HomeTech NewsMicrosoft's Project Silica Puts 4.8TB in Glass as Its Research Ends

Microsoft’s Project Silica Puts 4.8TB in Glass as Its Research Ends

Updated 28 October 2026: this article has been rewritten and checked against the Nature paper and Microsoft’s own statements. Capacity figures are corrected (4.8TB in fused silica and 2.02TB in borosilicate glass, not “4.8 to 7+ terabytes”), and the earlier line that Project Silica “remains in the research phase” is replaced: Microsoft says the research phase is complete and it has announced no product.

Microsoft researchers have written 4.8 terabytes of data into a 2mm-thick, 120mm-square plate of glass and read it back without errors, according to a paper published in Nature on 18 February 2026. The team also showed that a simpler version of the method works in borosilicate, the cheap glass used in oven doors and laboratory flasks, and that accelerated ageing tests suggest the written data should last more than 10,000 years.

The paper is the fullest public account of Project Silica, a programme Microsoft Research has run from Cambridge since 2017. It is also, in effect, a closing report. In a blog post published alongside it, Microsoft says “the research phase is now complete” and that it has published the work “so that others can build on them”. It has not announced a product, a price or an Azure service.

A laser writes hundreds of layers inside the glass

Silica stores data by firing femtosecond laser pulses, each lasting mere millionths of a billionth of a second, into the glass. Each pulse leaves a tiny permanent change, called a voxel, at a chosen depth, so data can be stacked in hundreds of layers. A microscope reads the layers back, and a machine-learning model turns the images into bits.

The Nature paper describes two ways of doing this. The first, “birefringent” voxels, changes how the glass treats polarised light and works only in high-purity fused silica. Using it, the team stored 4.84TB across 301 layers, a density of 1.59 gigabits per cubic millimetre, and wrote at 25.6 megabits per second per laser beam.

The second is new. “Phase” voxels alter the glass’s refractive index with a single pulse and can be written in borosilicate. That plate held 2.02TB across 258 layers. One beam wrote at 18.4 megabits per second, and splitting the laser into four beams raised that to 65.9 megabits per second without heat damage. The authors say thermal simulations suggest 16 or more beams should be possible.

Cheaper glass halves the capacity but simplifies the machines

The trade-off is plain in the numbers: borosilicate stores well under half as much per plate. Microsoft argues that it is still the more practical route. According to the company, the reader needs one camera instead of three or four, and the writer has fewer parts, making it easier to build and calibrate. Fused silica, by contrast, comes from “only a few sources”, the blog says.

In slides presented at a Library of Congress storage meeting in March 2026, the project’s research director put it more bluntly: capacity falls “from 4.8TB to 2TB”, but the “glass cost per TB is still lower”. The same slides describe the project as running from 2017 to 2025 and say the path towards commercialisation is “gated only by the laser”.

Speed is the other gap. Science reported that the four-beam writer remains several times slower than magnetic tape, and that the project’s research director believes more parallel beams could close the gap. The paper adds that off-the-shelf lasers running at 50MHz or more, combined with hundreds of beams, could scale writing further.

The 10,000-year figure comes from heating the glass

Nobody has kept glass for 10,000 years, so the lifetime is an extrapolation. The team heated written borosilicate samples, measured how quickly the voxels faded, and used the Arrhenius law, a standard way of relating decay rate to temperature, to project forward. The paper says the result points to a lifetime above 10,000 years even at 290°C, and longer at room temperature.

The authors are careful about the limits. The estimate “does not account for external influences, such as mechanical stress or chemical corrosion”, which the paper says are outside the study’s scope. Microsoft’s earlier demonstrations were practical rather than statistical: storing Warner Bros’ “Superman” film on glass, archiving music with the Global Music Vault in Svalbard, and a student-curated “Golden Record 2.0”. Science also reported that the team encoded map data for a flight simulator game in a borosilicate plate.

A British-built project with Southampton roots

Every listed author is at Microsoft Research in Cambridge, and the paper thanks the University of Southampton for collaborating “at an early stage of the project”. That makes Silica one of the more substantial pieces of storage research to come out of the UK in recent years, alongside the country’s growing photonics sector, where UK start-up Salience Labs recently raised $125 million for optical switches in AI data centres.

It also fits a pattern at Microsoft Research of publishing big results in the open. The lab’s recent work includes Magnetar, a system that proves software ran correctly, and a machine-learning method for weather forecasts two to six weeks ahead, also published in Nature.

Glass fixes the shelf life of archives, not the cost of running them

The strongest argument for Silica is not the headline lifespan; it is what the medium removes. The paper notes that today’s archives must keep migrating data to new media, a job that is “costly in time, equipment and energy”. A plate that needs no power, cannot be rewritten and does not need periodic copying attacks that cost directly. For records that must be kept unchanged for decades, such as financial, legal or regulatory archives, write-once glass is a natural fit, because tamper resistance is a physical property rather than a software setting.

The same property cuts the other way. Data written to glass cannot be erased, so any archive holding personal data would have to meet erasure duties under UK and EU data protection law in another way, most likely by encrypting records and destroying the keys. And a 10,000-year medium is only useful if readers and decoding software survive too; the paper itself notes that a full cloud archive would need a robotic glass library and “many other computer system design aspects”.

Silica is not the only long-term bet. Synthetic DNA storage promises even greater density, and French company Biomemory has launched DNA Shield, an offline DNA custody service. The difference is that Microsoft has now shown a fully automated write-read-decode system at terabyte scale, and then stepped back. The authors have filed several patents in Microsoft’s name, so the open question is what the company does with them: whether Microsoft quietly builds it into Azure archive storage, licenses it, or leaves it to others once faster, cheaper femtosecond lasers arrive.

Sources

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.