DeepSeek V4 and the Cost of Frontier AI Models

ZUSAMMENFASSUNG

In 30 Sekunden

In this April 2026 short, Matthew Berman argues that DeepSeek V4 could challenge US AI providers by offering capable open weights and lower prices. He cites a price-versus-performance chart and imagines a business choosing a cheaper model that it can customize. His claims about training hardware, comparative cost and broad enterprise suitability are not established by the short; model benchmarks and deployment costs vary.

1A new Chinese modelBerman presents DeepSeek V4 as a potential challenge to US model providers; the geopolitical outcome is his prediction.
2Price versus performanceHe uses an approximate chart to argue that DeepSeek offers a compelling price-performance tradeoff.
3The business decisionBerman argues that open weights and customization could make a lower-priced model attractive for enterprise workloads.

Das Wichtigste

DeepSeek announced V4 Pro and V4 Flash preview models and released model weights in April 2026.
Berman predicts that lower prices and the ability to customize weights could attract business customers.
A model’s performance on selected benchmarks does not prove it matches every competitor on every task.
API prices, training costs and the total cost of running a model are distinct measures.
The prediction that this release will end US AI leadership is an opinion about an uncertain future.

ThemenFoundation Models AI cost & performance Open-weight models Geopolitics

VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Transkript zeigen

Nur das Transkript

DeepSeek V4 and competition

Berman argues that DeepSeek’s release could challenge US AI leadership. He points to its open weights and claims about limited hardware and resources.

Matthew Berman argues that DeepSeek’s April 2026 V4 release could challenge US AI leadership. He describes the Chinese model as competitive with leading alternatives and released with open weights. DeepSeek announced separate V4 Pro and V4 Flash previews; the ultimate effect on national AI leadership is a prediction, not an established outcome.

Berman says DeepSeek achieved its result with fewer resources and constrained NVIDIA GPUs. The short does not document the complete training hardware, development expenditure or an equivalent comparison with competing models. DeepSeek’s own benchmark results show strengths and weaknesses across tasks rather than proof of matching every leading model in every use.

Price, performance and enterprise adoption

Berman compares model prices and asks why companies would pay more for closed models. The comparison is a historical snapshot and depends on individual tasks and deployment costs.

He presents a price-versus-performance chart and argues that businesses should consider the tradeoff between capability and cost. The graph shows approximate prices and a composite score; benchmark choices, model settings and pricing dates affect the result. API token prices are not the same as model training cost or the total expense of self-hosting.

Berman cites $30 per million output tokens for GPT-5.5 and a similar price for Claude Opus 4.7 in his comparison, then argues that DeepSeek could meet many business needs at a fraction of the cost. These are figures and assumptions from the video’s April 2026 context, not current quotes or proof that any model fits every use case.

For a company that does not need the most demanding frontier research capability, Berman says open weights provide another attraction: the ability to customize or fine-tune a model. He asks why such a buyer would pay more for a US provider’s hosted model. The answer depends on quality, security, infrastructure, support and the specific workload as well as token prices.