The rise of generative AI and large language models (LLMs) has brought incredible advancements, but also new security risks. Imagine training a cutting-edge AI model or querying it with sensitive data—you wouldn't want your cloud provider or anyone else snooping around, would you? That's where confidential computing comes in. Traditional methods using CPUs or integrated GPUs for trusted execution environments (TEEs) often fall short. They either don't provide enough protection or rely on components that broaden the attack surface. A new approach, called Ascend-CC, offers a fresh perspective. It focuses on using discrete NPU devices as the bedrock of trust, eliminating reliance on the potentially compromised host system. Ascend-CC wraps a shield of encryption around your AI models and data, keeping them safe from prying eyes, even the cloud provider's. This innovative approach uses a clever 'delegation-based memory' method. Think of it like a secure drop-off point: the host delivers the encrypted data and model to the NPU, then loses access. Only then does the NPU decrypt and process the information, encrypting the results before sending them back. Even trickier attacks, like injecting malicious commands into the model, are thwarted through task and binary attestation. This rigorous verification ensures that the model executes exactly as intended, preserving its integrity. The best part? Ascend-CC works seamlessly with current AI frameworks like PyTorch. No need for developers to overhaul their code—confidential computing becomes a simple, yet powerful upgrade. Tests with leading LLMs like Llama 2 and Llama 3 show that Ascend-CC adds minimal performance overhead, making it a practical solution for protecting your valuable AI assets. This research marks a significant step towards making confidential computing the norm in AI, opening doors for broader adoption and paving the way for a more secure AI-powered future.