One of the most consequential figures attached to DeepSeek was a dollar amount. Readers could reasonably ask whether a small laboratory had reproduced the work of much larger companies for less than six million dollars. Founder Liang Wenfeng’s history in quantitative finance explains where some of the machinery came from. The training paper explains which work the figure covers. Both are needed to understand what other laboratories might reproduce.
DeepSeek · High-Flyer
DeepSeek develops and serves language models. High-Flyer’s origins are in investment management. Early corporate records discussed in a 2025 US congressional report traced ownership through a holding company. A percentage held by an intermediary cannot be read as Liang’s personal stake, particularly after subsequent financing. US congressional report and corporate record references.
Capital and voting rights. Founder-backed development preceded outside fundraising. The Information subsequently reported a private round exceeding US$7 billion. Reuters reported that DeepSeek had brought in CITIC Securities for a prospective mainland flotation. Neither event supplies an exchange-traded share price or an allocation available to the public. The cited materials do not disclose the complete current voting register, Liang’s post-round percentage or a comparable founding cash figure. Reporting on the private financing. Reuters on the proposed listing.
The denominator behind US$5.576 million
The DeepSeek-V3 paper reports 2.788 million NVIDIA H800 GPU-hours across the recorded training stages. At an assumed US$2 per GPU-hour, the multiplication yields US$5.576 million. The authors explicitly exclude earlier research and experiments. Someone budgeting to repeat the work must also pay for unsuccessful runs, staff, infrastructure and the data pipeline. The paper gives engineers a useful comparison precisely because it states what was counted. DeepSeek-V3 technical report.
Which model is the customer using?
DeepSeek’s developer log records the V4.1 Flash update of 17 October and continuing access to V4 Pro. It dates changes to search and document handling separately. A purchasing team comparing responses needs the service, model version, date and tools enabled. A search tool that retrieves a recent document can improve an answer even while the underlying weights stay the same. DeepSeek’s dated API record.
A privacy finding with a named authority
South Korea’s Personal Information Protection Commission examined the DeepSeek service and documented transfers to a third party involving user inputs. It reported changes including an end to the transfer of prompts to that recipient and recommended corrective action. This is a finding about a deployed service. An organisation running released weights on its own infrastructure should map its own data flows before deciding whether the same exposure exists. Korean regulator’s findings and recommendations.
Information checked on 15 November 2026. Ownership figures and proposed transactions refer to the dates in the linked records.
Liked this post? Follow this blog to get more.