The Signal — April 8, 2026
Anthropic just signed a deal for 3.5 gigawatts of TPU capacity from Google and Broadcom, its revenue run rate has crossed $30 billion, and the compute arms race now has a price tag measured in power plants. Meanwhile, Samsung's Q1 profit forecast exceeds its entire 2025 annual earnings on AI chip demand alone. Microsoft quietly shipped three new in-house models, and the FCC banned Chinese-made routers with phones and cameras potentially next.
Anthropic Crosses $30B Revenue, Signs Multi-Gigawatt TPU Deal
Anthropic announced a compute agreement with Google and Broadcom for next-generation TPUs coming online in 2027. The deal covers roughly 3.5 gigawatts of TPU-based compute capacity, mostly US-based, aligned with the company's $50 billion domestic investment commitment.
The numbers backing the deal: Anthropic's annualized revenue run rate crossed $30 billion, more than tripling from $9 billion at the end of 2025. Over 1,000 enterprise customers now spend at least $1 million annually, a figure that doubled in under two months.
Three and a half gigawatts is roughly the output of three nuclear power plants, committed to a single company's compute needs for a single chip architecture. That's where frontier AI training sits now.
Sources: Anthropic, TechCrunch, Bloomberg
Samsung Q1 Profit Exceeds All of 2025
Samsung Electronics projected Q1 operating profit at roughly 57.2 trillion won (about $38 billion), an eightfold increase that would surpass its entire 2025 annual profit in a single quarter. SK Hynix shares jumped 15% on the news. Korea Investment raised SK Hynix's full-year profit estimate by 28% to 216 trillion won ($146.5 billion), which would be four times its 2025 levels.
AI infrastructure demand is stretching chip supply and pushing memory prices sharply higher. HBM (high-bandwidth memory) is the specific bottleneck. Every AI accelerator needs stacks of it, and Samsung and SK Hynix are two of the three companies in the world that can make it. When the factories selling shovels report profits like this, it tells you how much digging is actually happening.
Sources: Reuters, Reuters (SK Hynix)
Microsoft Ships Three In-House AI Models
Microsoft released three new models under its MAI brand: MAI-Transcribe-1 for speech-to-text ($0.36/hour, beating Whisper and Gemini Flash in 11+ languages), MAI-Voice-1 for text-to-speech ($22 per million characters), and MAI-Image-2 for image generation ($5/$33 per million tokens in/out). All are available on Microsoft Foundry.
The pattern here matters more than any individual model. Microsoft keeps building its own model stack across modalities while maintaining its OpenAI partnership. These prices undercut Google and OpenAI's comparable offerings. If you're an enterprise already on Azure, the switching cost to try Microsoft's in-house models is approximately zero. That competitive pressure is the point.
Sources: Microsoft AI
FCC Bans Chinese-Made Routers, Phones and Cameras Could Follow
The FCC has banned new consumer-grade Wi-Fi routers manufactured outside the US, with Chinese telecom and surveillance equipment as the primary target. Huawei and Hikvision could see final import orders cut within 30 days. The administration is considering expanding the ban to phones and cameras across critical infrastructure.
This is the broadest consumer tech restriction since the Huawei entity listing. Routers are the entry point, but the scope under consideration covers a much wider slice of consumer electronics. As US-China tech decoupling accelerates across hardware categories, the compute supply chain for AI training and inference gets squeezed on both sides.
Sources: CNET, TechRepublic
