Broadcom does a lot of low margin buisness like wifi chips, ethernet chips etc. Nvidia and apple are the high margin buisness.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
I think they are the clear leaders from China at this point. Deepaeek seems to have fallen behind Qwen, well, still waiting for their next model, maybe that might outperform Qwen.
Alibaba team is just quite productive. Another 6 releases expected today. And the omni model is just state of art.
< 250ms latency on audio in to audio out is extraordinary.
Their cadence is insane. Twice as fast compared to others. For months the only release I see are from this team. DeepSeek was on frontpage for two and half months, moonshot for a month, Zhipu only a month. Qwen is in news for a year hand half. Nobody has as much staying power as this team.
Alibaba's release cadence is very impressive, but it's worth keeping in mind that the likes of Deep Seek, Zhipu, and Moon Shot are making bigger models to try and compete with Open AI, Google, Anthropic, and xAI directly in the cutting edge LLM space. Alibaba is taking a different strategy of making a variety of small to medium models that are practical to run on the edge, which boast near-frontier performance while not breaking the bank, and can (though are not currently) be used in an ensemble offering to cut API costs.
In fact, this seems to be a strategy the Western AI industry is also pursuing - Open AI is also starting to make smaller models that they can redirect simple queries to, in order to cut down aggregate inference costs, and Google's models have reportedly been smaller from the start. The industry could be shifting to cost cutting after realizing how much money they've been burning and that AGI is not just around the corner. Although, US labs are still burning immense amounts of money:
My guess is, if AGI is not around the corner, and there's a decent chance that it is not, then the question will soon become who is able to offer the most bang for the buck while not setting investors' cash on fire. All that expensive infrastructure is not going to be easy to maintain if there's a race to the bottom on costs.
In such an environment, China should an advantage.
In fact, this seems to be a strategy the Western AI industry is also pursuing - Open AI is also starting to make smaller models that they can redirect simple queries to, in order to cut down aggregate inference costs, and Google's models have reportedly been smaller from the start. The industry could be shifting to cost cutting after realizing how much money they've been burning and that AGI is not just around the corner. Although, US labs are still burning immense amounts of money:
My guess is, if AGI is not around the corner, and there's a decent chance that it is not, then the question will soon become who is able to offer the most bang for the buck while not setting investors' cash on fire. All that expensive infrastructure is not going to be easy to maintain if there's a race to the bottom on costs.
In such an environment, China should an advantage.
keep in mind this has been Alibaba's strategy since the start. Both it and ByteDance can afford to do this strategy of developing across the board large model product, but only Alibaba has been open sourcing everything so aggressively. That is rapidly making it the model of choice for all the startups out there.
AGI is already here, or maybe its 30 years away, or maybe never. The problem is its very hard to define what AGI even means.
Moreover, there is not enough data to train models on every scenario. We have already passed the limit of text data.
No matter how much compute you have, you are still iterating on the same data.
The cost of intelligence is often overlooked when people look at benchmarks. 2025 appears to have been a year of acceleration.
The first model to breach the 60 barrier at Artificial Analysis index was OpenAI's o3 back in April of this year.
What's crazy is that o3's data point is the updated price from June. The original launch price for o3 was $10/$40 per million input/output tokens. OpenAI later lowered it to $2/$8, which is shown in the chart. Grok 4 Fast is now at $0.2/$0.5.
So we've gone from $10/$40 to $0.2/$0.5 in a matter of mere months for a top-performing model. The disappointment with GPT-5 is masking a much bigger change in the economics of AI. I still prefer local models due to privacy concerns, but for the masses, this will accelerate adoption even more.
The first model to breach the 60 barrier at Artificial Analysis index was OpenAI's o3 back in April of this year.
What's crazy is that o3's data point is the updated price from June. The original launch price for o3 was $10/$40 per million input/output tokens. OpenAI later lowered it to $2/$8, which is shown in the chart. Grok 4 Fast is now at $0.2/$0.5.
So we've gone from $10/$40 to $0.2/$0.5 in a matter of mere months for a top-performing model. The disappointment with GPT-5 is masking a much bigger change in the economics of AI. I still prefer local models due to privacy concerns, but for the masses, this will accelerate adoption even more.
The current approach to AGI, AFAIK, is not through gathering more text data, but via 1) creating more data through simulation, 2) self-improvement through introspection and reinforcement learning, 3) embodied data augmentation (multi-modality like GPT 5, Qwen Omni, etc). All three of these directions are compute-heavy, which is why hyper scaling is still going on with Western Big AI labs. But no one knows for sure what the right recipe is and it's possible we're either going in the wrong direction, or simply lack the hardware scale necessary to produce "sentience."
Any way, the next five years will be very interesting for the industry and the world. Hyper scalers will either find a path towards consistent self-improvement, or they'll stall. If they stall, the bubble will pop and investments will collapse. Then only the cost effective providers of AI solutions will survive while the hyper scaling AGI shops will be left holding the bag.
