Sure, they have huge GPU clusters, but there must be more going on - model optimizations, sharding, custom hardware, clever load balancing, etc.
What engineering tricks make this possible at such massive scale while keeping latency low?
Curious to hear insights from people who've built large-scale ML systems.
Hint: it's a money thing.
However I can share this written by my colleagues! You'll find great explanations about accelerator architectures and the considerations made to make things fast.
https://jax-ml.github.io/scaling-book/
In particular your questions are around inference which is the focus of this chapter https://jax-ml.github.io/scaling-book/inference/
Edit: Another great resource to look at is the unsloth guides. These folks are incredibly good at getting deep into various models and finding optimizations, and they're very good at writing it up. Here's the Gemma 3n guide, and you'll find others as well.
https://docs.unsloth.ai/basics/gemma-3n-how-to-run-and-fine-...
Doesn't OpenAI depend mostly on its relationship/partnership with Microsoft to get GPUs to inference on?
Thanks for the links, interesting book!
Even then, I think that their primary use case is going to be consumer grade good AI on phones. I dunno why Gemma QAT model fly so low on the radar, but you can basically get full scale Llamma 3 like performance from a single 3090 now, at home.
Google has already started the process of letting companies self-host Gemini, even on NVidia Blackwell GPUs.
Although imho, they really should bundle it with their TPUs as a turnkey solution for those clients who haven't invested in large scale infra like DCs yet.
They’ll definitely have the best model, but there is a chance they will f*up the product / integration into their products.
But then again even there, their reputation for abandoning products, lack of customer service, condescension when it came to large enterprises’ “legacy tech” lets Microsoft who is king of hand holding big enterprise and even AWS run rough shod over them.
When I was at AWS ProServe, we didn’t even bother coming up with talking points when competing with GCP except to point out how they abandon services. Was it partially FUD? Probably. But it worked.
there are few groups as talented at losing a head start as google.
That is, as a research person using our GPUs and TPUs I see first hand how choices from the high level python level, through Jax, down to the TPU architecture all work together to make training and inference efficient. You can see a bit of that in the gif on the front page of the book. https://jax-ml.github.io/scaling-book/
I also see how sometimes bad choices by me can make things inefficient. Luckily for me if my code/models are running slow I can ping colleagues who are able to debug at both a depth and speed that is quite incredible.
And because were on HN I want to preemptively call out my positive bias for Google! It's a privilege to be able to see all this technology first hand, work with great people, and do my best to ship this at scale across the globe.
> llama.cpp an other inference engines auto add a <bos> - DO NOT add TWO <bos> tokens! You should ignore the <bos> when prompting the model!
That makes the want to try exactly that? Weird
And folks at LMSys: https://lmsys.org/blog/
Large Model Systems (LMSYS Corp.) is a 501(c)(3) non-profit focused on incubating open-source projects and research. Our mission is to make large AI models accessible to everyone by co-developing open models, datasets, systems, and evaluation tools. We conduct cutting-edge machine learning research, develop open-source software, train large language models for broad accessibility, and build distributed systems to optimize their training and inference.Inference is (mostly) stateless. So unlike training where you need to have memory coherence over something like 100k machines and somehow avoid the certainty of machine failure, you just need to route mostly small amounts of data to a bunch of big machines.
I don't know what the specs of their inference machines are, but where I worked the machines research used were all 8gpu monsters. so long as your model fitted in (combined) vram, you could job was a goodun.
To scale the secret ingredient was industrial amounts of cash. Sure we had DGXs (fun fact, nvidia sent literal gold plated DGX machines) but they wernt dense, and were very expensive.
Most large companies have robust RPC, and orchestration, which means the hard part isn't routing the message, its making the model fit in the boxes you have. (thats not my area of expertise though)
"we do 1970s mainframe style timesharing"
there, that was easy
I ask because scaling an system that a substantially chunk of the population finds incredibly useful, including for the more efficient production of public goods (scientific research, for example) does seem like a problem that a) needs to be solved from a business point of view, and b) should be solved from a civic-minded point of view.
And if AI was truly the holy grail that it's being sold as then there wouldn't be 700M users per week wasting all of these resources as heavily as we are because generative AI would have already solved for something better. It really does seem like these platforms are, and won't be, anywhere as useful as they're continuously claimed to be.
Just like Tesla FSD, we keep hearing about a "breakaway" model and the broken record of AGI. Instead of getting anything exceptionally better we seem to be getting models tuned for benchmarks and only marginal improvements.
I really try to limit what I'm using an LLM for these days. And not simply because of the resource pigs they are, but because it's also often a time sink. I spent an hour today testing out GPT-5 and asking it about a specific problem I was solving for using only 2 well documented technologies. After that hour it had hallucinated about a half dozen assumptions that were completely incorrect. One so obvious that I couldn't understand how it had gotten it so wrong. This particular technology, by default, consumes raw SSE. But GPT-5, even after telling it that it was wrong, continued to give me examples that were in a lot of ways worse and kept resorting to telling me to validate my server responses were JSON formatted in a particularly odd way.
Instead of continuing to waste my time correcting the model I just went back to reading the docs and GitHub issues to figure out the problem I was solving for. And that led me down a dark chain of thought: so what happens when the "teaching" mode rethinks history, or math fundamentals?
I'm sure a lot of people think ChatGPT is incredibly useful. And a lot of people are bought into not wanting to miss the boat, especially those who don't have any clue to how it works and what it takes to execute any given prompt. I actually think LLMs have a trajectory that will be similar to social media. The curve is different and I, hopefully, don't think we've seen the most useful aspects of it come to fruition as of yet. But I do think that if OpenAI is serving 700M users per week then, once again, we are the product. Because if AI could actually displace workers en masse today you wouldn't have access to it for $20/month. And they wouldn't offer it to you at 50% off for the next 3 months when you go to hit the cancel button. In fact, if it could do most of the things executives are claiming then you wouldn't have access to it at all. But, again, the users are the product - in very much the same way social media played into.
Finally, I'd surmise that of those 700M weekly users less than 10% of those sessions are being used for anything productive that you've mentioned and I'd place a high wager that the 10% is wildly conservative. I could be wrong, but again - we'd know about that if it were the actual truth.
Very, very few problems _need_ to be solved. Feeding yourself is a problem that needs to be solved in order for you to continue living. People solve problems for different reasons. If you don't think LLMs are valuable, you can just say that.
1. prompt caching
2. some RAG to save resources
3. of course lots model optimizations and CUDA optimizations
4. lots of throttling
5. offloading parts of the answer that are better served by other approaches (if asked to add numbers, do a system call to a calculator instead of using LLM)
6. a lot of sharding
One thing you should ask is: What does it mean to handle a request with chatgpt? It might not be what you think it is.
source: random workshops over the past year.
They are also partnering with rivals like Google for additional capacity https://www.reuters.com/business/retail-consumer/openai-taps...
I don't think people realize the size of these compute units.
When the AI bubble pops is when you're likely to be able to realistically run good local models. I imagine some of these $100k servers going for $3k on eBay in 10 years, and a lot of electricians being asked to install new 240v connectors in makeshift server rooms or garages.
>A typical 1U or 2U server can accommodate 2-4 H100 PCIe GPUs, depending on the chassis design.
>In a 42U rack with 20x 2U servers (allowing space for switches and PDU), you could fit approximately 40-80 H100 PCIe GPUs.
Supermicro will sell you a full rack loaded with servers [1] providing 13.4 TB of GPU memory.
And with 132kW of power output, you can heat an olympic-sized swimming pool by 1°C every day with that rack alone. That's almost as much power consumption as 10 mid-sized cars cruising at 50 mph.
[1] https://www.supermicro.com/en/products/system/gpu/48u/srs-gb...
Unlike the investments in railways or telephone cables or roads or any other sort of architecture, this investment has a very short lifespan.
Their point was that whatever your take on AI, the present investment in data centres is a ridiculous waste and will always end up as a huge net loss compared to most other investments our societies could spend it on.
Maybe we'll invent AGI and he'll be proven wrong as they'll pay back themselves many times over, but I suspect they'll ultimately be proved right and it'll all end up as land fill.
I think we would get all this technology without going to the moon or Space Shuttle program. GPS, for example, was developed for military applications initially.
The servers will be replaced, the networking equipment will be replaced. The building will still be useful, the fiber that was pulled to internet exchanges/etc will still be useful, the wiring to the electric utility will still be useful (although I've certainly heard stories of datacenters where much of the floor space is unusable, because power density of racks has increased and the power distribution is maxed out)
Sure, datacenters will get rid of the hardware - but only because it's no longer commercially profitable run them, presumably because compute demands have eclipsed their abilities.
It's kind of like buying a used GeForce 980Ti in 2025. Would anyone buy them and run them besides out of nostalgia or curiosity? Just the power draw makes them uneconomical to run.
Much more likely every single H100 that exists today becomes e-waste in a few years. If you have need for H100-level compute you'd be able to buy it in the form of new hardware for way less money and consuming way less power.
For example if you actually wanted 980Ti-level compute in a desktop today you can just buy a RTX5050, which is ~50% faster, consumes half the power, and can be had for $250 brand new. Oh, and is well-supported by modern software stacks.
I think the existence of a pretty large secondary market for enterprise servers and such kind of shows that this won't be the case.
Sure, if you're AWS and what you're selling _is_ raw compute, then couple generation old hardware may not be sufficiently profitable for you anymore... but there are a lot of other places that hardware could be applied to with different requirements or higher margins where it may still be.
Even if they're only running models a generation or two out of date, there are a lot of use cases today, with today's models, that will continue to work fine going forward.
And that's assuming it doesn't get replaced for some other reason that only applies when you're trying to sell compute at scale. A small uptick in the failure rate may make a big dent at OpenAI but not for a company that's only running 8 cards in a rack somewhere and has a few spares on hand. A small increase in energy efficiency might offset the capital outlay to upgrade at OpenAI, but not for the company that's only running 8 cards.
I think there's still plenty of room in the market in places where running inference "at cost" would be profitable that are largely untapped right now because we haven't had a bunch of this hardware hit the market at a lower cost yet.
Even if you didn't have optimizations involved in terms of job scheduling, they would just build as many warehouses as necessary filled with as many racks as necessary to serve the required user base.
You can pick up a DGX-1 on Ebay right now for less than $10k. 256 GB vRAM (HBM2 nonetheless), NVLink capability, 512 GB RAM, 40 CPU cores, 8 TB SSD, 100 Gbit HBAs. Equivalent non-Nvidia branded machines are around $6k.
They are heavy, noisy like you would not believe, and a single one just about maxes out a 16A 240V circuit. Which also means it produces 13 000 BTU/hr of waste heat.
I just looked at a spec sheet for a 230V single-phase 12k BTU mini-split and the minimum circuit ampacity was 3A for the air handler and 12A for the condenser, add those together for 15A, divide by .8 is 18.75A, next size up is 20A. Minimum circuit ampacity is a formula that is (roughly) the sum of the full load amps of the motor(s) inside the piece of equipment times 1.25 to determine the conductor size required to power the equipment.
So the condensing unit likely draws ~9.5-10A max and the air handler around ~2.4A, and both will have variable speed motors that would probably only need about half of that to remove 12k BTU of heat, so ~5-6A or thereabouts should do it, which is around 1/3rd of the 16A server, or a COP of 3.
How useful is this Tesla-era hardware on current workloads? If you tried to run the full DeepSeek R1 model on it at (say) 4-bit quantization, any idea what kind of TTFT and TPS figures might be expected?
At 90°C you can do sous vide, so basically use that waste heat entirely.
For such temperatures you'd need a CO2 heat pump, which is still expensive. I don't know about gas, as I don't even have a line to my place.
In sane units: 3.8 kW
5.1 Horsepower
:D
Each of these NAND chips hundreds of dies of flash stacked inside, and they are hooked up to the same data line, so just 1 of them can talk at the same time, and they still achieve >1GB/s bandwidth. If you could hook them up in parallel, you could have 100s of GBs of bandwidth per chip.
You could probably make special NAND chips where all of them can be active at the same time, which means you could get 100GB+ bandwidth out of a single chip.
This would be useless for data storage scenarios, but very useful when you have huge amounts of static data you need to read quickly.
From what I understand, this hardware has a high failure rate over the long term especially because of the heat they generate.
Speculative decoding uses a smaller draft model to generate tokens with much less compute and memory required. Then the main model will accept those tokens based on the probability it would have generated them. In practice this case easily result in a 3x speedup in inference.
Another trick for structured outputs that I know of is "fast forwarding" where you can skip tokens if you know they are going to be the only acceptable outputs. For example, you know that when generating JSON you need to start with `{ "<first key>": ` etc. This can also lead to a ~3x speedup in when responding in JSON.
I'm not sure it improved the speed much
In my experiences I'd seen the calls to the target model reduced to a third of what they would have been without using a draft model.
You'll still get some gains on a local model, but they won't be near what they could be theoretically if everything is properly tuned for performance.
It also depends on the type of task. I was working with pretty structured data with lots of easy to predict tokens.
Also, you CAN run local models that are as good as GPT 4 was on launch on a macbook with 24 gigs of ram.
https://artificialanalysis.ai/?models=gpt-oss-20b%2Cgemma-3-...
Conversely, you can't do the same thing as a self hosted user, you can't really bank your idle compute for a week and consume it all in a single serving, hence the much more expensive local hardware to reach the peak generation rate you need.
I assume the former has massive overhead, but maybe it is worthwhile to keep responsiveness up for everyone.
First of all, they never “handle more requests than they have hardware.” That’s impossible (at least as I’m reading it).
The vast majority of usage is via their web app (and free accounts, at that). The web app defaults to “auto” selecting a model. The algorithm for that selection is hidden information.
As load peaks, they can divert requests to different levels of hardware and less resource hungry models.
Only a very small minority of requests actually specify the model to use.
There are a hundred similar product design hacks they can use to mitigate load. But this seems like the easiest one to implement.
1. You load all the weights of the model into GPU VRAM, plus the context.
2. You construct a data structure called the "KV cache" representing the context, and it hopefully stays in the GPU cache.
3. For each token in the response, for each layer of the model, you read the weights of that layer out of VRAM and use them plus the KV cache to compute the inputs to the next layer. After all the layers you output a new token and update the KV cache with it.
Furthermore, my understanding is that the bottleneck of this process is usually in step 3 where you read the weights of the layer from VRAM.
As a result, this process is very parallelizable if you have lots of different people doing independent queries at the same time, because you can have all their contexts in cache at once, and then process them through each layer at the same time, reading the weights from VRAM only once.
So once you got the VRAM it's much more efficient for you to serve lots of people's different queries than for you to be one guy doing one query at a time.
If you want a survey of intermediate level engineering tricks, this post we wrote on the Fin AI blog might be interesting. (There's probably a level of proprietary techniques OpenAI etc have again beyond these): https://fin.ai/research/think-fast-reasoning-at-3ms-a-token/
batching requests increase latency to first token, so it's tradeoff and MoE makes it more tricky because they are not equally used.
there was somewhere great article explaining deepseek efficiency that explained it in great detail (basically latency - throughput tradeoff)
But I also have to say 700M weekly users could mean 100M daily or 70k a minute (low ball estimate with no returning users...) is a lot, but achievable at startup scale. I don't have out current numbers but we are several orders of magnitude smaller of course :-)
The big difference to home use is the amount of VRAM. Large VRAM GPUs such as H100 are gated being support contracts and cost 20k. Theoretically you could buy a Mac Pro with a ton of RAM as an individual if you wanted to run auch models yourself.
Take a look at vLLM for an open source solution that is pretty close to the state of the art as far as handling many user queries:https://docs.vllm.ai/en/stable/
No, really. They just have entire datacenters filled with high end GPUs.
It's more cost effective to farm eggs from a hundred thousand chickens than it is for individuals to have chickens in their yard.
You CAN run a GPT-class model on your own machine right now, for several thousand dollars of machine... but you can get massively better results if you spend those thousands of dollars on API credits over the next five years or so.
Some people will choose to do that. I have backyard chickens, they're really fun! Most expensive eggs I've ever seen in my life.
I for one look forward to another 10 years of progress - or less - putting current models running on a laptop. I don’t trust any big company with my data
Consider a system were running a dozen queries at once is only marginally more expensive than running one query. What would you build?
https://www.seangoedecke.com/inference-batching-and-deepseek...
Here is an example of what happens
> The only way to do fast inference here is to pipeline those layers by having one GPU handle the first ten layers, another handle the next ten, and so on. Otherwise you just won’t be able to fit all the weights in a single GPU’s memory, so you’ll spend a ton of time swapping weights in and out of memory and it’ll end up being really slow. During inference, each token (typically in a “micro batch” of a few tens of tokens each) passes sequentially through that pipeline of GPUs
For OpenAI, I’d assume that a GPU is dedicated to your task from the point you press enter to the point it finishes writing. I would think most of the 700 million barely use ChatGPT and a small proportion use it a lot and likely would need to pay due to the limits. Most of the time you have the website/app open I’d think you are either reading what it has written, writing something or it’s just open in the background, so ChatGPT isn’t doing anything in that time. If we assume 20 queries a week taking 25 seconds each. That’s 8.33 minutes a week. That would mean a single GPU could serve up to 1209 users, meaning for 700 million users you’d need at least 578,703 GPUs. Sam Altman has said OpenAI is due to have over a million GPUs by the end of year.
I’ve found that the inference speed on newer GPUs is barely faster than older ones (perhaps it’s memory speed limited?). They could be using older clusters of V100, A100 or even H100 GPUs for inference if they can get the model to fit or multiple GPUs if it doesn’t fit. A100s were available in 40GB and 80GB versions.
I would think they use a queuing system to allocate your message to a GPU. Slurm is widely used in HPC compute clusters, so might use that, though likely they have rolled their own system for inference.
You can’t run GPT4 for yourself because the fixed costs are high. But the variable costs are low, so OAI can serve a shit ton.
Or equivalently the smallest available unit of “serving a gpt4” is more gpt4 than one person needs.
I think all the inference optimisation answers are plain wrong for the actual question asked?
https://www.tripadvisor.com/Restaurant_Review-g60763-d477541...
As already answered, AI companies use extremely expensive setups (servers with professional cards) in large numbers and all these things concentrated in big datcenters with powerful networking and huge power consumption.
Imagine - last time, so huge investments (~1.2% of GDP, and unknown if investments will grow or not) was into telecom infrastructure - mostly wired telephones, but also cable TV and later added Internet and cell communications and clouds (in some countries wired phones just don't cover whole country and they jumped directly into wireless communications).
Larger investments was into railroads - ~6% of GDP (and I'm also not sure, some people said, AI will surpass them as share of possible for AI tasks constantly grow).
So to conclude, just now AI boom looks like main consumer of telecom (Internet) and cloud infrastructure. If you've seen old mainframes in datacenters, and extremely thick core network cables (with hundreds wires or fibers in just one cable), and huge satellite dishes, you could imagine, what I'm talking about.
And yes, I'm not sure, will this boom end like dot-coms (Y2K), or such huge usage of resources will sustain. Why it is not obvious, because for telecoms (internet) also was unknown, if people will use phones and other p2p communications for leisure as now, or will leave phones just for work. Even worse, if AI agents become ordinary things, possible scenario, number of AI agents will surpass number of people.
They probably are using some interesting hardware, but there's a strange economy of scale when serving lots of requests for a small number of models. Regardless of if you are running single GPU, clustered GPU, FPGAs, or ASICs, there is a cost with initializing the model that dwarfs the cost of inferring on it by many orders of magnitude.
If you build a workstation with enough accelerator-accessible memory to have "good" performance on a larger model, but only use it with typical user access patterns, that hardware will be sitting idle the vast majority of the time. If you switch between models for different situations, that incurs a load penalty, which might evict other models, which you might have to load in again.
However, if you build an inference farm, you likely have only a few models you are working with (possibly with some dynamic weight shifting[1]) and there are already some number of ready instances of each, so that load cost is only incurred when scaling a given model up or down.
I've had the pleasure to work with some folks around provisioning an FPGA+ASIC based appliance, and it can produce mind-boggling amounts of tokens/sec, but it takes 30m+ to load a model.
[1] there was a neat paper at SC a few years ago about that, but I can't find it now
A second trick is to implement something called speculative decoding. Inference has two phases. One is prompt processing and another is token generation. They actually work the same way using what is called a forward pass, except prompt processing can do them in parallel by switching from matrix-vector to matrix-matrix multiplication and dumping the prompt’s tokens into each forward pass in parallel. Each forward pass will create a new token, but it can be discarded unless it is from the last forward pass, as that will be the first new token generated as part of token generation. Now, you put that token into the next forward pass to get the token after it, and so on. It would be nice if all of the forward passes could be done in parallel, but you do not know the future, so you ordinarily cannot. However, if you make a draft model that is a very fast model runs in a fraction of the time and guesses the next token correctly most of the time, then you can sequentially run the forward pass for that instead N times. Now, you can take the N tokens and put it into the prompt processing routine that did N forward passes in parallel. Instead of discarding all tokens except the last one like in prompt processing, we will compare them to the input tokens. All tokens up to and including the first token that differ, that come out of the parallel forward pass are valid tokens for the output of the main model. This is guaranteed to always produce at least 1 valid token since in the worse case the first token does not match, but the output for the first token will be equal to the output of running the forward pass without having done speculative decoding. You can get a 2x to 4x performance increase from this if done right.
Now, I do not work on any of this professionally, but I am willing to guess that beyond these techniques, they have groups of machines handling queries of similar length in parallel (since doing a batch where 1 query is much longer than the others is inefficient) and some sort of dynamic load balancing so that machines do not get stuck with a query size that is not actively being utilized.
As soon as you have enough users you can let your GPUs burn with a high load constantly, while your home solution would idle most of the time and therefore be way too expensive compared to the value.
I think the thing to remember is that the majority of chatGPT users, even those who use it every day, are idle 99.9% of the time. Even someone who has it actively processing for an hour a day, seven days a week, is idle 96% of the time. On top of that, many are using less-intensive models. The fact that they chose to mention weekly users implies that there is a significant tail of their user distribution who don't even use it once a day.
So your question factors into a few of easier-but-still-not-trivial problems:
- Making individual hosts that can fit their models in memory and run them at acceptable toks/sec.
- Making enough of them to handle the combined demand, as measured in peak aggregate toks/sec.
- Multiplexing all the requests onto the hosts efficiently.
Of course there are nuances, but honestly, from a high level last problem does not seem so different from running a search engine. All the state is in the chat transcript, so I don't think there any particular reason reason that successive interactions on the same chat need be handled by the same server. They could just be load-balanced to whatever server is free.
We don't know, for example, when the chat says "Thinking..." whether the model is running or if it's just queued waiting for a free server.
They are throwing money at this problem hoping you throw more money back.
If you try to run GPT4 at home, you'll still need enough VRAM to load the entire model, which means you'll need several H100s (each one costs like $40k). But you will be under-utilizing those cards by a huge amount for personal use.
It's a bit like saying "How come Apple can make iphones for billions of people but I can't even build a single one in my garage"
- the models are not too big for the cards. Specifically, they know the cards they have and they modify the topology of the model to fit their hardware well
- lots of optimisations. Eg the most trivial implementation of transformer-with-attention inference is going to be quadratic in the size of your output but actual implementations are not quadratic. Then there are lots of small things: tracing the specific model running on the specific gpu, optimising kernels, etc
- more costs are amortized. Your hardware is relatively expensive because it is mostly sitting idle. AI company hardware gets much more utilization and therefore can be relatively more expensive hardware, where customers are mostly paying for energy.
But software wise, they shard, load balance, and batch. ChatGPT gets 1000s (or something like that) of requests every second. Those are batched and submitted to one GPU. Generating text for 1000 answers is often the same speed as generating for just 1 due to how memory works on these systems.
So basically the main tricks are batching (only relevant when you have > 1 query to process) and MoE sharding.
IMO outfits like OpenAI are burning metric shit tonnes of cash serving these models. It pails in comparison to the mega shit tonnes of cash used to train the models.
They hope to gain market share before they start charging customers what it costs.
For storage, they also have massive amount of hard disks and SSD behind planet scale object file systems (like AWS's S3 or Tectonic at Meta or MinIO in prem) all connected by massive amount of switches and routers of varying capacity.
So in the end, it's just the good old Cloud, but also with GPUs.
Btw, OpenAI's infrastructure is provided and managed by Microsoft Azure.
And, yes, all of this requires billions of dollars to build and operate.
The largest chatgpt models are maybe 1-1.5tb in size and all of that needs to load into pooled vram. That sounds daunting, but a company like open ai has countless machines that have enough of these datacenter grade gpus with gobs of vram pooled together to run their big models.
Inference is also pretty cheap, especially when a model can comfortably fit in a pool of vram. Its not that the pool of gpus spool up each time someone sends a request, but whats more likely is that there’s a queue to f requests from someone like chatgpts 700 million users, and the multiple (I have no idea how many) pools of vram keep the models in their memory to chew through that nearly perpetual queue of requests.
https://ut.philkr.net/advances_in_deeplearning/
Especially the "Advanced Training" section to get some idea of tricks that are used these days.
Look at VLLM. It's the top open source version of this.
But the idea is you can service 5000 or so people in parallel.
You get about 1.5-2x slowdown on per token speed per user, but you get 2000x-3000x throughput on the server.
The main insight is that memory bandwidth is the main bottleneck so if you batch requests and use a clever KV cache along with the batching you can drastically increase parallel throughput.
Just illustrative guesses, not real numbers, I underestimate overheads here but anyway ...
Let's assume a $20k expert node can produce 500 tokens per second (15,000 per year). $5k a year for the machine per year. $5k overheads. 5 experts per token (so $50k to produce 15,000 megatokens with a 100% throughput). Say they charge up to $10 per million tokens ... yeah it's tight but I can see how it's doable.
Say they cost $100 per user per year. If it's $10 per million tokens (depends on the model) then they are budgeting 10 million tokens per user. That's like 100 books per year. The answer is that users probably don't use as much as the api would cost.
The real question is, how does it cost $10 per megatoken?
500 tokens per second per node is like 15,000 megatokens per year. So a 500 token node can bring in $150,000 per node.
Call it 5 live experts and a router. That's maybe $20k per expert per year. If it's a kilowatt power supply per expert, and $0.1 per kW power that's $1000 for power. The hardware is good for 4 years so $5k for that. Toss in overheads, and it's maybe $10k costs.
So at full capacity they can make $5 off $10 revenue. With uneven loads they make nothing, unless they have some optimisation and very good load balancing (if they can double the tokens per second then they make a decent profit).
I've had Google Gemini start talking to be in the context of someone else's chat, unrelated to mine.
ME: Does POSIX have any API for creating a filesystem object (such as a file) with a predetermined inode number supplied by the caller?
AI: <reformats my question into LaTeX syntax, out of the blue>
\documentclass{article}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amssymb}
\begin{document}
\section*{Does POSIX have any API for creating a filesystem object with a predetermined inode number supplied by the caller?}
\textbf{No, POSIX does not provide a standard API for creating a filesystem object (like a file or directory) with a caller-specified inode number.}
\subsection*{Inode Allocation in POSIX Filesystems}
[... SNIP]
AI: You are absolutely right! My apologies. That was a mistake on my part. I got carried away with the instruction to use LaTeX for mathematical and scientific notations and incorrectly applied it to the entire response. [... SNIP]
There was no such instruction. I've never chatted with any AI about LaTeX. it leaked from the tokens of someone else's chat.
Nope. That's not how it works. Attention doesn't work across multiple independent prompts queued in the same batch. It's not physically possible for the tokens of another chat to leak.
What most likely happened is that the model glitched out to the instructions in its (hidden) system prompt, which most likely does include instructions about using LaTeX for mathematical and scientific notation.
minimaxir•5h ago
That's a really, really big "sure."
Almost every trick to run a LLM at OpenAI's scale is a trade secret and may not be easily understood by mere mortals anyways (e.g. bare-metal CUDA optimizations)
v5v3•5h ago
With all the staff poaching the trade secrets may have now leaked?
minimaxir•2h ago
handfuloflight•1h ago