Apple Google Gemini deal for Siri
Apple has reportedly struck a landmark deal to integrate Google’s Gemini models into a revamped version of Siri, marking a significant shift in Apple’s AI strategy. The move would allow Gemini to power advanced reasoning, generative responses, and new on-device AI capabilities across iPhone, iPad, and Mac.
Rather than relying on a single provider, Apple appears to be positioning Siri as a multi-model system, where different AI engines are selected based on task, privacy needs, and performance constraints. If finalized, this would put Gemini in front of hundreds of millions of Apple users almost overnight.
Beyond the technical leap, the deal raises broader questions around data governance, platform control, and competition especially as Apple has previously leaned on OpenAI for generative AI experiments.
Why It Matters:
This isn’t just a Siri upgrade it’s a platform-level realignment. If Gemini becomes a core intelligence layer inside Apple’s ecosystem, it could reshape the mobile AI race and weaken the advantage of standalone AI apps. It also signals that even Apple is embracing a more modular, partner-driven approach to frontier AI.
OpenAI’s $10B Cerebras compute deal
OpenAI has signed a multi-year agreement with Cerebras valued at over $10 billion, securing access to as much as 750 megawatts of AI compute through 2028. The deal focuses heavily on high-throughput inference, a growing bottleneck as AI agents and real-time applications scale.
This marks one of OpenAI’s most decisive steps away from exclusive reliance on NVIDIA GPUs, validating wafer-scale architectures for real production workloads. Cerebras’ systems are designed to process massive models with fewer communication overheads making them especially attractive for inference-heavy use cases.
The sheer scale of the contract underscores a sobering reality: frontier AI is now constrained less by algorithms and more by power, silicon, and data center capacity.
Why It Matters:
This deal signals a diversification of the AI compute stack and a warning shot to GPU incumbents. As inference demand explodes, alternative architectures like wafer-scale chips may become essential, not optional. It also highlights just how energy-intensive the next wave of AI agents will be.
NVIDIA’s new LLM training method (TTT‑E2E)
NVIDIA has published TTT-E2E (Test-Time Training End-to-End), a new approach that allows large language models to learn from incoming context while keeping runtime latency constant. The technique directly addresses one of the biggest limitations of today’s LLMs: they can read long context, but they don’t truly adapt to it.
With TTT-E2E, models update internal representations on the fly without the usual performance penalties. This makes it especially relevant for streaming data, enterprise knowledge bases, and agentic systems that must continuously adjust to new information.
Rather than treating inference and learning as separate phases, NVIDIA is effectively blurring the line between the two.
Why It Matters:
If it is scalable, TTT-E2E can potentially unleash a whole new generation of adaptive AI models that can learn alongside information rather than simply remember it. This is a very important milestone on the road towards developing AI models that can persistently interact in real-time or remain contextually current on their own.
Liked this issue? Get the next one free.
One email on Tuesday. Unsubscribe in one click.