Notes from a DeepSeek HQ visit: 300 employees, no branding, and no interest in the singularity

Niko McCarty published a short thread of notes from a visit to DeepSeek's headquarters in Hangzhou, and the details are worth reading carefully because they cut against most of the narrative that built up around the company after it released R1 in January 2025. The claims here are observational, not benchmarked, so treat them as a field report rather than a technical disclosure. Still, the specifics are concrete enough to be useful.

{{IMAGE:1}}

What's claimed

The headline numbers: DeepSeek operates out of an unmarked 12-story building with no visible branding, employs around 300 people, and was founded in 2023 by Liang Wenfeng, running inside his quantitative hedge fund High-Flyer until relatively recently. McCarty met with the Head of Data and the Head of Infrastructure, the latter described as roughly 30 years old and one of the better AI buildout and energy specialists in the country. The labs reportedly felt young and energetic.

The more interesting claims are about posture. The team said they are content to remain roughly six months behind U.S. companies while keeping a low profile and a small team. When asked about their proudest achievement and their exit plans, they pointed backward to R1 rather than forward to some next model or grand vision. They also said they do no red teaming on their models, and that they had never met anyone from Anthropic.

What's actually new here

Nothing about the technology, and that's the point. There is no new model, benchmark, or architecture in these notes. What's new is the texture: a counterweight to the assumption that a lab punching at frontier-adjacent levels must be a scaled-up operation racing toward general intelligence.

The 300-employee figure is the load-bearing detail. If accurate, it puts DeepSeek at least an order of magnitude below Anthropic in headcount, which lines up with the company's earlier reputation for doing a lot with comparatively constrained resources. The R1 release was notable precisely because it demonstrated a reasoning model trained and served at a cost structure that embarrassed assumptions about how much capital frontier-style results require. A small, infrastructure-heavy team that openly accepts a six-month lag is consistent with that story. They are optimizing for efficiency and recruiting density, not for being first.

The competitive map in the notes is also useful context. McCarty lists Alibaba's Qwen, ByteDance, and Moonshot's Kimi as the main domestic rivals, with most Chinese users reaching for Kimi or DeepSeek, and younger users tunneling through VPNs to reach Claude despite Anthropic's regional access blocks. Big labs cluster in Beijing near Tsinghua and Peking University, with Hangzhou as the notable exception housing both DeepSeek and Qwen. Poaching between groups is described as common, which mirrors the U.S. market.

The parts that should make practitioners pause

Two claims deserve scrutiny rather than acceptance. First, the statement that DeepSeek does no red teaming. Taken literally, that's a meaningful safety gap, and it fits with the broader observation in the notes that Chinese AI regulation targets how models are deployed in products and services rather than the models themselves. The regulatory burden lands downstream, on applications, so an upstream lab has less external pressure to invest in adversarial testing. Whether "no red teaming" means none at all or simply none of the formal, named programs that Western labs publicize is unclear from a secondhand account, and the distinction matters.

Second, the cultural framing. The team reportedly brought up job loss, already a real problem given high youth unemployment in China, as their primary concern, and showed no interest in AGI-takeover scenarios. McCarty's read is that China treats AI as another technology rather than a singularity moment, with national attention still on infrastructure, basic needs, and medicine. That is a sociological observation, not a technical one, and it's easy to over-index on a single visit. But it's a reasonable corrective to the habit of assuming every frontier lab shares Silicon Valley's eschatology.

Why it matters

For anyone tracking the field through papers and model cards rather than press releases, these notes are a reminder that organizational shape tells you something. A lab that stays small, stays quiet, accepts a deliberate lag, and points to a year-old release as its proudest work is making different bets than one chasing the next capability jump every quarter. DeepSeek's published work has consistently emphasized training and inference efficiency, and a 300-person team with a standout infrastructure lead is the kind of operation that produces those results.

The limitations of the report are obvious and McCarty doesn't hide them: it's one visit, a few conversations, no verification of headcount or internal practices, and the team had reason to downplay their ambitions to visitors. None of it changes what R1 actually does on benchmarks. What it offers is grounding. The lab behind one of the more efficient reasoning models is, by this account, smaller and less apocalyptic than its reputation, and seems to want it that way.

{{IMAGE:2}}