MICROSOFT GIVES OPENAI AND ANTHROPIC THE BOOT — SWAPS THEM OUT FOR HOME-BUILT AI IN EXCEL AND OUTLOOK
Sources: TechCrunch | Bloomberg
So Microsoft, the company that has pumped something like $13 billion into OpenAI over the years, is now quietly replacing OpenAI’s models with its own home-built AI in Excel and Outlook. Tens of thousands of prompts a week that used to be handled by ChatGPT or Claude are now going through Microsoft’s internally built MAI models instead. That is not a test run. That is a business decision.
And the story gets better. Mustafa Suleyman, Microsoft’s AI chief, gave an interview where he basically said that a contract change about six months ago “set them free” from OpenAI to go after superintelligence themselves. Set free. Like they were in a relationship and finally got to keep the apartment. Microsoft is not abandoning OpenAI entirely but they are building something alongside it. Something that eventually might not need OpenAI at all.
This is the classic tech industry power play. You invest in a startup, it grows into a monster, and then you spend the next three years quietly building the thing that replaces it. IBM did it to everyone. Amazon did it to its own sellers. Now Microsoft is doing it to the company it owns a huge chunk of. OpenAI must be thrilled.
The MAI lineup now covers reasoning, code, image generation, transcription, and voice. It is not a toy anymore. When Microsoft starts routing real enterprise workloads through its own stack, that is a signal that the economics of paying third parties for AI are starting to look absurd, even to the companies doing the paying.
Amazon is doing the same thing. Uber has been experimenting with Chinese models to cut costs. The message from the industry this week is clear: nobody wants to keep paying the premium prices at OpenAI and Anthropic forever. The race is now about who can build the cheapest capable model, and the incumbents just figured out they cannot trust anyone else to win it for them.
For OpenAI, watching Microsoft route prompts away from GPT in Office 365 has got to sting. That is the product suite that probably drives more daily AI usage than anything else on earth. For Anthropic, losing Microsoft as a paying customer is just another reminder that even your biggest fans will replace you the moment they can do it themselves.
APPLE JUST CANCELLED ITS OWN M6 PRO AND MAX CHIPS TO BUILD SOMETHING BETTER — AND THE REASON IS ENTIRELY AI
Apple is not a company that cancels things. They are meticulous planners with roadmaps that stretch years into the future and teams of engineers working in secret on chips that will not ship for three years. Which is exactly why the news that Apple has scrapped its entire high-end M6 chip lineup is so remarkable.
The company was working on an M6 Pro and an M6 Max. They decided instead to skip those entirely and jump straight to the M7 series, which has dramatically upgraded AI performance built into its neural processing units. The acceleration was apparently significant enough that they decided it was worth the disruption and the delay.
What is interesting about this is not just the news itself, but what it signals about how Apple sees the next two to three years. They are not incrementally improving performance. They are restructuring their entire hardware roadmap around the assumption that AI capability will be the most important thing a chip can do, ahead of everything else. Raw performance gains, battery life improvements, display rendering — all of it takes a back seat to running better on-device AI.
The M7 Ultra is reportedly going to dramatically upgrade AI performance and may eventually power Apple Intelligence servers starting in 2029. That last part is the key. Apple is not just building better consumer chips. They are building chips that will run their AI cloud infrastructure. The consumer Mac and the server powering Siri will share silicon architectures. That is a massive vertical integration play that almost nobody else in the industry is attempting at this scale.
The base M6 MacBook Pro is still coming later in 2026. But the Pro and Max versions — the ones that demanding creative professionals and developers actually buy — are getting delayed and redesigned. If you were planning to buy a high-end Mac in the next twelve months, waiting for the M7 is now the obvious call.
Apple has been slower than everyone else to ship visible AI features. But their hardware team keeps quietly making decisions that suggest they believe AI will define the next decade of computing. Cancelling the M6 Pro to jump straight to AI-focused silicon is the clearest evidence yet that those two things are about to converge in a major way.
THE AI PRICE WAR GOES NUCLEAR: OPENAI, META, AND SPACEX RACE TO BUILD THE CHEAPEST SMART MODEL AS THE BIGGER-IS-BETTER ERA OFFICIALLY ENDS
For the past three years the AI industry operated on a single assumption: bigger is better. More parameters, more compute, more training data, better model. That era is now officially over.
Bloomberg reported this week on a direct three-way competition between OpenAI, Meta, and SpaceX/xAI all trying to build the most cost-efficient AI models rather than the most powerful ones. GPT-5.6 was specifically designed to use far fewer tokens while getting the same result. Meta’s new commercial model was priced to compete on cost. Grok 4.5 from xAI launched as Opus-class at a fraction of the cost. Even Google released Gemini 3.5 Flash at roughly one-third the price of its flagship model.
What changed? Two things. First, enterprise customers started looking at their AI bills and having quiet heart attacks. When you are running millions of API calls a day, the difference between paying OpenAI rates and paying for a Chinese model or a cheaper domestic alternative is not a rounding error. It is a budget line that can destroy a startup’s runway. Second, a wave of Chinese models proved you do not need to spend hundreds of millions on training to get something competitive on most tasks. DeepSeek V4 Pro costs $3.48 per million output tokens. Claude Fable 5 costs $50 for the same output. That math does not work in Anthropic’s favor.
The result is that the frontier labs are now competing on a dimension they explicitly said did not matter two years ago. They are racing to see who can give you the most intelligence per dollar, and that race is going to be brutal for everyone charging premium prices.
The interesting wrinkle is that this shift may actually help OpenAI in one way: they have the scale to absorb losses on cheap models and make it up on volume. Smaller labs do not have that luxury. So the era of cheap AI may end up being good for the biggest players and catastrophic for everyone in the middle trying to carve out a niche.
The other quiet winner here is the enterprise IT buyer who spent two years getting lectured about AI ROI and is now finally getting to negotiate on price. The market is moving their way faster than anyone expected.
AI DEMAND IS “ALMOST UNLIMITED” SAY EXECUTIVES — BUT WALL STREET IS STARTING TO WONDER IF THE MATH ACTUALLY ADDS UP
Source: CNBC
You could be forgiven for looking at AI stocks this week and thinking something is wrong. After a first half of 2026 that was genuinely historic, the major AI-adjacent names have seen some volatility. The chip stocks, the data center companies, the hyperscalers — all of them have had moments in July where you might wonder if the party is starting to wind down.
And then the executives open their mouths.
Speaking to CNBC this week, multiple AI executives described demand for AI chips, models, and data center capacity as almost unlimited. Not growing. Not strong. Almost unlimited. The kind of language you hear when someone is trying to convince investors, partners, and themselves simultaneously that the machine they are building will eventually be worth what they paid for it.
The interesting tension here is that everyone knows the current infrastructure build-out is staggering. The five biggest tech companies are collectively spending something in the range of a trillion dollars on AI hardware and data centers over the next few years. The Bank for International Settlements warned last week that this looks exactly like the structural patterns of every bubble in history. And yet the demand signals keep coming in hot.
Here is the thing though: both things can be true. You can have genuine and almost unlimited demand for AI capability while simultaneously having a pricing and supply dynamic that makes the economics unstable. The internet had unlimited demand in 1999 too. That did not mean every company building the pipes was going to survive. Most of them did not.
The executives are not lying when they describe the demand. They are, however, possibly confused about whether unlimited demand translates into unlimited profits. That distinction is going to matter a lot in twelve to eighteen months when the IPOs from OpenAI and Anthropic start hitting the market and the public is asked to pay for the infrastructure being built right now. Someone is eventually going to have to show their work.