alt Hacker NewsOpenAI unveils its first custom chip, built by Broadcom
Comments
No information on how significant the reduction in energy per token is. No information on amortized price per request. Increasingly its clear OpenAI must demonstrate order of magnitude reductions in cost to not die, this is investor story time without that information.
Although this seems to be for inference itself only and not training but inference is a recurring cost and training is a one time cost and so to me, even if Nvidia still gets moat on training, I don't think that it could ever justify its massive evaluations because for example, some chinese models are actually trained on Non-Nvidia models. The moat in that is incredibly thin.
(at the moment), I think that if I were Nvidia, I would be a bit terrified and I imagine the stock to not be doing super great as I can just imagine everyone online might start talking about it for better or for worse.
I am a bit impressed by OpenAI but is this what can be classified as a plan for OAI to salvage itself and all the commitments it has made nearing a 1.4 Trillion dollars from my memory and this article[0] is from 2025
But could OpenAI simply walk out of its commitments when necessary (for example to Nvidia) if this chip works out or what exactly might happen in the future as these commitments are asked to be paid for, its still smart for OAI to diversify with this chip and to have more deeper ways of revenue than just being a simple middleman but I imagine that Nvidia and others have also invested in OpenAI and they must not be happy with this change.
The thing with AI deals are that they have become so complicated that it is hard for me to find the first order impact of things, let alone second or third order impacts and financial accountability seems to be impacted quite heavily because of all of it and there is some sense that it is done so intentionally.
https://techcrunch.com/2025/11/06/sam-altman-says-openai-has...
> significantly better performance-per-watt than current state-of-the-art alternatives
An interesting example of how the current market dynamics incentivize low cost and therefore power efficiency and therefore lowering resource use.
im very excited that frontier models now have so much money and revenue they are releasing their own chips that could change the relationships and bottom line
hi
> built by Broadcom
AI is cooked bro. Broadcom is the death sentence of anything.
Is there any actual content on what the chips are?
You can't purchase Microsoft or AWS chips, but both of them do pretty good write-ups on what they've done. https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-...
This seems utterly empty of actual substance.
[dead]
[flagged]
[dead]
[dead]
[dead]
[dead]
[dead]
[dead]
lol
I call BS. It’s probably a white label around existing Broadcom IP, impossible to go from zero to this kind of chip in nine months. I doubt OpenAI had any significant contribution.
The similarities between the AI world and the crypto world are so much closer than any AI fanboy would ever admit.
This is why ram prices are fucked. Cause altman doesnt give a shit about normal people as long as openAI suceeds
Wow thats sounds tempting to use open ai newest chips
Big tech AI labs will develop LLM accelerators and hardware LLMs that increase frontier model output to tens of thousands-hundreds of thousands of TPS.
These chips will be used internally for their own business goals, giving them the capability to iterate at such an insane pace they will be able to clone every software product and software company on Earth. Meanwhile they'll trickle out 100-300 tps access to the rest of subscription users to drain them of their cash and keep the beast fed with fresh training data.
How can any individual company building a product, with access to 100-300 TPS behind-frontier security-gated, censored and capability gated models expect to compete with a company like Anthropic or OpenAI with frontier, unrestricted, unlocked models that can produce 100-1000x the output? 3-5 of their employees working to clone your 500 staff business will likely be easy pickings for them.
This should concern everyone.
The only reason they aren't 100% in on the strategy of replacing everyone is because they need us for training material and they needed the bootstrap. But the bootstrap problem is already gone, and they don't need to give us fair access to keep training data rolling.
One thing I don't like about California based companies is how cringe the names always are.
"Jalapeño" is such a bad name, having an "ñ" already makes it difficult and annoying to deal with in so many little ways. Good luck with that.
But also, theres the sort of "yes lets use Mexican related things because we're California" thought that I just really hate. I don't know, its like corporate Memphis to me. You see a product like this, you know it's an uppity califonia based firm that came up with it.
They tested on spark model, i bet it's a mix of that with focus on inference speed. Whatever it is, hopefully it shows up with current models as faster. Token/s is as big thing as anything else, and thats where they can really gain some edge over the competition.