Microsoft is changing both the models behind Copilot and the chips that run them. About a year after corporate customers complained about the assistant’s quality, Copilot now draws on Anthropic’s Claude models alongside OpenAI’s, and Microsoft has formally introduced Maia, its own AI chip tuned for Copilot work. At the same time, OpenAI is said to be negotiating to license Sora technology to a startup aimed at Hollywood, Apple is preparing an AI-centered home hub, and the Chan Zuckerberg Biohub has announced a $1.8 billion partnership for AI-driven medicine. These moves suggest that big tech is rethinking three questions at once: which models to use, what hardware to run them on, and which projects to keep.
How Claude ended up inside Copilot
Roughly a year ago, Copilot had a credibility problem. Business customers said its answers were often wrong, that it was buggy and slow, and that it struggled with tasks as basic as writing an Excel formula or building a usable PowerPoint deck.
Charles Lamanna, the Microsoft executive vice president who now oversees Copilot, proposed an unusual fix: license models from Anthropic, a direct rival of OpenAI. Given Microsoft’s deep, exclusive partnership with OpenAI, insiders saw the idea as close to heresy. Lamanna made his case to CEO Satya Nadella with side-by-side demonstrations, and on complex office workflows the Claude models clearly did better than the OpenAI models.
Six months later, Nadella put every back-end engineering team behind Copilot under Lamanna. He now runs the Office application groups as well as the AI systems that decide which model handles a task and how models are combined. Earlier this month, Nadella also handed him the Office product organization as a whole.
He has also merged the separate Copilot engineering teams and set what staff call a “no jerks” policy, a deliberate break from the stack-ranking culture Microsoft was known for a decade ago.
Copilot as the “Costco of AI”
Lamanna’s stated goal is to make Copilot the Costco of AI for business: the most useful work per dollar, for everyone from cash-strapped students to Fortune 500 executives, without large bills for tokens, the small chunks of text that AI models process and that providers use to set prices.
Competition explains the goal. OpenAI and Anthropic are fighting a price war for enterprise customers and urging them to build directly on their platforms rather than through Microsoft. Microsoft has to partner with both firms while selling against them. For a company whose grip on business software rested on Windows, Azure and Microsoft 365, generative AI has brought the toughest enterprise competition in decades, and with customers questioning the return on AI spending, its old pricing power may be harder to hold.
Lamanna also has engineers study outside tools: Copilot’s AutoPilot workflow feature was designed after the team examined OpenClaw, a popular open-source project.
Maia: Microsoft’s own AI chip
Microsoft has now officially unveiled Maia, its in-house line of AI accelerator chips. The Maia 300 series, whose development and internal rollout were first reported in August, is more efficient at running Copilot workloads than NVIDIA’s GPUs, the graphics processors widely used for AI, according to Microsoft. The company plans to deploy several gigawatts’ worth of Maia hardware across its global data centers starting next year, a scale that translates into millions of chips.
Microsoft expects three payoffs:
- Less reliance on NVIDIA: a smaller share of its enormous hardware budget goes to a single supplier.
- Cheaper Copilot: lower computing costs for its own AI services improve profit margins.
- More GPUs to rent out: moving internal work onto Maia frees high-demand NVIDIA chips for Azure customers who rent computing capacity.
For now, Microsoft is positioning Maia as a chip for its own software rather than a product for general Azure customers. It would welcome outside customers, but none has been confirmed. Over the summer, Microsoft also held talks aimed at getting Anthropic to use Maia, though no agreement has been made public.
Manufacturing is the open question. Microsoft has reportedly negotiated with the contract chipmaker TSMC to produce several hundred thousand Maia chips beginning next year. Because rivals are competing for the same factory capacity, how much of it Microsoft can secure may decide whether the plan succeeds.
▲ An in-house chip to cut reliance on NVIDIA
A new life for Sora in Hollywood
Earlier this year, OpenAI said it would scale back Sora, its video generation model, along with the related consumer video app, and redirect computing power and money to ChatGPT and its business products. Now OpenAI is reportedly negotiating to license core Sora technology to an outside company.
That company is a stealth startup, not yet named, co-founded by entertainment executive Jeffrey Katzenberg, a researcher who previously led Sora at OpenAI, and a former chief financial officer of Dropbox. The plan is to adapt Sora for Hollywood studios and sell to them directly. Earlier visions of Sora centered on consumer video tools or a social app; this venture is purely a business-to-business production play.
For OpenAI, the appeal is clear: it earns a return on the research and computing already poured into video AI and, through an equity stake, shares in any success without carrying ongoing development costs. The founders appear to believe that pairing Katzenberg’s studio relationships with deep Sora expertise gives them an edge rivals lack.
The obstacles are significant:
- Studios and Hollywood workers remain wary of AI because actors, writers and crews fear losing their jobs.
- Runway, an AI video company, first courted filmmakers but later shifted toward broader corporate customers after Hollywood proved hard to turn into a business.
- The startup must show that an AI production pipeline can deliver feature-length work at a quality studios will accept.
- The venture is young and has not decided whether to license tools or make its own films.
OpenAI is spreading out again
The Sora talks fit a wider pattern. OpenAI has restarted work on robotics, hired a Patreon co-founder to run a new products group for creators, and continues to develop personal hardware devices. That breadth looks like a return to an earlier habit of trying many products at once. It may not last: with rivals such as Anthropic exploring public stock offerings and investors pressing for profits, experimental projects that do not bring in money quickly could be cut once more.
AI servers in the garage
A growing number of engineers and AI enthusiasts in Silicon Valley are turning spare bedrooms, attics, garages, basements and even wine cellars into small AI computing rooms. They are buying Apple’s Mac mini and Mac Studio machines along with AMD and NVIDIA graphics cards, and some have had to upgrade their home electrical wiring to handle the power draw.
Nearly everything running on these setups is an open-weight model, meaning a model whose trained parameters are published so anyone can download and run it. Alibaba’s Qwen family is popular, along with newly released Western open-weight models. Closed models cannot be run on private hardware in the same self-contained way.
Two motives stand out:
- Cost: paying for cloud API tokens or monthly subscriptions adds up, while hardware is a one-time purchase. Even after counting high-end GPUs, Macs and electricity, heavy users find local computing far cheaper than cloud token prices.
- Privacy: developers building commercial prototypes do not want AI companies using their inputs to train future models.
The trend recalls the home media servers that hobbyists built in the late 1990s and early 2000s, but it may prove stickier, because local AI touches people’s private lives, daily workflows and intellectual property. It also demands serious IT skills, so most consumers are likely to stay with cloud services.
▲ A home AI server built in a garage
Apple’s home event on October 14
Apple is holding an event on October 14 focused on the smart home. It is expected to bring a unified home hub and smart home accessories after several years of work, an attempt to catch up with Google’s and Amazon’s ecosystems. Apple plans to put Siri and AI at the center of the new hub, even though Siri has long trailed competing voice assistants.
Over the past 18 months, Apple has also held events more often. Its traditional calendar, with software at WWDC in the summer and new iPhones in the fall, struggles to keep pace with AI startups that ship updates every couple of weeks, so Apple appears to be creating more moments to stay in the conversation.
Biohub’s $1.8 billion push into AI medicine
The Chan Zuckerberg Biohub, the biomedical research organization of Mark Zuckerberg and Priscilla Chan, has announced an expanded $1.8 billion partnership with government and industry partners. The goal is to use AI models to cure, prevent or manage all diseases. Partners include Meta, Google DeepMind and Isomorphic Labs, a drug discovery company spun out of Google DeepMind. The effort builds on a $500 million program Biohub launched in April to create predictive AI simulations of human cells.
Much of the work is about data: decades of scattered biological data from universities and government agencies must be cleaned and standardized so AI systems can learn from it, in the hope that frontier models will spot disease patterns and drug targets humans have missed.
The partnership also signals a shift in who leads American biomedical research. Since World War II, research universities and academic medical centers have driven it; now the government is working directly with big tech. Universities will stay important because of their large labs, but their dominance over foundational research is fading. AI companies are building labs of their own as well: Anthropic has set up a small wet lab, a facility for hands-on biological experiments. Drug companies are responding with joint ventures or their own AI models, and Biohub plans to bring several global drugmakers into its alliance.
One caveat matters: AI has not yet proven that it can independently discover new drugs faster or better than traditional methods. Tech companies also appear to see health investments as a way to counter public worries about AI risks and job losses. In the end, they will need to deliver real medicines for patients to earn lasting trust.
What to watch, and what to do now
| Player | Move | What to watch |
|---|---|---|
| Microsoft | Claude models in Copilot; Maia chips unveiled | How much TSMC capacity it secures |
| OpenAI | Talks to license Sora to a Hollywood startup | Whether a deal closes, and in what form |
| Home AI builders | Open-weight models on personal hardware | Power and maintenance demands |
| Apple | Home hub event on October 14 | The role of Siri and AI |
| Chan Zuckerberg Biohub | $1.8 billion AI medicine partnership | Drug results confirmed in the lab |
The common thread is that no one wants to depend on a single company, model or chip. Microsoft brought in a rival’s models and built its own silicon, OpenAI would rather license a costly side technology than run it, and power users are choosing their own hardware over the cloud. A few practical steps follow:
- If you choose AI tools for your organization, avoid relying on one model provider. Test several models side by side on your real work, such as spreadsheet formulas or slide decks. That kind of comparison is what brought Claude into Copilot.
- If you use AI heavily and worry about API bills or about your data being used for training, consider running an open-weight model such as Qwen on your own hardware. Check your electrical capacity and your team’s IT skills before you buy anything.
- To see where big tech’s AI strategy is heading, follow three signals: Maia production volumes, the outcome of the Sora licensing talks, and whether AI-generated drug candidates hold up in laboratory tests.