Entering 2025, the commercial question for AI was how model capability would translate into useful work. A business needed to connect performance on a test with the cost, reliability and oversight required in its own operations.

Published on 3 February 2025 and developed with OGroup CEO Maja Vujinovic, this report examines model development, computing infrastructure, agent-based applications, capital allocation and the effects of regulation and geopolitics. This introduction places those themes in their original time frame.

Read training-cost claims precisely

DeepSeek’s December 2024 technical report states that V3 trained on 2,048 Nvidia H800 GPUs. It reports 2.788 million GPU-hours across official training stages, with a cost estimate of $5.576 million at an assumed $2 per GPU-hour. The estimate excludes earlier research and ablation experiments. It therefore does not measure the company’s complete development cost. Source: DeepSeek-V3 Technical Report, version 1, Table 1 and section 3.1.

Match autonomy to the work

Anthropic’s agent-design guidance distinguishes fixed workflows from systems in which a model selects its next steps and tools. It recommends starting with a simple implementation and adding complexity when the task benefits, because agentic systems can increase latency and cost. Source: Building Effective Agents, originally published December 2024.

Our suggested commercial test follows one task from start to finish. Measure successful outcomes against the existing process, including the time spent checking and correcting results. That comparison makes model performance relevant to a purchasing or deployment decision.

Separate evidence from the outlook

The report’s published outline includes future predictions and potential investment opportunities. Those are dated analysis. They should be read alongside the evidence available at publication and reassessed as model capabilities, costs and rules change.

Sources and limits: This introduction uses the public report outline and the named technical references. DeepSeek’s figures are developer-reported, not independently reproduced here. Anthropic’s page has been updated since its original publication. This page is a historical report introduction, not a current model ranking.

Correction: The original public preview said DeepSeek-V3 was built without U.S. chips. DeepSeek’s technical report identifies Nvidia H800 GPUs. We have replaced that claim with the stated hardware and defined training-cost estimate.

Continue reading with your existing 51 PRO subscription.