I'm having 4 separate conversations with people today about rtx 6000 price increases so I wanted to make a post explaining my original reasoning and general ideas on pricing for 6000s in particular. I believe max q and workstation rtx pro 6000 blackwells will be the SINGLE LARGEST appreciating asset period the next two years. I have many reasons for this coming from years of buying used pc parts and laptops as well as following previous cycles personally, and using a lot of different inference hardware myself the last year or so. - They are the single most powerful gpu for inference aggregate tok/s you can run on north american 120v besides the dgx station which uses a whole circuit. You can run 3-4 on a 20A circuit. - They have triple the vram of 5090s with a binned version of the same chip, and the highest memory bandwidth for matmuls - nvfp4 and 4 bit precision tensor cores means highest flops of any silicon generally available on 6000 and 5090 - You can slap them in any existing machine with an open pcie slot Why 6000s over 5090s then? Supply is way way lower, far fewer were manufactured somewhere on the order of 10-100x, they also have triple the vram on the same wattage. This means many models aren't runnable locally without them, and that unlock requires 4-8 of them for 750B models. 4 minimum for stuff like dsv4.1. New huawei chips china is targeting models for are 144-288gb, essentially reverse engineered h200s. Which is 3 6000s and the 6000s will be 2x faster than whats available in China while being more power efficient. Why 6000s over m5 ultra studios? Cuda and the tensor cores, also designated memory (not shared with os) and nvfp4 means higher speed and precision even if its the same memory on paper for a spec sheet. Even with perfect mlx kernels and if it was the same bandwidth (it's 50% higher than max spec m5 ultra) the gpus will never be the same speed for flops needed for inference because of the way the ALU is laid out on 5090/6000s. Why 6000s over h200/b200/b300 or dgx station? Nvidia is still manufacturing b300s, enterprise demand is all going to b300s and rubin which is what nvidia will focus 100% of their production on for likely the next 2 years minimum due to demand and economies of scale. Consumers are a tiny fraction of the market its not worth their time to serve. DGX stations will be tied to the price of ripping a b300 out of a node and rigging it up which is already cheaper when you look at the price of an 8x node compared to a dgx station. Why not AMD? CUDA and ALU layouts. Same as MLX but ROCm is in even worse shape. Even with perfect kernels on both compared with the same model and vram/bandwidth it won't be as fast for aggregate tps. Why not 5000 or 5500s? Lower memory bandwidth than 5090s and 6000s. Due to all of these unique comparisons and attributes, combined with the extremely limited supply, nvidia is not manufacturing more of them, has no incentive to manufacture more of them, while demand will continue to increase, and the general market of buyers able to utilize them, of which many have no other option for many models such as dsv4.1 to even run them locally (maybe h200s with fans or h170x which are also limited supply, but again significantly lower tps (1/2 most likely). They are 2x the speed for most of the options and have the perfect tax on second hand markets like ebay on top of the limited supply and the power efficiency and speed for actual real world usecases. I had originally modeled these to go to $32-40k by next year based on potential users and supply in north america, but I think this will happen far sooner now as supply is out at retailers in the next few weeks and no more are likely to be manufactured ever. I do not believe it is unreasonable to see them go to auction on ebay etc for $100-$120k late next year-early 2028 with current demand curves given the limited supply. This would be a 15x from their price in april. Because of the relative *perceived* niche of the topend prosumer market for manufacturing planning, contrasted against the real world market allowing basically all SMBs to run models a few weeks behind frontier locally with them on regular 120v power on any machine with a pcie slot at 2x the speed and half the power draw of any other option, (besides maybe m5 ultra which is still 33% slower) these have truly unique characteristics compared to any other card. Extremely high utility, extremely high TAM, and extremely high potential demand for what they're actually able to run twice as fast as any alternative for far more people. At the same time as all of this, they have the lowest or tied for lowest supply of any blackwell gpu. Do with this information what you want.

My pleasure to come on Dwarkesh last week, I thought the questions and conversation were really good. I re-watched the pod just now too. First of all, yes I know, and I'm sorry that I speak so fast :). It's to my detriment because sometimes my speaking thread out-executes my thinking thread, so I think I botched a few explanations due to that, and sometimes I was also nervous that I'm going too much on a tangent or too deep into something relatively spurious. Anyway, a few notes/pointers: AGI timelines. My comments on AGI timelines looks to be the most trending part of the early response. This is the "decade of agents" is a reference to this earlier tweet Basically my AI timelines are about 5-10X pessimistic w.r.t. what you'll find in your neighborhood SF AI house party or on your twitter timeline, but still quite optimistic w.r.t. a rising tide of AI deniers and skeptics. The apparent conflict is not: imo we simultaneously 1) saw a huge amount of progress in recent years with LLMs while 2) there is still a lot of work remaining (grunt work, integration work, sensors and actuators to the physical world, societal work, safety and security work (jailbreaks, poisoning, etc.)) and also research to get done before we have an entity that you'd prefer to hire over a person for an arbitrary job in the world. I think that overall, 10 years should otherwise be a very bullish timeline for AGI, it's only in contrast to present hype that it doesn't feel that way. Animals vs Ghosts. My earlier writeup on Sutton's podcast . I am suspicious that there is a single simple algorithm you can let loose on the world and it learns everything from scratch. If someone builds such a thing, I will be wrong and it will be the most incredible breakthrough in AI. In my mind, animals are not an example of this at all - they are prepackaged with a ton of intelligence by evolution and the learning they do is quite minimal overall (example: Zebra at birth). Putting our engineering hats on, we're not going to redo evolution. But with LLMs we have stumbled by an alternative approach to "prepackage" a ton of intelligence in a neural network - not by evolution, but by predicting the next token over the internet. This approach leads to a different kind of entity in the intelligence space. Distinct from animals, more like ghosts or spirits. But we can (and should) make them more animal like over time and in some ways that's what a lot of frontier work is about. On RL. I've critiqued RL a few times already, e.g. . First, you're "sucking supervision through a straw", so I think the signal/flop is very bad. RL is also very noisy because a completion might have lots of errors that might get encourages (if you happen to stumble to the right answer), and conversely brilliant insight tokens that might get discouraged (if you happen to screw up later). Process supervision and LLM judges have issues too. I think we'll see alternative learning paradigms. I am long "agentic interaction" but short "reinforcement learning" I've seen a number of papers pop up recently that are imo barking up the right tree along the lines of what I called "system prompt learning" , but I think there is also a gap between ideas on arxiv and actual, at scale implementation at an LLM frontier lab that works in a general way. I am overall quite optimistic that we'll see good progress on this dimension of remaining work quite soon, and e.g. I'd even say ChatGPT memory and so on are primordial deployed examples of new learning paradigms. Cognitive core. My earlier post on "cognitive core": , the idea of stripping down LLMs, of making it harder for them to memorize, or actively stripping away their memory, to make them better at generalization. Otherwise they lean too hard on what they've memorized. Humans can't memorize so easily, which now looks more like a feature than a bug by contrast. Maybe the inability to memorize is a kind of regularization. Also my post from a while back on how the trend in model size is "backwards" and why "the models have to first get larger before they can get smaller" Time travel to Yann LeCun 1989. This is the post that I did a very hasty/bad job of describing on the pod: . Basically - how much could you improve Yann LeCun's results with the knowledge of 33 years of algorithmic progress? How constrained were the results by each of algorithms, data, and compute? Case study there of. nanochat. My end-to-end implementation of the ChatGPT training/inference pipeline (the bare essentials) On LLM agents. My critique of the industry is more in overshooting the tooling w.r.t. present capability. I live in what I view as an intermediate world where I want to collaborate with LLMs and where our pros/cons are matched up. The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code and humans are useless. For example, I don't want an Agent that goes off for 20 minutes and comes back with 1,000 lines of code. I certainly don't feel ready to supervise a team of 10 of them. I'd like to go in chunks that I can keep in my head, where an LLM explains the code that it is writing. I'd like it to prove to me that what it did is correct, I want it to pull the API docs and show me that it used things correctly. I want it to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I'm told works. I just think the tools should be more realistic w.r.t. their capability and how they fit into the industry today, and I fear that if this isn't done well we might end up with mountains of slop accumulating across software, and an increase in vulnerabilities, security breaches and etc. Job automation. How the radiologists are doing great and what jobs are more susceptible to automation and why. Physics. Children should learn physics in early education not because they go on to do physics, but because it is the subject that best boots up a brain. Physicists are the intellectual embryonic stem cell I have a longer post that has been half-written in my drafts for ~year, which I hope to finish soon. Thanks again Dwarkesh for having me over!

⚡ 532 tokens/s! DeepSeek V4.1 Flash is live on Inco, and it's the fastest provider on Artificial Analysis. #1# output speed. Not close. Try it: Results: