Nvidia walked into its fiscal second-quarter earnings this week with Wall Street expecting roughly $92 billion in revenue, nearly double a year earlier, the kind of number that has made it the most valuable company on Earth. On the same day it reported, its single largest customer showed off a chip built to need it less. OpenAI unveiled Jalapeño, its first in-house processor, designed with Broadcom and aimed squarely at inference. In benchmark tests OpenAI said the chip beat Nvidia's GB200 and GB300 systems, delivering 1.5 to 1.9 times more AI work per watt and up to 3.6 times lower latency. The timing was not subtle.
What makes that sting is where it lands. Inference, the day-to-day running of models rather than their training, is the fastest-growing slice of AI spending, and it grows every time an agent does more work. "A hyperscaler-designed chip can now match or beat Nvidia's Blackwell-class GPUs on inference efficiency," one analyst told CNBC, calling Jalapeño "a threat to Nvidia's inference margins, which is the field growing the most at the moment." The caveats are real: independent testers note the fair comparison is Nvidia's newer Rubin platform, which is shipping now, while Jalapeño is still engineering samples with only limited deployment planned by year-end.
OpenAI is not acting alone, and that is the deeper problem. Google has its own tensor chips, Meta has committed to a gigawatt of custom Broadcom silicon, and Anthropic pledged more than $100 billion to Amazon's Trainium hardware. Startups like Cerebras and Etched are circling too. Research firm Omdia expects custom chips to overtake GPUs in raw volume by 2028 and calls this "the biggest competitive threat to Nvidia," noting that about half of all AI infrastructure spending now comes from cloud giants that either build their own chips or easily could.
Nvidia's defence is that the moat is more than the die. It still owns the vast majority of AI compute and, just as importantly, the CUDA software ecosystem that developers are built around, and it insists that training and frontier work will stay on flexible, programmable GPUs. At the Hot Chips conference it pushed its Rubin platform hard and won a marquee endorsement, with Elon Musk's SpaceXAI adopting Nvidia's Vera CPU to run Grok's agents, becoming the second hyperscaler after Meta to buy the processor on its own. Nvidia is climbing into the orchestration layer that agents lean on, rather than ceding it.
Circling all of this is a financial question the earnings call could not dodge. Nvidia has entangled itself in a web of financing vehicles with Apollo, BlackRock, Blackstone and others, targeting more than $500 billion of third-party capital for AI infrastructure, and its credit has begun trading more like a BBB-rated borrower than the AA name it is. The company that sells the shovels is increasingly also helping to fund the diggers, an arrangement that looks brilliant while demand holds and precarious if it wobbles.
So Nvidia can post one of the best quarters in the history of semiconductors and still be right to worry. The threat is not that customers stop buying GPUs to train frontier models; they will not, for years. It is that the profitable, booming inference layer is exactly the part a well-funded OpenAI or Google can most plausibly carve out for itself. Dominance and erosion showed up in the same week, and both were telling the truth.