Microsoft Research's Project Silica has moved from laboratory curiosity to a demonstrable archival system capable of etching terabytes of data into ordinary borosilicate (Pyrex‑type) glass, promising longevity measured in
ten‑thousand‑year timescales while changing several of the practical trade‑offs that previously held the approach back.
Background / Overview
Project Silica, the Microsoft Research program that uses femtosecond lasers to write data as three‑dimensional voxels inside glass, has been public for several years. Early demonstrations required high‑purity fused silica and a relatively complex write/read stack; the new work shows the technique can be applied to common borosilicate glass, dramatically lowering material cost and supply constraints while simplifying the reader hardware.
This latest research demonstrates two distinct voxel regimes and a full write/read/decode pipeline: the older birefringent voxel approach (best in fused silica) and a newly developed
phase voxel approach (compatible with borosilicate). The team has shown a 120 mm square, 2 mm thick glass platter can hold multiple terabytes of usable data with fully automated writing and reading, and that written voxels survive accelerated ageing tests that extrapolate to lifetimes of more than 10,000 years at ambient temperatures.
In plain terms: researchers now can reliably record gigabits per cubic millimetre in ordinary oven‑grade glass using ultrashort laser pulses, read that data back with automated microscopes and machine‑learning decoders, and demonstrate that the marks are thermally stable for geological timespans under reasonable storage conditions.
How the technology works
Femtosecond laser direct writing and voxels
At the heart of Project Silica is
femtosecond laser direct writing. A tightly focused ultrashort laser pulse (on the order of 300–1,000 femtoseconds) deposits energy inside the transparent medium at a microscopic focal volume. That energy modifies the glass locally — changing its optical properties in ways that can later be probed by microscopy and computational decoding.
Each modification is a voxel (a three‑dimensional pixel). By stacking layers of voxels through the thickness of a 2 mm platter and controlling the properties of each voxel, researchers can encode multiple bits per voxel and build up terabytes of information in a compact, solid piece of glass.
Two voxel regimes: birefringent and phase voxels
- Birefringent voxels: These rely on polarization‑dependent refractive changes that act differently when light of different polarization is passed through them. Historically, writing birefringent voxels required multiple laser pulses and polarization control, and needed the purity of fused silica to form stable, low‑scatter structures. Birefringent voxels can achieve very high density and excellent voxel quality when manufactured in high‑purity silica.
- Phase voxels (new): These change the phase of transmitted light rather than polarization. Crucially, phase voxels can be formed with a single laser pulse and are compatible with borosilicate glass. Single‑pulse writing simplifies the optical hardware and increases per‑pulse energy efficiency. The trade‑off is that phase voxels exhibit greater three‑dimensional inter‑symbol interference (cross‑talk between nearby voxels), which the team mitigates using machine‑learning classifiers during decoding.
Parallel writing and system architecture
Researchers built a multibeam writer that can produce multiple voxel modifications in parallel, boosting throughput without thermally damaging the glass. The system design leverages:
- High‑repetition‑rate lasers (the experimental systems operate at multi‑MHz repetition rates).
- Polygon scanners, fast stage motion, and objective optics to place voxels at sub‑micron lateral pitch and micron‑scale layer spacing.
- Automated microscopes and camera-based readers to image layers and feed images into a decoder pipeline that uses classical error correction and machine learning to recover bits.
The result is an integrated pipeline: data receive → symbol encoding → automated, parallel laser writing → robotic glass handling → automated microscopy → ML‑assisted decoding → recovered user data.
The headline numbers (what the experiments achieved)
The project demonstrates concrete, repeatable metrics across both voxel regimes and glass types:
- Using birefringent voxels in fused silica:
- Data density: ≈ 1.59 Gbit/mm^3
- Usable capacity: ≈ 4.8 TB per 120 mm × 120 mm × 2 mm platter
- Layers: ~301
- Voxel pitch and spacing: ~0.500 μm × 0.485 μm lateral pitch; 6 μm layer spacing
- Write throughput per beam: ≈ 25.6 Mbit/s
- Write energy: ≈ 10.1 nJ per user bit
- Using phase voxels in borosilicate (Pyrex‑type) glass:
- Data density: ≈ 0.678 Gbit/mm^3
- Usable capacity: ≈ 2.02 TB per 120 mm × 120 mm × 2 mm platter
- Layers: ~258
- Voxel pitch and spacing: ~0.5 μm × 0.7 μm lateral pitch; 7 μm layer spacing
- Write throughput per beam: single beam ≈ 18.4 Mbit/s; four beams in parallel ≈ 65.9 Mbit/s
- Write energy: ≈ 8.85 nJ per user bit
- Accelerated ageing experiments for phase voxels in borosilicate extrapolate to lifetimes exceeding 10,000 years at room temperature under the assumptions of thermal Arrhenius behaviour and absence of external corrosion or mechanical stress.
- The reader hardware has been simplified for phase voxels: the team reduced the camera requirement from three or four cameras down to a single camera, lowering reader complexity and cost.
These figures represent lab demonstrations — controlled, repeatable experiments that validate the system concept and quantify key trade‑offs between density, throughput, and medium choice.
Why borosilicate matters (Practical advantage)
Moving from fused silica to
borosilicate glass is not a minor materials tweak — it addresses several practical barriers to adoption:
- Cost and availability: Fused silica of the required purity is expensive and sourced in limited volumes. Borosilicate glass (the same family used in ovenware and laboratory ware) is mass‑produced globally and far cheaper per unit volume.
- Manufacturing scale: Lower raw material cost and established supply chains make volumetric scaling more plausible, should the write/read hardware be industrialized.
- Simpler hardware: Phase voxels require only amplitude modulation and a single camera for reading, which reduces optics, modulators, and camera count.
In short, borosilicate lowers a real barrier between laboratory proof and any future attempt at building an industrial archival product.
Strengths: what makes glass archival compelling
- Durability and stability: Glass is chemically inert, non‑porous, and resistant to humidity, electromagnetic interference and many forms of degradation that plague magnetic and polymer‑based media.
- Energy‑free preservation: Once written, glass requires no power to preserve the data. No spinning motors, no refrigeration — ideal for cold, long‑term retention.
- Immutability: A glass platter is effectively write‑once, read‑many (WORM), which is an advantage for custodial archives that must preserve an incontrovertible record.
- High volumetric density for archival use: Tightly controlled femtosecond writing and multiple layers enable terabytes in an inch‑scale piece of glass, making dense physical archives possible.
- Automated verification: The research demonstrates fully automated write/read/verify cycles across billions of voxels, a key step for trustworthy archival systems.
- Compatibility with mass glass production: Demonstrating feasibility in borosilicate opens the door to established glass manufacturing practices and economies of scale.
Risks, open questions, and real‑world limitations
The laboratory achievements are notable, but the leap from validated lab demonstration to an operational, cost‑effective archival product remains nontrivial. Key risks and open questions include:
1. Throughput and operational scale
Write speeds per beam are modest relative to modern streaming needs. Even with multibeam systems, the demonstrated peak per‑platter writing rates mean scaling to petabytes per year — the operational load of research telescopes, hyperscale backups, or national archives — would require many writers working continuously or significant increases in beam parallelism and laser repetition rates.
2. Cost per terabyte (TCO)
The total cost of ownership hinges on:
- Capital cost of high‑precision femtosecond lasers, optics, and robotics.
- Cost of automation and reliability engineering needed for datacentre‑grade operation.
- Cost of reading hardware, image sensors, and compute for ML decoding.
Until a full cost model is released, claims about competitiveness with tape or cold cloud storage remain speculative.
3. Fragility and handling
Glass is brittle. A platter that survives centuries in a climate‑controlled vault can still shatter through mechanical shock. Any practical system must solve:
- Humidity and temperature control (glass is robust, but seals, labels and handling mechanisms are not).
- Protective enclosures and robotic handling to avoid operator damage.
- Redundancy strategies that cope with single‑platter loss.
4. Format longevity and decodeability
Preserving scratches in glass is one thing — preserving the
ability to decode them in a meaningful way centuries hence is another. Practical archival systems must ensure:
- Open, standardized encoding metadata stored redundantly.
- Preservation of decoder specifications and a set of readers (or blueprints) across generations.
- Secure methods of encryption, key escrow, and legal access that survive social and political change.
5. Environmental and chemical threats beyond thermal ageing
Accelerated ageing focused on thermal stability provides important bounds, but real‑world archives face events such as fire, flooding, chemical corrosion (e.g., acids), and mechanical abrasion. Glass resists many of these, but not all; vault and physical protection design remain essential.
6. Inter‑symbol interference and ML dependence
Phase voxels are more prone to 3D cross‑talk, increasing decoding complexity and reliance on machine‑learning classifiers. ML decoders work well in the lab, but long‑term archival strategies need:
- Robustness against distributional drift (e.g., unexpected changes in voxel appearance over millennia).
- Certifiable error bounds and verifiable, reproducible decoding procedures suitable for legal and scientific use.
7. Security and immutability trade‑offs
Immutability is a double‑edged sword. For some archives immutability is required, but it complicates corrections, takedown orders, and the ability to redact or erase data. Operational policies must be carefully designed.
Where glass storage fits in the archive ecosystem
Glass‑etched plates will not replace all storage. Instead, they are best thought of as a specialized tier of ultra‑cold, ultra‑durable archival media for:
- Cultural heritage (archives, national libraries, museum collections).
- Legal and governmental records that require long‑term non‑volatile retention.
- Scientific “golden copies” (datasets that must be preserved unchanged for centuries).
- Disaster‑hardened, provenance‑centric backups (where immutability and chemical stability are prioritized).
They are less suitable where frequent updates, random I/O, or cheap per‑GB active storage are required. For example, current tape libraries or cloud cold storage offer cheaper upfront cost and higher write throughput for active archival workflows. Glass storage's sweet spot is when longevity, immutability, and long‑term readability outweigh cost per gigabyte and streaming throughput.
Practical engineering challenges to commercializing the lab work
If Project Silica were to move into productization, engineering teams would need to solve at least the following areas:
- Robotic handling and glass library design to safely store and retrieve thousands of platters.
- Industrialization of multibeam lasers and optics to boost throughput and lower per‑bit write energy/cost.
- Reader miniaturization and ruggedization so that decoding can be performed in datacenter conditions with predictable throughput.
- Standardized file and metadata formats, with long‑lived schema and versioning to ensure future decodeability.
- Security primitives — strong, auditable encryption and key management workflows appropriate for centuries‑long custody.
- Comprehensive, audited error‑correction and redundancy strategies that translate single‑platter reliability into archive reliability targets.
- Manufacturing, quality control, and supply chain validation for sourcing consistent borosilicate media that meets optical and thermal specifications.
The durability claim: what “10,000 years” actually means
The 10,000‑year figure is derived from
accelerated ageing experiments and Arrhenius‑law extrapolation. The lab tests heat written voxels to elevated temperatures, measure their decay rates, and extrapolate back to estimate a characteristic lifetime at room temperature. That approach is a standard materials science technique and gives a scientifically grounded estimate, but it comes with caveats:
- Extrapolation assumes the same underlying decay mechanism at elevated temperatures holds at ambient temperature.
- The prediction presumes isolated, chemically inert storage conditions — external corrosion, mechanical stress, or catastrophic events are explicitly outside the scope of thermal lifetime measurements.
- Practical archive lifetimes require not only voxel stability but also continued availability of decoding hardware, software, and documentation.
So, while the thermal stability numbers are compelling and meaningful, they are
one dimension of archival durability. Real world archival practice must consider mechanical, social, legal and infrastructural durability alongside material stability.
Competitive landscape and alternatives
Glass storage competes conceptually with other long‑lived storage research areas:
- Magnetic tape: Mature, cheap per TB, established ecosystem, but limited lifetime (decades) and requires powered infrastructure for migration and verification.
- Cold cloud storage: Operationally convenient and scalable, but incurs ongoing energy and service costs and depends on service providers and migrations.
- DNA storage: Extremely high theoretical density and very long persistence under correct conditions, but write/read costs and latency remain challenging for practical, routine archival use.
- Other optical/quartz approaches: Various research groups and startups pursue laser inscription in quartz and similar media; each approach balances density, write speed, cost and longevity differently.
Glass storage's advantages — immediate readback with microscopy, low maintenance energy, and strong evidence for thermal permanence — make it a unique offering, but economical and operational viability will determine whether it becomes a mainstream archival tier.
Operational and policy implications
Archivists, enterprise IT and regulatory bodies should begin to consider what long‑lived physical media mean for policy:
- Governance: Long‑duration custody requires institutional and legal frameworks that outlast current organizational lifetimes — trusteeship models, multiple custodians, and open redundancy.
- Interoperability: Standardized encoding schemes, publicly archived decoders and metadata formats are essential to avoid vendor lock‑in or unreadable proprietary archives in the future.
- Digital provenance: Immutable physical media alter how provenance, chain‑of‑custody, and tamper evidence are implemented; archiving workflows must incorporate robust cryptographic verification stored both inside and outside the glass media.
- Disaster planning: Organizations should consider geographical dispersal, environmental protections and active migration plans even for “10,000‑year” media.
Verdict — what this research changes today and tomorrow
Microsoft’s Project Silica demonstrates two tangible and important advances in archival media research:
- It proves that ordinary borosilicate glass can host reliable, high‑quality phase voxels, widening the material base for glass storage beyond expensive fused silica.
- It shows how system‑level design — lasers, multibeam writeheads, automated microscopy and ML decoding — combine to create a usable end‑to‑end archival system, not just isolated materials experiments.
Those achievements reduce two critical barriers to future deployment: material cost/availability and reader complexity. However, major engineering, economic and operational gaps remain before glass‑etched platters become a mainstream archival product. Throughput, cost per TB, ruggedized handling, standardized formats and long‑term governance are still open problems.
For organizations with the need to preserve irreplaceable data for centuries — national libraries, cultural repositories, and certain scientific archives — glass storage is now firmly in the “serious contender” category. For everyday enterprise backups and high‑volume cold storage, more work is needed to make glass competitive with established tape and cloud economics.
What to watch next
- Improvements in multibeam and laser repetition technology that materially increase sustained write throughput.
- Industry partnerships or spinouts focused on industrializing writer/reader hardware and building glass libraries.
- Standardization efforts for encoding, metadata and long‑term decodeability.
- Cost models and pilot deployments demonstrating total cost of ownership versus tape and cold cloud.
- Robotic handling solutions and mechanical packaging that mitigate glass fragility in operational settings.
Conclusion
Project Silica’s step to etch data into ordinary borosilicate glass is a substantial technical milestone and a practical one: it turns a laboratory exotic into a materially feasible archival candidate. The work combines precise ultrafast optics, parallel beam engineering, automated microscopy and machine learning decoding into a coherent system that records terabytes inside a two‑millimetre slab of glass and argues that those marks will endure for ten millennia under benign storage conditions.
That combination — density, durability and a path to manufacturable media — makes glass‑based archival storage one of the most interesting long‑term data preservation technologies to emerge in recent years. The remaining challenges are engineering scale, cost, handling, and policy, not fundamental science. If the industry can solve those problems, a new tier of archival storage that requires no power to preserve data for generations could move from the lab into real world vaults — with profound implications for how institutions think about long‑term memory for the digital age.
Source: theregister.com
Microsoft's latest storage tech etches data into Pyrex glass