Topic
Covers the physical and technical foundations of AI systems — data centers, chips, power, cooling, networking, and the companies building compute capacity at scale.
Nvidia is not writing OpenAI a $250 billion check. It is co-signing a lease, because the bond market wouldn't lend to OpenAI on its own.
TSMC's $265 billion Arizona pledge lands the same day it books a record $22 billion quarter, with 2nm and CoWoS packaging at the center.
OpenAI's Jalapeño chip taped out in nine months, targets inference costs, and still leaves Nvidia running the training side.
Intel's Xe3P inference accelerator skips HBM entirely, packing more memory than any Nvidia or AMD data center GPU shipping in 2026.
Qualcomm's all-stock deal for Chris Lattner's Modular targets the software moat that has protected Nvidia's AI dominance more than any GPU spec.
Alphabet's $84.75 billion equity raise includes $10 billion from Berkshire Hathaway, but roughly $30 billion is earmarked for taxes, not GPUs.
Maia 200 ships with 216GB HBM3e, native FP4, and a potential Anthropic deal that would crack Nvidia's grip on frontier-model inference.
Record quarter, a 25× dividend hike, and an $80 billion buyback land alongside a China Data Center number that has collapsed to zero.
The WSE-3 is 57 times larger than an H100 and serves AI tokens from on-chip SRAM at 21 PB/s. Here's what that actually changes.
The Cerebras inference API runs Llama 3.3 70B at 2,314 tokens per second, but only if you configure streaming, model choice, and TCP warming correctly.
OpenAI, Microsoft, and NVIDIA quietly rebuilt Ethernet for frontier AI training. Here's how the Multipath Reliable Connection protocol actually works.
Anthropic just took over SpaceX's entire Colossus 1 data center, adding 220,000 GPUs to fix Claude's reliability problems and double Claude Code limits.