Microsoft and OpenAI break up (Amazon is pumped)

going to open this video a bit differently. I want to thank you for clicking on it because I know this is not the hottest topic for the developer audience. It's rare you guys care basically at all about anything with the word Microsoft in it. And I understand I've had to write code on Windows before myself and I would never wish my worst enemy to have to do the same thing. But the Microsoft OpenAI relationship is way more interesting than you think. And the direction it's going in as well as how OpenAI is going to follow through with this is fascinating. Their relationship started all the way back in 2019 and it is changing meaningfully now. As of a few days ago, they announced a change in their relationship, the next phase, and it's effectively a breakup. We'll go in depth on this relationship, the chaos that has ensued between them, and why OpenAI feels as though Microsoft has been holding them back for years now, and also very interestingly, what the role of this is with Anthropic. Because believe it or not, Anthropic is the reason all of this is happening. The details in this case are even crazier than I initially expected. And the more I dive in, the more absurd it all is. From heated fights between Microsoft and OpenAI executives about sharing details on reasoning models to the absurd nature of the exclusivity deal, the definition of AGI and how that was important for the agreement, and so much more. This one's going to be a bit of a wild ride, and I am so thankful you decided to come on it with us. But in order to pay the cost of doing all this, we got to take a quick break for today's sponsor. First, fun fact. Did you know that 85% of the web can't be accessed via APIs? I know that sounds crazy, but think about it. How many websites require you to sign in, click buttons, navigate, pass captions, and more just to get some basic information out of them. If that's frustrating you or blocking your agents, you really need to check out Browser Base. These guys get it. These guys built the browser that your agents need to do everything from searching to fetching to navigating real websites and doing real actions. You can use this to catch bugs in your own services or do tons of research across the entire internet or unblock your agents that are getting stuck on captions. If you want to see when prices change on some service that you rely on or your competitors are doing things on their websites, they can help you do that automatically. That's why companies like Microsoft, DeepMind, Lovable, Ramp, and more are all relying on Browserbase to run their browsers in the cloud. If you're building with agents, you've already hit a ton of limitations. I bet you'll be surprised how many browserbased can work around. Check them out now at soy.link/browserbase. For any of this to make any sense, we got to go to the start where Microsoft first invested in and partnered with OpenAI to support us building beneficial AGI. Remember way way back, the point of open AI was to try and make sure that when AGI happened, when we had general level intelligence where AI could do anything a human could do. When we hit that point, the goal of OpenAI was to make sure it wasn't just one company that was there and instead everyone could benefit from these advancements in AI. That's why Sam Alman, Greg Brockman, Elon Musk, and others decided to create and fund open AI. They wanted to make it possible for AI to exist outside of the bubble of specifically Google because they were the only company that seemed like they had any chance of doing it. They want to make sure everyone benefited. Obviously, OpenAI is in a very different state than they were when they started the company, and the only open- source thing they've done for a while has been GPTOSS, both 2120. Both pretty solid models, but it's been almost a year since then. Open a guy is a very different company than they were when they started as a nonprofit, but they do let everyone use their models at the very least. Anyways, let's go into how this all started because I think it'll make where they are now very interesting. Microsoft is investing $1 billion into open AI to support us building artificial general intelligence with widely distributed economic benefits. Remember, this is pre-Chat GPT. This is way, way back. We're partnering to develop a hardware and software platform within Microsoft Azure which will scale to AGI. We'll jointly develop a new Azure AI supercomputing technologies and Microsoft will become our exclusive cloud provider. So we'll be working hard together to further extend Microsoft Azure's capabilities in largecale AI systems. I hate that I'm fixating on this mdash here because this was before the models could write. So they actually just liked M dashes. The key piece here is Microsoft will become our exclusive cloud provider. What this means is that if you want to use OpenAI models, you have two places you can go. You can use the official OpenAI platform or you can use Azure because Azure is the only other place the weights will ever be sent. So if you want to use OpenAI models, you have two choices. It is also worth noting that earlier in the same year, they had the breakthrough with GPT2, which was the first breakthrough in text generation with large transformer models, now called LLMs. This is a big deal at the time, 2019, that nobody cared about outside of the research world, because again, it didn't seem like any of this was going particularly interesting places yet. I certainly didn't care yet. It was GPT3 and chat GPT that caused everything to start accelerating. Nobody cared at this point, but Microsoft had an instinct, an intuition that it might matter, so they decided to invest in it. This was the entire OpenAI team at the time. And I bet if we hunt in here, our boy Daario, beautiful. Knew I could find him if I hunted cuz remember Daario was the lead of the GPT3 training. He was the research lead when they trained GPT3. We'll have to talk about that in a bit. Open is producing a sequence of increasingly powerful AI technologies which require a lot of capital for computational power. The most obvious way to cover costs is to build a product, but that would mean changing our focus. Oh, how the tides have turned. Instead, we intend to license some of our preaggi technologies with Microsoft, becoming our preferred partner for commercializing them. So, the plan here was that OpenAI would make these models and make them work as well as they could and focus on research and Microsoft would take all of the things they built up until AGI was reached and they would be able to sell them to their customers in order to fund OpenAI and they would pay upfront a billion dollars. So they had the money they needed and the compute they needed to make new models and make AGI happen. And this piece is incredibly important here. We intend to license some of our pre-agi technologies. This eventually expanded in its breath to all OpenAI IP will be licensed to Microsoft to use how they want until AGI is achieved. So this wasn't a strict timeline. It wasn't we will do this for 5 years. It wasn't we'll do this till we hit a revenue target. It was, "We will share this until we reach the definition of AGI." There's a problem, though. I'm sure many of y'all know it. Now, what the [ __ ] is AGI? What is the definition of AGI? We don't have a very clear one. And this is the problem. They set a bar that was incredibly unclear. And that means there's no way for this deal to end. In 2023, Microsoft doubled down. Not really doubling down, they 10xed down. They put another $10 billion into OpenAI after Chat GPT started doing well and blowing up. They they knew that it was going to happen. So they decided to invest way more at that point. They also confirmed that they had done another quiet private investment in 2021 and that this collaboration would continue to share investment in the supercomputing at scale new AI powered experiences and most importantly exclusivity in the cloud provider. is OpenAI's exclusive cloud provider. Azure will power all OpenAI workloads across research products and API services. This deal was a big deal. This was why OpenAI and Microsoft were so tight together. But then something important happened. In September of 2024, OpenAI introduced 01, a new series of AI models designed to spend more time thinking before they respond. These were the first reasoning models where the model would generate text not to show you but to almost talk to itself to get to a better answer which has since become the way almost every single model works. But at the time it was really novel and back in September of 2024 this was a massive deal. OpenAI01 ranks in the 89th percentile on competitive programming questions. Places among the top 500 students in the US in a qualifier for the USA Math Olympiad. It exceeds human PhD level accuracy on a benchmark of physics, bio, and chemistry problems. Since then, it's become number one, number one, and number one in all of those. But the reason they had this huge leap at the time was that 01 by doing this reasoning was able to be way smarter. When you compare the scores with GBG40, which was their best model at the time, against 01 preview in 01, the final potential public version, the gap is hilarious. This was one of the biggest leaps in intelligence in the history of AI, and it changed everything. And they did a lot to keep the results private. They never shared the reasoning traces or the data. They just took a long time to generate a result instead. And the first company to even come close to catching up to this was Deepseek. But it was clear OpenAI had a massive lead at this point in time. Their models were smarter and better, and they had stumbled on this reasoning thing that made them way more powerful, but also the price was absurd. Oh, it's actually not too bad in retrospect, but it was $15 per 1 mil in and $60 per 1 mil out. But then they put out 01 mini, which was surprisingly cheap. And then they eventually put out 03 mini, which was a god tier model. I loved 03. It was one of the first models I fell like in love with. I really liked using it. But Deep Seek was so impressive it led to me building T3 Chat. The point I'm trying to make here is that OpenAI was really ahead and everybody who had invested heavily in AI up to that point almost was at a bit of a disadvantage because reasoning became the model. So Deepseeku hadn't done as big of investments, was more capable of flexing over to doing reasoning and that's why R1 was so groundbreaking. At the same time though, OpenAI had just gotten this huge investment from Microsoft and part of this was that IP deal where Microsoft would still get that exclusive access to their IP. Doesn't say here specifically, but you get the idea. And this means that very specifically all of the stuff OpenAI discovered around doing reasoning was stuff that they were supposed to share with Microsoft which they proceeded to not do. This article that came out a month after 01 says a lot. Microsoft and OpenAI bromance begins to fray. Apparently Microsoft wouldn't give them more money even though they wanted it after this breakthrough. But at the same time they were not happy. Alman once called OpenAs partnership with Microsoft the best bromance in tech. But the ties have started to fray. Microsoft had acquired a company that OpenAI saw as a competitor and they were very unhappy about that specifically because one of the employees from the company that was acquired was now part of the relationship between OpenAI and Microsoft. Dozens of Microsoft engineers work on site at OpenAI's offices in San Francisco and use laptops provided by OpenAI that are set up to maintain the startup security protocols. And here's where things fell apart. Some OpenAI staff recently complained that Sullman yelled at an OpenAI employee during a recent video call because he thought the startup was not delivering new technology to Microsoft as quickly as it should, according to two people familiar with the call. I can't find the source, but I have heard many times that this was about 01 and reasoning models. Since OpenAI had this massive breakthrough in model intelligence and Microsoft was still trying to train their own models, they felt as though they deserved that IP and any information that they had found as they did this research so they could build their own reasoning models as well. Remember this is still months before DeepSeek R1 dropped. Thank you chat for finding the information source here. This is important. At this point in time, Sman was trying to get more Microsoft stuff to be on Microsoft models instead of OpenAI models, which in retrospect is hilarious. like absolutely hilarious. This one calls out again that Microsoft wanted to know how 01 worked and OpenAI was not providing Microsoft with documentation about how it programmed 01 to think about user queries before answering them. The process known as chain of thought is a key ingredient in the secret recipe of any AI model. To be fair, this did not exist as a concept before. But yeah, raising his voice, Sman told OpenAI employees, including Mera Morati, who was a CTO at the time, the AI startup wasn't holding up its end of the wide-ranging deal that it had with Microsoft. The people familiar with the conversation said, "That's the key piece. This guy who was in charge of AI at Microsoft was pissed that OpenAI was not giving them the info on how 01 worked and how reasoning worked. And this was the start of the breakup. I would without question argue that the start of the end was this moment. And then we got an update in 2025, the next chapter of the OpenAI Microsoft partnership. Since 2019, Microsoft and Open have shared a vision to advance artificial intelligence responsibly and to make the benefits broadly available. yada yada yada. They formed a public benefit corporation and did recapitalization. Following the recapitalization, Microsoft holds an investment in OpenAI group PBC valued at approximately 135 billion representing roughly 27% on an asverted diluted basis. yada yada they own a big percentage. Part of the deal is that with their investment, it could only up to 10x in its value. So this 135 bill is largely because they put in 13.5 bill and 10x was the cap. And once they made it a real public benefit corporation, that converted to about 27% of their existing value. I don't want to go into all the details of the weirdness of that deal. I've covered it many times before, but here are the things that have evolved. Now, once AGI is declared by OpenAI, the declaration will now be verified by an independent expert panel because Microsoft was scared OpenAI would say we hit AGI early in order to get out of their shared IP deal. Microsoft's IP rights for both models and products are extended through 2032 and now include models post AGI with appropriate safety guardrails. Very big deal. Specifically, the 2032 part and now including models post AGI. Microsoft's IP rights to research defined as the confidential methods used in the development of models and systems will remain until either the expert panel verifies AGI or through 2030, whichever is first. Research IP includes, for example, models intended for internal deployment or research only. Beyond that, research IP does not include model architecture, model weights, inference code, fine-tuning code, and any IP related to data center hardware and software. Microsoft retains these non-ressearch IP rights. Microsoft IP rights now exclude OpenAI consumer hardware. OpenAI still doesn't have consumer hardware. They were just planning ahead for that. OpenAI can now jointly develop some products with third parties. API product developed with third parties will be exclusive to Azure. Non-API products may be served on any cloud provider. Very interesting. I would bet that something like Sora when hosting the videos didn't want to deal with the [ __ ] show that is storage on Azure and putting those videos on something like AWS would probably make a lot of sense. Microsoft can now independently pursue AGI alone or in partnership with third parties. Another big deal. Soon after this deal happened, Azure suddenly had support for anthropic models. I wonder why that was held up. Probably because this clause didn't exist. If Microsoft uses open AAIP to develop AGI prior to AGI being declared, the models will be subject to compute thresholds. Those thresholds are significantly larger than the size of systems used to train leading models today. Interesting. The revenue share agreement remains until the expert panel verifies AGI. Though payments will be made over a longer period of time. Again, there was a revenue share agreement where if OpenAI made money, they would have to share some percentage of it to Microsoft. Remember that detail as well. That's important. They contracted to purchase an incremental 2050 billion of Azure services and Microsoft will no longer have a right to first refusal to be OpenAI's compute provider. OpenAI can now provide API access to US government national security customers regardless of the cloud provider. That one becoming relevant is funny. And OpenAI is now able to release openweight models that meet requisite capability criteria. That's why they can do open weight models. A lot of this comes from internal frustration at Microsoft that Sachia was pissed their internal research teams were so far behind what OpenAI was doing. Peter Lee oversaw Microsoft's sprawling research efforts. And he told Nadella that Microsoft's researchers were blown away by the model's ability, the model being GBT4, by the way, to understand conversational language and generate humanlike answers. And they believed it showed sparks of AGI, capabilities on par with those of a human mind. Nadella abruptly cut off Lee mid-sentence, demanding to know how OpenAI managed to surpass the capabilities of the AI project that Microsoft's 1,500 person research team had been working on for decades. Quote, "Open built this with 250 people," Nadella said. "Why do we have Microsoft research at all?" He was pissed. And now we get to 2026, and I want to highlight something simple but funny. Here is the original announcement. Microsoft invested in partners with OpenAI. They have this fancy graphic that I'm pretty sure was built and generated with um what's it called? The Dolly model because they were really proud of that at the time. The next chapter of Microsoft OpenAI relationship has a silly little gradient with the logos here. The next phase of Microsoft open partnership. Nearly the exact same title here. No image, no fanciness. Amended agreement provides long-term clarity. Let's see what just changed. The rapid pace of innovation requires us to continue to evolve our partnership to benefit our customers in both companies. Today, we're announcing an amended agreement to simplify our partnership in the way we work together, grounded in flexibility, certainty, and a focus on delivering the benefits of AI broadly. What this means is that the AGI definition thing is out because no one can agree on what the term means. But also, it seems like Sam is one of the greatest negotiators of all time because he fleeced Microsoft. Point one, Microsoft remains OpenAI's primary cloud provider, not only primary. And OpenAI products will ship first on Azure unless Microsoft cannot and chooses not to support the necessary capabilities. OpenAI can now serve all of its products to customers across any cloud. We will have a lot to say about that momentarily. Second, Microsoft will continue to have licenses to open AIP for models and products through 2032. Microsoft's license will now be non-exclusive. So, if we go back to the previous renegotiation, you'll see the IP rights have been extended to 2032, but the confidential training method parts in the IP right to research would be AGI or 2030. So it seems like the only thing Microsoft got out of this was that the 2030 got extended to 2032 for some parts of the IP sharing. Not even clear what parts. Next piece is that Microsoft will no longer pay a revenue share to OpenAI because Microsoft whenever they served OpenAI models would have to pay some percentage of their revenue to OpenAI. I have so much to say about that part. We'll get there in a sec. We're still in history. And remember the revenue share goes both ways. So Microsoft has to pay openAI when they sell OpenAI models to their customers. But OpenAI also agreed to share their revenue with Microsoft. And they call out here that revenue share payments from OpenAI to Microsoft will continue through 2030 independent of OpenAI's technology progress at the same percentage but subject to a total cap. Microsoft continues to participate directly in OpenAI's growth as a major shareholder. I don't know where I saw this, but I am pretty sure OpenAI also is not doing traditional revenue share to Microsoft anymore. It is now a profit share model. And we all know how profitable Open AI is, which means, as far as I understand, they will no longer be paying jackfucking [ __ ] to Microsoft. I don't know why that's not here. I might be misremembering, but I'm almost positive that the profit share was a redefinition that was done that basically guaranteed OpenAI doesn't have to pay Microsoft [ __ ] I could be wrong on that. Everything else, no. This is like the one thing I might have been wrong about. One more really quick thing on this, though. This image is directly in violation of OpenAI's brand guidelines. This is the old logo. They use a different font now that doesn't have the seraps on it. So the dashes on the I shouldn't be there. But more importantly, they specifically call out when you have a partnership like this with the line between the two logos that you don't use the logo with the emblem. You just use the text logo. Again, they specifically say to not use the blossom with the partnership lockup. Hilarious that they just didn't follow the brand guidelines. So why does this all matter so much? Why did I bring you guys here? And what the hell does Anthropic have to do with it? This is the reason OpenAI and Amazon announced strategic partnership. AWS and OpenAI will co-create a stateful runtime environment powered by OpenAI models available on Amazon Bedrock for AWS customers to build generative AI applications and agents at production scale. The stateful runtime environment part here is particularly interesting. If you watch my video about the websocket changes to how the OpenAI API works, I go in detail on why statefulness is so useful in these environments. And when you combine that with how important caching is as we're trying to reduce compute and get better responses faster, you'll realize that having a semi-stateful environment is really, really, really powerful. And Microsoft is not the right cloud to build this on. Azure would be hellish to try and implement this on top of. So that's one of the big reasons here. They also call out that AWS will be the exclusive third party cloud distribution provider for OpenAI Frontier, which enables organizations to build, deploy, and manage teams of AI agents. OpenAI will consume 2 gawatt of tranium capacity through AWS infra to support demand for this new stful runtime environment as well as Frontier and other advanced workloads. OpenAI and Amazon will develop customized models available to power Amazon's consumerfacing applications and Amazon will invest 50 billion in OpenAI. This is the deal that OpenAI wanted to do that they were not able to do again because of that previous partnership. But there's another much deeper reason why OpenAI wanted to be working so closely with AWS. Just a few weeks ago, a leaked memo revealed OpenAI's open hostility towards one of their biggest investors. This is a leaked memo from internally at OpenAI. Our Microsoft partnership has been foundational to our success, but it's also limited our ability to meet enterprises where they are. For many, that is bedrock. Since we announced the partnership at the end of February, inbound demand from our customers for this offering has been frankly staggering. Just a few weeks ago, Anthropic shared that they hit 30 billion run rate because of enterprise demand. The growth of enterprise deals for Anthropic has been absurd, rivaling, if not surpassing growth for OpenAI. OpenAI does still make more money, but Anthropic was growing their enterprise revenue faster. This might be confusing to you if you've been watching my stuff for a while because it's pretty well established that outside of like front-end tasks, anthropic models are meaningfully worse than OpenAI's models for most code work. So, why the [ __ ] is Anthropic growing faster than OpenAI? Am I just wrong about the quality of the models? No, I was wrong about how much AWS matters, specifically Bedrock. Here's where we need to go down one of the weird tangents. If you're a startup and you apply for the AWS credits as a startup, they will give you up to $100,000 in credit to use on AWS. Google will go even further, giving you $350,000 in cloud credit for using the Google Cloud for your startup. Azure is well known for giving up to $500,000 of credit to startups, especially Y Combinator companies, and they'll actually go a bit further. I was able to squeeze a million dollars of credit out of them in January this year. This may seem like a lot because it is, but we're also spending upwards of $100,000 on inference every month across Google, OpenAI, and anthropic models. So, this burns up pretty quick. And here's where we get to the important detail. Almost every single one of these deals is earmarked. I don't know if this is public or not, but I don't [ __ ] care. This should be public information. If this cost me my credits, I'm going to be very sad and do a lot more sponsors. So, sorry in advance if that happens. A lot of those startup credits are not allowed to be used on anthropic models. That seems a little bit crazy, right? Like I can use my credit with Google to serve all of our Gemini traffic. I can use my credits on Azure to handle all my OpenAI traffic, but I can't use any of my credits anywhere to serve Sonnet or Opus. I cannot use my credits to serve anthropic models because Anthropic has very strict deals with the providers they work with, which are all of them by the way. You can use anthropic models in Google Vertex AI which is their AI provider for GCP stuff. You can use them with AWS Bedrock where a lot of people do and you can now use them with Azure as well. Quad models are available on all three major clouds. OpenAI models were only available on Azure though which was a huge competitive advantage for Azure but also kind of sucked for OpenAI because they couldn't sell models on AWS where a lot of people are. But there's an important detail here. The reason why I couldn't use my credits on these Anthropic models is because of the deal Anthropic struck with these cloud providers. They had a brutal revshare where they're getting a massive percentage of the money you are spending when you use the models on other clouds because they want to make their profits. So, I'm able to negotiate with these clouds to get better deals for bulk spend. None of them will budge on anthropic models. They just won't. From my experience, I've heard some people find success with it. I've had none. I cannot get Google or AWS or even Azure to give me any form of deal on anthropic models, much less use my credit so I don't have to pay for them. It seems like the deal is so egregious and so expensive that they would be taking too big of a loss because not only would they be giving me the compute worth up to a million dollars, they'd have to pay half a million fees to Anthropic. So, if they're giving, hypothetically speaking, 50% of the spend back to Anthropic and I use a million dollars of credit on Anthropic models, they just had to spend half a million dollars to just pay anthropic outright even though I didn't spend a single scent. That is my theory as to why this is. I don't know for sure. It could just be some weird term in the agreements, but the fact that I am not able to use my credits for quad models and I could use them for literally anything else should speak volumes as to the nature of these agreements that Anthropica has struck with the clouds. That said, almost all of my friends at Real Enterprise are using anthropic models via Bedrock and they might use cloud code, they might use cursor, they might use VS Code and Copilot or whatever else. Almost all of them when they're using Opus are doing it through Bedrock through a Bedrock API key. And I think this is one of the biggest reasons why Anthropic has seen the success they have in enterprises because Anthropic models are available in these other clouds, mostly AWS where a lot of people are. If you're looking for really powerful AI to power your developers, anthropic models allow you to do it in your existing cloud. Open AI ones require you to go strike a deal with Microsoft and Azure or you have to use the Open AI APIs directly. They're fine, but not what you want to do at your business. You want everything under one cloud if you can, and AWS is the cloud most people choose. AWS is key to Anthropic's ability to sell to enterprises. And since Anthropic effectively had an exclusive deal with AWS for models that are useful for code, if you wanted good code models and you were on AWS, you had to deal with Anthropic, which meant you were okay with to some extent the egregious nature of those deals and agreements. That ends today. That is over now. OpenAI will now expose their models through AWS. And my honest guess here is they're not going to be as egregious with those things. If OpenAI manages to get their models into GCP, I can probably use my credits there, too. I'd be really surprised if I wasn't able to use the startup credits I've gotten for the OpenAI models in those other clouds because I know for a fact I can already do that on Azure. So, that's my crazy theory here. This is specifically because OpenAI is petrified of the enterprise growth at at Anthropic outpacing their own enterprise growth and they had to do whatever they needed in order to get out of the exclusivity deal with Azure because any hope of Azure becoming the number one cloud is dead now. Azure and like I am thankful they were as generous with the terms of this deal as they were where they were very very clear to me that I could say whatever I want about Azure and be honest with my experience. We have spent zero dollars of this credit because trying to get Azure to work properly is hellish. If you're wondering why, there's many reasons, but the biggest by far is this. If you compare GBD 5.4's speeds on Azure, which is the blue line, versus OpenAI, which is the black line, I can't ship this. I literally can't ship this. Surely some days it does surprisingly okay and performs in a similar range like the 70 plus range, but sometimes it just gets.3 to two tokens per second. So if we were to use Azure to run our OpenAI models, there would just be some days where it goes under a token per second instead of the usual 70 plus. That is pathetic. And this has been the case with Azure hosting of OpenAI models since Azure started hosting OpenAI models. I know this because we tried really hard to use them early on in the history of T3 chat. We would just sometimes see these absurd performance regressions. No, that people in chat are saying, "Wait, that's not an outage." No, this has been happening every couple weeks for about 9 months to a year. And I have records of this. Sorry, did I say nine months? This post was from over a year ago. In March 10th of 2025, I called out that Azure is just absolutely screwing up 03 Mini performance. When you compare 03 mini speeds on OpenAI, which was regularly over 120 tokens per second, sometimes getting as high as 240 on Azure, it would randomly drop to the 20s. Somebody asked in chat, "So, how am I actually utilizing the 1 mil credits in Azure?" I'm not. I've spent 0 of it because I don't want my customers to have to experience this. And I hate to flame this video because they gave me a million dollars of credit and some employees there really stuck their heads out for me to do it. But I have been harassing them about this for over a year. I had dinner with the head of AI infrastructure at Microsoft and explained this in detail. And do you know what the [ __ ] did? He showed the charts to the team to which they responded, "Oh, those charts are pretty cool." And they copied all of the artificial analysis charts for their own internal docs and then never fixed the problem. So, I ended up not only not getting what I needed, I pissed off Artificial Analysis because I accidentally caused Microsoft to clone half their product in their internal docs. The Microsoft experience. I Yeah, every once in a while I have a moment where I realize like my channel is actually very unique. Like, who else has been in this position and makes YouTube videos about it? Like, it's hilarious, but here we are. I've been [ __ ] through it with Microsoft. I have a lot to say. I feel the same way OpenAI does that we are limiting ourselves trying to use them for [ __ ] I told one of my friends who also got that million-doll credit that we got the credit to which he responded oh yeah we saw one of my engineers shared it in our Slack and shared it with the caption who's going to tell him because they had gotten the million-dollar credit used about 5k of it and hated using Azure for inference so much they moved off and just went back to paying OpenAI because it was cheaper than dealing with Azure's [ __ ] And once again, Microsoft, I would like to remind you guys, I am more than happy to give this feedback privately as long as you [ __ ] deal with it. And as long as you continue to not do anything about it, I will have to do this. In a video that many of your potential enterprise customers are watching where they were going to possibly use Azure for OpenAI models, I am telling them and looking straight in the camera as I say this, don't use Azure for Open AI models. Now that they're on AWS, you should use AWS. I don't care how many credits you have. Your users deserve better than random spikes down to two TPS. And I will gladly amend the statement. The moment that Azure gets their [ __ ] together, but it's been over a year now. I have zero faith you'll do it. Hi, Theo from the future here. I couldn't stop thinking about the fact that I had a million dollars of Azure credits that I couldn't use. And I decided to put it to use by making a benchmark to highlight just how severely bad Azure's compute was. So, I made this bench, Azure bench. It's on my GitHub. You can see it at azure.t3. t3.gg and you can see how absurd the numbers are. It is 2.2x slower to use GPD55 on Azure than it is on OpenAI on average. But the worst case is 15x worse. You might notice that I'm not on Twitter here. I'm just looking at a screenshot. The reason is some of my friends at Microsoft nudged me to delete it, suggesting that I would have more leverage if I did. I don't like being told to delete things, so I did it. But I also gave them a notice that if this wasn't fixed within 15 days that I would be reposting much more aggressively and egregiously because this is unacceptable. And to their credit, they listened and it took them a lot less time than I expected. By noon the next day, the problem was resolved. And these charts are [ __ ] hilarious to look at. The rate at which they fixed it suggests there were just some really bad bugs in their implementation, which has mostly been confirmed to me at this point. And if you think that these numbers are just me for some reason, I don't know why you're here cuz you should trust me about these things at this point. You can go look at Open Router, which saw a massive jump where before the model would go as slow as eight tokens per second on Azure and now is consistently outperforming OpenAI's own endpoints. The latency has plummeted as well, going from 5 seconds at times down to 2 seconds or less, being lower than open AIS. And I even saw numbers as absurd as 200 second time to first tokens, and that's basically done now. Yeah, here's a 100 second TTF that just doesn't really happen anymore. They made real legitimate improvements. And if you go to Azure T3 GG, you'll see I even call out that they have fixed this and you'll see the numbers are pretty consistent. Azure is now as fast if not faster most of the time. So it took a lot of flaming, a lot of effort and burning a lot of compute and risking my million dollars of credit. But in the end, this was a win. I guess bullying works. And credit to Azure for listening. If you want to hear the full breakdown of this story, check out our podcast by the way, Nerd Snipe. We're on YouTube, Spotify, Apple, everything that you listen to your podcasts on. And the next episode, which might even be out by now, has me going indepth on the whole story here. So, check that out if you haven't yet. Really proud of the podcast. The last genuinely interesting piece here is the OpenAI trrenium usage. I've talked about this already extensively in the videos about why I think Claude models are getting worse. I can't say for sure, but Trrenium might be part of that. This is the first time OpenAI has had to make their models run on things that aren't Nvidia GPUs and I am genuinely very very curious how this goes because it seems like this was a huge problem for Anthropic and I don't know if OpenAI will succeed or not. Worst case, I do know that there's a lot of NVIDIA GPUs on Bedrock and I'm sure they will serve some amount of the traffic there. We will see how this all goes but hard to know for now. They already had a $ 38 billion multi-year agreement with AWS and that's been extended by a hundred billion over eight years. Again, the whole GDB by all the compute thing is showing. Literally everybody who's willing to sell them compute, they sign the deal and don't hesitate. If OpenAI fails, it's because they can't make good on these massive deals they're signing. But thus far, it seems to be working out. So, we'll see how that all goes. The commitment spans both Tranium 3 and the next generation Tranium 4 chips that aren't out yet. And it will power a broad range of advanced AI workloads. Tranium 4, which is expected to begin delivery in 2027, will provide another major performance gain, including significantly higher FP4 compute performance. Expanded memory bandwidth and increased high bandwidth memory capacity support increasingly capable AI systems to scale. Again, it seems like there's more RAM available on the chips from the other companies that aren't Nvidia. So, if you want bigger models and larger context windows, tranium and Google's compute seem to be good options. And this comment from Andy Jasse at Amazon says it all. We have lots of developers and companies eager to run services powered by OpenAI models on AWS. And our unique collaboration with OpenAI to provide a stateful runtime environment will change what's possible for customers building AI apps and agents. We continue to be impressed with what OpenAI is building. and we're excited not only about their choosing to go big on custom AI silicon with Tranium, but also our opportunity to invest in the company and partner over the long term. There you go. The weird partnerships happening across the industry are unbelievable, especially like the Google Metal one, man. But that's where a lot of the war is going to be fought. Right now, we are fighting the war between Anthropic, OpenAI, and Gemini. But I would not be surprised if by this time next year, the war has shifted to be Nvidia versus AMD, if either even able to survive all this, versus Tranium at AWS versus Google's TPUs. So, the only companies I see maintaining relevance in the wars are Google and Nvidia, but we'll see where this all ends up. I just wanted to dump my thoughts on this as I've been reading it and stressing out about it for a while. These are the things I have to deal with a lot as a CEO, and I am very surprised if any of you guys made it all the way through to here. This is not my usual dev video. We've barely even talked about code, but this stuff's important. This is what powers the future that we're all building into and we're all stuck with either way. I hope this is valuable for you guys. I know it helped me get some stuff off my chest. And maybe, just maybe, they'll finally [ __ ] fix Azure because there is literally no competitive advantage anymore. I I can't believe I accidentally made this another Azure crash out. But man, it's been rough. It's been really rough. Hopefully, they fix things. We'll see where it all goes. But until next time, peace ns.