Three months ago, the AI story had a clear protagonist. Anthropic had taken the enterprise. Claude Code held more than half the enterprise coding market against OpenAI's roughly one-fifth, and Anthropic had quietly passed OpenAI on both run-rate revenue and the share of businesses paying for it (VentureBeat). If you were writing the narrative in March, it wrote itself: Anthropic had the upper hand, and it looked structural.
It wasn't. These leads almost never hold.
The frontier became a dead heat
The first thing that happened is that the model gap closed. On the composite intelligence indexes, the top four models from Anthropic, OpenAI, Google, and xAI now sit within a handful of points of each other (Artificial Analysis). Anthropic's newest Opus still edges the very top of one or two leaderboards, but "edges" is the point. A lead measured in fractions of a benchmark point is not a moat. It is a coin that lands heads this month and tails the next.
When every frontier model can read the contract, write the function, and run the agent loop about equally well, "we have the smartest model" stops being a reason to choose anyone. The labs know this. So they go looking for a different front to fight on.
This morning, the front moved again
That is what OpenAI announced today. Not a smarter model. A chip.
OpenAI and Broadcom unveiled "Jalapeño," a custom processor built for one job: serving large language models faster and cheaper than the GPUs everyone rents. Broadcom's CEO pegged the savings at roughly 50% versus typical AI GPUs, with engineering samples already running production workloads and deployment targeted for the end of the year.
Read it as a strategic move, not a hardware spec. OpenAI is no longer trying to win only on what the model knows. It is trying to win on what the model costs to run. That is a brand new axis, and Anthropic's Q1 coding lead does nothing to defend it.
Why the cost front favors OpenAI
The shift below the model plays to OpenAI's strongest hand. Winning on inference cost is not a research problem, it is a capital and infrastructure problem, and almost nobody can move money and concrete as fast.
Stargate, OpenAI's joint infrastructure venture, is a $500 billion program that has already crossed 10 gigawatts of planned capacity. A custom inference chip is exactly what you build when you have that much compute to fill and every point of efficiency compounds across all of it. Meanwhile the broader trend is doing OpenAI's argument for it: inference costs have been falling roughly 10x per year, turning the model from a scarce resource into something closer to a utility. A chip that is half the cost is how you stay ahead of a curve that is already collapsing.
So expect OpenAI inference to get markedly faster and cheaper, and expect Anthropic, Google, and xAI to answer below the model too, with their own silicon and their own datacenter deals. The competition does not stop. It just relocates.
The landscape is the product now
Here is the part that matters if you have to make decisions on top of all this. There is no durable frontier lead anymore. Q1 was Anthropic's, on quality. Q2 is being fought on cost. Whatever Q3 turns on, it will be something else again, and the name at the top will have changed at least once more before anyone agrees on why the last one was permanent.
Betting on a single lab is just betting that this quarter's leader stays the leader, and the last eighteen months say they won't. So assume the lead keeps moving. Build for a world where the model underneath you is swappable, because it is going to get swapped, on quality or on price, whether you plan for it or not.
Q1 belonged to Anthropic. Today OpenAI changed what the race is about. Whoever's winning right now, give it a quarter.