Microsoft and OpenAI are breaking up?

Okay, let's break down what's going on with Microsoft and OpenAI based on the video.

Here's the summary:

Microsoft and OpenAI's cozy partnership, once fueled by billions in investment, seems to be hitting the rocks. The main drivers are the rapid commoditization of AI models, with alternatives (especially open source) catching up and getting way cheaper, and a major conflict over OpenAI allegedly refusing to share crucial technical details like "chain of thought" reasoning as per their agreement. This is pushing Microsoft to seriously explore a future where they rely on their own growing internal AI capabilities rather than OpenAI, leading to fascinating implications for the AI landscape.

Here are the key topics and details:

• The Microsoft/OpenAI Partnership:

  • Microsoft invested billions into OpenAI, providing crucial funding and infrastructure (Azure). It's hard to imagine OpenAI being where they are today without this help.
  • The investment structure for OpenAI is weird – it has a "capped return." This means investors don't get unlimited upside like in traditional early-stage investing; there's a limit (reportedly 10x). This structure made it hard for OpenAI to get traditional investors.
  • This weird structure made Microsoft a great partner because they needed the AI tech and had tons of money and Azure infrastructure, making the deal attractive despite the capped return.
  • Part of the deal was Microsoft getting access to OpenAI's cutting-edge models and developments. Azure is still the only way to use OpenAI's best models outside of OpenAI's own platform.
  • Access via Azure can sometimes be faster for certain models, though others can be worse (like O3 on Azure, which has "been tough").
  • The partnership has been very lucrative for both sides so far.

• AI Models Are Getting Commoditized:

  • As Microsoft CEO Satya Nadella puts it, AI models are getting commoditized, and I completely agree. This means the performance gap between top models is shrinking rapidly, and the focus is shifting to price/performance.
  • Previously, OpenAI had a massive lead in intelligence capabilities (felt like a 2x+ gap). Now, alternatives are getting closer and closer.
  • These alternatives are often much cheaper.
  • Sometimes, open-source alternatives are even surpassing OpenAI models in performance, as seen with DeepSeek's recent updates.
  • OpenAI's prices have been forced down somewhat recently but are still high compared to alternatives like Gemini. Some OpenAI models, like GPT 4.5, are incredibly expensive.

• Microsoft's Internal AI Efforts:

  • Microsoft is rumored to have developed its own internal AI model, known as PHY, which historically wasn't great.
  • However, they are reportedly adding their own reasoning models on top of PHY, and leaks suggest these are performing really well and are comparable to what they get from OpenAI. This makes them question their reliance on OpenAI.
  • They are developing competing models referred to as MAI.

• DeepSeek as a Disruptor:

  • DeepSeek is highlighted as a prime example of open-source models closing the gap.
  • Their V3 base model was groundbreaking; it was so impressive it motivated the creation of T3Chat.
  • DeepSeek V3 performs better than GPT-4O on some standard intelligence benchmarks while being drastically cheaper than models like GPT 4.5.
  • For example, GPT 4.5 is hundreds of times more expensive for tokens than DeepSeek V3, while performing the same or worse in benchmarks.
  • This shows the insane potential of open standards and open-source models – they are competing with the most expensive frontier models.
  • While speed on the official API isn't great yet, V3 is surprisingly small and efficient, suggesting others could optimize it for faster performance.

• The Shift in the AI Moat:

  • The valuable part ("the moat") in AI is shifting away from just the model itself towards owning the entire stack.
  • This full stack includes designing chips, hosting infrastructure, collecting/finding data, building the models, and creating the applications where users experience the AI.
  • Google is uniquely well-positioned because they have many pieces: data, science, hardware (designing their own chips), infra (GCP), and apps.
  • Microsoft has platforms/money and is trying to build science/models (mixed success), but they lack their own hardware design (relying on manufacturers like Qualcomm) and struggle with building good AI-consuming apps (Copilot integration issues are noted). They feel stuck being just one piece of the puzzle.
  • OpenAI is also reportedly trying to build out the full stack, including custom chips, which seems ambitious but indicates their desire to control the whole vertical.

• The Core Reason for Conflict (The Breakup Driver):

  • A key part of the Microsoft/OpenAI agreement was sharing technical innovations and development until a definition of AGI was met.
  • Rumors and reports suggest Microsoft stopped receiving this technical information from OpenAI.
  • Specifically, OpenAI allegedly refused to share details about how they achieve "chain of thought" reasoning – the process where a model thinks step-by-step before answering, which is crucial for performance.
  • This refusal angered Microsoft's AI lead, Suleiman, who felt OpenAI wasn't holding up their end of the deal, especially regarding O1 model details. This appears to be a major source of friction.

• Implications and Microsoft's Future Without OpenAI:

  • Microsoft feels they are learning more from freely available open-source research (like DeepSeek's PDFs) than from the technical insights provided by OpenAI, despite their $13 billion investment. I can see why they'd be upset about that.
  • Microsoft is considering releasing their own MAI models as an API, putting them in direct competition with OpenAI and other providers.
  • They are already experimenting internally with swapping out OpenAI models for their own MAI models in Microsoft Copilot. The fact that Microsoft allowed using Claude in Copilot was an early sign that the partnership was eroding.
  • This shift is happening extremely fast, likely accelerated by the pace of AI development.
  • The AGI definition in their contract (reportedly $100 billion in AI system profits) is unusual. OpenAI could potentially rush towards this definition to legally dissolve the partnership and stop sharing information, even though Microsoft now seems eager to reduce their reliance anyway.
  • Microsoft's position feels chaotic; they invested heavily to catch up, hired top talent, saw great progress externally (like DeepSeek), felt OpenAI wasn't accelerating them enough, and are now trying to figure out how to operate independently in this rapidly changing world.

• Recommended Resource: The AI Explained YouTube channel is a great source for deep AI news and provided much of the background and sources for this discussion. I highly recommend it if you like this kind of detailed analysis.

Overall, the situation is a chaotic but fascinating example of how quickly the AI landscape is evolving, challenging established partnerships and forcing major companies like Microsoft to adapt rapidly.