Open transcript
It's hard to imagine OpenAI where they are today without the help of Microsoft. From the funding to the people to the infrastructure, Microsoft's been all in on OpenAI for a while. So much so, they posted pretty much the exact same blog post two years later. I thought I had the same tab open here. Nope. But the reason we're talking about this isn't because the partnership was extended. It's actually the opposite. Microsoft is looking at what a future without OpenAI would be for them, which is fascinating. This kind of plays into my previous video about AI getting commoditized and the race to the floor to see how good a performance we can get for price. The lead that OpenAI had is closing. And the implications of that on the industry long term are genuinely fascinating. Enough so that I wanted to take the time to talk about it all with you. But as cheap as this AI stuff is getting, I still have to pay my team somehow. So let's get a quick word from today's sponsor before we dive in. Before we go too far, we need to make sure we understand the terms of the Microsoft x OpenAI partnership. It's also worth noting that the way you invest in OpenAI is really strange, which is OpenAI investments have a capped return. If I invest $1,000 into a company that's worth $100,000, then I own 1% of this company. If the company's valuation was to 1,000x, if it was to go from 100k to 100 mil, your investment would see the same impact. Your 100,000 would get those same extra three zeros because you have 1% of it. And 1% is still 1% regardless of how big the number is. This is why early stage investment works the way it does, because you're investing in a company at a very low valuation, expecting it to fail, with the small percentage chance it will go massive. Most early stage companies in the Y Combinator world are raising with valuations between $18 million and $30 million, roughly. It varies a lot over time, but this is the range for the valuation that you're investing against. You expect to lose the money, though. Most investors, like even the ones I have, the expectation is we'll go out of business and they get nothing back. But if they do get money back, it's not going to be a 50% win. It's not going to be an 80% win. It's a very high chance their return is going to be a 10 to 100x return. That's why you invest early stage is the likelihood of getting your money back is low, but the possibility that your money coming back ends up way more valuable is high enough that it's worth the risk. That's what early stage investing is. Open AI didn't want a bunch of investors to push them to do crazy things to get these massive multipliers. So a weird decision they made early on was to cap the potential return. The cap, if I recall, was a 10x return cap. So with this investment here, with 100k, if that ended up being a $100 million company, you wouldn't get back a million. You get back 10x of what you put in. So you would get $10,000 back instead, because you would hit that cap when you cash out. This meant that OpenAI's opportunity for traditional investments was weird. A lot of companies wanted to put their money in, but not a lot of companies wanted to meet these particular terms unless they got something else. This is why Microsoft ended up being a really good partner for them. The combination of Microsoft needing better AI tech, Microsoft having a bunch of infrastructure with Azure, and Microsoft having a ton of money they didn't know what to do with, that combo made them a really good partner for OpenAI. And that's why the partnership happened. Microsoft's needs for the things that OpenAI had done and had, and OpenAI's needs for the infrastructure and money that Microsoft had, it made a ton of sense. Especially if you consider the fact that at the time, back in 2023 especially, OpenAI stuff had not been even close to matched by other competitors. Certainly not open source ones like we have today with DeepSeq. Microsoft's multi-billion dollar investment into OpenAI was done to get them access to the best-in-class models. And to this day, the only way to use OpenAI's best models without doing the OpenAI platform is through Azure. On T3 chat, we have a couple OpenAI models, specifically 4.0 and 4.0 Mini, hosted through Azure instead of OpenAI. And that's awesome because we have a bunch of credit with Azure, but it's also awesome because the speeds for some of those models are way better. Some of them are way worse, though. I don't want to get, don't get me started about O3 on Azure. It's, it's been tough, okay? That aside, this partnership has been very lucrative for everyone involved, but some things have started to change in the industry. I could say this myself, but I'd rather quote Satya here. In his words, AI models are getting commoditized. And I absolutely agree here. As I covered in my video, The Race to the Bottom, most models are focused on best price to performance in meeting the current performance bars. I even made a chart during that video. Let me find it quick. Previously, the performance gap that OpenAI's models had compared to the competition was massive. It was, it felt like a 2x plus gap in the intelligence capabilities of their models. But over time, alternatives have been getting closer and closer to matching where OpenAI is at any given time. The gap between OpenAI and the alternatives is smaller than ever. And those alternatives are often way cheaper. And sometimes like in the case of DeepSeek R1, or especially in the case of the V3 overhaul that just shipped, they are surpassing what OpenAI is doing with open source alternatives. On top of that, the cost of a lot of these things has been going down too. Every once in a while, OpenAI does some crazy stuff like they drop GPT 4.5, which raises the price a bunch, or even worse, what they just did with 01 Pro over the API. The price of OpenAI stuff is being forced down finally, but it's not even close to what you get out of something like Gemini nowadays. There are also rumors that Microsoft now has their own internal model that they've been working on that takes advantage of all of these reasoning things that they've been seeing from other providers, especially in the open source world, that is comparable enough to the performance they're getting from OpenAI that they're questioning everything. Their model is known as PHY, which historically has not been particularly great or even worth talking about. But Microsoft is internally cooking their own reasoning models on top of it that appear, from the little bit of leaks that we've heard, to be really, really good. I have a whole dedicated video planned coming soon for the DeepSeq V3 update that just shipped. Quick TLDR, DeepSeq R1 and the R series are their reasoning models, but the reasoning models have to be based on a traditional LLM. With OpenAI, the O models have thus far been based on 4.0. That's part of why 4.5 is exciting, not because it's a good model, but because as a base for those reasoning models for an O4 or an O5, it is very, very promising. V3 is what R1 is based on. V3 was such a groundbreaking model when it dropped back in December that I kind of dropped everything to start playing with it, eventually leading to me building T3Chat. I don't think T3Chat would exist if I wasn't so impressed with the DeepSeq model when it came out, and also horrified by the state of the website, because that wasn't their focus. V3 was a very exciting release. And the performance you could get, especially considering the price, was insane. But look at this. Standard intelligence benchmarking, 3.7 Sonnet is underperforming compared to DeepSeq V3. For other benchmarks, like the speed, it's not great because they're benchmarking it on the official API, but it should be able to speed up if it's run a bit differently. This is what's so exciting about what DeepSeq is doing, is they're managing to get performance better than GPT-4O for prices that are a hell of a lot closer to Flash and 4.0 Mini than they are to 4.0. The fact that they are higher up in the intelligence while being closer to 4.0 Mini than they are to the models they're competing with is insane. O3 Mini High is still an incredible value, and I will continue saying this is the best model OpenAI has ever dropped in terms of all of the pieces. But V3 as a base model that is open source is super promising, and I am very excited to see what happens when the new V3 is used to train R2. It's going to be a very, very fun future. Performance speed-wise still isn't great, but I would hope to see other providers figure out how to make V3 faster because it's a surprisingly small and efficient model. They should be able to make it fly. But I'm bringing all of this up because it's apparent that the base primitives that we look at and use for all of these things have just gotten really good and are borderline commoditized. If I add one more model in here quick, 4.5, they don't even put it in the chart because it's too expensive to do the tests against. That's hilarious. I guess we have to trust DeepSeek's numbers because I can't use my trusted artificial analysis. Here is their benchmarks against GPT 4.5, which, if you don't remember, is one of the most expensive models ever released. 4.5 costs $75 per million input tokens and $150 per million output. Compared to the current DeepSeek V3 prices, $0.27 in, $1.10 out. 280 times more expensive for input tokens. 270 times more expensive for a model that consistently performs the same or worse. The fact that a model that doesn't think is comparing with a model that's that expensive shows that the open opportunities we have with these open standards are insane. And what DeepSeek has cooked with the V3 model and now with the R series as well is just unbelievable. And it's a fundamentally different world than it was when OpenAI and Microsoft started their partnership. Hell, it's a different world than it was when they evolved that partnership. Because OpenAI has managed to keep improving their models. Again, O3 Mini is incredible for what it is. But they are not improving fast enough to maintain the massive lead they previously had against their competition. Microsoft's ability to solve hard problems that aren't operating system level has historically been really bad. If even they are managing to figure out how to make these performant models, these great models themselves, it makes sense that they are questioning the billions of dollars they've been spending on OpenAI this whole time. I do want to call out the focus shift that's going on in the AI world now. The main piece brought up here that I think is worth thinking about is the go-to-market manager from OpenAI explaining that success depends on the bringing together of the pieces for successful AI experiences. I talked about this a little bit in a video I just put out, my Gemini video and why Google is in such a good state. The thing I tried to drive home there is that Google is uniquely well-positioned because they have the data part figured out. They have the science part figured out mostly and they have hardware at a level that nobody other than like NVIDIA and these other chip manufacturers have. But the idea of being the same company that builds the chips, that hosts the infra, that generates the models, that creates and finds the data to use for that model creation and training, that has the money to fund all of this, the platform to host it on with GCP, but also the apps that you can experience these things in. Google is the only company that currently has all of this figured out. Microsoft has this part roughly. They have the platforms and the money. They're starting to figure out the science side. It's been an interesting journey for them. They don't have the hardware in the same sense because they're not making their own chips. They're still at the whim of the manufacturers. If you don't believe me, look at the state of ARM for Windows. They are fully reliant on whatever the hell Qualcomm feels like doing, whereas Google is manufacturing and designing their own architecture now, which allows them to have a huge advantage compared to these other companies. I think this is the resurgence of the Microsoft versus Google wars. And OpenAI is trying specifically to compete. I have to actually reorchestrate this a tiny bit to make this make more sense. Give me one moment. There. Here. Main difference here. I swapped the science and the data parts because OpenAI wants to do this to this. OpenAI wants to set things up so they own all of these parts and they'll just deal with the hardware whenever. Google is here and they want to keep extending to the left. But if they can win here, hell, if they can just win here, they're still in a very good position and they know that. So they are being careful as they roll to the left. Microsoft's kind of fucked. Because they only have this. They are starting to extend here with mixed success. They couldn't figure out apps. The closest thing they had was GitHub Copilot, which they've now commandeered fully for their own branding because they don't even have their own platform that makes sense. Yeah. Don't get me started on Copilot on my Windows computers. It makes me feel like I am going insane. Also, apparently OpenAI is trying to do custom chips. So I guess I was wrong. OpenAI is also trying to extend all the way across this chart. They want to go all the way across. They want to have a fully integrated AI platform from the app you consume it with to the hardware and the chip architecture that is actually running the models. I wish them luck. CPU design and architecture is a chaotic world that I don't know if they're equipped to be in. Considering that their fumbles trying to get into web dev have been as terrifying as they are, and web dev is easy compared to architecture. I wish them luck. But I do think this is important that Microsoft sees the straight up risk they have being just here. They can be consumed from either side. So now as they realize that the science side is more and more commoditized and they can consume that themselves as they do whatever they can to collect data or they buy it for companies like Scale and Data Curve, as they try desperately to build their own apps that don't suck for consuming AI stuff, and maybe they finally make their own hardware. I wish them luck. Microsoft realized that their position here in relation to OpenAI is a small vertical competitor. They are terrified of losing their slice as OpenAI consumes that whole layer. So now they're trying to figure out how to succeed without OpenAI as core to their business in the AI world. And it's an absolutely fascinating shift of focus because, and I hope this isn't too controversial to take, more and more we're realizing the valuable part is here. If I can just swap out the models in my apps and users don't notice beyond maybe it's slightly faster or slightly smarter, or maybe it's just 5% dumber, who cares? The fact that it's gotten to the point where I can change the model on my backend and my users might not even notice, fundamentally shifts where the moats are in this world. Previously, OpenAI was the way you had good AI generation in your apps. Over time, that advantage has been lost. You can switch from an OpenAI model over to R1 and have an experience that's really, really good. Depending on which OpenAI model you were using before, the experience might be even better than what you were doing prior. That is incredible. That's a massive industry-wide shift that Microsoft is now feeling internally and starting to make decisions around. The core to the agreement between Microsoft and OpenAI wasn't just that they would invest this insane amount of money. It was that OpenAI would give them access to whatever innovations and development they had throughout the process of making better and better models. The rule was that once they hit a definition of AGI, that OpenAI would no longer have to give all of their learnings to Microsoft. The rumor that I've been hearing is that Microsoft stopped getting info from OpenAI, and there was a bit of a pissing match between the two because Microsoft and the researchers there asked OpenAI, what did they do to make O1 so good? Because this was before we understood reasoning as an open thing in the AI world, and OpenAI just refused to tell them. This appears to have started some fires in the relationship between these companies. Apparently, the Microsoft CTO wrote about this on LinkedIn. Not sure why this is news, but to summarize the obvious, we build big supercomputers to train AI models. Our partners at OpenAI use these supercomputers to train frontier-defining models, and then we both make these models available in the products and services so that lots of people can benefit from them. We rather like this arrangement. We've been at it for almost five years. We also, for years and years, have built AI models in Microsoft Research and in our product groups. Apparently, the Microsoft AI models are performing nearly as well as OpenAI and Anthropics on benchmarks. Microsoft's training reasoning models that could compete with models from OpenAI, as well as those from DeepSeek and Alibaba. Reasoning models are designed to simulate yada yada yada, chain of thought, cool. OpenAI was not sharing technical information about the chain of thought process as agreed, the report noted, which became a source of conflict between the partners. This is the key. OpenAI has been refusing to share the technical info about how they're doing chain of thought stuff with Microsoft, which is preventing them from taking advantage of their partnership to make their models better. And now this whole thing is starting to fall apart. Microsoft's considering releasing MAI later this year as an application program interface, API. It would put Microsoft into direct competition with APIs from rivals and partners. It also said that the company's already experimenting with swapping out MAI models for OpenAI's models in Microsoft Copilot. I was sus of this starting to happen as soon as Microsoft allowed for you to use Claude in Copilot. But that's when I realized this partnership was starting to erode. And it feels like it's moving much faster than anyone would have predicted, certainly myself included. AI definitely accelerates things. It might not be the speed that we ship products, but it's certainly accelerating the speed at which these types of partnerships erode. Also, the Stargate project, which was the fancy OpenAI announcement with Trump and Oracle, where they're partnering to build a bunch of AI infrastructure in the U.S., definitely doesn't involve Microsoft heavily. OpenAI claims that they're part of the partnership, but we'll see how that goes. Credit to the information for this groundbreaking report. Nothing else has this level of detail. This is clearly a story they figured out and broke themselves. Apparently, last fall, during a call with senior leaders at OpenAI and Microsoft, Suleiman, who is the person at Microsoft in charge of AI, was fighting with OpenAI staffers to explain how O1 worked. He was peeved that OpenAI wasn't providing Microsoft with documentation about how it had programmed O1 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. Raising his voice, Suleiman told OpenAI employees, including Mira, who was the OpenAI CTO at the time, that the AI startup wasn't holding up its end of the wide-ranging deal it had with Microsoft, the people familiar with the conversation said. The more I've read through these sources, the more I'm really starting to understand why Microsoft feels the way they do. My honest take is that the feeling they have internally is that the AI teams at Microsoft are learning more from DeepSeq than they are from OpenAI. And when you consider the fact that Microsoft spent $13 billion investing into OpenAI, and they are getting more out of DeepSeq's generic open stuff that they can just read the PDFs for for free online, they're getting more from that than they are from spending $13 billion on OpenAI. I see why they're starting to get upset. Before we wrap up, I want to do one last shout-out. This channel is making me think about these things a hell of a lot more and has a ton of insight info that's been super beneficial. The AI Explained channel on YouTube has become one of my favorite channels to watch. I'm actually going to hit the bell on it. I recommend you do the same if you like this deep news stuff. Very, very good source. I would not have covered all of this if it wasn't for the things that were covered in the most recent AI Explained video. Give them a sub if you haven't. Phenomenal channel. Huge shout-out to AI Explained for cluing me in on a lot of this and giving me some of the sources to start digging deeper. It was a really good starting point for me to dive in and figure out what the hell was going on. Oh, I don't know if we formalized what the AGI definition is from OpenAI. Microsoft and OpenAI's deal immediately ends when AGI is achieved. But since it's hard to define, they defined it as $100 billion in profits from AI systems. So AI systems need to be able to generate $100 billion in profits, at which point we will now, in the definition of their agreement, call it AGI, and the partnership can now dissolve. And OpenAI doesn't have to give Microsoft access to all of this stuff anymore. The reason that matters is because Microsoft wants to get out of the partnership. So they're going to rush to this definition of AGI simply to not have to give Microsoft access to things anymore. One other thing of note is that definition of AGI was determined, at least publicly, after this interview with the CEO. This was December 9th, 2024. And they even say in here that the definition of AGI is weird. And then the definition was decided on pretty soon after. So this shows just how much they're flying by the seat of their pants right now. All of these things are happening on the fly as they figure stuff out. It's kind of chaotic. And Microsoft as a company is not built to move at this speed. And they're trying their hardest to figure out how to and what parts are even worth moving this way. I suspect we're going to be getting a hell of a lot more info. What happened here was Microsoft putting a lot of money in to try and catch up on something they were behind on, which was all the science of AI stuff. They then hired one of the guys who made DeepMind to take over AI at Microsoft, watched all of these things happen outside, realized that OpenAI isn't helping them accelerate the way they wanted to. And all of these other things in the public are helping more than OpenAI is. How the hell do we get out of this chaos? And I think this makes all the sense in the world. And I am incredibly thankful more than ever that I am not the CEO of a model frontier company, because this is all chaos. And I am very happy to continue hanging out in my peaceful little world, building the best AI chat up ever. If you haven't already tried it, it's only eight bucks a month. Give T3 chat a shot. Let me know what you guys think on this one. Am I overreacting to this weird news between Microsoft and OpenAI? Is this something that will even matter or that we'll remember in a few years? Curious what you think. Until next time. Peace, nerds.
