Theses›Datacentre & AI compute›Optical memory

Optical memory

The restraints are winning: HBM is scarce, but standalone optical memory is a roughly $50M research market and nothing ships at volume. HBM4 and HBM5 supply should catch up around 2029. The window is early.

  • Early

The call, a draftHBM scarcity opens room for memory that holds bits in light, or in media addressed by light, but most approaches store the bit in the light itself, which caps how small a cell can get. Memory is a commodity layer, so even a working platform tends to hand its value to Micron, SK Hynix and Samsung or to the photonic compute system around it, as Marvell's $3.25B purchase of Celestial AI shows.

D1D2D3R1R2R3C1C2foragainst D1D2D3R1R2R3C1C2foragainst
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. D1HBM is sold out into 2027 and growing from about $34B toward $98B by 2030, which creates demand for any memory tier beside it.
  2. D2Foundry-made photonic memory exists: a photonic latch on GlobalFoundries' Fotonix process was shown at IEDM 2025, and a 209-state phase-change cell in Dec 2025.
  3. D3Magneto-optical memory stores the bit in a magnetic domain and has passed trillions of write cycles, retiring endurance as an objection.
  4. R1Most optical memories store the bit in light, which fixes the cell at micron scale, roughly three orders of magnitude larger than electronic memory.
  5. R2Memory is a commodity layer where the profit goes to the incumbent or the compute system, and Marvell's $3.25B purchase of Celestial AI shows the layer being absorbed.
  6. R3Different optical compute designs encode data in wavelength or in brightness, so a memory solved for one needs conversion for another and the market splits into smaller prizes.
  7. C1An optical memory platform reaches a pilot with a memory maker or a hyperscaler.31 Dec 2030
  8. C2A memory maker signs a co-development agreement with an optical memory company.31 Dec 2029
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
2027202820292030today12watch 2027202820292030today12watch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Whether HBM stays sold out through 2027.Dec 2027 · open
  2. 2Whether HBM4 and HBM5 supply catches up before any optical memory reaches volume manufacturing.Dec 2029 · open
Expressionslisted, private, technologies
  • MUMicron
    • cheap
     HBM3E/HBM4 — one of three suppliers (vs Samsung, SK Hynix).
  • 000660.KSSK hynix
    • cheap
     Makes HBM stacked DRAM for AI accelerators (the socket optical memory would have to displace), sold to NVIDIA, AMD and hyperscalers.
  • 005930.KSSamsung Electronics
    • cheap
     Makes HBM and DRAM for AI accelerators and servers, sold to chip designers and hyperscalers.
  • MRVLMarvell Technology (owns Celestial AI)
    • cheap
     Sells custom AI silicon and, via Celestial AI, the Photonic Fabric optical link that pools memory for accelerators.
  • GFSGlobalFoundries
    • cheap
     Makes silicon-photonics wafers (Fotonix) on which photonic SRAM cells have been fabricated for research and chip designers.
Private · 6 private companies, 2 of them in the US.

memQ · MemStera · Light Standard Technology · Celestial AI · Cerabyte · ewigbyte

7 more private companies tracked, not shown

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

$525M raised by 4 companies, $261M 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 strategic investor: Celestial AI, Cerabyte.

What changedthe record, newest first
  • 29 Jul 2026I asked the researcher who co-authored the leading paper on optical memory whether it could hold what an AI model needs. He said no, and not in a "not yet" way. Worth noting he now runs a company whose architecture stores nothing, so he has a side. But when the people publishing the best results in a field tell you the mainstream application is not there, that is the cheapest diligence you will ever get.
  • 21 Jul 2026A complication nobody else raised: solving optical memory once may not help everyone, because rival designs encode data differently — some in the colour of the light, some only in its brightness — so each would still need its own translation step. The prize may be several smaller ones rather than a single winner.
  • 21 Jul 2026A photonics investor makes the strong claim that memory is not just an unsolved piece but the reason companies abandoned optical computing altogether. He thinks it is solvable, and more by clever architecture than by new materials. If he is right, the first useful optical chips do not need it at all, because you can compute in light and only convert once at the very end to store the answer.
3 earlier
  • 23 Jun 2026Optical memory has two branches, classical (which leaks value to electronic incumbents) and quantum (which cannot be routed around with electronics); the quantum branch is the structurally defensible one.
  • 22 Jun 2026The "optical memory cannot get dense" wall is only true when the bit is stored in light; magneto-optical stores it in a magnetic domain and can in principle miniaturise, so the real gate is the integrated laser and addressing, not a physical density ceiling.
  • 18 Jun 2026The long-promised "optical RAM" that would unlock all-optical packet switching is still blocked by basic physics and the market routed around it; the place optical memory is actually a funded product today is quantum networking, a completely different use.