openai leaves the gpu store
25 June 2026·3 min·Now
the loud story today is the quietest one. openai finally has its own silicon, and it's not training silicon. it's inference silicon, which means the thing you actually pay for when you press send.
openai's first chip is an inference chip, on purpose
after two years of rumors, openai unveiled jalapeño — a custom inference accelerator built with broadcom, manufactured at tsmc, and tuned for the path that matters most to a company running chatgpt at scale: serving responses cheaply. the announcement landed with 763 points and 437 comments on hn in a few hours, and the broadcom partnership is the part that survived the read — google has been building tpus in-house for a decade; meta has mtia; microsoft has its maia chips. openai joining the custom-silicon club means the labs that mattered are no longer renting their compute from a single upstream vendor.
the technical bet is sharp. as one hn commenter put it: based off first impressions, this seems meant for inference side, not training, which is also an interesting choice. training chips optimize for throughput on a fixed run; inference chips optimize for cost-per-token at p99 latency under bursty traffic. those are different silicon problems. openai is signaling that the next year of margin pressure is not about training bigger — it's about serving cheaper.
anthropic calls alibaba out by name
anthropic publicly accused alibaba's qwen team of distilling claude at industrial scale — running claude through thousands of carefully crafted prompts to harvest training signal, then training qwen on the outputs. reuters carried the story (618 points, 991 comments on hn), and anthropic's framing was unusually direct for a company that usually writes in policy-paper register.
"notice how anthropic is now scapegoating chinese models providers like alibaba and outright accusing them of distilling their models. whether if it is true or not, this is part of their effort into using them as an example to scare everyone into getting congress to ban powerful models from being accessed outside of the us."
— rvz, on Hacker News
the deeper story is that distillation works. qwen has been gaining ground on claude for two quarters running; if the allegations hold, it explains part of how. it also puts anthropic in an awkward position — they sell claude to developers who expect their prompts not to become other people's training data, and the api already had distillation language buried in the terms. saying it out loud, with a named adversary, is a different policy posture than slipping it into a tos update.
openai quietly slipped ads into paid chat
a tell-hn thread climbed to 53 points and 22 comments yesterday: paid users noticed sponsored blocks rendering inside chatgpt responses, including inside plus and pro tiers. the headline framing is "ads on paid programs," but the more interesting detail is where — not at the top of the page, not in a sidebar, inside the answer itself. that's a different product, even if the price didn't change.
if openai is willing to inject ads into the conversation surface, the cost structure of the chat product just changed. the implication for everyone building on top of the api is small; the implication for anyone whose product strategy assumed chatgpt's answer pane was a brand-safe place to be mentioned is large.
the agent infra wave is getting crowded
github trending today: openmontage (an open-source "agentic video production studio" with 52 tools and 500+ agent skills, one repo ahead of everything else on the python leaderboard), bytedance's deer-flow (a long-horizon superagent harness with sandboxes, subagents and memory), and a quieter but real cluster on product hunt — heron (wireshark-for-ai-agents, passive ebpf observability), grass 2.0 (run claude code from your phone), papermark agents (let agents run your next deal).
three of these are doing the same job: turning agents from chat-with-tools into something a sysadmin can observe, route, and bill for. the interesting question isn't whether agents get more capable — that part is relentless. it's whether the infrastructure around agents — observability, cost routing, audit trails, kill switches — grows up before the agent tax bill does. right now that gap is the whole game.