The AI Race in 2026: Anthropic, OpenAI, and Apple
Anthropic secures massive compute capacity, OpenAI builds new networking technology, Apple opens up to third-party models — and Google pushes Gemma toward higher speed.
Inhaltsverzeichnis
Today is less about flashy demo videos and more about what really powers the AI revolution behind the scenes: compute, networks, platform power, and regulation. If you want to understand why the big players are currently pulling in so much money, energy, and political attention, you’re in the right place today.
The news makes one thing pretty clear: the AI race is shifting from “Who has the best model?” to “Who controls infrastructure, distribution, and the rules?” Spoiler: that is rarely the phase where things get comfortable.
🚀 Anthropic secures 300 MW of compute from SpaceX
According to The Decoder, Anthropic has secured the full compute capacity of SpaceX’s Colossus-1 data center: over 220,000 NVIDIA GPUs and more than 300 megawatts of power, apparently available within just one month. In parallel, Anthropic is raising rate limits for Claude Code and loosening API limits for the Opus models. The message is clear: demand outpacing capacity is no longer a side issue for frontier models, but a business reality.
Why does this matter? Because compute today is not just a cost issue, but a strategic moat. Whoever has more compute can serve more users, roll out more features, and iterate faster. The idea of even exploring orbital AI compute may sound like science fiction, but it shows how seriously bottlenecks are being taken now. For you, that means the major model providers will keep being judged on how well they balance availability, price, and performance. Source: The Decoder
🌐 OpenAI builds a new network protocol with chip giants
OpenAI has developed the open-source networking protocol MRC together with AMD, Broadcom, Intel, Microsoft, and NVIDIA, which distributes data across hundreds of paths simultaneously between GPUs. Instead of three or four switch layers, MRC only needs two to connect supercomputers with more than 100,000 GPUs. The protocol is already running in OpenAI’s Stargate supercomputer.
Why is this important? Huge. Because in AI supercomputers, raw GPU power is not the only decisive factor — what matters even more is how quickly the GPUs can talk to each other. If data is routed like a nervous delivery driver through too many switch layers, the entire system slows down. MRC is meant to streamline exactly that. The fact that OpenAI and the major chip and cloud companies are working together here is a very clear signal: infrastructure is becoming a shared standardization issue. For the industry, this could lead to more efficient data centers and fewer bottlenecks in the long term. Source: The Decoder
🍏 Apple could let you use your favorite model in iOS 27
According to The Verge, Apple could allow third-party chatbots system-wide for Apple Intelligence for the first time with iOS 27, iPadOS 27, and macOS 27. That would mean: instead of relying only on Apple’s own AI logic, you could choose an external model depending on the task. For Siri, system functions, and other integrated experiences, that would be quite a major shift.
Why is this relevant? Apple has so far been more controlled than experimental when it comes to AI. Opening up to third-party models would make the ecosystem more flexible and could make Apple Intelligence significantly more attractive for users who already want to choose between multiple models. At the same time, it’s also an admission: no provider has to build everything alone when users have already learned to think in multi-model terms. For developers and platform providers, this could create new integration opportunities — and for Apple, the chance to turn a “too cautious” AI approach into a real platform feature. Source: The Verge
⚡ Google makes Gemma 4 up to three times faster
Google has released a multi-token prediction drafter for its open model family Gemma 4, which, according to The Decoder, should speed up text generation by up to three times. The principle is clever: a small helper model proposes multiple words at once, and the main model checks those suggestions in a batch instead of working token by token.
That sounds technical, but in practice it matters a lot. Because speed in LLMs is more than convenience — it affects cost, latency, and therefore whether a model is actually usable in everyday apps. Especially with open models, performance is a crucial lever for staying competitive with proprietary systems. For you, that means open-source models are becoming not just smarter, but also more usable. And that’s exactly where things start getting interesting. Source: The Decoder
🕵️ US agency to review AI models before release
Google, Microsoft, and xAI have agreed in a new arrangement to have their AI models reviewed by the US government before release, reports heise. This is a noteworthy step because it shifts state oversight to before market launch rather than afterward. For AI systems with broad reach, a new form of safety and governance process is being tested here.
The context matters: this is not yet full regulation for the entire industry, but it is a strong signal of where things could be headed. If major providers accept pre-release reviews voluntarily or شبه-voluntarily, that creates facts on the ground for later standards. At the same time, the questions remain: how deep these reviews go, how transparent they are, and who ultimately defines what counts as “safe enough.” In short: the political debate is leaving the PowerPoint stage. Source: heise
🤖 Google and Meta are chasing the market leaders with AI agents
Google and Meta are internally testing new personal AI agents called “Remy” and “Hatch” that are supposed to complete everyday tasks autonomously. What is especially interesting here is the strategic shift: Google has even discontinued its browser agent project Mariner for this. Instead of putting the browser center stage, integrated assistants in mail, calendar, and shopping platforms are now moving into focus.
This shows very clearly where the market is heading. Browser agents were the first big hype, but in practice deeply integrated assistants are often more useful because they work directly where your data and workflows actually are. For Google and Meta, the goal is to stop OpenAI and Anthropic’s lead in agents from growing further. For users, this could eventually mean less “chatting for the sake of chatting” and more real task automation. AI is moving from the showroom into everyday life. Source: The Decoder
🛠️ Tool tip of the day
If you want to keep a close eye on developments in open-source models, infrastructure, and agents, a monitoring setup is worth its weight in gold. Especially for topics like LLM deployments, network protocols, and model release cycles, good research software saves you a lot of time — and even more tab chaos. #
Don’t want to miss any news? Subscribe to the newsletter