Daily Briefing — April 7, 2026

OpenAI won't turn a profit until 2030 at the earliest. Anthropic gets there a bit sooner, maybe 2028, but still burns through cash for years before that. These aren't rumours or analyst guesses. This is what the companies themselves are telling potential investors as they prep for IPOs, and the Wall Street Journal got a look at the numbers.

Here's the part that stops you mid-scroll. OpenAI expects to spend $121 billion on computing power in 2028 alone. That's not total operating costs. That's just the compute bill for training AI models. For context, they'll spend around $25 billion on that same thing this year. So we're talking about a near fivefold increase in roughly three years. Anthropic's numbers are smaller but the trajectory is the same direction: up, aggressively, for a long time.

What makes this genuinely interesting is that revenue is also surging. OpenAI projects nearly $275 billion in revenue by 2030, spending faster than it earns, on purpose, because the bet is that whoever builds the most capable models owns the market. The losses are the strategy, not a bug in it.

Seven robotaxi companies, including Waymo, Tesla, and Amazon's Zoox, were asked by Senator Ed Markey to disclose how often their remote operators have to step in and help their so-called autonomous vehicles. The companies mostly stonewalled. What little they did reveal was already enough to raise eyebrows: Waymo uses remote agents based in the Philippines, and Tesla is the only company openly admitting that its remote operators can actually take control of the vehicle. Most of the others would not even confirm basic intervention rates.

Here is why this matters beyond the headlines. The "autonomous" in autonomous vehicles has always been doing a lot of heavy lifting. What we are increasingly seeing is that the operational backbone of these systems relies on human labor, often cheap offshore labor, that the industry had zero interest in disclosing voluntarily. A senator had to drag it out of them, and even then the answers were incomplete.

There was also a safety incident in Austin where a Waymo vehicle drove past a school bus with its stop sign extended. That is not a minor software hiccup. That is a scenario with real consequences for real kids, and it happened under a system the company was actively obscuring from public scrutiny. When companies fight this hard against transparency, it usually means the numbers underneath are not flattering.

A startup called Freestyle just launched on Hacker News with a product that spins up virtual machines in under 700 milliseconds. That's not a typo. From API request to a fully ready machine, you're looking at sub-second provisioning. The pitch is aimed squarely at AI coding agents, the kind that need fresh, isolated environments to run code, test things, and blow up safely without taking anything else down with them.

The features tell a clear story about where the market is heading. Live VM cloning without pausing the machine, hibernation that costs you nothing while idle, and tight GitHub integration with per repo webhooks filtered by branch or path. These are not features built for a developer sitting at a laptop. They are built for agents running hundreds of parallel tasks where cold start time and compute cost actually eat into whether the whole thing is economically viable.

This is the infrastructure layer catching up to the AI agent hype.