence on AI tools; combined with weekend overnight usage this led to sudden load spikes.
- Suspected model iteration testing: Some analysts speculate DeepSeek may have been gray‑testing a V4 model (which had been delayed several times). Compatibility issues between old and new architectures in the underlying storage aggregation layer might have arisen. The new version's enhanced context handling and code capabilities could have unexpectedly triggered instability during testing. Signs of code capability upgrades on the web client in early March further support this hypothesis.
- Infrastructure shortcomings: As a fast‑growing AI startup, DeepSeek's rapid iteration strategy relied on quick deployments, but its self‑built compute and operations systems have not yet fully matched global‑scale traffic. These three days of anomalies exposed stability weaknesses for domestic large models under high load, highlighting gaps compared with international giants.
Impact: user churn, industry wake‑up call and subtle shifts in competitive landscape
For users: Many consumer users (students, programmers, professionals) were left stranded at critical moments, with some conversations lost, triggering dissatisfaction. Topics such as "DeepSeek crashed" and "DeepSeek became DeepSleep" trended, and users complained: "It suddenly disconnected while I was using it—who can tolerate that?" Although the company fixed the issue promptly, the three‑day sequence of anomalies still caused some users to temporarily switch to competitors such as Tongyi Qianwen and Kimi, resulting in short‑term churn.
For DeepSeek itself: The incident set a record for the longest outage and damaged brand reputation in the short term. The company will need to invest more resources to strengthen operations and redundancy design. On the positive side, the swift response and the "fully restored" status indicate that emergency capabilities remain in place. In the long run, this may accelerate DeepSeek's efforts to improve infrastructure and shift toward more stable enterprise‑level services.
For the AI industry as a whole: The incident again put "AI infrastructure stability" in the spotlight. China's large AI models are in a period of rapid growth, with user scale and compute demand exploding in tandem, while bottlenecks in supply chains, power and operations have surfaced. DeepSeek's consecutive outages serve as a wake‑up call: high‑performance models must be matched by reliable service capability, otherwise retaining users will be difficult. Competitors have seized the opportunity to expand market share, and some platforms have seen noticeable traffic return.
For capital markets and policy: As a domestic AI representative, DeepSeek's stability directly affects investor confidence. The consecutive anomalies may prompt regulators and companies to jointly consider how to enhance AI service resilience and promote optimized allocation of compute resources.
Conclusion: recovery is not the end; stability becomes the new battleground in AI competition
DeepSeek has confirmed that all services have been fully restored and said it will continue to monitor operations. Although the incidents were handled, the exposed compute bottlenecks, architectural compatibility issues and operational pressures will require long‑term solutions. For users, this has been a lesson in availability; for the industry, it signals a shift from competing solely on model capabilities to competing on service stability.
If DeepSeek can use this event to upgrade its infrastructure and roll out V4 and other products at the right time, it may turn crisis into opportunity. The domestic AI sector must also confront the shared challenge of building more reliable and more inclusive AI services amid tight compute supply and exploding demand. Stability has become a new front in AI competition.
▶️ Enjoy this article? Creating content is hard work!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com