Inside a piece of borosilicate glass 120 millimetres square and just 2 millimetres thick — about the size of a drink coaster and thinner than a stack of two credit cards — Microsoft researchers have written terabytes of data across hundreds of layers of laser-etched voxels, and they say the marks will still be readable in the year 12,026. The result appears in a Nature paper from Project Silica, and it marks the first end-to-end demonstration of glass archival storage that runs on the same material used to make Pyrex baking dishes.
The medium is passive. No electricity. No cooling. No refresh cycles. A shelf will do.
What a voxel actually is
A voxel is a 3D pixel — a permanent, microscopic deformation buried inside solid glass. To create one, the Project Silica team fires a femtosecond laser, which delivers pulses so brief they are measured in quadrillionths of a second. Each pulse slightly rearranges the atomic structure of the glass at a precise depth, leaving behind a mark that light can later detect.
Stack those marks in three dimensions and a coaster-sized slab becomes a filing cabinet. The team writes with femtosecond lasers at speeds measured in megabits per second per beam, at an energy cost measured in nanojoules per bit. Hundreds of layers fit within the 2-millimetre thickness of the plate.
Once the voxel is written, it is done. The research director for Project Silica at Microsoft Research noted: The nice thing about the glass is, once it’s written, it’s immutable. You’re done.
The kitchenware breakthrough
Earlier versions of the technology needed fused silica — an ultra-pure glass sold by only a handful of specialty manufacturers. That was fine for a lab demonstration. It was terrible for a commercial archive.
The recent paper changes the material. The team showed that the same encoding trick works on borosilicate glass, the cheap, tough stuff used in cookware, oven doors, and laboratory beakers. New Atlas noted that this single substitution is what makes the technology plausibly deployable at scale.
Borosilicate shrugs off water, dust, magnetic fields, and heat that would liquefy a hard drive. Drop it in a flood. Leave it in an attic. Store it beside a magnet. The voxels don’t care.
Two ways to mark the glass
The Nature paper describes two encoding methods. The first, birefringent voxels, changes how the glass interacts with polarised light. It works well in fused silica but not reliably in borosilicate. The team refined it into what they call a pseudo-single-pulse technique — one laser pulse split so that it begins one voxel and finishes another simultaneously, which lets the beam scan faster.
The second method is new. Phase voxels alter the physical structure of the glass itself, changing how light waves travel through it. One pulse, one voxel. Simpler hardware, faster writing, and — critically — it works in borosilicate.
There is a catch. Phase voxels sit closer together and interfere with each other in three dimensions, producing a noisy read. The team solved that with a machine learning classification model that untangles the interference and reconstructs the original bits. Ars Technica’s writeup describes this as the piece that finally makes the phase-voxel approach practical.
How you read a glass book
To retrieve the data, an automated microscope scans the plate layer by layer. Earlier Project Silica readers needed multiple cameras working in concert. The new system uses one.
A machine learning model then decodes the captured images. If a voxel is faint, mis-shaped, or partially obscured by an inclusion in the glass, forward error correction fills the gap. The system is designed to assume imperfection and recover from it — the same principle that lets a scratched CD still play.
The reduction to a single camera matters more than it sounds. It shrinks the reader, cuts the cost, and makes it feasible to imagine racks of these things inside a data centre, each one pulling terabytes off a stack of glass slides.
Ten thousand years, tested by baking
The obvious question is how anyone knows the glass will last ten millennia when the experiment has been running for less than a decade. The answer is accelerated ageing.
Microsoft’s team heated written borosilicate samples and measured how the voxels degraded under stress. The data remained stable even at high temperatures, which the researchers extrapolate to a room-temperature lifetime of more than 10,000 years. They also developed a nondestructive optical method to measure voxel degradation in place, so a plate can be checked without being consumed.
For scale: an LTO-10 magnetic tape cartridge, the current enterprise benchmark for cold storage, holds tens of terabytes but is rated for just decades, and it needs a climate-controlled room. Sony’s Optical Disc Archive, the main optical competitor, was rated at 100 years and several terabytes per cartridge — and was discontinued in 2025, leaving the field wide open.
Glass, by contrast, sits on a shelf.
What has already been written to it
Project Silica has been running experiments for roughly a decade. In 2019, the team encoded the entire 1978 Superman movie on a coaster-sized slide of fused silica. They have since pushed capacity to multiple terabytes on similar-sized plates. They partnered with the Global Music Vault to preserve recordings beneath the Arctic ice on Svalbard.
The recent paper is the moment the substrate became ordinary. Gizmodo compared it to reading the past off stone tablets and old parchment — except a future civilisation would need only a microscope and a decent laser to read us.
The write speeds are still slow by modern standards. A single Silica laser beam manages speeds measured in megabits per second. The team is working on splitting the beam to write many voxels simultaneously, and using the tiny flashes of light produced during writing as a real-time calibration signal, but glass will not replace the SSD in your laptop.
The digital dark age problem
Most of what humanity has ever recorded digitally is written on media that decay within a human lifetime. The global datasphere is growing rapidly. Very little of that data has a home that will outlast its author.
Photographic plates outlasted the people who made them — Rosalind Franklin’s Photo 51, exposed in a King’s College basement in the early 1950s, is still readable today. A hard drive from 2005 mostly is not. The archives that hold medical records, financial ledgers, film masters, and scientific datasets are locked into an expensive migration loop: copy everything to a new generation of media every five to ten years, forever, or lose it.
Archival estates rarely fail because cartridges chemically degrade on schedule. They fail because compatibility windows close, drive generations evolve, firmware support sunsets, and robotics require refresh. The medium is not the problem. The treadmill is.
A glass plate that needs no power, no cooling, and no rewrite for a hundred centuries breaks the treadmill.
Where it fits, and where it doesn’t
Glass is unlikely to displace tape any time soon. LTO tape remains the benchmark for enterprise cold storage. The media cost per terabyte remains low, the ecosystem is mature, and enterprises have decades of operational experience managing refresh cycles. Glass, if commercialised, is more likely to serve what could be called “ultra-long retention” — the vault beneath the vault.
Think of the categories: national medical archives, court records, film negatives, seed-bank documentation, the working data of long-duration space missions. The kind of information that has to survive the institutions that produced it.
The immutability of a written voxel also has a security dimension. A ransomware attack cannot overwrite a phase voxel; the atomic structure of the glass has already been rearranged. But compliance is a system property, not a substrate property. A 10,000-year medium does not remove the need for encryption keys, metadata catalogues, and audit trails around it. Someone still has to know which plate holds what.
Microsoft has said the research phase of Project Silica is officially complete, and has not published a commercialisation roadmap. The paper is the milestone. What happens next — a product, a licensing deal, a hand-off to a specialist archival firm — remains open.
What a reader in the year 12,000 would see
Picture the plate in cross-section. Two millimetres of glass. Inside, hundreds of layers of voxels, each a permanent atomic dimple, arranged in a grid so dense that a coaster holds the text of two million printed books. Light passes through the top surface, bends slightly as it meets each voxel, and emerges carrying a signal that a microscope can capture and a neural network can decode.
No moving parts. No battery. No refresh. The plate does not know how old it is.
If a civilisation ten thousand years from now finds one on a shelf, they will need a camera, a light source, and the algorithms to interpret what they see. The glass itself will not have changed. It will still be a 120-millimetre square, two millimetres thick, cool to the touch, made of the same material as a Pyrex measuring jug — with the Superman soundtrack, or a hospital’s cancer registry, or every episode of a television series, held motionless inside it, waiting to be read.
