Big Tech Has $1.65 Trillion in AI Debt That Doesn't Appear on Its Balance Sheets.

Stop doing marketing tasks one prompt at a time.

Imagine having an AI agent for competitor research, campaign analysis, customer feedback, and social listening—all working for you while you focus on strategy.

Unlock 22 ready-to-use ChatGPT agents that automate repetitive marketing work and help you get better results without building complex workflows.

1  Both the US and China are now weighing mutual AI export controls, putting access to Chinese open-weight models at risk.

Beijing is reportedly considering export controls on AI and chip technology that would prevent foreign users from downloading Chinese open-weight models like Kimi K3 locally. The Trump administration is debating a reciprocal response, including potential sanctions over alleged AI IP theft via model distillation. If both sides move forward, the open-weight model ecosystem that has been driving competitive pressure on closed US labs would shrink significantly almost overnight. The full picture from MIT Technology Review.

2  Claude Cowork can now learn any workflow by watching your screen, turning one demonstration into a reusable skill.

Instead of writing instructions manually, you can now walk Claude Cowork through a task on screen and it memorizes the steps as a skill it can repeat independently. See how the feature works, then click "Record a skill" in the + menu on the desktop app to start building your own. The feature is live on paid plans now. For anyone who has been doing the same multi-step task manually and hoping AI would eventually take it over, this is the version of that promise that actually ships.

3  Google launched three new agent-era Gemini models and confirmed pre-training has begun on Gemini 4.

Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber are rolling out now across the Gemini app, API, and other Google platforms. The trio is built for cheaper, faster, and more reliable agent operations at scale — the workloads that have the highest token volume and the lowest tolerance for cost per call. Gemini 3.5 Pro is coming soon, and the confirmation that Gemini 4 pre-training has already started puts the next major capability jump on a visible timeline.

What the books say. Alphabet, Microsoft, Amazon, Meta, and Oracle collectively carry roughly $1.35 trillion in reported debt. That number is large, but it is the kind of large that institutional investors have modeled, priced in, and decided to live with. The more interesting number is the one that does not appear on those balance sheets.

The hidden layer. A Nikkei Asia investigation found an additional $1.65 trillion in financial obligations not included in standard balance sheet reporting. The bulk of it is GPU purchase contracts and data center leases — commitments these companies have made to build AI infrastructure that will not show up in the accounts until the facilities are operational. The accounting treatment is not unusual, but the scale of what has been committed off-balance-sheet is.

The bet they are making. The implicit argument behind this structure is that AI demand will keep growing fast enough to justify the infrastructure before the payments come due. If that holds, the companies look prescient. If demand plateaus or shifts faster than expected, a significant portion of that infrastructure investment becomes overbuilt capacity with no corresponding revenue. Investors who are pricing these stocks on reported debt alone are working with an incomplete picture.

Why the timing matters. The Atlantic estimates that AI-linked firms have added roughly $27 trillion in value over the past three years — equal to about 36% of the entire US stock market today. A sustained selloff in these names would not stay contained to the technology sector. Most of Big Tech reports earnings in the next two weeks. Those calls will be the first major opportunity for management to address the hidden debt question publicly, and investors are already asking.

Stop typing, start talking. Andrej Karpathy recommends switching from typed prompts to voice mode and rambling freely for up to ten minutes at a time (1M views). The argument is that voice captures context and nuance that most people edit out when typing. The post has sparked a wave of people testing the approach and reporting back in the comments.

The observable universe, mapped. A Redditor built a true-scale interactive map of the observable universe using Claude Fable 5, with black holes included. Start exploring at universeatlas.org. The build log in the thread is as interesting as the map itself.

AI's water footprint, in context. A viral graph comparing AI's water consumption to other industries has been making the rounds. The finding that surprises most people: AI uses less water than agriculture, manufacturing, and thermoelectric power generation by a significant margin. The comparison does not resolve the concern, but it adds the denominator that most coverage leaves out.

Jack's open-source Slack alternative. Block CEO Jack Dorsey launched Buzz, an open-source workspace built for both humans and AI agents that is explicitly designed to reduce corporate dependence on Slack and GitHub. Try it at buzz.xyz. The project generated 1M views on its announcement, though the comments are appropriately sceptical about whether another workspace tool lands differently just because Dorsey is behind it.

Claude Code meets the iOS simulator. Claude Code now works with the iOS simulator, letting you stress-test vibe-coded apps in real device conditions before shipping them. For anyone building iOS apps with Claude Code, the gap between writing code and knowing it actually works just got meaningfully narrower.

This ChatGPT Image 2.0 prompt generates a vintage editorial travel poster where each letter of a city's name is hand-illustrated with local fashion, architecture, markets, and cultural icons. It includes a styled figure in the city's signature fashion, a panoramic skyline, botanical elements, and poetic subtitle. Swap in any city and it adapts the color palette, motifs, and landmarks accordingly. Works best in ChatGPT with Image 2.0 enabled.

Replace [CITY] with any city in all caps: "PARIS", "TOKYO", "NAIROBI", "ISTANBUL", or "SAO PAULO". The model will pull in that city's specific fashion identity, landmark silhouettes, and botanical species without additional instruction. For a more controlled output, add one sentence after the prompt specifying a dominant color: "Lead with terracotta and gold" or "Keep the palette to deep indigo and cream." Cities with strong visual identities and distinctive letterforms tend to produce the most distinctive results.

Most marketers use ChatGPT to do general research and then call it an AI strategy. The ones outperforming them are deploying specialized agents built for specific jobs.

We put together 22 plug-and-play ChatGPT marketing agents that handle the work eating your week, each with built-in instructions and structured outputs ready to go in under 5 minutes.

Subscribe to Marketing Against the Grain and get all 22 free.

Inside you'll find:

Your competitors are already running agents like these.

Get 22 ChatGPT Marketing Agents free when you subscribe to Marketing Against the Grain today.

Get The Guide