Theses›Datacentre & AI compute›Computational CXL memory

Computational CXL memory

Computational CXL memory (near-data processing pushed INTO the pooled-memory tier: KV-cache management, preprocessing, vector search running on cores embedded in CXL memory devices, the XCENA and Dnotitia pattern) is a distinct lane from the CXL pooling/switching silicon already recorded as…

  • Bear
  • Closing
  • $12B by 2035

The callRestraints win until a hyperscaler proves it: Astera, Marvell, Samsung and SK hynix can add compute to pooled memory as a feature, and both funded entrants are priced.

D1D2R1R2R3C1foragainst D1D2R1R2R3C1foragainst
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 restraints.
  1. D1AI inference is becoming a memory problem: every long conversation and agent holds a KV cache that grows with context, and HBM cannot be bought in the volumes wanted.
  2. D2Plain CXL pools add capacity but haul every byte back to the processor, so compute inside the pool saves link bandwidth and power.
  3. R1Astera, Marvell, Samsung and SK hynix can add near-data processing as a feature, and if they do the category folds into the controller market they already hold.
  4. R2The two funded entrants are growth-stage, XCENA on $185M and Dnotitia on about $61M, there is no European entrant, and the best exit is Enfabrica's acqui-hire for over $900M.
  5. R3CPU support for CXL has slipped for years, and if HBM supply normalises after 2028 the cost case for a capacity tier weakens.
  6. C1A hyperscaler discloses production KV-cache offload on computational CXL memory with total-cost economics, XCENA or a peer, showing people pay for near-memory compute as a product.end-2027
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
202720282029today123predictionswatch 202720282029today123predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1By end-2027 computational CXL memory ships in at least one named hyperscaler or AI-cloud production deployment with disclosed workload economics (XCENA or a peer).Dec 2027 · open
  2. 2A named hyperscaler or AI cloud runs computational CXL in production with disclosed economics; that would justify a proper sizing.Dec 2027 · open
  3. 3HBM supply after 2028, and any European device-layer entrant.Dec 2028 · open
  • ·An Astera, Marvell, Samsung or SK hynix acquisition or in-house product absorbs the function, confirming the bear case.
Expressionslisted, private, technologies
  • ALABAstera Labs
    • full
     Aries retimers, Taurus smart cable modules, Scorpio switches.
  • MRVLMarvell Technology
    • cheap
     Inphi optical DSP + custom AI ASIC for AWS.
  • MUMicron
    • cheap
     HBM3E/HBM4 — one of three suppliers (vs Samsung, SK Hynix).
  • 005930.KSSamsung Electronics
    • cheap
     CXL memory expander modules (CMM-D) built from its own DRAM, sold to server OEMs and hyperscalers; also fabs XCENA's MX1.
  • 000660.KSSK hynix
    • cheap
     CXL DRAM expander modules (CMM-DDR5) and a computational memory solution prototype, sold to server OEMs and hyperscalers.
  • 688008.SSMontage Technology
    • not screened
     CXL memory expander controllers (MXC) supplied to DRAM module makers.
  • MCHPMicrochip Technology
    • cheap
     SMC 2000 series CXL smart memory controllers sold to memory-module and server makers.
Private · 6 private companies, 3 of them in South Korea; none in continental Europe.

XCENA · Dnotitia · Primemas · Panmnesia · Enfabrica · MemVerge

Technologies
Who has already bet on itprivate rounds, from public sources

$648M raised by 6 companies, $471M of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a top-tier venture firm: Enfabrica. Backed by a strategic investor: Enfabrica, MemVerge.

What changedthe record, newest first
  • 16 Jul 2026The interesting move in memory this quarter was compute sneaking INTO the memory pool: two well-funded teams are betting that managing an AI model's working memory is better done by the memory device itself than by the GPU. Whether that becomes a product category or a feature the giants absorb is the whole question.
Reading
Show the reading