Theses›Datacentre & AI compute›Optical DRAM

Optical DRAM

Magneto-optical is the one optical bit that scales; the moat is a cornerable integrated stack

  • Early
  • $247B by 2035

The call, a draftThe restraints are winning. Magneto-optical memory can in principle reach DRAM density, but the readout ring is about a million times bigger than the bit, and nobody has addressed a dense array on a chip. The window is early.

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 through 2027 while Micron, SK Hynix and Samsung pour capex into more of the same, opening room for a different memory tier.
  2. D2Endurance is no longer an objection: magneto-optical cells have been shown at trillions of cycles against 2.4 billion in the original paper.
  3. D3Imperial researchers switched permalloy nanomagnets of sub-diffraction size with a continuous-wave laser at 2.74 mW.
  4. R1Each 50 by 50 micron readout ring addresses one cell, about six orders of magnitude off DRAM, and integrated sub-diffraction addressing on a photonic chip is unshown.
  5. R2No memory-medium company has reached €1B on its product in 25 years; Everspin earns about $55M after 17 years and Optane died on the incumbent cost curve.
  6. R3The cornerstone all-optical switching patent expired on 24 May 2026, so the only moat left is process and capital, where memory incumbents are strongest.
  7. C1An on-chip sub-diffraction magneto-optical write and address scheme is demonstrated on a photonic chip.31 Dec 2027
  8. C2A magneto-optical memory shows DRAM or storage-class density on a manufacturable process with a named foundry or partner.31 Dec 2028
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
20272028202920302031today12345predictionswatch 20272028202920302031today12345predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1By end-2027 an integrated, on-chip sub-diffraction magneto-optical write/address scheme is demonstrated, the load-bearing 'solve the laser speed' milestone.Dec 2027 · open
  2. 2By end-2028 at least one magneto-optical memory demonstrates DRAM- or storage-class areal density at an HVM-credible, cornerable process with a named foundry or commercialization partner.Dec 2028 · open
  3. 3Whether HBM4 and HBM5 supply catches up with demand first.Dec 2029 · open
  4. 4Contingent capture: through 2030 no standalone magneto-optical memory startup reaches a durable >$1B outcome ON THE PRODUCT; the pattern stays acqui-hire by a memory incumbent or photonic-compute box.Dec 2030 · open
  5. 5Whether any magneto-optical memory company passes $1B on its product rather than an acqui-hire.Dec 2030 · open
Expressionslisted, private, technologies
  • MUMicron
    • cheap
     HBM3E/HBM4 — one of three suppliers (vs Samsung, SK Hynix).
Private · 1 private company.

MemStera

1 more private company tracked, not shown

Technologies
What changedthe record, newest first
  • 29 Jul 2026A useful lesson in doing your own homework. I spent two weeks working out which university held the best patents in a corner of optical memory, concluded it was one particular group, and went looking for the researchers. They had spun the company out six months earlier, and I had already had a call with the founder in January. The interesting part is what they built with it: not memory, but a learning chip. That is now the second world-class group in this field to look at optical memory and choose to build compute instead.
  • 29 Jul 2026The senior author on the leading optical-memory paper told me optical memory will never store what AI needs, and gave the number that explains why: the magnetic bit is nanometres, but the ring you need to read it is fifty microns across, so the thing you can address is about a million times bigger than the thing you store. He also said endurance, which used to be the objection, is now effectively unlimited. The interesting part is that the constraint moved from the physics to the plumbing.
  • 15 Jul 2026Ran the patent search that was meant to confirm a risk and it refuted it, then handed me a worse one. The foundational patent in a field I am betting on quietly entered the public domain seven weeks ago, and the university that invented the effect has filed nothing in over a decade. My instinct was to read an unencumbered estate as good news. It is not: if no patent blocks anyone, the only moat left is process and capital, and that is the incumbent's home ground, not a startup's. A dead patent is not a clear runway, it is a removed toll gate — and the incumbent drives through it too.
3 earlier
  • 15 Jul 2026Swept a field I thought I knew and found my map was drawn from the wrong search term. The two companies I had as a cohort look like one company renamed. The research leaders I had as the centre of gravity own only half the product; the other half sits on a different continent in labs that never co-author with them. The uncomfortable read: I had been sampling who walked into my inbox and calling it a landscape.
  • 15 Jul 2026Passed a magneto-optical memory company on team, not tech. Second time in this space. When a whole cohort keeps failing the same non-technical test, the thesis is not wrong, the sourcing is: the job is to find or build the entrant that clears it, not to lower the bar for the ones in front of you.
  • 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 question is the integrated laser and addressing, not a physical density ceiling, and whether a startup can hold the process the way a memory maker holds a process.