Same GPU, 50% More Output: Why the Operator Matters — Oliver Kennon, SemiAnalysis

Thank you to everyone who Thank you to everyone who took the time to be here took the time to be here took the time to be here today. My name is today. My name is today. My name is Oliver Cannon. I am a Oliver Cannon. I am a Oliver Cannon. I am a research analyst at research analyst at research analyst at SemiAnalysis. Hmm, time and SemiAnalysis. Hmm, time and SemiAnalysis. Hmm, time and attention are the most valuable attention are the most valuable attention are the most valuable things we have. Um, things we have. Um, things we have. Um, unless, of course, you're an unless, of course, you're an unless, of course, you're an early- early- early- stage investor in OpenAI or Anthropic stage investor in OpenAI or Anthropic . Um, with that, uh, let's . Um, with that, uh, let's . Um, with that, uh, let's get down to business. get down to business. get down to business. Um, not all GPUs Um, not all GPUs are created are created equal. The equal. The equal. The GPU is GPU is GPU is at the operator level. at the operator level. at the operator level. Um, buying Um, buying Um, buying equipment gives you equipment gives you equipment gives you potential. Um, potential. Um, potential. Um, operational operational operational capacity capacity capacity determines how much determines how much determines how much useful work you useful work you useful work you can actually can actually can actually do. Um, that do. Um, that do. Um, that brings us to the brings us to the brings us to the question that I'm question that I'm question that I'm going to try to answer with going to try to answer with going to try to answer with all of you all of you all of you today. Who can today. Who can today. Who can create create create the intelligence you need at the the intelligence you need at the the intelligence you need at the lowest price? lowest price? lowest price? Cost Cost Cost matters. Um, who, um, matters. Um, who, um, matters. Um, who, um, in today's in today's in today's presentation I want to presentation I want to presentation I want to show you guys show you guys show you guys why and how operational why and how operational why and how operational capability changes capability changes capability changes this value and why it this value and why it this value and why it matters in the commodity matters in the commodity matters in the commodity market, which is market, which is market, which is computing computing computing resources. What is important resources. What is important resources. What is important in the commodity market? in the commodity market? in the commodity market? Um, many Um, many Um, many technology technology technology investors or investors or investors or technologists may technologists may technologists may not be familiar with not be familiar with not be familiar with this structure. Um, this structure. Um, this structure. Um, so a little so a little so a little crash course. Um, in the crash course. Um, in the crash course. Um, in the commodity market, commodity market, commodity market, producers compete producers compete to provide to provide to provide comparable output comparable output comparable output . Um, . Um, . Um, lower lower lower production costs give you production costs give you production costs give you more opportunities more opportunities more opportunities to compete when to compete when to compete when prices fall. Uh, that's what the prices fall. Uh, that's what the cost curve shows. Producers cost curve shows. Producers are ranked by are ranked by are ranked by how much they how much they how much they spend to spend to spend to do the same do the same do the same useful work. If useful work. If useful work. If two producers two producers two producers earn the same earn the same earn the same price, the one who price, the one who price, the one who spends less on spends less on spends less on doing the work doing the work doing the work gets more of it for gets more of it for gets more of it for themselves. Pretty simple, themselves. Pretty simple, themselves. Pretty simple, right? For us, right? For us, right? For us, intelligence is an intelligence is an intelligence is an economic economic economic outcome that outcome that outcome that interests us. Ultimately,

interests us. Ultimately, interests us. Ultimately, our our North Star metric will be North Star metric will be North Star metric will be the price for a successfully the price for a successfully the price for a successfully completed task with the completed task with the completed task with the required quality. Er required quality. Er , unfortunately, we're not , unfortunately, we're not , unfortunately, we're not quite there yet. This is quite quite there yet. This is quite quite there yet. This is quite difficult difficult difficult to standardize. Um, to standardize. Um, to standardize. Um, today the price per today the price per today the price per token is a useful token is a useful token is a useful practical practical practical metric, especially metric, especially metric, especially for comparable for comparable for comparable workloads workloads workloads and quality requirements, and and quality requirements, and and quality requirements, and gets us closer to the gets us closer to the gets us closer to the delivered delivered delivered result, much result, much result, much closer than the often closer than the often closer than the often quoted, uh, all you quoted, uh, all you quoted, uh, all you finance bros here finance bros here finance bros here have heard this, uh, GPU price have heard this, uh, GPU price have heard this, uh, GPU price in dollars per hour. Er in dollars per hour. Er , or as we , or as we , or as we consider it, consider it, consider it, cost. Um, we cost. Um, we cost. Um, we view view view GPU hours as a measure of GPU hours as a measure of GPU hours as a measure of input, not input, not input, not output, output, output, the difference between the difference between the difference between cost and price. cost and price. cost and price. Customer prices Customer prices Customer prices differ from the differ from the differ from the manufacturer's costs of manufacturer's costs of manufacturer's costs of performing the work. performing the work. GPU hours are the GPU hours are the input, intelligence input, intelligence is the output. Uh, is the output. Uh, is the output. Uh, on this slide we on this slide we on this slide we can visualize the can visualize the can visualize the input data input data input data needed to needed to needed to create useful create useful create useful intelligence. Um, we'll intelligence. Um, we'll intelligence. Um, we'll start with the start with the start with the hardware and the hardware and the hardware and the electrical, the electrical, the electrical, the GPUs, uh, for GPUs, uh, for GPUs, uh, for this presentation.

this presentation. this presentation. Software Software Software and and and operations can be operations can be operations can be thought of as a thought of as a thought of as a multiplier on these first multiplier on these first multiplier on these first two inputs, with the two inputs, with the two inputs, with the end product being end product being end product being useful intelligence. useful intelligence. The operator is the one who The operator is the one who selects and selects and selects and implements the implements the implements the software that software that software that serves the model and serves the model and serves the model and ensures that ensures that ensures that requests are passed through the requests are passed through the requests are passed through the system. The machines system. The machines system. The machines may be may be may be available, but you available, but you available, but you still need still need still need someone to make them someone to make them someone to make them provide answers provide answers provide answers that customers that customers that customers can actually can actually can actually use. What use. What use. What matters is how much matters is how much matters is how much useful work you useful work you useful work you get and whether get and whether get and whether it meets it meets it meets your requirements for the your requirements for the your requirements for the given task. given task. given task. The conversion may The conversion may The conversion may vary depending vary depending vary depending on the chip type. on the chip type. on the chip type. Let's look at Let's look at Let's look at this with a measured this with a measured this with a measured example. Here is a example. Here is a example. Here is a specific example from specific example from specific example from our public our public inference calculator. inference calculator. You all can You all can You all can check it out, check it out, check it out, semi-analysis.com. Um, both semi-analysis.com. Um, both semi-analysis.com. Um, both estimates estimates estimates use use B200 GPUs serving the B200 GPUs serving the same model same model same model with about 8000 with about 8000 with about 8000 input tokens and 1000 input tokens and 1000 input tokens and 1000 output tokens, which output tokens, which output tokens, which is a pretty standard is a pretty standard is a pretty standard ratio. ratio. ratio. Target P50–100 output Target P50–100 output Target P50–100 output tokens per second tokens per second tokens per second for each for each for each user. Um, the user. Um, the user. Um, the service stack service stack is the is the is the software that software that software that processes processes processes the model's responses. In the P90 scenario, the model's responses. In the P90 scenario, the model's responses. In the P90 scenario, one of these stacks one of these stacks one of these stacks produces about 200 produces about 200 produces about 200 generated generated generated output tokens per output tokens per output tokens per second, while second, while second, while the other produces about the other produces about the other produces about 300. This is a big difference 300. This is a big difference 300. This is a big difference that depends on that depends on that depends on the the the software, software, software, settings, and settings, and settings, and the number of the number of the number of requests being executed.

requests being executed. requests being executed. All choices are made by All choices are made by All choices are made by the operator. Same the operator. Same the operator. Same chip, same target chip, same target chip, same target interactivity, interactivity, interactivity, about 50% more about 50% more about 50% more raw data. That's a raw data. That's a raw data. That's a lot. So what does this lot. So what does this lot. So what does this mean for the mean for the mean for the buyer's account? How much do buyer's account? How much do these intelligences actually cost the client? these intelligences actually cost the client? For this example, For this example, For this example, let's assume the let's assume the let's assume the hourly bill is hourly bill is hourly bill is the same for both the same for both the same for both deployments, and both deployments, and both deployments, and both meet your meet your meet your requirements. With requirements. With requirements. With about 1.5 times the about 1.5 times the performance, your performance, your cost to rent cost to rent cost to rent one eligible one eligible one eligible token is proportionally token is proportionally token is proportionally lower. That is, an hour lower. That is, an hour lower. That is, an hour buys you more buys you more buys you more useful and useful and useful and useful output useful output useful output . What is considered . What is considered . What is considered useful intelligence? useful intelligence? useful intelligence? Useful intelligence Useful intelligence Useful intelligence should do should do should do the job you the job you the job you hired it to do. Pretty hired it to do. Pretty hired it to do. Pretty simple, right? simple, right? simple, right? The answer should be The answer should be The answer should be good enough good enough good enough for the task, for the task, for the task, come come come quickly enough, and be quickly enough, and be quickly enough, and be available when available when available when you need it. Your you need it. Your you need it. Your information information information must also remain within the must also remain within the must also remain within the access limits access limits access limits you agreed upon. In you agreed upon. In you agreed upon. In other words, it other words, it other words, it must be safe.

must be safe. must be safe. This is an underrated This is an underrated This is an underrated requirement that is not requirement that is not requirement that is not factored into factored into chip performance metrics, chip performance metrics, but it is essential but it is essential but it is essential for the chip to be user- for the chip to be user- for the chip to be user- friendly and useful. friendly and useful. In our In our published published published security work, we discovered security work, we discovered monitoring access that monitoring access that exposed logs and exposed logs and exposed logs and metrics from another metrics from another metrics from another customer. This is a case, and customer. This is a case, and unfortunately common, unfortunately common, unfortunately common, when when when GPUs were GPUs were GPUs were fast, and in the fast, and in the fast, and in the case I'm case I'm case I'm talking about, very, very talking about, very, very talking about, very, very fast, which makes fast, which makes fast, which makes it all the more sad. Uh it all the more sad. Uh , and the service still , and the service still , and the service still failed miserably failed miserably failed miserably . Um, . Um, . Um, ensuring ensuring ensuring cluster security requires cluster security requires cluster security requires constant work, just like constant work, just like ensuring ensuring promised promised promised performance. In performance. In performance. In most cases, it is the most cases, it is the most cases, it is the operator who bears this operator who bears this operator who bears this responsibility. responsibility. Someone has to Someone has to make the system make the system make the system work in work in work in production. production. production. Let's assume that a Let's assume that a Let's assume that a faster faster faster update becomes available. Someone update becomes available. Someone update becomes available. Someone should test should test should test it on a it on a it on a customer workload customer workload , verify that , verify that , verify that the result meets the result meets the result meets the requirements, and then the requirements, and then the requirements, and then decide when, where, and decide when, where, and decide when, where, and how to deploy it. how to deploy it. how to deploy it. This means validating This means validating This means validating actual responses actual responses , testing under , testing under , testing under load, and load, and load, and providing a way to providing a way to providing a way to revert to a revert to a revert to a previous version previous version previous version if the update if the update if the update causes problems causes problems causes problems after it's after it's after it's deployed. Once deployed. Once deployed. Once deployed, someone deployed, someone deployed, someone needs to maintain needs to maintain needs to maintain it and respond to it and respond to it and respond to issues as they arise.

issues as they arise. This work can be This work can be outsourced to your outsourced to your outsourced to your supplier, your supplier, your supplier, your own team, or own team, or own team, or both. In any both. In any both. In any case, you need case, you need case, you need to know who is responsible to know who is responsible to know who is responsible and whether they can and whether they can and whether they can do it. So, how do we do it. So, how do we do it. So, how do we evaluate an operator? Uh evaluate an operator? Uh , Semi Analysis ClusterMax, , Semi Analysis ClusterMax, , Semi Analysis ClusterMax, we're currently we're currently we're currently at version at version at version 3.0, uh, it evaluates 3.0, uh, it evaluates 3.0, uh, it evaluates the operational the operational the operational capability of capability of capability of managed managed managed GPU clusters, several hundred, uh, GPU clusters, several hundred, uh, GPU clusters, several hundred, uh, their number is growing. their number is growing. their number is growing. Please Please Please contact us contact us contact us if you would like to if you would like to if you would like to upgrade to ClusterMax 4.0. Uh upgrade to ClusterMax 4.0. Uh , we're looking at , we're looking at , we're looking at provisioning, provisioning, provisioning, reliability, reliability, reliability, monitoring, and monitoring, and monitoring, and security. Uh, security. Uh, security. Uh, hands-on testing hands-on testing , documentation, and , documentation, and , documentation, and customer feedback customer feedback customer feedback help us help us help us understand how a understand how a understand how a service works, not just on a service works, not just on a service works, not just on a good day, but also good day, but also good day, but also when the machine breaks when the machine breaks when the machine breaks down. I want to down. I want to down. I want to take this opportunity take this opportunity take this opportunity to emphasize again that to emphasize again that to emphasize again that ClusterMax ratings ClusterMax ratings ClusterMax ratings apply to apply to apply to managed clusters, managed clusters, managed clusters, er, different from er, different from er, different from bare metal, very bare metal, very bare metal, very different. Uh, with different. Uh, with different. Uh, with bare metal, more of the bare metal, more of the bare metal, more of the operational work is operational work is operational work is usually done by usually done by usually done by the customer. Er, the customer. Er, the customer. Er, applying the applying the applying the managed managed managed service rating here doesn't service rating here doesn't service rating here doesn't automatically cover automatically cover automatically cover every product that every product that every product that the vendor sells.

the vendor sells. the vendor sells. Levels are Levels are Levels are analogs. A bit of a tongue analogs. A bit of a tongue analogs. A bit of a tongue twister. Eh, I twister. Eh, I twister. Eh, I managed, quite managed, quite managed, quite happy with it. And happy with it. And happy with it. And pricing and pricing and pricing and availability are also availability are also availability are also important factors important factors important factors in this assessment. Uh, these are in this assessment. Uh, these are in this assessment. Uh, these are metrics that metrics that metrics that don't directly don't directly don't directly measure, uh, the volume of measure, uh, the volume of measure, uh, the volume of token production, token production, token production, uh, but they're incredibly uh, but they're incredibly uh, but they're incredibly important. Obviously, important. Obviously, important. Obviously, very important. Uh, very important. Uh, very important. Uh, the assessment helps us the assessment helps us the assessment helps us understand who can understand who can understand who can do it. Now we do it. Now we do it. Now we can move on to the can move on to the can move on to the next question. next question. next question. How much does it really How much does it really How much does it really cost to produce cost to produce cost to produce useful intelligence? useful intelligence? useful intelligence? Let's start with the Let's start with the Let's start with the calculation bill and other calculation bill and other calculation bill and other costs that manufacturers costs that manufacturers costs that manufacturers still have to still have to still have to cover. The cost cover. The cost cover. The cost of rent and the cost of of rent and the cost of of rent and the cost of production production production are different. In are different. In are different. In our example, the our example, the GPU vendor GPU vendor rents hours, the rents hours, the rents hours, the output manufacturer output manufacturer output manufacturer buys those hours and buys those hours and buys those hours and sells the responses. The sells the responses. The sells the responses. The rental bill is rental bill is rental bill is revenue for revenue for revenue for the supplier and the supplier and the supplier and input cost for the input cost for the input cost for the output manufacturer. The output manufacturer. The output manufacturer. The GPU owner GPU owner GPU owner pays for the pays for the pays for the hardware, hardware, hardware, electricity, electricity, electricity, space, and space, and space, and support. The output support. The output support. The output producer's total cost producer's total cost producer's total cost includes includes includes the rent plus the rent plus the rent plus any any any software, software, software, personnel, and other personnel, and other personnel, and other services it services it services it pays for in addition to that pays for in addition to that pays for in addition to that rent. If the rent. If the rent. If the output manufacturer owns its output manufacturer owns its output manufacturer owns its own IT own IT equipment, we equipment, we equipment, we spread spread spread the investment in the the investment in the the investment in the equipment over its equipment over its equipment over its service life and service life and service life and add add add operating operating operating costs. What a costs. What a costs. What a producer charges and what he producer charges and what he producer charges and what he spends are separate spends are separate spends are separate numbers. To understand numbers. To understand numbers. To understand the value of useful the value of useful the value of useful intelligence, we intelligence, we intelligence, we need to know need to know need to know how much useful how much useful how much useful work work work the costs actually produce. Er the costs actually produce. Er , better operations , better operations , better operations reduce this cost of reduce this cost of reduce this cost of useful intelligence. The useful intelligence. The useful intelligence. The B200 comparison B200 comparison B200 comparison showed why production volume showed why production volume matters. More matters. More acceptable responses acceptable responses acceptable responses for the same billable for the same billable for the same billable hours can hours can hours can allocate a

allocate a allocate a manufacturer's manufacturer's manufacturer's equipment rental costs to more equipment rental costs to more equipment rental costs to more useful work. For a useful work. For a useful work. For a manufacturer that manufacturer that manufacturer that owns its own owns its own GPUs, the same GPUs, the same logic spreads the logic spreads the logic spreads the investment in investment in investment in hardware and hardware and hardware and operating operating operating costs over a larger costs over a larger costs over a larger volume of output. volume of output. volume of output. Calculate all Calculate all Calculate all the costs required the costs required the costs required to provide the service to provide the service to provide the service during the same during the same during the same period, and then period, and then period, and then divide by the volume of divide by the volume of divide by the volume of products that products that products that customers can customers can customers can actually actually actually use. This use. This use. This gives us the cost of an gives us the cost of an gives us the cost of an acceptable token acceptable token acceptable token for a comparable for a comparable workload. The best workload. The best operations also operations also operations also require investment require investment . Testing . Testing . Testing an update, an update, an update, deploying it safely, deploying it safely, deploying it safely, and maintaining its and maintaining its and maintaining its performance performance performance requires people and requires people and requires people and money. The additional money. The additional money. The additional volume of production must volume of production must volume of production must justify the justify the justify the additional costs. This additional costs. This additional costs. This advantage can be advantage can be advantage can be maintained when maintained when maintained when the team iterates on this the team iterates on this the team iterates on this work as the work as the software and software and workloads workloads workloads change. But one change. But one change. But one good benchmark doesn't good benchmark doesn't good benchmark doesn't establish it. You establish it. You establish it. You have to earn it have to earn it have to earn it every day. Each new every day. Each new every day. Each new workload workload workload still needs still needs still needs validation, but a validation, but a validation, but a repeatable repeatable repeatable operational process, from what operational process, from what operational process, from what we have we have we have observed, can observed, can observed, can sustain a long-term sustain a long-term sustain a long-term cost advantage cost advantage cost advantage and price advantage.

and price advantage. and price advantage. However, even if you However, even if you However, even if you continue continue continue to improve, to improve, to improve, your competitors are not your competitors are not your competitors are not sitting still. Your sitting still. Your sitting still. Your position on the position on the position on the cost curve is relative. cost curve is relative. cost curve is relative. This slide of all This slide of all This slide of all slides should clearly slides should clearly slides should clearly demonstrate demonstrate demonstrate the importance of the product the importance of the product the importance of the product model in all of this model in all of this . Your position on the . Your position on the . Your position on the cost curve cost curve cost curve depends on what depends on what your competitors have to offer, not your competitors have to offer, not just you. This is a just you. This is a just you. This is a relative indicator. relative indicator. relative indicator. Here is a simple hypothesis. Here is a simple hypothesis. Here is a simple hypothesis. We compare the We compare the We compare the same task, the same task, the same task, the same required same required same required quality, quality, quality, using using using fictitious fictitious fictitious cost units. I cost units. I cost units. I like to imagine like to imagine like to imagine gold bars. Your gold bars. Your gold bars. Your total cost of a total cost of a total cost of a successful task successful task successful task drops from 10 to 8. Over the drops from 10 to 8. Over the drops from 10 to 8. Over the same period, same period, same period, your your your competitors' cost drops from 6 competitors' cost drops from 6 competitors' cost drops from 6 to 12. You have improved to 12. You have improved to 12. You have improved your results. This is your results. This is your results. This is great. But you both great. But you both great. But you both did it. And in this did it. And in this did it. And in this example, you started as a example, you started as a example, you started as a cheaper manufacturer and cheaper manufacturer and cheaper manufacturer and ended as a more ended as a more ended as a more expensive one. This gives expensive one. This gives expensive one. This gives the competitor more the competitor more the competitor more opportunity to opportunity to opportunity to lower the lower the lower the selling price, provided that selling price, provided that selling price, provided that customers find customers find customers find both services both services both services equally acceptable equally acceptable . Improving your . Improving your . Improving your results doesn't results doesn't results doesn't automatically move automatically move automatically move you forward.

you forward. you forward. Applying the Applying the Applying the commodity model commodity model commodity model forces you forces you forces you to realize that you to realize that you to realize that you must compare must compare must compare your costs of your costs of your costs of useful work on a useful work on a useful work on a relative basis. relative basis. relative basis. Comparing your Comparing your Comparing your own product to what own product to what your your competitors have to offer. This competitors have to offer. This competitors have to offer. This comparison is easy comparison is easy comparison is easy to miss at to miss at to miss at times like today, times like today, times like today, when demand far when demand far when demand far exceeds supply exceeds supply , and will continue to be so for the , and will continue to be so for the , and will continue to be so for the foreseeable foreseeable foreseeable future. Two or three future. Two or three future. Two or three years. Visibility. years. Visibility. years. Visibility. Good. Deficits can Good. Deficits can Good. Deficits can hide these hide these hide these differences. When differences. When differences. When demand is strong and demand is strong and demand is strong and supply is limited, supply is limited, supply is limited, producers with very producers with very producers with very different costs different costs different costs can find can find can find customers. High customers. High customers. High income may make these income may make these income may make these differences less differences less differences less pressing. If you're a pressing. If you're a pressing. If you're a CEO CEO , why should you , why should you , why should you focus focus focus your valuable your valuable your valuable time, energy, and time, energy, and time, energy, and money on solidifying money on solidifying money on solidifying your place on the your place on the your place on the cost curve when you cost curve when you cost curve when you could be building a could be building a could be building a new, massive new, massive new, massive data center? data center? data center? Uh, at times like these, it's Uh, at times like these, it's Uh, at times like these, it's important to take a step important to take a step important to take a step back and look back and look back and look a little into perspective. a little into perspective. a little into perspective. Uh, look at the Uh, look at the Uh, look at the historical historical historical perspective a little. Uh, perspective a little. Uh, perspective a little. Uh, every commodity in every commodity in every commodity in human history has had human history has had human history has had fluctuations in supply and fluctuations in supply and fluctuations in supply and demand. Just because we demand. Just because we demand. Just because we don't yet see when it don't yet see when it don't yet see when it will happen doesn't mean will happen doesn't mean this cycle will be any this cycle will be any this cycle will be any different. Uh, unless you different. Uh, unless you different. Uh, unless you 're a complete AI maxi, 're a complete AI maxi, 're a complete AI maxi, which some of us are.

which some of us are. which some of us are. So I am very prone to So I am very prone to So I am very prone to making mistakes. I making mistakes. I making mistakes. I make mistakes every day. Uh, make mistakes every day. Uh, make mistakes every day. Uh, but I would say that but I would say that but I would say that if you have a if you have a if you have a leveraged business, you leveraged business, you leveraged business, you need to be need to be need to be especially conscientious about especially conscientious about especially conscientious about it. Being it. Being it. Being proactive in your proactive in your proactive in your creation, building creation, building creation, building your your your cost advantage, uh, may cost advantage, uh, may cost advantage, uh, may not be as not be as not be as attractive as attractive as attractive as increasing profits, increasing profits, increasing profits, but it's crucial but it's crucial but it's crucial if you want to if you want to if you want to survive the business cycle. survive the business cycle. survive the business cycle. Uh, that's why you Uh, that's why you Uh, that's why you want to develop want to develop want to develop your operational your operational your operational capabilities and capabilities and capabilities and cost advantage cost advantage cost advantage while you have while you have while you have the opportunity, uh, when the opportunity, uh, when the opportunity, uh, when the sun is shining, when the sun is shining, when the sun is shining, when you can raise you can raise you can raise capital. Er, they capital. Er, they matter even more when matter even more when customers have more customers have more customers have more options and prices are options and prices are options and prices are under pressure. So how do we under pressure. So how do we under pressure. So how do we clearly define what, uh clearly define what, uh , the buyer is actually getting , the buyer is actually getting , the buyer is actually getting ? Comparative ? Comparative ? Comparative calculations calculations calculations require clear require clear require clear estimates. Now let's estimates. Now let's estimates. Now let's look at look at look at the computation, uh, that the the computation, uh, that the intelligence manufacturer buys. We have already intelligence manufacturer buys. We have already seen that the seen that the seen that the same chip can same chip can same chip can provide different provide different provide different results depending results depending results depending on how it is on how it is on how it is operated. Er, operated. Er, operated. Er, two two B200 GPU hours offers may include B200 GPU hours offers may include different services and different services and different services and leave different leave different leave different operational tasks to operational tasks to operational tasks to the customer. The the customer. The the customer. The product structure needs a product structure needs a product structure needs a way to describe these way to describe these way to describe these differences so that differences so that differences so that buyers know which buyers know which buyers know which services they can services they can services they can consider consider consider comparable. This is an comparable. This is an comparable. This is an assessment task.

assessment task. The rating describes a The rating describes a specific specific specific computing computing computing service and what service and what service and what its buyers can expect its buyers can expect its buyers can expect . It does not . It does not . It does not make intelligence make intelligence make intelligence interchangeable for interchangeable for interchangeable for every task. every task. every task. Once we have Once we have Once we have identified this service, identified this service, identified this service, we can ask we can ask we can ask what its what its what its rental price tells us. What does the rental price tells us. What does the rental price benchmark measure? Er, the rental price benchmark measure? Er, the rental price benchmark rental price benchmark rental price benchmark summarizes the price of summarizes the price of GPU hours GPU hours within a defined within a defined within a defined market. This is not the market. This is not the manufacturer's production costs and is not a measure of the manufacturer's production costs and is not a measure of the information provided. information provided. information provided. We need to know We need to know We need to know where these where these where these prices are coming from. An advertised prices are coming from. An advertised prices are coming from. An advertised offer and a offer and a offer and a completed deal are completed deal are completed deal are not the same thing. not the same thing. The scope of services, the terms of The scope of services, the terms of the contract, whose prices the contract, whose prices are selectively selected, and how are selectively selected, and how they are combined – they are combined – all of this forms this all of this forms this all of this forms this starting point. The quality starting point. The quality starting point. The quality of operation can of operation can of operation can affect these prices affect these prices affect these prices without being separately without being separately without being separately measured by an index measured by an index . Performing a . Performing a . Performing a benchmark benchmark benchmark rental calculation rental calculation rental calculation does not show does not show does not show how much useful how much useful how much useful work your deployment will produce work your deployment will produce work your deployment will produce . To . To . To use it use it use it as a hedge, as a hedge, as a hedge, you need to understand how you need to understand how you need to understand how it tracks the it tracks the it tracks the rent you rent you rent you actually pay actually pay actually pay or earn. What or earn. What or earn. What can can computing futures hedge?

computing futures hedge? Recently, a lot of attention has been Recently, a lot of attention has been Recently, a lot of attention has been paid paid paid to to to computing futures. For anyone computing futures. For anyone computing futures. For anyone who doesn't know, a who doesn't know, a who doesn't know, a futures contract futures contract futures contract has a standardized has a standardized has a standardized deadline for deadline for deadline for future delivery future delivery future delivery or cash or cash or cash settlement. Monetary settlement. Monetary settlement. Monetary calculation depends calculation depends calculation depends on the price benchmark. on the price benchmark. on the price benchmark. Buyers and sellers Buyers and sellers Buyers and sellers face face face uncertainty about uncertainty about uncertainty about rental prices. rental prices. rental prices. Remember the Remember the Remember the service stacks and service stacks and service stacks and monitor failure monitor failure monitor failure monitoring. Evaluating monitoring. Evaluating monitoring. Evaluating who runs who runs who runs the system helps the system helps the system helps describe the services that describe the services that describe the services that customers receive. This customers receive. This customers receive. This helps us helps us helps us determine the rating so that determine the rating so that determine the rating so that buyers know which buyers know which buyers know which computing computing computing services are services are services are comparable. comparable. comparable. The benchmark should then The benchmark should then The benchmark should then track the track the track the rent they rent they rent they actually pay actually pay actually pay or earn. That's or earn. That's or earn. That's why I make this why I make this why I make this argument. argument. argument. Operator rating is an important Operator rating is an important Operator rating is an important building block building block building block for building a for building a for building a reliable reliable computing computing technology futures market. The rating technology futures market. The rating technology futures market. The rating determines this work. determines this work. determines this work. It does not, by itself, It does not, by itself, It does not, by itself, determine a futures determine a futures determine a futures rating, rating, rating, establish a reliable establish a reliable establish a reliable benchmark, or generate benchmark, or generate benchmark, or generate enough trades for enough trades for enough trades for easy buying and easy buying and easy buying and selling. It also doesn't selling. It also doesn't selling. It also doesn't necessarily make necessarily make necessarily make hedging hedging hedging effective. effective. effective. Rent hedging Rent hedging Rent hedging can help can help can help offset offset offset rent changes. It rent changes. It rent changes. It does not fix a failed does not fix a failed does not fix a failed deployment and does not deployment and does not deployment and does not guarantee that your guarantee that your guarantee that your customer will pay.

customer will pay. Setting a price Setting a price higher than the index higher than the index higher than the index does not in itself does not in itself does not in itself make it a bad make it a bad make it a bad hedge. What matters is hedge. What matters is hedge. What matters is whether changes in the reference whether changes in the reference whether changes in the reference value value value reflect changes in reflect changes in reflect changes in your rent. your rent. your rent. Unexpected changes in Unexpected changes in Unexpected changes in these relationships these relationships these relationships leave a real leave a real leave a real price risk. These price risk. These price risk. These tools can tools can tools can help you help you help you understand what you're understand what you're understand what you're buying and manage your buying and manage your buying and manage your rental risk rental risk rental risk . For an analytics producer, . For an analytics producer, rent is rent is one part of the one part of the one part of the cost. The operator cost. The operator cost. The operator still has to still has to still has to convert these convert these convert these calculations into work calculations into work calculations into work that the customer can that the customer can that the customer can actually actually actually use. This use. This use. This closes the circle. So, the closes the circle. So, the closes the circle. So, the GPU GPU GPU falls to the falls to the falls to the operator level. So, who operator level. So, who operator level. So, who can produce the can produce the analytics you need at a analytics you need at a lower cost? At the lower cost? At the lower cost? At the lowest price. lowest price. lowest price. A manufacturer that can A manufacturer that can A manufacturer that can repeatedly repeatedly repeatedly meet your meet your meet your requirements at the lowest requirements at the lowest requirements at the lowest total price for a total price for a total price for a successfully completed successfully completed successfully completed task. Calculate task. Calculate task. Calculate all the costs of all the costs of all the costs of providing the service, and providing the service, and providing the service, and then divide by the then divide by the then divide by the volume of products that volume of products that volume of products that the customer can actually the customer can actually the customer can actually use. use. use. Operating Operating Operating capacity affects capacity affects capacity affects both sides of both sides of both sides of this equation: this equation: this equation: how much you how much you how much you spend and spend and spend and how much more useful how much more useful how much more useful work you can work you can work you can do. Additionally, do. Additionally, do. Additionally, your position your position your position depends on the position of depends on the position of depends on the position of your competitors. You your competitors. You your competitors. You can improve can improve can improve your results and your results and your results and still fall behind if still fall behind if still fall behind if your competitors your competitors your competitors are improving are improving are improving faster than you. This is faster than you. This is faster than you. This is why why why operator understanding is important operator understanding is important operator understanding is important throughout the throughout the throughout the cycle, not just at a cycle, not just at a cycle, not just at a specific point in time. Uh specific point in time. Uh , especially when , especially when , especially when customers have more customers have more customers have more choice and prices choice and prices choice and prices are under are under are under more pressure. In more pressure. In more pressure. In conclusion, if you're conclusion, if you're conclusion, if you're just getting started and want to just getting started and want to just getting started and want to take full take full take full control, this is exactly what control, this is exactly what control, this is exactly what you need. The you need. The you need. The GPU GPU GPU drops to the

drops to the drops to the operator level. operator level. operator level. Help Help Help build a more build a more build a more informed computing market informed computing market . Ladies and . Ladies and gentlemen, thank gentlemen, thank gentlemen, thank you very much for your time you very much for your time you very much for your time today. At SemiAnalysis, we today. At SemiAnalysis, we today. At SemiAnalysis, we know that markets do not know that markets do not know that markets do not become informed become informed become informed and efficient on their own and efficient on their own and efficient on their own . We strive . We strive . We strive to be an active to be an active to be an active partner in this partner in this partner in this work. However, it work. However, it work. However, it takes two to tango. takes two to tango. takes two to tango. Uh, ideally, Uh, ideally, Uh, ideally, actually, more than actually, more than actually, more than two. Uh, if you two. Uh, if you two. Uh, if you want to collaborate want to collaborate want to collaborate with us, contribute, with us, contribute, with us, contribute, or join the or join the or join the team, we'd team, we'd team, we'd love to hear from you. We love to hear from you. We love to hear from you. We always hire always hire always hire great people. great people. great people. Please contact Please contact Please contact us. If this resonates with you, us. If this resonates with you, us. If this resonates with you, please please please email email email [email protected] regarding [email protected] regarding partnership opportunities. And if partnership opportunities. And if you are interested in you are interested in you are interested in potentially potentially potentially joining the joining the joining the team, my team, my team, my email address is email address is email address is [email protected], k e n n o n. [email protected], k e n n o n. [email protected], k e n n o n. Thank you all for your attention and Thank you all for your attention and Thank you all for your attention and have a nice day.