Glass Data Storage

The useful question is whether an archive can retain readable information without repeatedly replacing its recording medium. Glass research attacks that media problem; it does not remove the need to preserve a complete retrieval system. This card covers laser-written glass, not the glass substrate beneath an HDD's magnetic coating.

In the 2026 experiments, fused silica stores birefringent voxels read through polarization optics; borosilicate stores phase voxels read with phase-contrast microscopy. The latter encode refractive-index changes, not the reversible state switching of phase-change memory. [1]

The table keeps the authors' reported usable capacities for 120 x 120 x 2 mm plates. These are calculated values, not demonstrations of writing every stated terabyte; throughput is peak per-beam writing. The Methods specify an engineering-overhead factor of 0.747. [1]

Capacity accounting needs clarification. Directly applying that factor to the reported densities gives 1.59 x 120 x 120 x 2 x 0.747 / 8000 = 4.28 TB and 0.678 x 120 x 120 x 2 x 0.747 / 8000 = 1.82 TB, rather than 4.84 and 2.02 TB. The divisor converts Gbit to decimal TB. This is an unresolved accounting difference, not a replacement capacity measurement; the table retains the published values. [1]

Silica's 2023 architecture stages input, writes a platter, and reads it back before discarding the staged copy. Images are decoded with machine learning and error correction. Its library separates writers from readers and prevents stored platters returning to writers; that protects against rewriting through the read path, not every possible loss of access. The paper combines a prototype with full-system simulation. [2]

Glass composition also matters. Lawson et al. screened 30 commercial glasses for phase-voxel recording: 23 completed single-layer screening and 17 completed the multilayer stage. Five rejected glasses could form voxels but lacked the emission response needed by the tested closed-loop energy controller. That is a control-system constraint, not proof that those materials cannot store optical information. The comparison used 400 fs pulses and a fixed 0.5 x 0.7 micrometre lateral pitch; it does not rank all glasses under every writing method. [4]

The 10,000-year claim comes from accelerated thermal aging and Arrhenius extrapolation of signal decay, not millennia of service observation. Mechanical damage and chemical corrosion are outside that lifetime model. Absolute read throughput was not evaluated in the 2026 study. [1]

Microsoft's February 2026 update describes its research phase as complete and makes the published findings available for others to build on. That dated milestone is not a product availability announcement. The research label here describes the cited evidence, not a claim about every glass-storage venture. [3]

For archive planning, retain independent copies, file-format documentation, decoder specifications and key-recovery procedures. Treat continued access as a separate acceptance test from media retention. A deployment decision also needs measured ingest and restore performance, operating costs, service commitments and a migration path.