Coping and seething from an OpenAI research scientist:
Every American lab has shown their hand post R1, and none managed to convince. At this point, the entire $10T+ house of cards is going to start coming down once R2 is released in March/April (likely April).
OpenAI is screwed the most. Their largest investor is Microsoft, whose largest source of revenue is from Azure (cloud). Azure is already serving DeepSeek V3/R1 and will have to implement DeepSeek's latest optimizations as a matter of basic market economics. But doing so would increase DeepSeek's market share and damage the perception of Microsoft as an AI leader, which is what their ~$3T valuation is based off of. GPT-4.5 is a flop, and OpenAI were caught lying on the model card multiple times. First, they compared GPT-4.5 to o1-mini, but referred to it as the full o1. They also claimed a 10x increase in training efficiency, but removed the claim once people pointed out it was incompatible with their token costs. o3 was so computationally expensive ($3000 per query on highest thinking time) that they cannot release it as a standalone model, instead having to fold it into GPT-5. They were also caught cheating the FrontierMath benchmark last month.
Anthropic is in a similar situation. Claude-3.7 is impressive, but too expensive for its performance. I assume they are also bleeding billions per year. Their biggest investor Amazon, and the point about Azure and OpenAI applies equally to AWS and Anthropic. Also, Anthropic's CEO Dario Amodei claimed that DeepSeek smuggled GPUs and were lying about their numbers. He's now caught with his pants down given today's release. He also has a history of sketchy behavior. He admitted that he discovered scaling laws while working at Baidu, but when he moved to OpenAI and published a paper on scaling laws, he didn't cite Baidu's work. The mythology that the founders of OpenAI and Anthropic (who are ex-OpenAI) discovered scaling laws in one of the reasons they are still perceived to be leading in AI.
Meta Generative AI will also throw in the towel soon. Llama 4 was delayed due to R1, and if they don't get it out before R2, it's game over. Even if they do, it's unlikely it will outperform R1. Their ads team is already using DeepSeek as opposed to LLama.
x and Google will stay around longer because of their comparatively larger financial resources. They are willing to bleed money to serve their models at low cost. Gemini isn't impressive, and Grok barely outperformed DeepSeek on benchmarks which they specifically trained for. Also, while all American labs are bleeding cash, I suspect x is bleeding the most relative to its market share. x has the largest cluster and training costs, and they have nowhere near the most users. I also have inside information that x is paying many people thousands per week to solve math/coding problems so their solutions can be used to train Grok. This isn't factored into the publicly revealed costs, and also means that their algorithmic improvements are quite poor. I don't see how they can scale this up either.
Nvidia is also screwed. Jevon's paradox doesn't apply when existing GPUs can meet all inference needs. They will still have a market for training, but even then, the only major purchasers of Blackwell are the above 5 companies, the exact 5 companies who are bleeding billions while still struggling against DeepSeek.
For Chinese companies, as mentioned earlier in this thread, 4 companies are alive in the race for AGI: DeepSeek, Moonshot (makers of kimi), ByteDance, and Alibaba. These are the only labs with reasoning models. All of them have a positive outlook, and it's worth noting that Alibaba is the largest investor of Moonshot. On the other hand, the head of Qwen left for ByteDance, which indicates that ByteDance has more resources and/or desirability. There will also be a Qwen release next week.
Of course, other Chinese companies have very strong models for specialized tasks, but I'm only referring to AGI, since China dominates those categories already.
