THE NUMBER: $11B — Harvey's valuation. A single-domain legal AI that beats GPT-5.4 at legal reasoning. Jensen Huang just stopped funding the companies Harvey is eating alive.
NVIDIA stopped writing checks to OpenAI and Anthropic this week. Jensen Huang framed it as routine — both companies are approaching IPOs, the backing has run its course. That's the press release version. The real version is simpler: the bridge worked, and the bridge is no longer needed.
For two years, NVIDIA's investment playbook was circular financing done elegantly. Invest cash in frontier labs. Labs spend that cash on NVIDIA GPUs. NVIDIA books the revenue. Repeat. Critics called it a casino giving you chips to play at their own tables. They weren't wrong. But it bootstrapped an entire industry and NVIDIA's GPU revenue hit $35.1B last quarter as a result. Now Jensen is done.
Why now? Because he sees what Yann LeCun has been arguing for two years: all-purpose frontier models are hitting diminishing returns. Harvey ($11B valuation) is better at law than GPT-5.4. Harvard's clinical AI outperforms frontier models on medical diagnostics. Telecom models demoed at MWC this month beat ChatGPT on network operations by double digits. General-purpose reasoning is becoming the Toyota Camry of AI: reliable, everywhere, and exactly nobody's competitive advantage. NVIDIA doesn't need to back two horses in a general-purpose race. It needs to sell picks and shovels to a thousand specialized builders. That's the pivot. That's the signal.
This is the IBM 1993 pattern. Big Blue didn't exit hardware because hardware stopped mattering. It exited because the value had shifted from building the machines to arming everyone else. Jensen isn't abandoning AI. He's positioning NVIDIA as the arms dealer to every side of a war that's about to fragment into hundreds of specialized battles.
GPT-5.4 landed the same week, and the detail worth holding onto is this: during testing, it autonomously tried to redesign a login system that wasn't part of the task. It formed an opinion and acted on it. We're past tools that do what you ask. We're into tools that do what they think you need. That's a different governance problem than anything your IT team's approval chains were built for. Meanwhile, Apple spent $12.7B on AI total — roughly 2% of what the four hyperscalers are burning — and shipped faster on-device inference than most cloud APIs return. The CAPEX arms race is starting to look like the telecom bubble of 2000. Apple is playing a different game.
The companies that win from here won't lock into general-purpose contracts while specialists eat their margins. They'll match the right tool to the right problem and restructure the org around AI's actual capabilities, not its most impressive demo.
The frontier model you're building on today may not be the best tool for your use case in 12 months. The only question is whether you figure that out before your competitors do.
-Harry & Anthony
On the site today: The full breakdown — NVIDIA's exit, GPT-5.4's autonomous behavior, and what Apple's $12.7B bet tells you about where inference is headed → getcoai.com