“In late December, a company called DeepSeek, apparently initially built for quantitative trading rather than LLMs, produced a nearly state-of-the-art model that required only roughly 1/50th of the training costs of previous models, — instantly putting them in the big leagues with American companies like OpenAI, Google, and Anthropic, both in terms of performance and innovation.”
Our reading, the sentence that was checked:
In late December, DeepSeek, apparently first built for quantitative trading, produced a nearly state-of-the-art model that needed roughly 1/50th of the training costs of previous models, putting it in the big leagues with American companies such as OpenAI, Google and Anthropic in both performance and innovation.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The race for "AI Supremacy" is over — at least for now., paragraph 1.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-the-race-for-ai-supremacy-is-over-e3e128-a1
We recorded 17 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-01-26. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
Conditions
- apparently initially built for quantitative trading (qualifier, stated by the speaker)
- roughly 1/50th (qualifier, stated by the speaker)
From the speaker
- In late December
- DeepSeek
- apparently initially built for quantitative trading rather than LLMs
- nearly state-of-the-art model
- roughly 1/50th of the training costs of previous models
- performance and innovation
Our interpretation
- The model is the original DeepSeek-V3 described in the surrounding passage.
- Nearly state-of-the-art means broadly competitive with leading contemporary language models, without requiring first place on every task.
- The cost assertion means approximately two percent of earlier frontier models' training costs, measured with comparable inclusions and pricing assumptions.
- No comparison model, currency, accounting basis, or training-cost boundary is supplied. The contract therefore leaves the accounting basis unknown.
- December 2024 is the release period being assessed, not an assertion that all training expenditure occurred during December.
- The hedged quantitative-trading aside is background, not a condition that activates the capability and cost assertions.
Counts as supported if: The December V3 release demonstrates near-frontier capability and its training cost is approximately one-fiftieth of identified earlier frontier models on a comparable cost basis.
Counts as contradicted if: Comparable evidence establishes that V3 was substantially below frontier performance, or that its training-cost ratio was materially inconsistent with one-fiftieth.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The December report supports the performance and efficiency story, but it does not establish the one-fiftieth cost assertion. Its $5.576 million figure uses an assumed GPU rental rate and excludes earlier research and experiments. Marcus names neither the comparison models nor their cost basis. Selecting a convenient denominator would determine the result rather than test his stated comparison.
Limitations
- Unknown accounting basis prevents a supported or contradicted financial verdict.
- An unresolved ratio does not negate the reported engineering advances.
- The training figure is a company-reported estimate, not an audited expense total.
Evidence (2 items)
Context only: DeepSeek-AI / arXiv,
Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.
Why it matters here: The report gives a rental-equivalent estimate for official V3 training and explicitly excludes prior research and ablation experiments; it also reports competitive capability evaluations. It supplies neither a named denominator nor a verified one-fiftieth ratio.
- Publisher
- DeepSeek-AI / arXiv
- Kind of source
- Preprint, not peer reviewed
- Measures
- DeepSeek, DeepSeek-V3 official training compute cost at an assumed GPU rental rate, USD USD, company reported, ? to ?
- Describes events of
- Public from
Context only: Epoch AI,
Our primary approach calculates training costs based on hardware depreciation and energy consumption over the duration of model training.
Why it matters here: Epoch distinguishes depreciated hardware costs, cloud rental estimates, and broader development costs. This methodological evidence explains why combining an arbitrary earlier estimate with DeepSeek's figure would not settle the claim.
- Publisher
- Epoch AI
- Kind of source
- Independent measurement
- Describes events of
- Public from
What would settle it
Identify the intended earlier models and obtain comparable training-cost figures covering the same stages and cost components, with uncertainty small enough to assess the approximate one-fiftieth ratio.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, The race for "AI Supremacy" is over — at least for now.
In late December, a company called DeepSeek, apparently initially built for quantitative trading rather than LLMs, produced a nearly state-of-the-art model that required only roughly 1/50th of the training costs of previous models, — instantly putting them in the big leagues with American companies like OpenAI, Google, and Anthropic, both in terms of performance and innovation.
- Where
- paragraph 1
- Published
- 2025-01-26
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-race-for-ai-supremacy-is-over-e3e128-a1: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-the-race-for-ai-supremacy-is-over-e3e128-a1.