Google's TPU v7 beats NVIDIA on price.

Anthropic just posted numbers that validate the AI business model. Their revenue grew 13 times, while compute expenses only grew 4.3 times. They now have over 11,900 customers spending $100K or more annually.

This is the Jevons Paradox in action. As AI gets cheaper and more efficient to run, businesses are finding exponentially more valuable work for it to do. Anthropic isn't just surviving the compute cost curve; they are crushing it by turning efficiency gains into huge revenue multiples.

If you are building an AI product, stop worrying about the baseline cost of intelligence. The market is clearly willing to pay six figures for tools that actually work. Anthropic is winning because they focus on enterprise value, not just raw model size. Their ability to decouple compute costs from revenue growth is the blueprint every other AI lab needs to follow.

OpenAI has officially entered the custom AI hardware competition, disrupting the market with their new Jalapeno chips. This is a direct challenge to the existing GPU supply chain and a clear signal that the leading AI lab is tired of paying external markups.

By moving into custom silicon, OpenAI is indicating that relying on third-party hardware is a strategic vulnerability. They need to control the stack from the silicon to the API to keep their pricing competitive as models get larger. The Jalapeno chips give them a path to internalize their biggest cost center.

NVIDIA should be sweating. When your biggest customer decides to build their own hardware, you stop selling them as many chips. For developers, OpenAI's vertical integration will likely translate to lower API costs and faster inference down the line, as they pass those hardware savings directly to the end user.

United Airlines has officially restricted unauthorized third-party AI agents from using its app, site, and booking through travel agencies. They are citing security risks and a desire to avoid middlemen who sit between the airline and the passenger.

This is the first major strike in the war between consumer AI assistants and enterprise walled gardens. Airlines prioritize direct customer relationships, upselling, and loyalty data. An AI agent that bypasses the interface destroys that entire business model by stripping away the airline's ability to market directly to the traveler.

If you are building a B2C agent that navigates other companies' platforms, expect to be blocked. The future of AI agents isn't rogue scraping; it's negotiated API access. United just showed that convenience for the user doesn't trump the airline's bottom line. Builders need to secure official partnerships, or their agents will simply be locked out of the transaction.

Elon Musk announced that xAI has completely abandoned JAX in favor of a custom C and C++ stack. Grok 4.8 is currently undergoing reinforcement learning on this new architecture ahead of a soft rollout this week.

Pivoting away from standard AI frameworks like JAX is a huge engineering risk, but it signals a demand for absolute control over hardware performance and memory management. Standard tools are great for research, but xAI is optimizing for sheer speed at scale, cutting out the middle layers to talk closer to the metal.

Builders take note: the abstraction layers are getting stripped away at the frontier. If xAI's custom stack delivers a faster, cheaper Grok 4.8, expect other labs to start questioning their reliance on off-the-shelf training frameworks. When you are operating at this scale, saving a few percentage points on overhead by writing custom C++ is worth the intense engineering effort.