Nvidia opens NVLink Fusion, and pays $2B to keep the fabric its own
NVLink licensed to third-party CPUs and accelerators (MediaTek, Marvell, Astera, Lightmatter); a $2B Marvell stake reads as a toll booth, not an investment
Summary
Nvidia is countering the open-fabric push with controlled openness. NVLink Fusion licenses its scale-up interconnect to third-party CPUs and accelerators, partners include MediaTek, Marvell, Alchip, Astera Labs, Synopsys, Cadence and photonics startup Lightmatter, letting custom silicon attach to Nvidia's rack-scale stack (Vera CPUs, ConnectX, BlueField, Spectrum-X) via NVLink-C2C chiplets. On 31 March 2026 Nvidia put $2B into Marvell, which will supply custom XPUs and Fusion-compatible networking. Critics read the stake as "a toll booth, not an investment": rivals' chips may join, but only on Nvidia's interconnect, a defensive move against open UALink and Ultra Ethernet.
By the numbers
- $2B, Nvidia's investment in Marvell (announced 31 March 2026).
- 7+, named Fusion silicon partners (MediaTek, Marvell, Alchip, Astera Labs, Synopsys, Cadence, Lightmatter).
- C2C, NVLink chip-to-chip path extending the fabric to custom dies.
- 2, open rivals it answers (UALink scale-up, Ultra Ethernet scale-out).
Why it matters
Whoever owns the scale-up fabric owns where the next decade's AI margin pools. By letting partners' silicon plug into NVLink rather than defect to UALink, Nvidia keeps the interconnect, and its standards gravity, proprietary while appearing open. The Marvell cash binds a would-be open-fabric supplier into the Nvidia orbit.
What to watch
- Whether Fusion partners also ship UALink/UEC parts or commit to Nvidia's fabric.
- Antitrust/lock-in scrutiny of the Marvell stake and Fusion licensing terms.
- Lightmatter's optical NVLink as the photonics-interconnect wildcard.