OpenAI just released the first public benchmarks for JalapeƱo, its custom inference chip, at Hot Chips on Tuesday. The numbers look good: 1.5 to 1.9 times higher throughput per kilowatt than Nvidia's Blackwell systems, and 1.7 to 3.6 times lower latency. The chip ships at 700 watts, designed to fit modern rack density. Small volumes in late 2026. Broader volume in 2027.
But here's the angle that matters: OpenAI didn't pick this venue and this benchmark to prove it beats Nvidia. It picked them to prove it can finance billions in data center debt.
The venue is a semiconductor conference with a room full of hardware engineers. The benchmark is SemiAnalysis' InferenceX, run with a third party present and published independently. That's not a marketing comparison. That's proof of concept for loan underwriting.
As Jon Markman pointed out in Forbes, the timing is suspicious. This month, OpenAI is arranging "nine figures of construction credit per gigawatt." A published performance-per-watt figure isn't interesting to engineers. It's evidence to a lender. Infrastructure capital prices on unit economics: output per megawatt-hour, throughput per kilowatt, cost per inference. JalapeƱo's benchmarks exist to show that cost curve is real.
The obvious problem: OpenAI is comparing JalapeƱo (which won't ship until December) against Nvidia Blackwell (which already exists). When JalapeƱo actually deploys at volume, Nvidia's Rubin will be on the market. Analysts are already flagging that Rubin might look different. JalapeƱo also uses newer HBM4 memory, which adds to its advantage. The comparison is incomplete and everyone knows it.
OpenAI doesn't care. The benchmark isn't meant for Nvidia. It's meant for a bank.
This move also tells you something about the hardware business right now. When inference margins are the fastest-growing profit center in AI, when the real money isn't in training models but in serving them cheap and fast, every company big enough wants its own silicon. The old pattern (design in-house, outsource to TSMC, buy packaging from ASICs vendors) now happens at scale. OpenAI with Broadcom. Google with TPUs. Meta with custom silicon. Alibaba's building theirs. Amazon's building theirs.
That puts Nvidia in a weird spot. Blackwell still owns the vast majority of AI compute and CUDA still locks customers in. But for inference, the workload that actually makes money, Nvidia no longer has a monopoly. It has an incumbent advantage. Those are different things.
The funny part: JalapeƱo's existence doesn't threaten Nvidia's inference market so much as it validates that inference margins are worth a $10+ billion custom silicon bet. OpenAI proved the market worth defending. Now everyone's defending it.
JalapeƱo isn't the real story here. The real story is that OpenAI just convinced a bank that in-house silicon makes sense at multi-gigawatt scale. Everything else follows from that.
