Subscribe

The players · lab

DeepSeek

It publishes its model weights for anyone to inspect and reuse, which is more than most rivals do, but the headline number that made it famous has never matched independent accounting of what training it actually took.

The 30-second read

The Hangzhou lab that rattled Silicon Valley with a cut-price reasoning model, gives its weights away for free, and is now racing to build the chip supply chain to keep training the next one without Nvidia.

2 weak · 3 mixed · 1 solid

Watch next

Whether DeepSeek's push onto Huawei's Ascend chips for its new Inner Mongolia data centre can actually train a frontier model at the quality and cost it claims once it stops leaning on Nvidia hardware.

Checked 24 Sept 2026 · Open each finding for dated sources

How did the old BS index work?

The retired formula gives 2 Weak × 100 + 3 Mixed × 50 + 1 Solid × 0, divided by 6 rated questions = 58 after rounding. Unknowns are excluded. It is an editorial shorthand, not a measured probability, so this dossier leads with individual findings instead.

The six questions

  1. WeakLaunch claims vs independent tests

    When they say a model is better, faster or cheaper, do independent evaluations agree?

    A paper co-authored by DeepSeek's founder and published in Nature said the R1 reasoning model's final training run cost $294,000 on 512 Nvidia H800 chips. An independent breakdown of DeepSeek's total server capital expenditure put the real infrastructure cost at $1.3 billion, hundreds of times the figure that first made headlines.

    2 receipts
    • The Nature article, which listed Liang as one of the co-authors, said DeepSeek's reasoning-focused R1 model cost $294,000 to train and used 512 Nvidia H800 chips. The Hindu, 18 Sept 2025 ↗
    • Of the $1.3 bn CapEx calculated, much of it is directed toward operating and maintaining the expensive GPU clusters, which forms the backbone of DeepSeek's computational power. Deccan Herald, 31 Jan 2025 ↗

  2. MixedMoney claims vs the filing

    Do their valuation, revenue and user numbers survive contact with primary disclosures?

    DeepSeek closed its first outside funding round at more than $50 billion in valuation in June 2026, using a structure that kept founder Liang Wenfeng in control. Within a month it was reportedly seeking a fresh round at a $74 billion valuation ahead of a possible IPO, both figures resting on unaudited private reports rather than filed disclosures.

    2 receipts
    • DeepSeek has reportedly raised $7.4 billion at a valuation above $50 billion in a founder-controlled funding round. Moneycontrol, 15 June 2026 ↗
    • China's DeepSeek plans a new funding round at a $74 billion valuation, aims to raise up to 50 billion yuan, and is considering a Shanghai IPO. Hindustan Times, 15 July 2026 ↗

  3. MixedPromises kept

    Did the things they announced with a date actually ship, on time, as described?

    DeepSeek's next flagship model, V4, was expected around the turn of the year but was still described as an unreleased, hidden-from-US-chipmakers preview in late February 2026 and had still not shipped by early April, before finally launching that month.

    2 receipts
    • DeepSeek, the Chinese artificial intelligence lab whose low-cost model rattled global markets last year, has not shown U.S. chipmakers its upcoming flagship model for performance optimization. Mint, 25 Feb 2026 ↗
    • Since then, there has been a great deal of interest in DeepSeek-V4, a next-generation model that has yet to be released. The Hindu, 3 Apr 2026 ↗

  4. MixedSafety and incidents

    When something went wrong, did they disclose it quickly and honestly?

    Security researchers at Wiz found a DeepSeek database sitting open on the internet with more than a million lines of chat logs and software keys, fully accessible with no authentication. DeepSeek was not the one to catch it, but once Wiz alerted the company it secured the database within about an hour.

    2 receipts

  5. SolidTransparency

    Do they publish what an outsider needs to check them?

    DeepSeek released its R1 reasoning model under an open MIT license, letting anyone download, inspect, run and modify the weights rather than access them only through a paid API. Outside commentary has pointed to that openness as letting researchers examine and scrutinise the model in ways closed labs do not allow.

    2 receipts

  6. WeakLegal and regulatory record

    What do courts and regulators say about them?

    Italy's data protection authority, the Garante, ordered DeepSeek to block its chatbot in the country after the company failed to satisfactorily answer questions about what personal data it collects, where it is stored and how users are told. Ireland followed with its own block, and regulators in both countries opened investigations into DeepSeek's data practices.

    2 receipts
    • The authority, called Garante, expressed dissatisfaction with DeepSeek's response to its initial query about what personal data is collected, where it is stored and how users are notified. The Hindu, 30 Jan 2025 ↗
    • DeepSeek has been blocked in Italy and Ireland, with regulators investigating its data practices. Moneycontrol, 29 Jan 2025 ↗

Their models: what they said vs what others found

DeepSeek V4.1 Flash open-weights · 10 Sept 2026

They said552-billion-parameter model built to outperform key rivals while costing dramatically less to run.

Independent testsNothing independent yet. Treat the launch numbers as marketing until someone else runs them.

DeepSeek V4 Pro (GA) reasoning · 13 Aug 2026

They saidGeneral-availability V4 Pro scores 62.7 on the DeepSWE coding benchmark, up from 12.8 in the April preview.

Independent testsIndependent evaluator Artificial Analysis ranks it below Moonshot's Kimi K3 and Anthropic's Claude Opus 5 on its Intelligence Index despite the price gap.

Latest · 4 Sept 2026 DeepSeek is building a data centre in Inner Mongolia meant to house at least 160,000 Huawei Ascend 950DT AI chips, one of the largest known clusters of Chinese-made AI accelerators. Moneycontrol ↗

Each question is rated separately against its receipts. Ratings can change as evidence improves. Compare every player →