The CFO AI Digest: September 17
Welcome to your roundup. OpenAI launched GPT-5-Codex and signed a $300B cloud deal with Oracle. Microsoft expanded access to Copilot, while Google released a new LLM with stronger security protocols and a payment protocol for agents.
Here’s what you need to know:
1. OpenAI upgrades AI coding agent
OpenAI released GPT‑5-Codex, a version of GPT-5 optimized for agentic coding. It’s faster, more reliable, and trained on real-world engineering tasks like building full projects, doing large-scale refactors, and reviewing code. It can run independently for hours, debug, test, and iterate until tasks are complete. Codex works across the terminal, IDE, GitHub, and mobile.
CFO takeaway: Codex is becoming an AI teammate. And with Sam Altman comparing the product to the early days of ChatGPT, Codex might be on track to become OpenAI’s next breakout product. This could speed up tool-building and reduce the engineering resources finance needs: think bots, dashboards, scripts, and integrations built with less engineering help.
2. OpenAI signs $300B cloud deal with Oracle
OpenAI has signed a $300 billion cloud contract with Oracle for computing power over roughly five years. The deal, which begins in 2027, is one of the largest cloud contracts ever signed and would require 4.5 gigawatts of power over its term, more than what two Hoover Dams produce.
CFO takeaway: OpenAI is betting that demand for ChatGPT and AI infrastructure will keep rising, and Oracle is betting on OpenAI. As the Wall Street Journal notes, it’s a high-stakes deal for both customers. It’s also a continuing signal of how capital-intensive AI has become and how much faith is riding on a massive future payoff.
3. Microsoft brings Copilot Chat to 365 apps
Microsoft has rolled out Copilot Chat and agents across Word, Excel, PowerPoint, Outlook, and OneNote. All users get free access to in-app AI chat that’s context-aware — it can understand the file you're working in, pull in recent documents, and tailor responses accordingly. Premium features, like GPT-5-powered responses and agents such as Researcher and Analyst, require a paid Copilot license.
CFO takeaway: Microsoft is standardizing the use of AI agents across its core productivity apps as part of a broader shift among vendors to embed AI in their most popular tools. This will expose a new group of individuals to AI. It should also result in less friction across tools, more consistent governance. This wider availability could also increase focus on measuring the ROI of AI tools, both in tracking the impact of free versions and in justifying the additional spend on paid licenses.
4. Google launches VaultGemma
Google released a new LLM called VaultGemma, calling it the first model to combine strong privacy protections with state-of-the-art utility. Using a method called differential privacy, the model is designed to learn from data without memorizing or leaking sensitive information, making it a safer option for working with regulated data. It’s trained on public and synthetic data and meets U.S. government privacy standards.
CFO takeaway: Privacy-first AI is maturing, opening up the possibility of using LLMs to analyze sensitive and personal data and use these tools in highly regulated industries with strict data security rules. This matters for finance teams exploring AI use on sensitive internal data, like contracts or customer records. Other enterprise AI vendors might soon follow Google’s lead.
5. YouTube adds AI highlights for livestreams
YouTube Live launched a suite of livestreaming upgrades, including a new AI feature that spots the most engaging parts of a livestream, such as spikes in chat, reactions, or emotional moments, and turns them into Shorts automatically. It’s designed to help creators stay visible in the algorithm, extend the life of their content, and grow their audiences with less manual editing.
CFO takeaway: Auto-editing points to a broader trend: generative AI taking over labor-intensive tasks, like editing and repackaging long-form video into short clips optimized for discovery and monetization. Expect this kind of auto-editing to show up more in enterprise video, sales enablement, and even internal comms, potentially reducing production costs and expediting timelines.
6. Google announces payment protocol for AI agents
Google has introduced the Agent Payments Protocol (AP2), a standard that enables AI agents to securely make purchases on behalf of users. It uses advanced digital contracts to verify user intent and authorize transactions. Buyers either approve the purchase details in real time via a checkout confirmation, or pre-authorize through an “intent mandate” that sets clear rules, like spend limits. AP2 supports real-time transactions across cards, bank transfers, and even crypto. It’s backed by more than 60 industry partners, including Mastercard, PayPal, Intuit, and Salesforce.
CFO takeaway: Google is building the payment rails for AI agents to transact autonomously. That means preparing for a reality where software can initiate purchases, trigger approvals, and move money with minimal human input. Procurement, expense management, and audit workflows will all need to evolve to track, authorize, and verify AI-driven spend in real time.
See you next week.
Recommended reading
- Anthropic Economic Index report: Uneven geographic and enterprise AI adoption (Anthropic)
- Defeating Nondeterminism in LLM Inference (Horace He in collaboration with others at Thinking Machines)
