Tag

#inference

Articles tagged inference.

Microsoft Maia 300: Specs, Rack Design and the 2027 Deployment Plan
Hardware & Devices

Microsoft Maia 300: Specs, Rack Design and the 2027 Deployment Plan

Microsoft unveiled Maia 300, its third in-house AI accelerator, built on TSMC N2P with eight HBM4 stacks and a 950-watt power draw. A rack holds 96 of them, production deployment starts in 2027, and Microsoft says Maia will become its primary inference system.

Maya Iverson
Qualcomm Is Paying Amazon in Warrants to Buy Up to $60B of Its Chips
Markets & Money

Qualcomm Is Paying Amazon in Warrants to Buy Up to $60B of Its Chips

Amazon can buy up to $60 billion of Qualcomm data-center chips and optics over 10 years. Qualcomm issued it a warrant on 25 million shares at $161.26 that vests as orders and payments happen, with 3.75 million shares vested on the initial commitment.

Sam Whitaker
Anthropic Walked Away From a $6B Decart Deal After Doing the Diligence
Big Tech

Anthropic Walked Away From a $6B Decart Deal After Doing the Diligence

Anthropic ended talks to acquire Israeli startup Decart for roughly $6 billion after completing due diligence, according to Bloomberg. It would have been the company's largest known acquisition, and Decart builds the cross-silicon inference optimization that Anthropic's lease book makes valuable.

Sam Whitaker
Nvidia Is Talking to Rebellions, a $2.3B Korean Inference-Chip Startup
Hardware & Devices

Nvidia Is Talking to Rebellions, a $2.3B Korean Inference-Chip Startup

Jensen Huang met Rebellions CEO Sunghyun Park in Santa Clara to discuss a partnership, an investment, or a possible acquisition. Rebellions builds data-center NPUs tuned for inference and is backed by SK hynix, Samsung Ventures, and Arm.

Maya Iverson
Setting AI in Stone
Big Tech

Setting AI in Stone

Etched raised at $10.3 billion for transformer-only chips. AMD just bought Taalas, which etches model weights into transistors. Hank Reed on an industry betting that AI has stopped moving.

Hank Reed
AMD Bought Taalas, Whose Chips Burn Model Weights Straight Into Transistors
Hardware & Devices

AMD Bought Taalas, Whose Chips Burn Model Weights Straight Into Transistors

AMD agreed to acquire Toronto-based Taalas, whose chips hardwire an AI model's weights permanently into silicon to eliminate the memory reads that cap GPU inference speed. Taalas claims 17,000 tokens per second on Llama 3.1 8B at a tenth of an H200's power.

Maya Iverson
Etched Raised $300M at $10.3B for a Chip That Only Runs Transformers
Startups & Funding

Etched Raised $300M at $10.3B for a Chip That Only Runs Transformers

Etched closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia with a16z and SK hynix. Its Sohu chip hardcodes the transformer architecture and claims a large inference speedup over Nvidia, with over $1 billion in signed contracts.

Jordan Beck
Fireworks Raised $1.5B at $17.5B While Serving 40 Trillion Tokens a Day
Startups & Funding

Fireworks Raised $1.5B at $17.5B While Serving 40 Trillion Tokens a Day

Fireworks closed a $1.5 billion Series D at a $17.5 billion valuation with Nvidia participating. It says it passed $1 billion in annualized revenue, and 95% of its traffic runs specialized models.

Jordan Beck