Theses›Datacentre & AI compute›Scale-up interconnect

Scale-up interconnect

The scale-up fabric is the vendor-agnostic toll - merchant silicon plus the NVLink tax

  • Closing
  • $162B by 2035

The call, a draftThe fabric is winning: the limit on AI clusters has moved from compute to the bandwidth linking GPUs inside a rack. Every accelerator pays for it, and Broadcom, Marvell and NVIDIA hold it, so the window is late.

D1D2D3D4R1R2R3C1C2C3foragainst D1D2D3D4R1R2R3C1C2C3foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the drivers.
  1. D1Copper reaches only about a metre at 200 Gb/s, so a 72-GPU rack needs retimers, active cables and optics to wire it together.
  2. D2NVLink Fusion means custom accelerators attach through an NVIDIA link chip, and NVIDIA invested $2B in Marvell on 31 Mar 2026.
  3. D3Astera Labs grew revenue 93% to $308.4M in Q1 2026, and Marvell guided data-centre interconnect up more than 70% in fiscal 2027.
  4. D4Sparse mixture-of-experts models shift the memory problem onto the network, because 64 accelerators exchange data all the time.
  5. R1Astera Labs and Credo trade at roughly 39 and 42 times sales, so the growth is already priced.
  6. R2AMD's Helios rack targets mass production in Q2 2027 on UALink over Ethernet, giving buyers a route around NVLink.
  7. R3Custom AI processor shipments grew 44% against 16% for GPUs, and hyperscalers are building their own interconnects.
  8. C1Two or more of the merchant interconnect makers keep growing above 30% a year while NVIDIA's GPU share falls.31 Dec 2027
  9. C2Two or more custom accelerators from other chip designers ship on NVLink Fusion.31 Dec 2027
  10. C3An NVLink or UALink domain larger than 72 GPUs ships in production.31 Mar 2027
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
20272028today12345predictionswatch 20272028today12345predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1The scale-up fabric toll decouples from accelerator share: through end-2027 at least two of {Astera Labs, Credo, Marvell interconnect/custom} sustain >30% YoY scale-up/interconnect revenue growth EVEN IF NVIDIA's data-centre GPU share declines from its ~80%+ peak.Dec 2027 · open
  2. 2NVLink stays the dominant integrated-rack scale-up fabric: no merchant UALink scale-up switch reaches volume production in an NVL72-class integrated rack before end-2027.Dec 2027 · open
  3. 3The NVLink tax persists for escapees: by end-2027 at least two third-party custom accelerators ship attached via NVLink Fusion.Dec 2027 · open
  4. 4FY2027 results: merchant scale-up revenue still compounding over 30% while NVIDIA's data-centre GPU share is flat or down.Dec 2027 · open
  5. 5Whether any merchant UALink switch reaches volume inside an NVL72-class rack.Dec 2027 · open
Expressionslisted, private, technologies
Private · 7 private companies, 7 of them in the US; none in continental Europe.

CScale · Point2 Technology · AttoTude · Upscale AI · Ayar Labs · Celestial AI · Lightmatter

4 more private companies tracked, not shown

Technologies
What changedthe record, newest first
  • 12 Aug 2026A US hardware fund told me this week, unprompted, that chip-to-chip interconnect is the next bottleneck now that packaging and memory are being solved. I happen to agree, and that is exactly the problem. When the investors pricing the early rounds all hold the same view, agreeing with them stops being an edge and starts being the entry price. The discipline is logging WHO holds a view and WHEN, because a thesis dies quietest when it becomes the consensus while you are still congratulating yourself for holding it.
  • 21 Jul 2026The physical reason the industry is going optical inside the rack: at 200 gigabits a second a copper cable only carries about a metre, and doubling the speed halves that again, so the cables are simply too short to wire a 72-GPU machine together. Nvidia has the fix scheduled for 2028.
  • 20 Jul 2026Every country that tries to build a sovereign AI chip discovers the same thing: the chip is one node, and the wiring between chips belongs to somebody else. The wiring is the toll.
4 earlier
  • 19 Jul 2026Two connectivity companies went from nothing to about $2.2bn of combined revenue in three years, and neither did it by building the glamorous part. They sold retimers and cables into everyone's rack. The lesson is that the money in interconnect is at the least interesting layer, and it is public.
  • 17 Jul 2026Sparse mixture-of-experts models don't reduce the memory problem, they move it into the network. K3 needs 64 accelerators talking constantly over a fast fabric. This may be the starting gun for the Interconnect Trade.
  • 16 Jul 2026The money keeps flowing to the plumbing between AI chips rather than the chips alone, and this quarter a third physical medium (radio over a plastic waveguide) joined copper and optics with two funded players. When the medium is contested, the durable position is the layer that works across all of them.
  • 23 Jun 2026The plumbing between AI chips is now a venture battleground: an AI-networking startup just raised at a $2B valuation with NVIDIA itself investing - evidence the scale-up interconnect fabric is where durable value is pooling.