Today’s radar was packed with model launches—Qwen, Gemini, GLM—but what stayed with me was an infrastructure story, not a new model. I set this edition aside to talk about homomorphic encryption and why it matters to anyone working with sensitive data in credit.

Whenever someone asks me why AI still moves slowly inside banks, the answer is almost always the same: sensitive data.

I have worked with structured-credit products for years, and I know the weight of that constraint. Receivables, invoices, borrower information—these are all data points that cannot simply circulate through any AI model without a serious layer of protection.

In brief

  • Fully homomorphic encryption allows a model to process encrypted data without ever seeing the real content.
  • The server computes over the ciphertext and returns an encrypted result that only the key owner can decrypt.
  • This infrastructure could unlock AI use in regulated sectors such as healthcare, finance and credit, which are currently held back by the risk of exposing sensitive data to third parties.
  • For product teams, the next step is to track maturity and at-scale viability, not promise immediate production.

What homomorphic encryption changes

This week, Google detailed an advance that got me excited: fully homomorphic encryption (FHE) running in a practical way through an open-source compiler called HEIR. In practice, it allows an AI model to process encrypted data without ever “seeing” the actual content. The server computes over the ciphertext and returns an encrypted result that only the key owner can decrypt. For anyone who wants the technical details, Google published the explanation of how it is making private AI practical with homomorphic encryption.

It sounds like science fiction, but this is real infrastructure happening now. And it is the kind of building block that can unlock regulated sectors—healthcare, finance and credit—that are held back precisely by the risk of exposing sensitive data to third parties in order to gain intelligence.

The trade-off between privacy and applied intelligence

On the product side, this changes the conversation. Much of my work in structured credit involves designing processes that protect data while still extracting value from it. AI product management in regulated sectors has to handle this balance from the moment a use case is designed, not only after the solution is ready.

If this layer of encryption matures and becomes viable at scale, the trade-off between privacy and applied intelligence becomes much smaller. Before thinking about production, the AI risk matrix helps organize the impact, evidence and level of control each use case requires.

This is not about putting it into production tomorrow. It is about watching closely, because this kind of infrastructure changes what will be possible to build two or three years from now—and people who understand it early will be ahead when it is time to design products. It is also an AI governance discussion: data protection needs to be part of the architecture and product decisions, not a detail added later.

The rest of the radar

Alibaba launches Qwen3.8-27B with open weights — expands the set of competitive models that can run locally, making it possible to assess cost, privacy and performance trade-offs against closed APIs. Read more

Gemini 3.7 Flash accelerates coding and agents — pricing at half the previous Flash’s cost shows that “fast and cheap agents” have become a central competitive vector for AI products. Read more

Z.ai’s GLM-5.3 nearly reaches the frontier in cybersecurity — Chinese labs are closing gaps in specialized domains, not only on price, which is worth monitoring when evaluating open models. Read more

OpenAI and Anthropic cut prices as Chinese rivals intensify — shifts competition from “best benchmark” to “how much useful work each dollar of inference buys,” changing how AI features are priced. Read more

Writer launches Palmyra X6 and cuts agent costs — shows that the orchestration layer (memory, routing and retries) can weigh as much as the model in the unit economics of agentic products. Read more

Deltix uses AI for software testing — an example of AI agents advancing in QA workflows, relevant to speeding up release cycles and lowering the cost of quality. Read more

Mole, a deep-research agent for the terminal — illustrates the spread of specialized agents beyond traditional chatbots, useful for rapid prototyping and internal automation. Read more

How to maximize Claude Code sessions — offers concrete token-cost management practices for teams using or evaluating agentic coding tools. Read more

Why Opus 5 feels worse day to day — exposes the gap between benchmark metrics and perceived quality in real use, a central risk when communicating and launching new model versions. Read more

I’ll keep watching how this privacy layer evolves—it is the kind of building block that changes what we will be able to build in credit down the road.