Will Your Data be Sealed in Kitchen Glass for 10000 Years?

On February 18, 2026, Microsoft Research published a landmark paper in Nature detailing the most significant advance yet in Project Silica — a decade-long effort to store digital data inside glass using ultrafast femtosecond lasers. The headline finding: the technology no longer requires rare, expensive fused silica. It now works on borosilicate glass, the humble material behind your Pyrex baking dish.

For data center architects and archivists, this distinction matters enormously. Fused silica has long been a technical bottleneck — costly, difficult to manufacture at scale, and commercially fragile as a supply chain. Borosilicate glass, by contrast, is manufactured by the billion square meters annually, used in everything from pharmaceutical vials to telescope mirrors to kitchen cookware. Shifting to this substrate removes two of the biggest barriers that have kept Project Silica in the laboratory: cost and availability of media.

Two Types of Voxels, Two Roads to Permanence

The Silica system now employs two fundamentally different approaches to encoding data. Both use femtosecond laser pulses — bursts of light lasting just one quadrillionth of a second — to create microscopic three-dimensional data points called voxels inside glass. But the physical mechanism, and the glass required, differ substantially.

Phase Voxels — the Borosilicate Breakthrough

The new phase voxel method works by altering the isotropic refractive index of the glass — essentially changing how light waves travel through it — rather than its polarization. This approach requires only a single laser pulse per voxel, a key simplification over earlier methods that needed many pulses. Because the laser no longer needs to pause and reposition between pulses, it can scan continuously across the glass, making high-speed parallel writing far more practical.

The result: a 120 mm square, 2 mm thick borosilicate plate stores 2.02 TB across 258 layers, at a density of 0.678 Gbit/mm³. The reading hardware has also been radically simplified — where the old system required three or four cameras to reconstruct data, the phase voxel approach needs only one. A machine learning model handles the three-dimensional inter-symbol interference that makes reading phase voxels challenging.

Birefringent Voxels — Higher Density, Higher Purity

The original fused silica approach, using birefringent voxels, has also advanced. These voxels encode data by altering how glass interacts with polarized light — the refractive index changes depending on the direction of light propagation. Previously, forming a single voxel required many laser pulses. The new pseudo-single-pulse technique reduces this to just two pulses by splitting a single pulse: one pulse seeds a void, a second elongates it. Critically, the two pulses can be delivered simultaneously to adjacent voxels in a scanning motion, enabling continuous writing without pausing the laser.

This technique pushes fused silica capacity to 4.8 TB per 120 mm plate across 301 layers — the highest density yet demonstrated, at 1.59 Gbit/mm³. Single-beam write speed is 25.6 Mbit/s. The trade-off: birefringent voxels demand high-purity fused silica, limiting material options and keeping costs elevated.

“The laser focus can be swept rapidly across the glass, enabling writing speed limited only by the speed of the femtosecond laser itself.”

— Richard Black, Partner Research Manager, Microsoft Research

Parallel Beams and the Throughput Problem

A persistent weakness of glass storage has been write speed. To address this, the research team developed a multi-beam irradiation system that splits the laser into independently modulated channels. Using four beams in parallel, they demonstrated a sustained write throughput of 65.9 Mbit/s without causing thermal damage to the glass — a figure made possible by a mathematical model that calculates real-time temperature changes and keeps simultaneous laser spots from interfering with each other.

The team believes scaling to 16 or more beams is feasible, which would imply write speeds approaching 263 Mbit/s — roughly 33 MB/s. For context, that remains significantly slower than LTO-10 magnetic tape, which writes at 400 MB/s uncompressed. Glass storage is not designed to compete with tape for speed; its value proposition is longevity and zero-maintenance permanence.

Proving 10,000 Years Without Waiting 10,000 Years

One of the more elegant contributions of the paper is a new non-destructive aging test. The team developed a method to read minute physical changes inside a glass plate — changes that accumulate as the voxels age — without destroying the sample. By combining this readout with standard accelerated degradation protocols (repeatedly heating inscribed plates to 500°C to simulate long-term aging at lower temperatures), they modeled the data retention curve and confirmed data integrity at room temperature for over 10,000 years. The glass is resistant to water, heat, electromagnetic fields, and dust.

This matters because conventional media testing typically destroys samples. The new non-destructive approach means a production system could periodically check the health of stored glass platters in situ — a significant operational advantage.

Why This Changes the Calculus for Long-Term Storage

The economics of enterprise data archiving are defined by one inescapable problem: every storage medium in widespread use today degrades. Magnetic tapes, which dominate cold storage, have a practical lifespan of roughly 30 years under ideal conditions. HDDs fare worse under continuous operation. The industry response — periodically migrating all data to new media — consumes enormous quantities of energy, labor, and capital, and creates a continuous risk of data loss during migration.

Glass changes this calculation entirely. A Silica plate is write-once, read-many (WORM) and requires no power to maintain its data. There are no moving parts, no magnetic coatings, no organic binders to degrade. Once written, a glass platter can sit in a room-temperature warehouse for centuries without refreshing. For the specific workloads where data is written once and read rarely — regulatory archives, scientific datasets, cultural heritage preservation, AI training corpora, medical records — this is a transformative property.

Microsoft has signaled that practical deployment could involve robotic glass libraries in data centers, with erasure coding to allow data recovery even if a portion of a plate is damaged. The company has already demonstrated real-world proof-of-concept deployments: a collaboration with Warner Bros. stored the 1978 Superman film on a fused silica plate, and a partnership with a music archive preservation project is storing recordings under Arctic ice.

What Remains to Be Solved

The write speed gap relative to magnetic tape is the most significant remaining technical challenge. Richard Black, Project Silica’s research director, has indicated the femtosecond laser itself — not the glass or encoding method — is the primary bottleneck. Scaling from four to 16 or more beams is seen as the near-term path, but manufacturing high-repetition-rate femtosecond lasers at acceptable cost adds its own engineering complexity.

The unresolved choice between voxel types is also significant. Phase voxels offer simpler hardware and lower material cost but sacrifice density. Birefringent voxels offer superior density and performance but require fused silica. A future production system will have to commit to one path — and that choice has not been made. As Tom’s Hardware noted, this means “nothing is ready for prime time yet.”

What Microsoft has achieved with Project Silica is not a product — not yet. It is something arguably more important: proof that the underlying physics works end-to-end, at practical densities, with cheap materials, verified to last longer than recorded history. The engineering path to a commercial system is long, but the scientific foundations are now solid.

In a world where the volume of data generated annually now exceeds the total capacity of every storage medium manufactured to date, the question of what happens to data in a century is no longer academic. Project Silica’s glass plates may prove to be one of the few credible answers.