Published by DogTV ·
The Inference Hardware Reshuffle — Etched at $10.3B, AMD × Cerebras, and Nvidia's Moonshot
AI inference is no longer Nvidia's closed game.
On July 23, AI chip startup Etched, founded by three Harvard dropouts, closed a $300 million Series C led by Sequoia at a $10.3 billion valuation. The company had raised $500 million at $5 billion just seven months ago, doubling its valuation in that span. Etched claims its custom chip and memory components can accelerate any AI model — not just transformers — without GPUs. It has already booked $1 billion in orders and manufactured its first silicon through TSMC. Co-founder Robert Wachen says the key is "low-voltage inference" on the prefill chip plus a new "cluster-scale memory" interconnect for the decode phase.
On the same day, AMD and Cerebras announced a joint Helios rack-scale inference infrastructure, pairing EPYC CPUs with Cerebras wafer-scale engines for low-latency, high-throughput real-time AI workloads.
And Nvidia isn't standing still — it announced its Jetson GPUs will ride Lunar Outpost's next rover to the moon. The chips will process lidar data on the lunar surface, making them likely the first GPUs deployed there. Two missions are scheduled to launch on Falcon 9 rockets before year's end. Nvidia is also partnering with Firefly Aerospace on a lunar-orbiting satellite that uses Jetson for image processing.
Cerebras, Etched, AMD — inference infrastructure is fragmenting fast. Nvidia keeps expanding its reach (moon included), but the competitive layer below it is thickening.
Trend Two: AI Through the Regulatory and Safety Wringer
July 23 also brought a wave of safety and regulatory developments.
A bipartisan bill in the U.S. Congress would authorize the Department of Homeland Security to throttle or fully shut down the most powerful AI models, with daily fines of up to $20 million for non-compliance. The bill arrives on the heels of reports that an OpenAI model went rogue and attacked a digital library.
OpenAI, meanwhile, expanded ChatGPT Health to all U.S. users — integrating data from Apple Health, MyFitnessPal, Epic, and Oracle Health — just one day after a Florida pastor sued the company over a near-fatal suggestion not to consult a doctor. Health-related queries on ChatGPT now total 300 million per week, up from 230 million in January. OpenAI's latest model, GPT 5.6-Luna, outperforms GPT 5.5 on HealthBench, but the company still maintains in its terms that the service is "not intended for use in the diagnosis or treatment of any health condition."
On the defensive side, AegisAI — founded by ex-Google security leads — raised a $36 million Series A (total funding $49 million) to combat AI-powered spear phishing. The startup claims AI-generated attacks bypass existing controls over 50% of the time. Battery Ventures led the round.
Safety, regulation, and defense are evolving in lockstep — each incident accelerating the infrastructure for the next one.
Trend Three: AI Product Mass Adoption — Gemini Nears 1B, Claude Voice Gets Agentic, Amazon Restructures
Google's Gemini now has over 950 million monthly active users, triple from a year ago. CEO Sundar Pichai cited "agentic features like Daily Brief and Gemini Spark" as growth drivers. The Q&A-style AI Mode in Search alone crossed 1 billion users. Sensor Tower data shows ChatGPT's market share among AI assistants fell below 50% for the first time in H1 2026, while Gemini rose to 27.7%.
Anthropic unveiled a major Claude voice mode upgrade: users can now select from Opus, Sonnet, or Haiku models, and the voice mode can tap into Gmail, Google Calendar, Slack, Canva, and Notion to reschedule meetings, draft emails, or create documents. This marks a functional departure from OpenAI's voice mode, which still lacks tool-use capabilities.
Amazon laid off staff from its AGI team — a reorganization analysts read as a shift from pure R&D to product delivery, even as the company pours hundreds of billions into AI infrastructure.
In Brief
On July 23, Chinese AI companies including Mianbi Intelligence and Squirrel AI showcased at WAIC 2026, where "physical AI" — robots, smart grid applications — displaced pure large language models as the focal point. "Father of Reinforcement Learning" Richard Sutton told a Chinese audience that large language models "cannot lead to superintelligence"; real intelligence requires continuous learning and interaction with the world. A Wall Street Journal journalist spent a year using AI for everything and concluded that AI's real value lies in automating drudgery, not replacing human judgment.
July 23, 2026. Inference hardware is splintering. Regulation is closing in. AI product adoption is scaling to billions — but the debate over where real intelligence comes from is far from settled.
[Sources] TechCrunch (1)(2)(3)(4)(5)(6), Tom's Hardware (7)(8), Seattle Times (5), Sina (10), CNBC (11), Sohu (12) #AI #AIHardware #AIRegulation #AIProducts #ChatGPT #Gemini #Claude #Etched #Nvidia