PLUS: Meta undercuts OpenAI's pricing, running 70B models locally, and an AI that invents molecules
Happy reading
The high-profile partnership between Apple and OpenAI has completely imploded, with Apple now filing a federal lawsuit against the AI giant. The suit alleges OpenAI stole trade secrets in its push to develop its own competing hardware.
This legal battle signals a major strategic shift in the industry, where AI software leaders are beginning to challenge their hardware partners directly. As the lines between platforms and devices blur, is this the start of a new competitive era for big tech?
In today’s Next in AI:
Apple's lawsuit reveals OpenAI's secret hardware plans
Meta's new model undercuts competitor pricing
How new mini-PCs run 70B models locally
An AI system that invents new molecules
Apple's AI War
Next in AI: Apple has filed a federal lawsuit against OpenAI, alleging the company stole trade secrets to build competing hardware. This stunning move marks a complete reversal of their high-profile partnership from 2024.
Explained:
The lawsuit claims OpenAI’s hardware chief, a former Apple VP, directed job candidates to bring Apple hardware parts to interviews for "show and tell" sessions.
The partnership, which integrated ChatGPT into iOS, soured after OpenAI signaled its hardware ambitions by acquiring former Apple designer Jony Ive's startup for $6.4 billion.
In a clear sign of the fallout, Apple's upcoming Siri update will now use Google's Gemini models, sidelining OpenAI as the tech giants battle for dominance in the AI War.
Why It Matters: This legal battle goes beyond a simple dispute, signaling a new era where AI software leaders are directly challenging their hardware partners. For professionals, this means the lines between AI platforms and the devices they run on are about to get much more competitive and complex.
Meta Crashes the AI Party
Next in AI: Meta’s stock surged after revealing its AI data centers cost half of Wall Street’s estimates, while simultaneously launching a new, aggressively priced AI model designed to undercut competitors and attract developers.
Explained:
A BofA analysis reveals Meta builds its AI compute capacity for $22 billion per gigawatt—roughly half the previously estimated $45 billion—easing investor concerns over its massive spending plans.
The company launched its new Muse Spark 1.1 model with pricing that drastically undercuts competitors, charging just $1.25 per million input tokens compared to Anthropic's $5.
CEO Mark Zuckerberg also confirmed Meta is exploring a new business line: renting its AI compute power to other companies, positioning itself as a potential challenger in the cloud market.
Why It Matters: Meta is leveraging its immense scale to compete on both infrastructure cost and model pricing simultaneously. This dual-front strategy puts significant pressure on competitors and could accelerate AI adoption by making powerful tools more affordable for developers.
Your Desktop AI Server
Next in AI: A new category of powerful mini-PCs, led by the recently released AMD's Ryzen AI Halo, is emerging to run huge 70B+ parameter language models locally. This shifts development from expensive cloud servers directly onto your desktop.
Explained:
The key is unified memory, which combines system RAM and VRAM into a single, massive pool of up to 128GB, allowing these compact machines to load models that are too large for even a 32GB high-end consumer GPU.
There's a trade-off between capacity and speed, as the systems have a lower memory bandwidth than discrete GPUs, resulting in usable but not blazing-fast generation speeds of around 5-6 tokens/second for large dense models.
These devices are being positioned as a turnkey AI appliance, shipping with pre-configured software hubs that let developers immediately start using tools like LM Studio and ComfyUI without a complex setup process.
Why It Matters: This new hardware class gives developers and researchers the ability to experiment with large-scale AI privately and without ongoing cloud costs. It marks a significant step toward decentralizing AI power, which could fuel a new wave of personalized and secure AI applications.
AI Invents a Molecule
Next in AI: Researchers have unveiled CoCoGraph, a new AI system that designs entirely new, chemically-valid molecules. This breakthrough promises to dramatically speed up the discovery of new drugs and advanced materials.
Explained:
The system, detailed in Nature Machine Intelligence, uses a diffusion model to learn how to build molecules while ensuring every creation is chemically valid from the start.
CoCoGraph operates more efficiently than competing models, using fewer parameters and less computing power while creating molecules that are more realistic across dozens of physicochemical properties.
In a test against 121 chemists, the AI-generated molecules were so convincing that the experts were fooled about 40% of the time, highlighting the system's ability to produce highly plausible results.
Why It Matters: This technology moves molecular discovery from slow trial-and-error to rapid, targeted creation. It opens the door for scientists to explore a vast, uncharted chemical universe, accelerating the development of next-generation medicines and materials.
AI Pulse
SK Hynix plans a massive $26.5 billion American depositary receipt (ADR) offering in the U.S., capitalizing on soaring investor demand for AI-related hardware stocks.
Palo Alto Networks' CEO argued that AI token prices must fall by 90% before large-scale enterprise adoption becomes practical, highlighting the significant cost barriers still facing businesses.
The University of Chicago banned laptops and phones in first-year law classes to ensure students learn to think independently without over-relying on AI tools.
Meta reversed course on its new Muse Image tool, pulling the feature that generated AI images from public Instagram profiles after significant user backlash over privacy.