OpenAI, NVIDIA And Anthropic Just Split. Here's How I'd Spend $20, $60 Or $200.

Open AI just showed us habanero, its Open AI just showed us habanero, its first chip for running AI models. And by first chip for running AI models. And by first chip for running AI models. And by the end of this video, you're going to the end of this video, you're going to the end of this video, you're going to understand why that chip, the fight over understand why that chip, the fight over understand why that chip, the fight over cursor, and Jensen Huang's response from cursor, and Jensen Huang's response from cursor, and Jensen Huang's response from Nvidia reveal three separate camps all Nvidia reveal three separate camps all Nvidia reveal three separate camps all fighting over you and the future of AI. fighting over you and the future of AI. fighting over you and the future of AI. And you're also going to know how I And you're also going to know how I And you're also going to know how I would personally navigate all of that to would personally navigate all of that to would personally navigate all of that to spend my 20, my 60, or my $200 a month spend my 20, my 60, or my $200 a month spend my 20, my 60, or my $200 a month so that no one provider controls my so that no one provider controls my so that no one provider controls my memory or my files. And actually, memory or my files. And actually, memory or my files. And actually, Anthropic's approach is the one I would Anthropic's approach is the one I would Anthropic's approach is the one I would copy as an individual. Look, I can't copy as an individual. Look, I can't copy as an individual. Look, I can't design a chip, I can't call Jensen to design a chip, I can't call Jensen to design a chip, I can't call Jensen to finance a data center, but I can stop finance a data center, but I can stop finance a data center, but I can stop one AI company from holding the only one AI company from holding the only one AI company from holding the only copy of my memory and my files and my copy of my memory and my files and my copy of my memory and my files and my instructions and my work. I talked about instructions and my work. I talked about instructions and my work. I talked about that in the Apple video on Monday, 3 that in the Apple video on Monday, 3 that in the Apple video on Monday, 3 days after the habanero news. Open AI days after the habanero news. Open AI days after the habanero news. Open AI said it would stop giving future models said it would stop giving future models said it would stop giving future models to cursor after SpaceX bought cursor. to cursor after SpaceX bought cursor. to cursor after SpaceX bought cursor. Then, we're not done. Jensen used Then, we're not done. Jensen used Then, we're not done. Jensen used Nvidia's earnings call to explain why Nvidia's earnings call to explain why Nvidia's earnings call to explain why Open AI will still need Nvidia for a Open AI will still need Nvidia for a Open AI will still need Nvidia for a very long time to come. These are not very long time to come. These are not very long time to come. These are not separate stories, guys. Together, they separate stories, guys. Together, they separate stories, guys. Together, they show AI splitting into three kind of show AI splitting into three kind of show AI splitting into three kind of loose camps or alliances. Not announced loose camps or alliances. Not announced loose camps or alliances. Not announced alliances to be clear, but read between alliances to be clear, but read between alliances to be clear, but read between the lines. First, Open AI wants to own the lines. First, Open AI wants to own the lines. First, Open AI wants to own more of your system overall. Nvidia more of your system overall. Nvidia more of your system overall. Nvidia wants to sell something to everybody.

wants to sell something to everybody. wants to sell something to everybody. And then the third is Anthropic wanting And then the third is Anthropic wanting And then the third is Anthropic wanting enough suppliers and partners that it enough suppliers and partners that it enough suppliers and partners that it can switch whenever it needs to and get can switch whenever it needs to and get can switch whenever it needs to and get through its fundamental compute through its fundamental compute through its fundamental compute constraint. By the way, if you're new constraint. By the way, if you're new constraint. By the way, if you're new here, I'm Nate B Jones. I've spent the here, I'm Nate B Jones. I've spent the here, I'm Nate B Jones. I've spent the last 20 years in tech and the last few last 20 years in tech and the last few last 20 years in tech and the last few helping leaders use AI inside real helping leaders use AI inside real helping leaders use AI inside real businesses. Habanero matters because it businesses. Habanero matters because it businesses. Habanero matters because it shows how far Open AI will go to control shows how far Open AI will go to control shows how far Open AI will go to control costs and to vertically integrate. But costs and to vertically integrate. But costs and to vertically integrate. But the larger question here is where you the larger question here is where you the larger question here is where you would put your money when the companies would put your money when the companies would put your money when the companies selling AI are starting to pick sides. selling AI are starting to pick sides. selling AI are starting to pick sides. So, if we start with habanero, that will So, if we start with habanero, that will So, if we start with habanero, that will help us understand Open AI's plan, and help us understand Open AI's plan, and help us understand Open AI's plan, and then we can understand the other two then we can understand the other two then we can understand the other two alliances or loose groupings, and then alliances or loose groupings, and then alliances or loose groupings, and then we can figure out where to spend. OpenAI we can figure out where to spend. OpenAI we can figure out where to spend. OpenAI says "Jalapeno beat Nvidia GB200 and says "Jalapeno beat Nvidia GB200 and says "Jalapeno beat Nvidia GB200 and GB300 systems on latency and throughput GB300 systems on latency and throughput GB300 systems on latency and throughput per kilowatt across three different open per kilowatt across three different open per kilowatt across three different open weight model tests." The chip went from weight model tests." The chip went from weight model tests." The chip went from a blank page to being taped out, which a blank page to being taped out, which a blank page to being taped out, which means fully designed, in just 9 months, means fully designed, in just 9 months, means fully designed, in just 9 months, which is super fast for a chip. For which is super fast for a chip. For which is super fast for a chip. For selected parts of that work of design, selected parts of that work of design, selected parts of that work of design, code generated with OpenAI's own models code generated with OpenAI's own models code generated with OpenAI's own models ran 1 and 1/2 to 1.8 times faster than ran 1 and 1/2 to 1.8 times faster than ran 1 and 1/2 to 1.8 times faster than versions written by human experts. And versions written by human experts. And versions written by human experts. And this is not the only time we've seen this is not the only time we've seen this is not the only time we've seen this. Deep Seek has also optimized their this. Deep Seek has also optimized their this. Deep Seek has also optimized their kernels with AI models, and sometimes kernels with AI models, and sometimes kernels with AI models, and sometimes you don't know how it works, it just you don't know how it works, it just you don't know how it works, it just works. But there are important limits works. But there are important limits works. But there are important limits here. Jalapeno is not the answer to here. Jalapeno is not the answer to here. Jalapeno is not the answer to everything. Jalapeno does run models, everything. Jalapeno does run models, everything. Jalapeno does run models, it's an inference chip, it doesn't it's an inference chip, it doesn't it's an inference chip, it doesn't replace the giant Nvidia systems used to replace the giant Nvidia systems used to replace the giant Nvidia systems used to train models. OpenAI compared it with train models. OpenAI compared it with train models. OpenAI compared it with GB200 and GB300 and with Nvidia's newer GB200 and GB300 and with Nvidia's newer GB200 and GB300 and with Nvidia's newer Vera Rubin generation. Now, the best

Vera Rubin generation. Now, the best Vera Rubin generation. Now, the best AI-written code results cover only AI-written code results cover only AI-written code results cover only selected parts of a model, not every selected parts of a model, not every selected parts of a model, not every single workload. OpenAI is also still a single workload. OpenAI is also still a single workload. OpenAI is also still a huge Nvidia customer, and Nvidia says huge Nvidia customer, and Nvidia says huge Nvidia customer, and Nvidia says OpenAI has roughly 12 gigawatts of OpenAI has roughly 12 gigawatts of OpenAI has roughly 12 gigawatts of Nvidia systems planned or committed Nvidia systems planned or committed Nvidia systems planned or committed through 2030. That sounds plausible. So, through 2030. That sounds plausible. So, through 2030. That sounds plausible. So, to cut a long story short here, let's to cut a long story short here, let's to cut a long story short here, let's look at it. OpenAI is still buying look at it. OpenAI is still buying look at it. OpenAI is still buying enormous amounts of Nvidia hardware enormous amounts of Nvidia hardware enormous amounts of Nvidia hardware while also using Jalapeno to lower the while also using Jalapeno to lower the while also using Jalapeno to lower the cost of work chat GPT and Codex repeat cost of work chat GPT and Codex repeat cost of work chat GPT and Codex repeat billions of times as they serve all of billions of times as they serve all of billions of times as they serve all of us AI models. The hard part of a custom us AI models. The hard part of a custom us AI models. The hard part of a custom chip is not only the design part, it's chip is not only the design part, it's chip is not only the design part, it's also that you need software that tells also that you need software that tells also that you need software that tells the chip exactly how to run each part of the chip exactly how to run each part of the chip exactly how to run each part of a model. Now, Nvidia has spent decades a model. Now, Nvidia has spent decades a model. Now, Nvidia has spent decades making that easier through CUDA. A making that easier through CUDA. A making that easier through CUDA. A company can build a faster chip on paper company can build a faster chip on paper company can build a faster chip on paper and still lose because its engineers and still lose because its engineers and still lose because its engineers cannot program that chip quickly enough. cannot program that chip quickly enough. cannot program that chip quickly enough. OpenAI is using its own coding models to OpenAI is using its own coding models to OpenAI is using its own coding models to attack exactly that problem in Nvidia's attack exactly that problem in Nvidia's attack exactly that problem in Nvidia's traditional strong suit. It says Codex traditional strong suit. It says Codex traditional strong suit. It says Codex and GPT Astra took three different model and GPT Astra took three different model and GPT Astra took three different model families that were not in the original families that were not in the original families that were not in the original chip plan and got them running well in chip plan and got them running well in chip plan and got them running well in just two months on the jalapeno chip just two months on the jalapeno chip just two months on the jalapeno chip design. Open AI can test the code on design. Open AI can test the code on design. Open AI can test the code on jalapeno, see what worked and have the jalapeno, see what worked and have the jalapeno, see what worked and have the AI try many more versions faster than a AI try many more versions faster than a AI try many more versions faster than a human team could test by hand. Open AI human team could test by hand. Open AI human team could test by hand. Open AI wants to own more compute and own the wants to own more compute and own the wants to own more compute and own the whole inference stack so they can whole inference stack so they can whole inference stack so they can affordably serve you a lot of AI. Chat affordably serve you a lot of AI. Chat affordably serve you a lot of AI. Chat GPT and Codex show where things are GPT and Codex show where things are GPT and Codex show where things are headed with customers and Open AI is headed with customers and Open AI is headed with customers and Open AI is going to keep building out the back end

going to keep building out the back end going to keep building out the back end of the stack to serve customers better. of the stack to serve customers better. of the stack to serve customers better. So, they're going to plan on owning more So, they're going to plan on owning more So, they're going to plan on owning more of the model stack, owning more of the of the model stack, owning more of the of the model stack, owning more of the software that runs them and jalapeno software that runs them and jalapeno software that runs them and jalapeno gives them a chip to build on for some gives them a chip to build on for some gives them a chip to build on for some of that inference work. The company of that inference work. The company of that inference work. The company obviously is investing in data centers, obviously is investing in data centers, obviously is investing in data centers, buying power, developing devices. buying power, developing devices. buying power, developing devices. They're going all in on owning the full They're going all in on owning the full They're going all in on owning the full ecosystem that allows them to deliver AI ecosystem that allows them to deliver AI ecosystem that allows them to deliver AI to customers. That creates one single to customers. That creates one single to customers. That creates one single loop that Open AI wants to own as a loop that Open AI wants to own as a loop that Open AI wants to own as a whole. Chat GPT shows Open AI which jobs whole. Chat GPT shows Open AI which jobs whole. Chat GPT shows Open AI which jobs cost the most so the model team can cost the most so the model team can cost the most so the model team can change the model. Then the team running change the model. Then the team running change the model. Then the team running Chat GPT can change how requests reach Chat GPT can change how requests reach Chat GPT can change how requests reach the chips. Codex can improve the chip's the chips. Codex can improve the chip's the chips. Codex can improve the chip's software. The chip team can use all of software. The chip team can use all of software. The chip team can use all of that to design the next version of that to design the next version of that to design the next version of jalapeno. So, Open AI doesn't need to jalapeno. So, Open AI doesn't need to jalapeno. So, Open AI doesn't need to own every single chip it runs to make own every single chip it runs to make own every single chip it runs to make that loop work. It only needs to own that loop work. It only needs to own that loop work. It only needs to own enough of the expensive repeated enough of the expensive repeated enough of the expensive repeated inference work to change the unit inference work to change the unit inference work to change the unit economics of the costs. Now, let's move economics of the costs. Now, let's move economics of the costs. Now, let's move to Open AI and Cursor. The Open AI to Open AI and Cursor. The Open AI to Open AI and Cursor. The Open AI versus Cursor decision shows that same versus Cursor decision shows that same versus Cursor decision shows that same plan at a product level. Cursor was plan at a product level. Cursor was plan at a product level. Cursor was obviously an independent coding tool for obviously an independent coding tool for obviously an independent coding tool for a while and then SpaceX bought it. So, a while and then SpaceX bought it. So, a while and then SpaceX bought it. So, Cursor now calls Grok and Composer as Cursor now calls Grok and Composer as Cursor now calls Grok and Composer as its own models and SpaceX AI owns one of its own models and SpaceX AI owns one of its own models and SpaceX AI owns one of the main places where developers choose the main places where developers choose the main places where developers choose which coding model to use and that is a which coding model to use and that is a which coding model to use and that is a tremendous advantage to the XAI team tremendous advantage to the XAI team tremendous advantage to the XAI team long term. It's one of the reasons Elon long term. It's one of the reasons Elon long term. It's one of the reasons Elon bought Cursor. OpenAI says it will stop bought Cursor. OpenAI says it will stop bought Cursor. OpenAI says it will stop giving future models to Cursor on the giving future models to Cursor on the giving future models to Cursor on the 12th of November because the new 12th of November because the new 12th of November because the new ownership created trust and contractual ownership created trust and contractual ownership created trust and contractual problems. Now, the larger lesson here is

problems. Now, the larger lesson here is problems. Now, the larger lesson here is clear. If an AI company can remove a clear. If an AI company can remove a clear. If an AI company can remove a model from the tool where you work when model from the tool where you work when model from the tool where you work when ownership changes and the companies ownership changes and the companies ownership changes and the companies become rivals, you have to ask yourself, become rivals, you have to ask yourself, become rivals, you have to ask yourself, are you trapped somewhere? Are you are you trapped somewhere? Are you are you trapped somewhere? Are you trapped somewhere where they can make a trapped somewhere where they can make a trapped somewhere where they can make a deal and suddenly you don't have the AI deal and suddenly you don't have the AI deal and suddenly you don't have the AI model you're used to. This is not the model you're used to. This is not the model you're used to. This is not the only time this has happened, either. only time this has happened, either. only time this has happened, either. Anthropic famously withdrew their Anthropic famously withdrew their Anthropic famously withdrew their support for Windsurf. This is an example support for Windsurf. This is an example support for Windsurf. This is an example of OpenAI really focusing full loop and of OpenAI really focusing full loop and of OpenAI really focusing full loop and being committed to making sure that being committed to making sure that being committed to making sure that OpenAI intelligence is available across OpenAI intelligence is available across OpenAI intelligence is available across all OpenAI surfaces, but they're not all OpenAI surfaces, but they're not all OpenAI surfaces, but they're not going to necessarily make it available going to necessarily make it available going to necessarily make it available in places where they feel like the model in places where they feel like the model in places where they feel like the model could potentially be distilled or their could potentially be distilled or their could potentially be distilled or their valuable training and usage data could valuable training and usage data could valuable training and usage data could somehow be extracted and used by other somehow be extracted and used by other somehow be extracted and used by other companies. Now, look at Nvidia. Jensen's companies. Now, look at Nvidia. Jensen's companies. Now, look at Nvidia. Jensen's answer to Hal Tan was not that custom answer to Hal Tan was not that custom answer to Hal Tan was not that custom chips will fail. His answer is that chips will fail. His answer is that chips will fail. His answer is that Nvidia sells much more than one chip. Nvidia sells much more than one chip. Nvidia sells much more than one chip. Nvidia sells the systems used to train a Nvidia sells the systems used to train a Nvidia sells the systems used to train a model, the systems to improve it, the model, the systems to improve it, the model, the systems to improve it, the systems to connect thousands of systems to connect thousands of systems to connect thousands of computers together in data centers, and computers together in data centers, and computers together in data centers, and to run many different kinds of work. And to run many different kinds of work. And to run many different kinds of work. And it sells those systems through every it sells those systems through every it sells those systems through every single major cloud provider. And that single major cloud provider. And that single major cloud provider. And that makes Nvidia the core of our second makes Nvidia the core of our second makes Nvidia the core of our second camp. Nvidia wants to sell something to camp. Nvidia wants to sell something to camp. Nvidia wants to sell something to everybody.

everybody. everybody. OpenAI can move some repeated work onto OpenAI can move some repeated work onto OpenAI can move some repeated work onto Hal Tan, but remember they still need Hal Tan, but remember they still need Hal Tan, but remember they still need Nvidia for training, they'll need them Nvidia for training, they'll need them Nvidia for training, they'll need them for unusual jobs, they're not moving off for unusual jobs, they're not moving off for unusual jobs, they're not moving off Nvidia anytime soon, nor are they Nvidia anytime soon, nor are they Nvidia anytime soon, nor are they planning to. Google can obviously sell planning to. Google can obviously sell planning to. Google can obviously sell its own TPUs while offering Nvidia its own TPUs while offering Nvidia its own TPUs while offering Nvidia systems through Google Cloud. Again, systems through Google Cloud. Again, systems through Google Cloud. Again, they can't move off Nvidia, either. They they can't move off Nvidia, either. They they can't move off Nvidia, either. They don't have enough TPUs. Anthropic can don't have enough TPUs. Anthropic can don't have enough TPUs. Anthropic can buy Amazon and Google chips while also buy Amazon and Google chips while also buy Amazon and Google chips while also using Nvidia capacity through Microsoft using Nvidia capacity through Microsoft using Nvidia capacity through Microsoft and SpaceX. It all ends with Jensen. And and SpaceX. It all ends with Jensen. And and SpaceX. It all ends with Jensen. And this is why Jensen can invest in this is why Jensen can invest in this is why Jensen can invest in companies that want to reduce their companies that want to reduce their companies that want to reduce their Nvidia bill. He is betting specifically Nvidia bill. He is betting specifically Nvidia bill. He is betting specifically that the whole AI market will grow so that the whole AI market will grow so that the whole AI market will grow so fast that custom chips can take pieces fast that custom chips can take pieces fast that custom chips can take pieces of it and Nvidia is still an overall of it and Nvidia is still an overall of it and Nvidia is still an overall winner. Open AI can move one kind of winner. Open AI can move one kind of winner. Open AI can move one kind of work away from Nvidia while buying more work away from Nvidia while buying more work away from Nvidia while buying more Nvidia systems in total as AI grows Nvidia systems in total as AI grows Nvidia systems in total as AI grows overall. Nvidia also benefits when the overall. Nvidia also benefits when the overall. Nvidia also benefits when the work changes. A custom chip may be work changes. A custom chip may be work changes. A custom chip may be excellent at a given model family, but excellent at a given model family, but excellent at a given model family, but then a new model arrives, a different then a new model arrives, a different then a new model arrives, a different workload, or maybe the company needs to workload, or maybe the company needs to workload, or maybe the company needs to train instead of running the model. train instead of running the model. train instead of running the model. Nvidia sells the general system that Nvidia sells the general system that Nvidia sells the general system that handles jobs that nobody planned for 6 handles jobs that nobody planned for 6 handles jobs that nobody planned for 6 months ago, and they're basically months ago, and they're basically months ago, and they're basically betting they can stay that adaptable.

betting they can stay that adaptable. betting they can stay that adaptable. Jensen doesn't need to win every single Jensen doesn't need to win every single Jensen doesn't need to win every single individual battle in this war as long as individual battle in this war as long as individual battle in this war as long as all three camps need Nvidia somewhere. all three camps need Nvidia somewhere. all three camps need Nvidia somewhere. The third camp is less tidy. The third camp is less tidy. The third camp is less tidy. Anthropic sits somewhere in the middle Anthropic sits somewhere in the middle Anthropic sits somewhere in the middle of all this, but no single company owns of all this, but no single company owns of all this, but no single company owns the whole thing. Anthropic uses Amazon's the whole thing. Anthropic uses Amazon's the whole thing. Anthropic uses Amazon's Trainium chips. It has a multi-gigawatt Trainium chips. It has a multi-gigawatt Trainium chips. It has a multi-gigawatt agreement for Google TPUs built with agreement for Google TPUs built with agreement for Google TPUs built with Broadcom. Microsoft provides Nvidia Broadcom. Microsoft provides Nvidia Broadcom. Microsoft provides Nvidia capacity that Anthropic uses. Anthropic capacity that Anthropic uses. Anthropic capacity that Anthropic uses. Anthropic also now says it's using all of SpaceX's also now says it's using all of SpaceX's also now says it's using all of SpaceX's Colossus 1 data center with more than Colossus 1 data center with more than Colossus 1 data center with more than 220,000 Nvidia GPUs for Claude. And 220,000 Nvidia GPUs for Claude. And 220,000 Nvidia GPUs for Claude. And notably, Claude is not getting pulled notably, Claude is not getting pulled notably, Claude is not getting pulled from Cursor, which is also owned by the from Cursor, which is also owned by the from Cursor, which is also owned by the same company. So, that mix gives same company. So, that mix gives same company. So, that mix gives Anthropic a lot of strategic choices. Anthropic a lot of strategic choices. Anthropic a lot of strategic choices. Different work can run on different Different work can run on different Different work can run on different systems. Suppliers have to compete on systems. Suppliers have to compete on systems. Suppliers have to compete on price and capacity to earn Anthropic's price and capacity to earn Anthropic's price and capacity to earn Anthropic's business. And if one company can't business. And if one company can't business. And if one company can't provide enough chips for them, Anthropic provide enough chips for them, Anthropic provide enough chips for them, Anthropic is able to move some work somewhere is able to move some work somewhere is able to move some work somewhere else. They've done a lot of work to else. They've done a lot of work to else. They've done a lot of work to their credit on the model to get it to their credit on the model to get it to their credit on the model to get it to run on those different chips. Anthropic run on those different chips. Anthropic run on those different chips. Anthropic gives up some of the close coordination gives up some of the close coordination gives up some of the close coordination Open AI is getting from owning more of Open AI is getting from owning more of Open AI is getting from owning more of the stack, but it's less dependent on the stack, but it's less dependent on the stack, but it's less dependent on any given supplier. Now, throw in any given supplier. Now, throw in any given supplier. Now, throw in Cursor. Cursor adds Anthropic Cursor. Cursor adds Anthropic Cursor. Cursor adds Anthropic distribution into that overall camp.

distribution into that overall camp. distribution into that overall camp. SpaceX AI can put Grok and Composer in SpaceX AI can put Grok and Composer in SpaceX AI can put Grok and Composer in front of developers while Cursor still front of developers while Cursor still front of developers while Cursor still sells access to Claude and actually also sells access to Claude and actually also sells access to Claude and actually also to Gemini. So, Anthropic and Google get to Gemini. So, Anthropic and Google get to Gemini. So, Anthropic and Google get customers inside a tool owned by a customers inside a tool owned by a customers inside a tool owned by a company with a competing model, but company with a competing model, but company with a competing model, but which Anthropic is also buying capacity which Anthropic is also buying capacity which Anthropic is also buying capacity from. We're walking away because of the from. We're walking away because of the from. We're walking away because of the risk of usage data getting loose. Now, risk of usage data getting loose. Now, risk of usage data getting loose. Now, these are loose camps, right? They're these are loose camps, right? They're these are loose camps, right? They're not formal teams. Google competes with not formal teams. Google competes with not formal teams. Google competes with Nvidia in chips and sells Nvidia systems Nvidia in chips and sells Nvidia systems Nvidia in chips and sells Nvidia systems in the cloud. Microsoft works closely in the cloud. Microsoft works closely in the cloud. Microsoft works closely with OpenAI and also helps Anthropic get with OpenAI and also helps Anthropic get with OpenAI and also helps Anthropic get Nvidia capacity. And SpaceXAI competes Nvidia capacity. And SpaceXAI competes Nvidia capacity. And SpaceXAI competes with Anthropic while Cursor sells with Anthropic while Cursor sells with Anthropic while Cursor sells Claude. So, the same companies are Claude. So, the same companies are Claude. So, the same companies are competing in one place and working competing in one place and working competing in one place and working together in another. This reminds me of together in another. This reminds me of together in another. This reminds me of when I was at Amazon at Prime Video and when I was at Amazon at Prime Video and when I was at Amazon at Prime Video and we knew that Netflix was running on AWS. we knew that Netflix was running on AWS. we knew that Netflix was running on AWS. But within that context, their methods But within that context, their methods But within that context, their methods are different enough that we ought to are different enough that we ought to are different enough that we ought to pay attention and I think we should pay attention and I think we should pay attention and I think we should consider them as separate user consider them as separate user consider them as separate user groupings. OpenAI really does want to groupings. OpenAI really does want to groupings. OpenAI really does want to own the full loop. Nvidia wants to make own the full loop. Nvidia wants to make own the full loop. Nvidia wants to make sure the ecosystem is open enough to sure the ecosystem is open enough to sure the ecosystem is open enough to sell to everyone and Anthropic wants sell to everyone and Anthropic wants sell to everyone and Anthropic wants several ways to make sure that they have several ways to make sure that they have several ways to make sure that they have compute available to meet especially compute available to meet especially compute available to meet especially business users' needs. And the Cursor business users' needs. And the Cursor business users' needs. And the Cursor fight makes that risk really easy for me fight makes that risk really easy for me fight makes that risk really easy for me to see, right? Imagine you spend a year to see, right? Imagine you spend a year to see, right? Imagine you spend a year inside an AI app, your projects, your inside an AI app, your projects, your inside an AI app, your projects, your chat history, your rules, your saved chat history, your rules, your saved chat history, your rules, your saved memories, they all live there. And then memories, they all live there. And then memories, they all live there. And then another company buys that up or two another company buys that up or two another company buys that up or two suppliers fight and suddenly the model suppliers fight and suddenly the model suppliers fight and suddenly the model you rely on just vanishes from that app.

you rely on just vanishes from that app. you rely on just vanishes from that app. You can open another app, but can you You can open another app, but can you You can open another app, but can you continue the job without rebuilding continue the job without rebuilding continue the job without rebuilding everything in a new app? Like does the everything in a new app? Like does the everything in a new app? Like does the intelligence change how the compute intelligence change how the compute intelligence change how the compute works for you? It does. And that's why I works for you? It does. And that's why I works for you? It does. And that's why I use Open Brand. I keep the important use Open Brand. I keep the important use Open Brand. I keep the important memory systems in a system that I memory systems in a system that I memory systems in a system that I control and then I let ChatGPT or Claude control and then I let ChatGPT or Claude control and then I let ChatGPT or Claude or Gemini or Cursor or any model on my or Gemini or Cursor or any model on my or Gemini or Cursor or any model on my own computer read and manage it. My own computer read and manage it. My own computer read and manage it. My documents can stay in normal files, my documents can stay in normal files, my documents can stay in normal files, my code stays in repos that I control, and code stays in repos that I control, and code stays in repos that I control, and my important processes and skills stay my important processes and skills stay my important processes and skills stay in instructions that I can move. If you in instructions that I can move. If you in instructions that I can move. If you want to make sure you're not beholden to want to make sure you're not beholden to want to make sure you're not beholden to different models, a service like different models, a service like different models, a service like Openrouter, which was bought by Stripe Openrouter, which was bought by Stripe Openrouter, which was bought by Stripe recently, can help send work to a bunch recently, can help send work to a bunch recently, can help send work to a bunch of different models depending on what of different models depending on what of different models depending on what you need. Even with Openrouter, you you need. Even with Openrouter, you you need. Even with Openrouter, you don't want Openrouter to manage your don't want Openrouter to manage your don't want Openrouter to manage your entire computing experience. What would entire computing experience. What would entire computing experience. What would I actually buy today? I actually buy today? I actually buy today? At around 20 bucks a month, I would pay At around 20 bucks a month, I would pay At around 20 bucks a month, I would pay for one main provider based on the work for one main provider based on the work for one main provider based on the work I do every single week. Now, I don't I do every single week. Now, I don't I do every single week. Now, I don't actually pay for 20, so I will get to actually pay for 20, so I will get to actually pay for 20, so I will get to what I pay for, but that's where I would what I pay for, but that's where I would what I pay for, but that's where I would start. I would keep a free account with start. I would keep a free account with start. I would keep a free account with at least one serious rival, and I would at least one serious rival, and I would at least one serious rival, and I would use it often enough to know how it use it often enough to know how it use it often enough to know how it works. I would not send every prompt to works. I would not send every prompt to works. I would not send every prompt to five different models looking for a tiny five different models looking for a tiny five different models looking for a tiny improvement at that spend level. One improvement at that spend level. One improvement at that spend level. One company would get most of my daily work, company would get most of my daily work, company would get most of my daily work, and my memory and files would stay and my memory and files would stay and my memory and files would stay outside it because it's really trivial outside it because it's really trivial outside it because it's really trivial at this point to set up something like at this point to set up something like at this point to set up something like OpenBrain. Like, I think I did it again, OpenBrain. Like, I think I did it again, OpenBrain. Like, I think I did it again, like I did a fresh install, it was like like I did a fresh install, it was like like I did a fresh install, it was like 5 minutes. It was very easy. If I had a 5 minutes. It was very easy. If I had a 5 minutes. It was very easy. If I had a $60 budget, my general setup would be $60 budget, my general setup would be $60 budget, my general setup would be direct access to OpenAI at a $20 level, direct access to OpenAI at a $20 level, direct access to OpenAI at a $20 level, direct access to Claude at a $20 level,

direct access to Claude at a $20 level, direct access to Claude at a $20 level, and Cursor Pro if I code every day. and Cursor Pro if I code every day. and Cursor Pro if I code every day. OpenAI would handle ChatGPT, Codex, OpenAI would handle ChatGPT, Codex, OpenAI would handle ChatGPT, Codex, research, agents, and connected tools. research, agents, and connected tools. research, agents, and connected tools. With Claude, I'm going to handle a lot With Claude, I'm going to handle a lot With Claude, I'm going to handle a lot of the documents and the writing with of the documents and the writing with of the documents and the writing with the right Claude model. I probably the right Claude model. I probably the right Claude model. I probably wouldn't use Opus 5 right now. Uh and I wouldn't use Opus 5 right now. Uh and I wouldn't use Opus 5 right now. Uh and I would also get a second answer on would also get a second answer on would also get a second answer on important decisions. And Cursor would be important decisions. And Cursor would be important decisions. And Cursor would be great cuz I would get multiple model great cuz I would get multiple model great cuz I would get multiple model options inside the same subscription. options inside the same subscription. options inside the same subscription. Now, you can reverse those jobs. If most Now, you can reverse those jobs. If most Now, you can reverse those jobs. If most of your work lives in Gmail, in Docs, in of your work lives in Gmail, in Docs, in of your work lives in Gmail, in Docs, in Drive, in Google Workspace, maybe Google Drive, in Google Workspace, maybe Google Drive, in Google Workspace, maybe Google AI Pro can replace one of the direct AI Pro can replace one of the direct AI Pro can replace one of the direct plans for you. Replace, not plans for you. Replace, not plans for you. Replace, not automatically add, right? Like, I'm automatically add, right? Like, I'm automatically add, right? Like, I'm trying to think within a budget here cuz trying to think within a budget here cuz trying to think within a budget here cuz all of us have budgets. I think the all of us have budgets. I think the all of us have budgets. I think the principle is really simple. Your main AI principle is really simple. Your main AI principle is really simple. Your main AI paid plan should position you to get the paid plan should position you to get the paid plan should position you to get the most out of your week. If it doesn't, most out of your week. If it doesn't, most out of your week. If it doesn't, what are you doing? Now, if you're going what are you doing? Now, if you're going what are you doing? Now, if you're going to spend over $200, which spoiler alert, to spend over $200, which spoiler alert, to spend over $200, which spoiler alert, that is what I do, then my bar changes. that is what I do, then my bar changes. that is what I do, then my bar changes. As a serious user, I pay for more than As a serious user, I pay for more than As a serious user, I pay for more than one $200 a month frontier plan, and I do one $200 a month frontier plan, and I do one $200 a month frontier plan, and I do it expecting savings back from every it expecting savings back from every it expecting savings back from every single one individually. I push them on single one individually. I push them on single one individually. I push them on that. So, I pay that. So, I pay that. So, I pay uh the $200 plan for Anthropic, I pay it uh the $200 plan for Anthropic, I pay it uh the $200 plan for Anthropic, I pay it for Codex, and I pay it for Grok. And in for Codex, and I pay it for Grok. And in for Codex, and I pay it for Grok. And in each case, my bar to them is the same.

each case, my bar to them is the same. each case, my bar to them is the same. You will find a way to pay for this for You will find a way to pay for this for You will find a way to pay for this for me. You will find a way through me. You will find a way through me. You will find a way through subscriptions, you will find a way subscriptions, you will find a way subscriptions, you will find a way through bug bounties, you will find a through bug bounties, you will find a through bug bounties, you will find a way through saving me time. You will way through saving me time. You will way through saving me time. You will find a way through expanding my ability find a way through expanding my ability find a way through expanding my ability to code so I don't have to spend as much to code so I don't have to spend as much to code so I don't have to spend as much money on coding. Whatever it is, you're money on coding. Whatever it is, you're money on coding. Whatever it is, you're going to find time to give me my money going to find time to give me my money going to find time to give me my money back. And if you don't have that bar, I back. And if you don't have that bar, I back. And if you don't have that bar, I don't think that you are holding a don't think that you are holding a don't think that you are holding a serious enough expectation of AI work to serious enough expectation of AI work to serious enough expectation of AI work to use that $200 a month plan productively. use that $200 a month plan productively. use that $200 a month plan productively. And And I have a very high expectation And And I have a very high expectation And And I have a very high expectation when I put that investment down that I when I put that investment down that I when I put that investment down that I will be using those models every single will be using those models every single will be using those models every single day, that I have defined work for each day, that I have defined work for each day, that I have defined work for each of the models, that I am pushing them of the models, that I am pushing them of the models, that I am pushing them hard. And if I don't, then I'm going to hard. And if I don't, then I'm going to hard. And if I don't, then I'm going to wind back that plan over time. And I wind back that plan over time. And I wind back that plan over time. And I think that's a really reasonable think that's a really reasonable think that's a really reasonable expectation. And so, I fully admit, I am expectation. And so, I fully admit, I am expectation. And so, I fully admit, I am not everybody. I am a rare user in being not everybody. I am a rare user in being not everybody. I am a rare user in being willing to spend that much. But I want willing to spend that much. But I want willing to spend that much. But I want to open it up and tell you this is how I to open it up and tell you this is how I to open it up and tell you this is how I think about it. If it doesn't justify, think about it. If it doesn't justify, think about it. If it doesn't justify, if it doesn't pay for itself, I'm not if it doesn't pay for itself, I'm not if it doesn't pay for itself, I'm not paying for it. And increasingly, these paying for it. And increasingly, these paying for it. And increasingly, these models are good enough to do that. In models are good enough to do that. In models are good enough to do that. In fact, OpenAI is moving to outcome-based fact, OpenAI is moving to outcome-based fact, OpenAI is moving to outcome-based pricing for that reason with select pricing for that reason with select pricing for that reason with select enterprise customers, where the enterprise customers, where the enterprise customers, where the enterprise customer only pays when an enterprise customer only pays when an enterprise customer only pays when an outcome is achieved. That's basically outcome is achieved. That's basically outcome is achieved. That's basically what I am challenging these models on what I am challenging these models on what I am challenging these models on when I pay $200 a month in plans. I am when I pay $200 a month in plans. I am when I pay $200 a month in plans. I am looking specifically to make sure that I looking specifically to make sure that I looking specifically to make sure that I can invest in high-value frontier models can invest in high-value frontier models can invest in high-value frontier models across multiple places in the ecosystem, across multiple places in the ecosystem, across multiple places in the ecosystem, that I learn to not be dependent on any that I learn to not be dependent on any that I learn to not be dependent on any one model, that I keep my memory

one model, that I keep my memory one model, that I keep my memory separate, and that I insist on each of separate, and that I insist on each of separate, and that I insist on each of them individually, earning their place. them individually, earning their place. them individually, earning their place. Regardless of how much you're spending, Regardless of how much you're spending, Regardless of how much you're spending, I think you should be asking yourself I think you should be asking yourself I think you should be asking yourself this. Whether you're spending 20 or 60 this. Whether you're spending 20 or 60 this. Whether you're spending 20 or 60 or 200 or more. Look, I can't design a or 200 or more. Look, I can't design a or 200 or more. Look, I can't design a chip, I can't call Jensen to finance a chip, I can't call Jensen to finance a chip, I can't call Jensen to finance a data center, but I can stop one AI data center, but I can stop one AI data center, but I can stop one AI company from holding the only copy of my company from holding the only copy of my company from holding the only copy of my memory and my files and my instructions memory and my files and my instructions memory and my files and my instructions and my work. If your main model and my work. If your main model and my work. If your main model disappeared tomorrow, would you miss it? disappeared tomorrow, would you miss it? disappeared tomorrow, would you miss it? Would you feel it? Would the switch Would you feel it? Would the switch Would you feel it? Would the switch hurt? hurt? hurt? Because your goal actually is that you Because your goal actually is that you Because your goal actually is that you would find a way to continue working would find a way to continue working would find a way to continue working with the rest of the intelligence that's with the rest of the intelligence that's with the rest of the intelligence that's available to you relatively smoothly available to you relatively smoothly available to you relatively smoothly because your memory, your skills, and because your memory, your skills, and because your memory, your skills, and all of that can be picked up by the new all of that can be picked up by the new all of that can be picked up by the new model relatively quickly. It may not be model relatively quickly. It may not be model relatively quickly. It may not be perfect. I'm not saying that it would be perfect. I'm not saying that it would be perfect. I'm not saying that it would be seamless. When I switch between Codex seamless. When I switch between Codex seamless. When I switch between Codex and Claude, I do feel that even though I and Claude, I do feel that even though I and Claude, I do feel that even though I have my memory all shared and everything have my memory all shared and everything have my memory all shared and everything else. But, it's way more seamless else. But, it's way more seamless else. But, it's way more seamless because of the memory piece. Let's because of the memory piece. Let's because of the memory piece. Let's ladder out where we were. We've talked ladder out where we were. We've talked ladder out where we were. We've talked about the three big camps in AI, OpenAI, about the three big camps in AI, OpenAI, about the three big camps in AI, OpenAI, Nvidia, and Anthropic. We've talked Nvidia, and Anthropic. We've talked Nvidia, and Anthropic. We've talked about them making three different bets.

about them making three different bets. about them making three different bets. We've talked about how the Anthropic bet We've talked about how the Anthropic bet We've talked about how the Anthropic bet is kind of analogous to us, to our is kind of analogous to us, to our is kind of analogous to us, to our option as consumers. And I've told you option as consumers. And I've told you option as consumers. And I've told you this is how I think about sort of this is how I think about sort of this is how I think about sort of picking tools and how I pick them and picking tools and how I pick them and picking tools and how I pick them and why I pick them. I would like to know why I pick them. I would like to know why I pick them. I would like to know from you what are you willing to spend from you what are you willing to spend from you what are you willing to spend on AI? What is your bar for making money on AI? What is your bar for making money on AI? What is your bar for making money back? Which model are you picking and back? Which model are you picking and back? Which model are you picking and why right now? What is earning any kind why right now? What is earning any kind why right now? What is earning any kind of value for you that you're like, I of value for you that you're like, I of value for you that you're like, I will pay for this? And are you at 20, will pay for this? And are you at 20, will pay for this? And are you at 20, are you at 60, or 200? are you at 60, or 200? are you at 60, or 200? Also, finally, are there models that Also, finally, are there models that Also, finally, are there models that have lost your trust? We haven't talked have lost your trust? We haven't talked have lost your trust? We haven't talked about that in this video, but I'd be about that in this video, but I'd be about that in this video, but I'd be really curious to hear. Are there models really curious to hear. Are there models really curious to hear. Are there models where you have said, "Nope, I'm done. I where you have said, "Nope, I'm done. I where you have said, "Nope, I'm done. I don't want to hear from this model don't want to hear from this model don't want to hear from this model again. I don't want to hear from this again. I don't want to hear from this again. I don't want to hear from this company again. I'm going over here." company again. I'm going over here." company again. I'm going over here." Cuz that's something that I've seen and Cuz that's something that I've seen and Cuz that's something that I've seen and I'd be curious to hear what your opinion I'd be curious to hear what your opinion I'd be curious to hear what your opinion is and where you have lost trust in is and where you have lost trust in is and where you have lost trust in models. Okay. That's what we've got. models. Okay. That's what we've got. models. Okay. That's what we've got. Three camps. Which camp are you in? How Three camps. Which camp are you in? How Three camps. Which camp are you in? How do you think about it? Look, I can't do you think about it? Look, I can't do you think about it? Look, I can't design a chip. I can't call Jensen to design a chip. I can't call Jensen to design a chip. I can't call Jensen to finance a data center, but I can stop finance a data center, but I can stop finance a data center, but I can stop one AI company from holding the only one AI company from holding the only one AI company from holding the only copy of my memory and my files and my copy of my memory and my files and my copy of my memory and my files and my instructions and my work. I'll see you instructions and my work. I'll see you instructions and my work. I'll see you next time. Cheers.