Jonathan Ross on DeepSeek, Open Models and AI Infrastructure

Groq co-founder and CEO Jonathan Ross joins Harry Stebbings shortly after DeepSeek's model release to discuss its significance, training costs, open models, AI infrastructure and geopolitics. Groq sells inference hardware and services and hosts DeepSeek models, so Ross has a direct commercial interest in the market. Several allegations and forecasts in the conversation are not independently substantiated by the interview.

Ross calls the release a major event and says it showed strong model performance could be achieved with less computing expenditure than many assumed. He cautions that a reported roughly $6 million training-compute figure is not the full cost of developing a model. The interview does not provide complete audited costs, and the figure should not be presented as the cost of the entire project.

He explains model distillation by comparing a smaller model learning from outputs of a stronger one to a student receiving expert tutoring. Ross alleges DeepSeek used OpenAI outputs and incurred additional costs generating data, while also praising its reinforcement-learning engineering. He provides no direct evidence for the alleged sourcing, expenditure or authorization; these claims remain allegations, not established findings.

Ross says scaling results depend on data quality as well as token quantity, and uses AlphaGo and self-play to illustrate how training can produce better examples. He argues high-quality generated data can improve a model with fewer tokens. This is a conceptual explanation, not a full accounting of DeepSeek's training pipeline or proof of its data sources.

He praises an approach that checks answers automatically for tasks with verifiable results, reducing the need for people to grade every response. When Stebbings asks about reward-model details, Ross says he has not studied that area deeply. The exchange does not establish that every DeepSeek training task was automatically verifiable.

Asked about hardware restrictions, Ross argues remote access to cloud compute can complicate controls and says Groq blocks some IP addresses from China, although intermediaries can obscure location. He does not know whether a particular cloud use would violate export law. His remarks are policy and operational views, not a legal determination or proof of a specific transaction.

Ross expresses concern that sensitive data entered into a hosted AI service could be accessible to a government, and speculates about Chinese state access and influence. He offers no evidence that DeepSeek transferred particular user data to authorities or that all services retain deleted data in the way he describes. Privacy risk depends on the service's actual architecture, policies and legal context.

He argues that political constraints can shape responses from services operating in China and gives sensitive political questions as an example. Asked whether DeepSeek is an instrument of the Chinese government, he says companies operating there face pressure; his broader assertions about motives and access are opinions and allegations, not findings established in the interview.

Ross says Groq had previously declined to run Chinese models but made DeepSeek available through its API after the model attracted users. He says Groq does not store their data and presents its hosting as an alternative. This is a provider's self-description and commercial positioning; the interview does not independently audit retention or security.

He thinks DeepSeek made model commoditization more apparent and distinguishes model capabilities from possible competitive advantages such as brand, network effects and scale. He credits OpenAI's brand while interpreting Stargate as an effort to build infrastructure scale. These are strategic judgments, not proof that model differences or switching costs have disappeared.

Stebbings questions whether a proposed $500 billion infrastructure effort looks excessive after a more efficient model. Ross argues demand for inference may make substantial spending worthwhile, recalling that at Google inference cost multiples of training. He estimates inference could eventually account for 95% of computing use, a prediction with no defined time or measured baseline in the interview.

On a possible US response, he considers both restrictions and incentives, but says he does not know the best solution. He speculates that free services could be subsidized for strategic reasons and suggests reciprocal responses to subsidies. The interview supplies no evidence for the proposed subsidy or a legal basis for any particular restriction.

Ross argues open models can spread widely and predicts that openness will prevail. He recommends that OpenAI consider releasing models openly to compete for users and rely on brand and distribution. That is his proposed strategy for another company, not evidence of its plans or a guarantee that open models always win.

He sees Meta's network effects as potentially complementary to open model development and says model switching can be easier than changing cloud infrastructure. His claim that models have no switching costs is an opinion; differences in integration, behavior, evaluation and deployment can matter. He again speculates about competitors' data practices without presenting direct evidence.

The speakers discuss Europe, competition with China and alleged intellectual-property theft. Ross calls for greater European risk-taking and fair competition, while criticizing what he believes are unfair government advantages. His broad accusations about state conduct are his views and are not independently documented in this conversation.

He worries AI could change the relationship between population and economic output if chips and power can augment a workforce. As a proposal for Europe, he suggests rapidly creating many startup hubs akin to Station F to concentrate ambitious entrepreneurs. Both the economic scenario and the proposed results are speculative.

He distinguishes generative AI from the information-distribution technologies that preceded it and emphasizes products built on models. Discussing Perplexity, he thinks it could benefit if model hallucinations fall enough to enable more applications. His examples involving medical and legal work are predictions; the interview does not validate safety or remove the need to check outputs.

Ross argues lower inference prices can increase usage, invoking Jevons' paradox. He says Groq sees developer interest rise when model costs fall and predicts greater demand for compute. This is the experience and forecast of an inference provider; lower prices do not by themselves prove aggregate spending or demand will always rise.

Stebbings says he bought Nvidia shares after a decline and asks whether Ross shares his optimism. Ross believes efficiency improvements can increase the value of compute demand, while noting markets do not immediately reflect fundamentals. Their investment views are not independent valuations or a recommendation to buy a stock.

Ross expects training to remain high margin but argues inference will be a larger, higher-volume market with lower margins. He suggests specialized inference vendors such as Groq could take that work while Nvidia remains strong in training. This is a competitor's commercial thesis, not a verified market-size or margin forecast.

Explaining mixture-of-experts models, he says only selected experts are activated for a given input, allowing a large parameter count without using all parameters for every token. He gives approximate figures for DeepSeek and explicitly expresses uncertainty about the number of experts selected. Those estimates are illustrative rather than a verified specification or benchmark.

He points to a smaller Meta model that he says improved substantially through fine-tuning as an illustration of data quality. He predicts other labs will apply sparse architectures, synthetic data and more training, making tradeoffs between training costs and inference costs. The comparisons and forecasts are his interpretation, without a controlled benchmark in the interview.

Asked about DeepSeek limiting some sign-ups, Ross attributes the restriction to insufficient compute and says serving each additional query requires capacity. He also argues the project's reported compute figure received effective publicity. His explanation of DeepSeek's operational constraints and marketing intentions is speculative in the absence of direct evidence.

On Stargate, he says he has heard conflicting accounts about financing and interprets the project as a bid for infrastructure scale. He thinks building power and data centers takes time and predicts OpenAI's brand may strengthen. The $500 billion discussion does not establish committed financing, deployment or future returns.

Ross is both excited and nervous about faster AI development. He worries models could automate discovery and exploitation of cybersecurity vulnerabilities, creating a cycle of attack and defense with difficult attribution. The scenario is a security concern, not proof of a particular attack or attribution.

He expects natural-language tools to make software creation faster and describes an application built by a Groq engineer. Asked where value remains if many people can build similar tools, he argues quality and craftsmanship can differentiate products. He distinguishes an incomplete early release from one that fails basic reliability, though his claims about future productivity are predictions.