Three years after banning ChatGPT over a source code leak, Samsung just became one of OpenAI's biggest enterprise customers

Good morning 👋 The company that banned ChatGPT and made leaking data into it a fireable offense just became one of OpenAI's largest enterprise customers. The reversal is more telling than the deal.

In today's issue:

• 🔭 Samsung goes from ChatGPT ban to full enterprise deployment
• 🔬 Research: agent identity, fault recovery, and diffusion model interpretability
• 🏛️ Policy: the EU picks a team to build a 400B-parameter open-source frontier model
• 🔗 Quick hit: NVIDIA rewrites data-center cooling specs

🔭 THE ONE THING

🔄 Three years after banning ChatGPT over a source code leak, Samsung just became one of OpenAI's biggest enterprise customers

OpenAI announced it's deploying ChatGPT Enterprise and Codex to all Samsung Electronics employees in South Korea and to all employees in its Device eXperience (DX) division globally. The DS semiconductor unit stays under tighter restrictions outside Korea, so "worldwide" is doing a bit of work in the framing. Still: this is a company that made ChatGPT a fireable offense in 2023 after engineers leaked chip design code into it, built a mandatory AI security training gate before handing out credentials. The reversal is the real signal. Enterprise AI is past the pilot phase, and the holdouts are becoming the outliers.

🔬 RESEARCH HIGHLIGHTS

• Sovereign Execution Brokers — authorizing an agent's identity isn't the same as authorizing its actions. This paper proposes a broker layer that sits between agentic reasoning and production mutation calls, requiring certificate-bound authority per action. An agent with a valid IAM token can still hallucinate a destructive DELETE. Identity is necessary, not sufficient. link

• Hierarchical Recovery for Cross-Device Agents — when an agent fails mid-task (browser times out, API drops), the standard response is full replanning. Expensive and usually unnecessary. Local recovery at the failure point outperforms global replanning on both success rate and latency. If you're building multi-device orchestration, subtask-level fault handling beats a full restart. link

• How Transparent is DiffusionGemma? — diffusion LMs generate in a continuous latent space, meaning more computation happens somewhere you can't read. This paper finds that does hurt interpretability relative to autoregressive models. If "why did it do that" matters for your use case, that's a real cost to weigh against the speed argument. link

🏛️ POLICY & REGULATION

• European Commission selected EUROPA, led by Italian startup Domyn, as winner of the Frontier AI Grand Challenge: build a 400B+ parameter open-source model across all 24 official EU languages. Challenge launched February 2026; winner announced June 19. Most concrete EU move yet toward a homegrown frontier model, and the Commission framed it explicitly as a sovereignty play. Whether Domyn has the compute and coordination to actually train at that scale is the question the announcement didn't answer. Source

🔗 QUICK LINKS

• NVIDIA raised the liquid cooling threshold on its AI server hardware to 45°C, which means data centers can run a direct warm-water loop without a chiller — simpler plumbing, lower operating cost. link

See you tomorrow.

Pradeep Perugu