🕶️ Meta just bought 5 years of GPUs

SiFive announced $400 million funding led by Atreides and Nvidia, valuing the company at $3.65 billion. The money backs its push to build high‑performance data‑center CPUs based on RISC‑V, an open instruction set architecture (the common language between software and chips). SiFive sells processor IP —licensed blueprints other companies use to make their own silicon— and says hyperscalers want customizable CPUs for AI data centers.

This points to a broader shift toward open standards in the data center as “agentic AI” grows. SiFive also highlighted closer ties with Nvidia interconnects, linking to plans that detail NVLink Fusion, which aims to move data faster between chips.

The least obvious ripple: software vendors and OS distributions may accelerate RISC‑V support, since ports of CUDA (Nvidia’s GPU programming platform), Red Hat, and Ubuntu already exist. Foundries and design houses could see more demand for custom cores as hyperscalers test alternatives to proprietary ISAs like x86 and Arm. Incumbent CPU providers may face pricing and lock‑in pressure if open, lower‑power designs meet performance targets.

If you build chips, compilers, or data‑center orchestration software, watch buyer interest in “customizable CPU solutions in IP form.” For AI infrastructure founders, opportunities sit in toolchains, runtime schedulers, and observability that optimize CPU‑GPU coordination and power use. Systems‑software teams should budget for RISC‑V builds, CI, and performance profiling now, so you’re ready when early access hardware lands.

Costs are rising. Clients are paying slower. Hiring feels riskier than ever.

And every day brings another hit.

The Survival Hub gives you practical, in-the-trenches support to respond:

Built for leaders navigating uncertainty.

Staying standing isn’t about doing more. It’s about knowing what to do next.

Explore BELAY’s Survival Hub.

Meta will spend an additional $21 billion with CoreWeave for AI cloud capacity from 2027 to 2032, on top of a prior $14.2 billion agreement through 2031. CoreWeave also plans to bolster its balance sheet and raises $3 billion in new convertible debt. The deal signals Meta’s intent to scale training and inference on GPUs while it ramps its own buildout.

This locks in multi-year supply during a GPU shortage and shows even “hyperscalers” — very large cloud buyers — are hedging with outside partners. At the same time, Meta is not going all-in on outsourced compute; it will also build Texas data center capacity, pointing to a hybrid model of own-data-centers plus specialized cloud.

If you’re an AI infrastructure or foundation-model founder, assume reserved capacity is the new normal and plan for multi-cloud portability to avoid training bottlenecks. If you build tools for scheduling, observability, or cost control on GPU clusters, demand should rise as buyers juggle in-house and third-party capacity. For application startups, model roadmap risk now includes compute access and timing,  design features, and release plans around when you can actually get GPUs, not just when models are ready.