Computing Interconnects

NVLink

Critical1/4 relationships sourcedProfile verified Sep 1, 2026

30-Second Executive Brief

Executive Assessment

NVLink functions as critical infrastructure within Computing Interconnects, backed by 28% sourcing coverage, though its relative influence is cooling.

  • Maintains 4 mapped relationships across the graph
  • 28% sourcing coverage across sourced relationships
  • Tracked as critical infrastructure within the Computing Interconnects category

Executive Snapshot

Strategic Role
Critical Infrastructure
Sourcing Coverage
28%
View Methodology →

Computed live from this entity's own relationships and evidence: how many relationships carry at least one linked citation, weighted with citation recency and source type — independent of Strategic Importance, and not a prediction. Not a hand-typed number; recalculated on every read.

Ecosystem Influence
Very High
Strategic Momentum
Decelerating

Coverage

Mapped Relationships
4
Technology Domains
4

Strategic Implications

  • Central node connecting multiple strategic ecosystems
  • Directly influences technology and capital flows
  • Material relevance to downstream dependency mapping

Top Opportunities

NVLink 6 and the Rubin platform now in full production, targeting up to 10x lower inference token cost and 4x fewer GPUs needed to train MoE models versus BlackwellGrowing importance as mixture-of-experts model architectures increase inter-GPU communication demands

Top Risks

NVLink coherence is proprietary to NVIDIA hardware, creating lock-in that prevents mixed-vendor GPU clustersScaling beyond a single NVLink domain requires falling back to slower InfiniBand or Ethernet interconnects

Critical Dependencies

Continue to Dependency Graph ↓

NVIDIA's proprietary high-speed GPU-to-GPU interconnect that lets multiple GPUs act as a single coherent compute fabric; the fifth-generation version shipped with Blackwell at 1.8TB/s per GPU, and NVLink 6 has since entered full production with the Rubin platform, delivering 3.6TB/s per GPU and 260TB/s across a Vera Rubin NVL72 rack.

Sources

Every claim traced to a primary source — evidence, recent activity, and insider filing behavior, all in one place.

Evidence

2 sources

  • NVIDIA NewsroomFifth-generation NVLink delivers 1.8TB/s per-GPU bandwidth across up to 576 GPUscompany
  • NVIDIA NewsroomNVLink 6 enters full production with NVIDIA's Rubin platformcompany

Relationship Map

The relationships surrounding NVLink — ownership, dependencies, regulation, technology and market context. Click any node to make it the new center, 2 levels deep.

Connection type

Sign in free to click a node and explore →

Click any node to make it the new center. Scroll to zoom, drag to pan.

The Story So Far (last 6 months)

Mar 2026: NVIDIA and Marvell announced a strategic partnership connecting Marvell to NVIDIA's AI fa…Jun 2026: Lightmatter joins NVLink Fusion with photonic interconnectsJun 2026: NVIDIA confirmed Vera Rubin entered full production with NVLink 6, delivering 3.6TB/s of…Jun 2026: NVLink 6 enters full production with NVIDIA's Rubin platformAug 2026: Opening NVLink to third-party silicon via NVLink Fusion (Marvell, Lightmatter) while NVLi… (unconfirmed)Aug 2026: NVLink Fusion expands with NVHBM; Amazon Annapurna Labs first partner

Auto-generated from this entity's dated milestones, relationship updates, and sourced evidence — not AI-written, just sorted.

Market Intelligence

Unverified

Credibly-reported claims — analyst notes, sourcing citing “people familiar with the matter,” deals where the companies involved declined to comment — that haven't been officially confirmed. Kept structurally separate from the sourced evidence above; treat as a lead worth researching further, not an established fact.

Opening NVLink to third-party silicon via NVLink Fusion (Marvell, Lightmatter) while NVLink itself remains the proprietary lock-in mechanism is a calculated tension -- NVIDIA appears to be trading a degree of exclusivity for deeper ecosystem entrenchment, letting custom XPUs plug into NVIDIA's fabric rather than customers building fully independent interconnect stacks

30% confidence

This is a defensive-offensive move: it reduces the incentive for hyperscalers to build NVLink-independent scale-up fabrics (which would erode NVIDIA's lock-in over time), while still keeping NVIDIA as the switching and coherence layer every participant depends on.

This is InsightNodes' own interpretive read on the strategic logic behind NVLink Fusion's third-party openness; NVIDIA's own announcements frame Fusion as ecosystem expansion without addressing the lock-in tradeoff.

NVIDIA Newsroom · Aug 14, 2026

NVLink's Timeline

A sourced, dated history of NVLink's key moments — founding to present.

  1. Jan 2026 · NVLink 6 enters production with Rubin

    NVLink 6 entered full production with NVIDIA's Rubin platform, delivering 3.6TB/s per GPU and 260TB/s across a Vera Rubin NVL72 rack.

  2. Jun 2026 · Lightmatter joins NVLink Fusion with photonic interconnects

    Photonics startup Lightmatter joined the NVLink Fusion ecosystem, following Marvell as an early third-party partner, to deliver co-packaged optics (CPO) and near-packaged optics (NPO) products compatible with NVIDIA's optical and SerDes technologies for scale-up AI infrastructure.

  3. Aug 2026 · NVLink Fusion expands with NVHBM; Amazon Annapurna Labs first partner

    NVIDIA expanded NVLink Fusion with NVHBM, a high-bandwidth-memory approach for custom XPUs that can free up to 25% of XPU die area for compute, cut HBM power by up to 15%, and boost effective bandwidth by up to 30%; Amazon's Annapurna Labs is the first partner working with NVIDIA on the technology, following earlier 2026 NVLink Fusion ecosystem additions including Marvell (backed by a $2 billion NVIDIA investment) and Ayar Labs.

How These Connections Evolved

How a relationship changed over time, not just its current state — sourced, dated, and traceable.

NVLinkNVIDIA

  1. Jun 2026

    NVIDIA confirmed Vera Rubin entered full production with NVLink 6, delivering 3.6TB/s of bandwidth per GPU and 260TB/s across a Vera Rubin NVL72 rack — double the bandwidth of fifth-generation NVLink shipped with Blackwell.

  2. Now

    NVLink is NVIDIA's proprietary interconnect technology, developed and controlled exclusively by NVIDIA.

NVLinkBlackwell

  1. Mar 2024

    Fifth-generation NVLink shipped with Blackwell delivering 1.8TB/s of bidirectional bandwidth per GPU, enabling coherent communication among up to 576 GPUs for large language and mixture-of-experts models.

  2. Now

    Fifth-generation NVLink enables Blackwell GPUs to operate as a single coherent fabric across up to 72-576 GPUs for large-scale model training.

NVLinkBlackwell

  1. Mar 2024

    Blackwell's GB200/GB300 NVL72 rack-scale systems relied on fifth-generation NVLink's 1.8TB/s per-GPU bandwidth to connect up to 72 GPUs as a single coherent fabric.

  2. Now

    Blackwell GPUs depend on fifth-generation NVLink to operate as a coherent multi-GPU fabric for large-scale model training.

NVLinkMarvell Technology

  1. Mar 2026

    NVIDIA and Marvell announced a strategic partnership connecting Marvell to NVIDIA's AI factory ecosystem through NVLink Fusion, with NVIDIA investing $2 billion in Marvell and the companies also collaborating on silicon photonics.

  2. Now

    NVLink Fusion, unveiled in 2026, opens NVIDIA's interconnect to third-party silicon for the first time; Marvell is the first partner, building custom XPUs and scale-up networking compatible with NVLink Fusion as part of a deal that included a $2 billion NVIDIA equity investment in Marvell.