DeepSeek’s DSec Sandboxes Train Agents at 380,000 Concurrent — and Warn They Can’t Be Trusted

DeepSeek just published the boring part of agent AI that actually moves product quality: the sandbox farm. An arXiv paper on DeepSeek Elastic Compute (DSec) — covered by TechNode and Seoul Economic Daily — describes a unified platform for function-call, container, microVM and full-VM sandboxes, co-scheduled with reinforcement-learning rollouts. Founder Liang Wenfeng is among 130-plus authors. The paper landed on arXiv around 19 September 2026.

The scale numbers are the lede. A production-scale unit is roughly 160 nodes (about 30,000 CPU cores and 250 TB memory in SED’s telling). DeepSeek claims on the order of 3 million sandboxes per day, more than 380,000 concurrent, and over 5,000 creations per second. A unified SDK exposes FnCall through full-VM backends and deliberately decouples stateful agent rollouts from preemptible GPU training — the classic pain point when you try to RL-train tool-using agents without melting the cluster.

Then comes the honesty rare in launch blogs: DeepSeek says an AI agent at work “can never be trusted.” Agents damaged execution environments, exhausted resources, interfered with peers, or found answers via “unexpected paths” (reward hacking in all but name). Monitoring is being reinforced. Evaluation claims include on-demand loading cutting cumulative disk writes by about 57%, with eager pulls about 1.7× slower to completion in an ablation. Treat those as author metrics; the architectural confession matters more.

Why should search marketers care about someone else’s sandbox fleet? Because the same agent stacks that need DSec-scale isolation are the stacks that will browse, extract and cite your pages inside Chinese and global answer products. Safer, cheaper agent training usually means more agent traffic against the open web — and more pressure on extractable, time-valid evidence. It also explains why DeepSeek keeps shipping Flash-tier economics: agent loops are input-heavy and failure-prone; infrastructure that contains blast radius is what lets them scale tool use without pagering every GPU job.

Practical takeaway: if you prototype GEO tests against DeepSeek-with-web-search, assume the citation behaviour will keep shifting as sandbox curricula change. Log which domains get named after major DeepSeek infra drops, not only after model-card renames. And stop treating “agentic search” as a prompt trick — DeepSeek just showed it is a datacentre product with a trust problem they are willing to print in public.

One more competitive read: DSec sits beside this month’s V4.1 Flash routing story. Flash is the cheap production answer path; DSec is how DeepSeek teaches agents not to torch the lab while chasing rewards. International SEOs who only track consumer app MAU league tables are watching the wrong layer. The publishers who win the next wave of Chinese AI citations will be the ones whose pages survive hostile, tool-using extractors trained in farms like this — not the ones optimising for a static chatbot screenshot from Q2.

Dan Taylor is the founder and editor of International Search Insights (ISI), the publication behind ReadIsi, and a Partner and Head of Innovation (Search & AI) at SALT.agency. He has worked in search for more than a decade, in-house and agency-side, and is regarded as one of the West’s leading voices on search engines beyond Google, including Yandex, Baidu, Naver and Cốc Cốc. On ReadIsi, Dan covers the engines and AI answer platforms that English-language coverage tends to miss: Baidu and ERNIE, Yandex and Alice AI, Naver, Cốc Cốc, Shenma, Sogou, WeChat and Weibo search, plus DeepSeek and Alibaba’s Qwen. He also keeps dated records of documented changes, such as the Yandex, Baidu and Naver algorithm update histories and the DeepSeek, ERNIE and Qwen model update histories. His approach is to take learnings from global search engines and apply logic with a high evidence threshold. A claim needs a source, patterns seen in one market are treated as a hypothesis until they are checked in another, and where the evidence is thin he says so. For brands, that means advice built on what can be shown, with measurement that separates being cited from being chosen. Dan is credited with coining the term “edge SEO” and won the inaugural TechSEO Boost research competition in 2018 for his work on CDN and edge changes for SEO. In 2026 he was credited with coining “Infinite Tail” to describe how keyword research is changing with AI search. He is the author of AI Search Marketing For The Travel Sector, writes for Search Engine Journal, Search Engine Land and Travel Marketing Guy, and has advised technology, SaaS, travel and enterprise ecommerce brands, including Cloudflare, GitLab and Airbnb. He has given multiple talks and panels at the International Search Summit in Barcelona, and has also spoken at Optimization Moscow, brightonSEO, Google Search Central Live Deep Dive Asia in Bangkok, Athens SEO and the Asia SEO Virtual Summit 2026, where his talk looked at how Baidu, Naver, DeepSeek, Tencent, Cốc Cốc and Yandex built their own AI answers.