I work at Google on these systems everyday (caveat this is my own words not my employers)). So I simultaneously can tell you that its smart people really thinking about every facet of the problem, and I can't tell you much more than that.
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.
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.
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)
It would take talent for them to mess up hosting businesses who want to use their TPUs on GCP.
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.
Yeah honestly. They could just try selling solutions and SLAs combining their TPU hardware with on-prem SOTA models and practically dominate enterprise. From what I understand, that's GCP's gameplay too for most regulated enterprise clients.
Googles bread and butter is advertising, so they have a huge interest in keeping things in house. Data is more valuable to them than money from hardware sales.
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.
It’s my understanding that google makes bulk of ad money from search ads - sure they harvest a ton of data but it isn’t as valuable to them as you’d think. I suspect they know that could change so they’re hoovering up as much as they can to hedge their bets. Meta on the other hand is all about targeted ads.
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.
Im a research person building models so I can't answer your questions well (save for one part)
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.
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.
An H100 is a $20k USD card and has 80GB of vRAM. Imagine a 2U rack server with $100k of these cards in it. Now imagine an entire rack of these things, plus all the other components (CPUs, RAM, passive cooling or water cooling) and you're talking $1 million per rack, not including the costs to run them or the engineers needed to maintain them. Even the "cheaper"
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.
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.
Fair warning: the BMCs on those suck so bad, and the firmware bundles are painful, since you need a working nvidia-specific container runtime to apply them, which you might not be able to get up and running because of a firmware bug causing almost all the ram to be presented as nonvolatile.
Heat pump sure, but how is gas furnace more efficient than resistive load inside the house? Do you mean more economical rather than more efficient (due to gas being much cheaper/unit of energy)?
Cooling BTUs already take the coefficient of performance of the vapor-compression cycle into account. 4w of heat removed for each 1w of input power is around the max COP for an air cooled condenser, but adding an evaporative cooling tower can raise that up to ~7.
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.
Are you talking about the guy in Temecula running two different auctions with some of the same photos (356878140643 and 357146508609, both showing a missing heat sink?) Interesting, but seems sketchy.
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?
My personal sneaking suspicion is that publicly offered models are using way less compute than thought. In modern mixture of experts models, you can do top-k sampling, where only some experts are evaluated, meaning even SOTA models aren't using much more compute than a 70-80b non-MoE model.
Even is the AI bubble does not pops, your prediction about those servers being available on ebay in 10 years will likely be true, because some datacenters will simply upgrade their hardware and resell their old ones to third parties.
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.
Off topic, but I bought my (still in active use) 980ti literally 9 years ago for that price. I know, I know, inflation and stuff, but I really expected more than 50% bang for my buck after 9 whole years…
Someone's take on AI was that we're collectively investing billions in data centers that will be utterly worthless in 10 years.
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.
If it is all a waste and a bubble, I wonder what the long term impact will be of the infrastructure upgrades around these dcs. A lot of new HV wires and substations are being built out. Cities are expanding around clusters of dcs. Are they setting themselves up for a new rust belt?
The servers may well be worthless (or at least worth a lot less), but that's pretty much true for a long time. Not many people want to run on 10 year old servers (although I pay $30/month for a dedicated server that's dual Xeon L5640 or something like that, which is about 15 years old).
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)
They probably are right, but a counter argument could be how people thought going to the moon was pointless and insanely expensive, but the technology to put stuff in space and have GPS and comms satellites probably paid that back 100x
I don’t mean to invalidate your point (about genuine value arising from innovations originating from the Apollo program), but GPS and comms satellites (and heck, the Internet) are all products of nuclear weapons programs rather than civilian space exploration programs (ditto the Space Shuttle, and I could go on…).
Reality is that we don’t know how much of a trope this statement is.
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.
Sure, but what about the collective investment in smartphones, digital cameras, laptops, even cars. Not much modern technology is useful and practical after 10 years, let alone 20. AI is probably moving a little faster than normal, but technology depreciation is not limited to AI.
What I wonder is what this means for Coreweave, Lambda and the rest, who are essentially just renting out fleets of racks like this. Does it ultimately result in acquisition by a larger player? Severe loss of demand? Can they even sell enough to cover the capex costs?
To piggyback on this, at enterprise level in modern age, the question is really not about "how are we going to serve all these users", it comes down to the fact that investors believe that eventually they will see a return on investment, and then pay whatever is needed to get the infra.
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.
I wonder if it's feasible to hook up NAND flash with a high bandwidth link necessary for inference.
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.
NAND is very, very slow relative to RAM, so you'd pay a huge performance penalty there. But maybe more importantly my impression is that memory contents mutate pretty heavily during inference (you're not just storing the fixed weights), so I'd be pretty concerned about NAND wear. Mutating a single bit on a NAND chip a million times over just results in a large pile of dead NAND chips.
No it's not slow - a single NAND chip in SSDs offers >1GB of bandwidth - inside the chip there are 100+ wafers actually holding the data, but in SSDs only one of them is active when reading/writing.
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.
Why stop at 80 H100s for a mere 6.4 terabytes of GPU memory?
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.
You have thousands of dollars, they have tens of billions. $1,000 vs $10,000,000,000. They have 7 more zeros than you, which is one less zero than the scale difference in users: 1 user (you) vs 700,000,000 users (openai). They managed to squeak out at least one or two zeros worth of efficiency at scale vs what you're doing.
Also, you CAN run local models that are as good as GPT 4 was on launch on a macbook with 24 gigs of ram.
You can knock off a zero or two just by time shifting the 700 million distinct users across a day/week and account for the mere minutes of compute time they will actually use in each interaction. So they might no see peaks higher than 10 million active inference session at the same time.
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.
During times of high utilization, how do they handle more requests than they have hardware? Is the software granular enough that they can round robin the hardware per token generated? UserA token, then UserB, then UserC, back to UserA? Or is it more likely that everyone goes into a big FIFO processing the entire request before switching to the next user?
I assume the former has massive overhead, but maybe it is worthwhile to keep responsiveness up for everyone.
Inference is essentially a very complex matrix algorithm run repeatedly on itself, each time the input matrix (context window) is shifted and the new generated tokens appended to the end. So, it's easy to multiplex all active sessions over limited hardware, a typical server can hold hundreds of thousands of active contexts in the main system ram, each less than 500KB and ferry them to the GPU nearly instantaneously as required.
I think the most direct answer is that at scale, inference can be batched, so that processing many queries together in a parallel batch is more efficient than interactively dedicating a single GPU per user (like your home setup).
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/
700M weekly users doesn't say much about how much load they have.
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.
I'm sure there are countless tricks, but one that can implemented at home, and I know plays a major part in Cerebras' performance is: speculative decoding.
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.
To measure the performance gains on a local machine (or even standard cloud GPU setup), since you can't run this in parallel with the same efficiency you could in a high-ed data center, you need to compare the number of calls made to each model.
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.
It is not just engineering. There are also huge, very huge, investments into infrastructure.
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.
At the heart of inference is matrix-vector multiplication. If you have many of these operations to do and only the vector part differs (which is the case when you have multiple queries), you can do matrix-matrix multiplication by stuffing the vectors into a matrix. Computing hardware is able to run the equivalent of dozens of matrix-vector multiplication operations in the same time it takes to do 1 matrix-matrix multiplication operation. This is called batching. That is the main trick.
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.
The short answer is "batch size". These days, LLMs are what we call "Mixture of Experts", meaning they only activate a small subset of their weights at a time.
This makes them a lot more efficient to run at high batch size.
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"
I'm pretty much an AI layperson but my basic understanding of how LLMs usually run on my or your box is:
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.
First off I’d say you can run models locally at good speed, llama3.1:8b runs fine a MacBook Air M2 with 16GB RAM and much better on a Nvidia RTX3050 which are fairly affordable.
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.
The idea that a GPU is dedicated to a single inference task is just generally incorrect. Inputs are batched, and it’s not a single GPU handling a single request, it’s a handful of GPUs in various parallelism schemes processing a batch of requests at once. There’s a latency vs throughput trade off that operators make. The larger that batch size the greater the latency, but it improves overall cluster throughput.
Your model keeps the weights on slow memory and needs to touch all of them to make 1 token for you. By batching you make 64 tokens for 64 users in one go. And they use dozens of GPUs in parallel to make 1024 tokens in the time your system makes 1 token. So even though the big system costs more, it is much more efficient when being used by many users in parallel. Also, by using many fast GPUs in series to process parts of the neural net, it produces output much faster for each user compared to your local system. You can't beat that.
- 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.
1. They have many machines to split the load over
2. MoE architecture that lets them shard experts across different machines - 1 machine handles generating 1 token of context before the entire thing is shipped off to the next expert for the next token. This reduces bandwidth requirements by 1/N as well as the amount of VRAM needed on any single machine
3. They batch tokens from multiple users to further reduce memory bandwidth (eg they compute the math for some given weights on multiple users). This reduces bandwidth requirements significantly as well.
So basically the main tricks are batching (only relevant when you have > 1 query to process) and MoE sharding.
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)
Baseten serves models as a service, at scale. There’s quite a lot of interesting engineering both for inference and infrastructure perf. This is a pretty good deep dive into the tricks they employ: https://www.baseten.co/resources/guide/the-baseten-inference...
Lots of good answers that mention the big things (money, scale, and expertise). But one thing I haven’t seen mentioned yet is that the transformer math is probably against your use case. Batch compute on beefy hardware is currently more efficient than computing small sequences for a single user at a time, since these models tend to be memory bound and not compute bound. They have the users that makes the beefy hardware make sense, enough people are querying around the same time to make some batching possible.
TL;DR: It's massively easier to run a few models really fast than it is to run many different models acceptably.
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
If the explanation really is, as many comments here suggest, that prompts can be run in parallel in batches at low marginal additional cost, then that feels like bad news for the democratization and/or local running of LLMs. If it’s only cost-effective to run a model for ~thousands of people at the same time, it’s never going to be cost-effective to run on your own.
Sure, but that's how most of human society works already.
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.
50 years ago general computers were also time shared. Then the pendulum swing to desktop, then back to central.
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
Well, you can also batch your own queries. Not much use for a chatbot but for an agentic system or offline batch processing it becomes more reasonable.
Consider a system were running a dozen queries at once is only marginally more expensive than running one query. What would you build?
Well, their huge GPU clusters have "insane VRAM". Once you can actually load the model without offloading, inference isn't all that computationally expensive for the most part.
Multi-tenancy likely explains the bulk of it. $10k vs. $10b gives them six orders of magnitude more GPU resources, but they have 9 orders of magnitude more users. The average user is probably only running an active ChatGPT query for a few minutes per day, which covers the remaining 3 orders of magnitude.
You and your engineering team might be able to figure it out and purchase enough equipment also if you had received billions of dollars. And billions and billions. And more billions and billions and billions. Then additional billions, and more billions and billions and even more billions and billions of dollars. They have had 11 rounds of funding totaling around $60 billion.
> 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
Once you have enough GPUs to have your whole model available in GPU RAM you can do inference pretty fast.
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.
Look for positron.ai talks about their tech, they discuss their approach to scaling LLM workloads with their dedicated hardware. It may not be what is done by OpenAI or other vendors, but you'll get an idea of the underlying problems.
The first step is to acquire hardware fast enough to run one query quickly (and yes, for some model size you are looking at sharding the model and distributed runs). The next one is to batch request, improving GPU use significantly.
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/
Have you looked at what happens to tokens per second when you increase batch size? The cost of serving 128 queries at once is not 128x the cost of serving one query.
This. the main trick, outside of just bigger hardware, is smart batching. E.g. if one user asks why the sky is blue, the other asks what to make for dinner, both queries go though the same transformer layers, same model weights so they can be answered concurrently for very little extra GPU time. There's also ways to continuously batch requests together so they don't have to be issued at the same time.
Complete guess, but my hunch is that it's in the sharding. When they break apart your input into its components, they send it off to hardware that is optimized to solve for that piece. On that hardware they have insane VRAM and it's already cached in a way that optimizes that sort of problem.
At the end of the day, the answer is... specialized hardware. No matter what you do on your local system, you don't have the interconnects necessary. Yes, they have special software, but the software would not work locally. NVIDIA sells entire solutions and specialized interconnects for this purpose. They are well out of the reach of the standard consumer.
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.
Huge batches to find the perfect balance between compute and memory banthwidth, quantized models, speculative decoding or similar techniques, MoE models, routing of requests on smaller models if required, batch processing to fill the GPUs when demand is lower (or electricity is cheaper).
They also don’t need one system per user. Think of how often you use their system over the week, maybe one hour total? You can shove 100+ people into sharing one system at that rate… so already you’re down to only needing 7 million systems.
not affiliated with them and i might be a little out of date but here are my guesses
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.
I work at a university data center, although not on LLMs. We host state of the art models for a large number of users. As far as I understand, there is no secret sauce. We just have a big GPU cluster with a batch system, where we spin up jobs to run certain models. The tricky part for us is to have the various models available on demand with no waiting time.
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.
I work at Google on these systems everyday (caveat this is my own words not my employers)). So I simultaneously can tell you that its smart people really thinking about every facet of the problem, and I can't tell you much more than that.
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-...
Same explanation but with less mysticism:
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)
Doesn't google have TPU's that makes inference of their own models much more profitable than say having to rent out NVDIA cards?
Doesn't OpenAI depend mostly on its relationship/partnership with Microsoft to get GPUs to inference on?
Thanks for the links, interesting book!
Yes. Google is probably gonna win the LLM game tbh. They had a massive head start with TPUs which are very energy efficient compared to Nvidia Cards.
The only one who can stop Google is Google.
They’ll definitely have the best model, but there is a chance they will f*up the product / integration into their products.
It would take talent for them to mess up hosting businesses who want to use their TPUs on GCP.
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.
Google employees collectively have a lot of talent.
Google bought Wiz for $32B to try to pull through GCP into the Wiz customer roster. The effort is there, success remains to be seen.
But they’re ASICs so any big architecture changes will be painful for them right?
Yeah honestly. They could just try selling solutions and SLAs combining their TPU hardware with on-prem SOTA models and practically dominate enterprise. From what I understand, that's GCP's gameplay too for most regulated enterprise clients.
Googles bread and butter is advertising, so they have a huge interest in keeping things in house. Data is more valuable to them than money from hardware sales.
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.
It’s my understanding that google makes bulk of ad money from search ads - sure they harvest a ton of data but it isn’t as valuable to them as you’d think. I suspect they know that could change so they’re hoovering up as much as they can to hedge their bets. Meta on the other hand is all about targeted ads.
https://www.cnbc.com/2025/04/09/google-will-let-companies-ru...
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.
Im a research person building models so I can't answer your questions well (save for one part)
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.
Why does the unsloth guide for gemma 3n say:
> 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
> Another great resource to look at is the unsloth guides.
And folks at LMSys: https://lmsys.org/blog/
An H100 is a $20k USD card and has 80GB of vRAM. Imagine a 2U rack server with $100k of these cards in it. Now imagine an entire rack of these things, plus all the other components (CPUs, RAM, passive cooling or water cooling) and you're talking $1 million per rack, not including the costs to run them or the engineers needed to maintain them. Even the "cheaper"
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.
What do you mean 10 years?
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.
Fair warning: the BMCs on those suck so bad, and the firmware bundles are painful, since you need a working nvidia-specific container runtime to apply them, which you might not be able to get up and running because of a firmware bug causing almost all the ram to be presented as nonvolatile.
It's not waste heat if you only run it in the winter.
Opt if you ignore that both gas furnaces and heat pumps are more efficient than resistive loads.
Heat pump sure, but how is gas furnace more efficient than resistive load inside the house? Do you mean more economical rather than more efficient (due to gas being much cheaper/unit of energy)?
> 13 000 BTU/hr
In sane units: 3.8 kW
> In sane units: 3.8 kW
5.1 Horsepower
You mean 1.083 tons of refrigeration
You’ll need (2) 240V 20A 2P breakers, one for the server and one for the 1-ton mini-split to remove the heat ;)
Matching AC would only need 1/4 the power, right? If you don't already have a method to remove heat.
Cooling BTUs already take the coefficient of performance of the vapor-compression cycle into account. 4w of heat removed for each 1w of input power is around the max COP for an air cooled condenser, but adding an evaporative cooling tower can raise that up to ~7.
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.
Well I don't know why that unit wants so many amps. The first 12k BTU window unit I looked at on amazon uses 12A at 115V.
Well, get a heat pump with a good COP of 3 or more, and you won't need quite as much power ;)
Just air freight them from 60 degrees North to 60 degrees South and vice verse every 6 months.
Are you talking about the guy in Temecula running two different auctions with some of the same photos (356878140643 and 357146508609, both showing a missing heat sink?) Interesting, but seems sketchy.
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?
My personal sneaking suspicion is that publicly offered models are using way less compute than thought. In modern mixture of experts models, you can do top-k sampling, where only some experts are evaluated, meaning even SOTA models aren't using much more compute than a 70-80b non-MoE model.
Even is the AI bubble does not pops, your prediction about those servers being available on ebay in 10 years will likely be true, because some datacenters will simply upgrade their hardware and resell their old ones to third parties.
Would anybody buy the hardware though?
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.
Off topic, but I bought my (still in active use) 980ti literally 9 years ago for that price. I know, I know, inflation and stuff, but I really expected more than 50% bang for my buck after 9 whole years…
Except their insane electricity demands will still be the same, meaning nobody will buy them. You have plenty of SPARC servers on Ebay.
There is also a community of users known for not making sane financial decisions and keeping older technologies working in their basements.
But we are few, and fewer still who will go for high power consumption devices with esoteric cooling requirements that generate a lot of noise.
Someone's take on AI was that we're collectively investing billions in data centers that will be utterly worthless in 10 years.
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.
If it is all a waste and a bubble, I wonder what the long term impact will be of the infrastructure upgrades around these dcs. A lot of new HV wires and substations are being built out. Cities are expanding around clusters of dcs. Are they setting themselves up for a new rust belt?
The servers may well be worthless (or at least worth a lot less), but that's pretty much true for a long time. Not many people want to run on 10 year old servers (although I pay $30/month for a dedicated server that's dual Xeon L5640 or something like that, which is about 15 years old).
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)
They probably are right, but a counter argument could be how people thought going to the moon was pointless and insanely expensive, but the technology to put stuff in space and have GPS and comms satellites probably paid that back 100x
I don’t mean to invalidate your point (about genuine value arising from innovations originating from the Apollo program), but GPS and comms satellites (and heck, the Internet) are all products of nuclear weapons programs rather than civilian space exploration programs (ditto the Space Shuttle, and I could go on…).
Reality is that we don’t know how much of a trope this statement is.
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.
Sure, but what about the collective investment in smartphones, digital cameras, laptops, even cars. Not much modern technology is useful and practical after 10 years, let alone 20. AI is probably moving a little faster than normal, but technology depreciation is not limited to AI.
Utterly? Moores law per power requirement is dead, lower power units can run electric heating for small towns!
What I wonder is what this means for Coreweave, Lambda and the rest, who are essentially just renting out fleets of racks like this. Does it ultimately result in acquisition by a larger player? Severe loss of demand? Can they even sell enough to cover the capex costs?
These are also depreciating assets.
To piggyback on this, at enterprise level in modern age, the question is really not about "how are we going to serve all these users", it comes down to the fact that investors believe that eventually they will see a return on investment, and then pay whatever is needed to get the infra.
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.
I wonder if it's feasible to hook up NAND flash with a high bandwidth link necessary for inference.
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.
NAND is very, very slow relative to RAM, so you'd pay a huge performance penalty there. But maybe more importantly my impression is that memory contents mutate pretty heavily during inference (you're not just storing the fixed weights), so I'd be pretty concerned about NAND wear. Mutating a single bit on a NAND chip a million times over just results in a large pile of dead NAND chips.
No it's not slow - a single NAND chip in SSDs offers >1GB of bandwidth - inside the chip there are 100+ wafers actually holding the data, but in SSDs only one of them is active when reading/writing.
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.
Four H100 in a 2U rack didn't sound impressive, but that is accurate:
>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.
Why stop at 80 H100s for a mere 6.4 terabytes of GPU memory?
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...
And the big hyperscaler cloud providers are building city-block sized data centers stuffed to the gills with these racks as far as the eye can see
This isn’t like how Google was able to buy up dark fiber cheaply and use it.
From what I understand, this hardware has a high failure rate over the long term especially because of the heat they generate.
You have thousands of dollars, they have tens of billions. $1,000 vs $10,000,000,000. They have 7 more zeros than you, which is one less zero than the scale difference in users: 1 user (you) vs 700,000,000 users (openai). They managed to squeak out at least one or two zeros worth of efficiency at scale vs what you're doing.
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-...
You can knock off a zero or two just by time shifting the 700 million distinct users across a day/week and account for the mere minutes of compute time they will actually use in each interaction. So they might no see peaks higher than 10 million active inference session at the same time.
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.
During times of high utilization, how do they handle more requests than they have hardware? Is the software granular enough that they can round robin the hardware per token generated? UserA token, then UserB, then UserC, back to UserA? Or is it more likely that everyone goes into a big FIFO processing the entire request before switching to the next user?
I assume the former has massive overhead, but maybe it is worthwhile to keep responsiveness up for everyone.
Inference is essentially a very complex matrix algorithm run repeatedly on itself, each time the input matrix (context window) is shifted and the new generated tokens appended to the end. So, it's easy to multiplex all active sessions over limited hardware, a typical server can hold hundreds of thousands of active contexts in the main system ram, each less than 500KB and ferry them to the GPU nearly instantaneously as required.
During peaks they can kick out background jobs like model training or API users doing batch jobs.
One clever ingredient in OpenAI's secret sauce is billions of dollars of losses. About $5 billion dollars lost in 2024. https://www.cnbc.com/2024/09/27/openai-sees-5-billion-loss-t...
I think the most direct answer is that at scale, inference can be batched, so that processing many queries together in a parallel batch is more efficient than interactively dedicating a single GPU per user (like your home setup).
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/
700M weekly users doesn't say much about how much load they have.
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.
I'm sure there are countless tricks, but one that can implemented at home, and I know plays a major part in Cerebras' performance is: speculative decoding.
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.
gpt-oss-120b can be used with gpt-oss-20b as speculative drafting on LM Studio
I'm not sure it improved the speed much
To measure the performance gains on a local machine (or even standard cloud GPU setup), since you can't run this in parallel with the same efficiency you could in a high-ed data center, you need to compare the number of calls made to each model.
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.
a 6:1 parameter ratio is too small for specdec to have that much of an effect. You'd really want to see 10:1 or even more for this to start to matter
It is not just engineering. There are also huge, very huge, investments into infrastructure.
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.
At the heart of inference is matrix-vector multiplication. If you have many of these operations to do and only the vector part differs (which is the case when you have multiple queries), you can do matrix-matrix multiplication by stuffing the vectors into a matrix. Computing hardware is able to run the equivalent of dozens of matrix-vector multiplication operations in the same time it takes to do 1 matrix-matrix multiplication operation. This is called batching. That is the main trick.
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.
The short answer is "batch size". These days, LLMs are what we call "Mixture of Experts", meaning they only activate a small subset of their weights at a time. This makes them a lot more efficient to run at high batch size.
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"
I wonder then if its possible to load the unused parts into main memory, while the more used parts into VRAM
I'm pretty much an AI layperson but my basic understanding of how LLMs usually run on my or your box is:
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.
First off I’d say you can run models locally at good speed, llama3.1:8b runs fine a MacBook Air M2 with 16GB RAM and much better on a Nvidia RTX3050 which are fairly affordable.
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.
The idea that a GPU is dedicated to a single inference task is just generally incorrect. Inputs are batched, and it’s not a single GPU handling a single request, it’s a handful of GPUs in various parallelism schemes processing a batch of requests at once. There’s a latency vs throughput trade off that operators make. The larger that batch size the greater the latency, but it improves overall cluster throughput.
AFAIK main trick is batching, GPU can do same work on batch of data, you can work on many requests at the same time more efficiently.
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)
Your model keeps the weights on slow memory and needs to touch all of them to make 1 token for you. By batching you make 64 tokens for 64 users in one go. And they use dozens of GPUs in parallel to make 1024 tokens in the time your system makes 1 token. So even though the big system costs more, it is much more efficient when being used by many users in parallel. Also, by using many fast GPUs in series to process parts of the neural net, it produces output much faster for each user compared to your local system. You can't beat that.
I think it’s some combination of:
- 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.
1. They have many machines to split the load over 2. MoE architecture that lets them shard experts across different machines - 1 machine handles generating 1 token of context before the entire thing is shipped off to the next expert for the next token. This reduces bandwidth requirements by 1/N as well as the amount of VRAM needed on any single machine 3. They batch tokens from multiple users to further reduce memory bandwidth (eg they compute the math for some given weights on multiple users). This reduces bandwidth requirements significantly as well.
So basically the main tricks are batching (only relevant when you have > 1 query to process) and MoE sharding.
You also can't run a Google search. Some systems are just large!
> Sure, they have huge GPU clusters
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)
Trade secret?
With all the staff poaching the trade secrets may have now leaked?
That's half the reason tech companies poach.
Baseten serves models as a service, at scale. There’s quite a lot of interesting engineering both for inference and infrastructure perf. This is a pretty good deep dive into the tricks they employ: https://www.baseten.co/resources/guide/the-baseten-inference...
Lots of good answers that mention the big things (money, scale, and expertise). But one thing I haven’t seen mentioned yet is that the transformer math is probably against your use case. Batch compute on beefy hardware is currently more efficient than computing small sequences for a single user at a time, since these models tend to be memory bound and not compute bound. They have the users that makes the beefy hardware make sense, enough people are querying around the same time to make some batching possible.
TL;DR: It's massively easier to run a few models really fast than it is to run many different models acceptably.
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
Isn’t the answer to the question just classic economies of scale?
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?
It’s the same principle as:
https://www.tripadvisor.com/Restaurant_Review-g60763-d477541...
If the explanation really is, as many comments here suggest, that prompts can be run in parallel in batches at low marginal additional cost, then that feels like bad news for the democratization and/or local running of LLMs. If it’s only cost-effective to run a model for ~thousands of people at the same time, it’s never going to be cost-effective to run on your own.
Sure, but that's how most of human society works already.
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.
50 years ago general computers were also time shared. Then the pendulum swing to desktop, then back to central.
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
Well, you can also batch your own queries. Not much use for a chatbot but for an agentic system or offline batch processing it becomes more reasonable.
Consider a system were running a dozen queries at once is only marginally more expensive than running one query. What would you build?
Simple answer: they are throwing billions of dollars at infrastructure (GPU) and losing money with every user.
You’re not losing money if money flows in faster than it flows out
Well, their huge GPU clusters have "insane VRAM". Once you can actually load the model without offloading, inference isn't all that computationally expensive for the most part.
Multi-tenancy likely explains the bulk of it. $10k vs. $10b gives them six orders of magnitude more GPU resources, but they have 9 orders of magnitude more users. The average user is probably only running an active ChatGPT query for a few minutes per day, which covers the remaining 3 orders of magnitude.
You and your engineering team might be able to figure it out and purchase enough equipment also if you had received billions of dollars. And billions and billions. And more billions and billions and billions. Then additional billions, and more billions and billions and even more billions and billions of dollars. They have had 11 rounds of funding totaling around $60 billion.
I think this article can be interesting:
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
Once you have enough GPUs to have your whole model available in GPU RAM you can do inference pretty fast.
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.
How does a billion dollar company scale in a way that a single person cannot?
Look for positron.ai talks about their tech, they discuss their approach to scaling LLM workloads with their dedicated hardware. It may not be what is done by OpenAI or other vendors, but you'll get an idea of the underlying problems.
The first step is to acquire hardware fast enough to run one query quickly (and yes, for some model size you are looking at sharding the model and distributed runs). The next one is to batch request, improving GPU use significantly.
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/
Have you looked at what happens to tokens per second when you increase batch size? The cost of serving 128 queries at once is not 128x the cost of serving one query.
This. the main trick, outside of just bigger hardware, is smart batching. E.g. if one user asks why the sky is blue, the other asks what to make for dinner, both queries go though the same transformer layers, same model weights so they can be answered concurrently for very little extra GPU time. There's also ways to continuously batch requests together so they don't have to be issued at the same time.
Complete guess, but my hunch is that it's in the sharding. When they break apart your input into its components, they send it off to hardware that is optimized to solve for that piece. On that hardware they have insane VRAM and it's already cached in a way that optimizes that sort of problem.
They have more than 700mX your computing budget?
batching & spread of users over time will get you there already
The marginal value of money is low. So it's not linear. They can buy orders of magnitude more GPUs than you can buy.
At the end of the day, the answer is... specialized hardware. No matter what you do on your local system, you don't have the interconnects necessary. Yes, they have special software, but the software would not work locally. NVIDIA sells entire solutions and specialized interconnects for this purpose. They are well out of the reach of the standard consumer.
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.
Azure servers
Huge batches to find the perfect balance between compute and memory banthwidth, quantized models, speculative decoding or similar techniques, MoE models, routing of requests on smaller models if required, batch processing to fill the GPUs when demand is lower (or electricity is cheaper).
Money. Don't let them lie to you. just look at nvidia.
They are throwing money at this problem hoping you throw more money back.
Because they spend billions per year on that.
Data centers, and use of client hardware, those 700M clients' hardware are being partially used as clusters.
They also don’t need one system per user. Think of how often you use their system over the week, maybe one hour total? You can shove 100+ people into sharing one system at that rate… so already you’re down to only needing 7 million systems.
By setting billions of VC money on fire: https://en.wikipedia.org/wiki/OpenAI
No, really. They just have entire datacenters filled with high end GPUs.
not affiliated with them and i might be a little out of date but here are my guesses
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.
https://en.wikipedia.org/wiki/Autoscaling
Finally, some1 with the important questions!
Hint: it's a money thing.
They rewrote it in Rust/Zig the one you have is written in Ruby. :-p
They are hosted on Microsoft Azure cloud infrastructure and Microsoft owns 49%
They are also partnering with rivals like Google for additional capacity https://www.reuters.com/business/retail-consumer/openai-taps...
In fact logout gpt I found it hosted on azure
I work at a university data center, although not on LLMs. We host state of the art models for a large number of users. As far as I understand, there is no secret sauce. We just have a big GPU cluster with a batch system, where we spin up jobs to run certain models. The tricky part for us is to have the various models available on demand with no waiting time.
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.