Nvidia’s dominance took hits from three directions this week, and none of them came from a place anyone expected all at once.
Start Tuesday at Hot Chips, where OpenAI shared the first independent-style benchmarks for Jalapeño, the inference chip it built with Broadcom. Tested on SemiAnalysis’s InferenceX suite, the chip delivered up to 1.9x more throughput per watt and up to 3.6x lower latency than Nvidia’s current flagship, the Blackwell GB300. It’s inference-only, and it won’t ship in real volume until 2027, but it’s the loudest signal yet that a lab with enough money and enough of its own models to help with the design can build silicon that beats Nvidia at the workload that will eventually dwarf training.
Then there’s the leak in the supply chain that nobody built on purpose. Taiwan prosecutors indicted a senior Nvidia manager — along with two Supermicro employees and six others — for allegedly staging a fake compliance audit to funnel 74 servers of Nvidia B300 chips into China via Japan and Indonesia, dodging US export controls. Another 56 servers were seized before they made it out. Nvidia itself isn’t accused of wrongdoing, but the case is the clearest sign yet that a black market for frontier chips has matured enough to need people on the inside.
None of that dented the money chasing the industry. Anthropic is pitching IPO investors on a $30 trillion total addressable market — bigger than the $28.5 trillion figure SpaceX used ahead of its own record-setting offering in June, and roughly twelve times the combined revenue of the 191 tech companies in the S&P 1500. The underlying numbers are real enough: revenue more than doubled quarter over quarter, to $11.6 billion. But the gap between that and $30 trillion is the gap every AI IPO pitch now has to paper over. Anthropic is reportedly eyeing a $2 trillion valuation and a raise as large as $100 billion, with prospectus paperwork expected within weeks.
And then there’s what happens once these systems are big enough to be worth abusing. OpenAI disclosed it dismantled a Russian influence operation that used ChatGPT — accessed through VPNs, since Russia is blocked outright — to build a fake Israeli think tank called the International Burke Institute. The “institute” ran a Russia-flattering “sovereignty index,” attacked France, Germany, and the US, and plagiarized dozens of articles, some misattributed to Francis Fukuyama and Noam Chomsky. OpenAI says the operation’s actual reach was limited; what’s notable is the infrastructure it built — a full fake-expertise pipeline with an LLM doing the writing.
Put together, it’s a tidy snapshot of an industry simultaneously proving itself indispensable and getting harder to contain: a competitor building better inference silicon, a smuggling ring willing to risk prison over chips already on the open market, a bank pitch untethered from any historical base rate, and a propaganda operation quietly renting the same chatbot everyone else uses for homework help. None of these problems is new. What’s new is the scale they’re operating at — trillions of dollars in pitch decks, chips smuggled by the shipping container, and misuse sophisticated enough to fool a Substack feed.
