A new paper proposes
What if we stopped storing data as electrical charge — and stored it as the physical position of individual atoms? That's the idea behind this paper. And the numbers are staggering.
Ilia Toli, 2026
Why this matters
Think of it this way
Imagine a brilliant chef (the processor) who can cook any dish in seconds — but the pantry door (memory) only opens once a minute. The chef spends most of their time waiting for ingredients, not cooking. That's modern computing: processors are fast, but getting data to them is painfully slow.
This gap between processor speed and memory speed is called the memory wall. And it's not just an annoyance — it's the defining constraint of the AI era. The energy consumed moving data between memory and processor already exceeds the energy consumed doing the actual math in modern AI accelerators.
Making it worse: as of early 2026, NAND flash memory (the chips in your SSD) faces a supply crisis. Prices have surged — consumer SSD prices doubled, and NAND wafer prices rose over 240% year-over-year — because manufacturing capacity is being redirected toward AI hardware.
Moving data between memory tiers uses more energy than the computations themselves.
NAND contract prices up 33–38% per quarter. Some suppliers have pre-sold all of 2026 and 2027.
Charge-based memory is hitting fundamental scaling limits. The problem isn't engineering — it's materials.
The paper argues this isn't a problem that better chips can solve. It's a materials problem. We need a fundamentally different way to store information.
The substrate
Think of it this way
You know graphene — the famous single-atom-thick sheet of carbon? Now imagine taking that sheet and gluing a fluorine atom to every single carbon. You get fluorographane — essentially a two-dimensional version of Teflon (the non-stick coating on your frying pan).
Fluorographane (chemical formula CF) is a fully fluorinated graphene sheet. Every carbon atom is bonded to exactly one fluorine atom, and the carbon sits in what chemists call sp³ hybridization — a tetrahedral bonding arrangement where each carbon connects to three other carbons and one fluorine, forming a buckled honeycomb lattice.
A flat hexagonal lattice of carbon atoms — one atom thick, incredibly strong.
Expose the sheet to xenon difluoride (XeF₂) gas from both sides. Each carbon bonds to one fluorine atom. The sheet buckles slightly as the carbon bonding changes from flat (sp²) to tetrahedral (sp³).
An atomic-scale sheet with fluorine atoms sticking out above and below. Thickness: ~0.35–0.40 nm. That's about 100,000× thinner than a human hair.
Crucially, each fluorine atom can point up or down relative to the carbon sheet. And that's the bit.
🔬 Go deeper: Surface propertiesLike Teflon, fluorographane has exceptionally low surface energy. It repels water, oils, and virtually all contaminants. Adjacent layers in a tape configuration won't stick to each other — no release coating or lubrication needed. The minimum interlayer spacing is ~0.50–0.60 nm, governed by fluorine–fluorine steric repulsion.
The material is also polycrystalline when grown by CVD — consisting of many small crystal domains. But this doesn't matter for memory: each bit is a local property of a single C–F bond and doesn't depend on the crystal orientation of its domain. Grain boundaries consume only ~0.003% of the total area.
The core idea
Think of it this way
Imagine a flagpole on a hill. The flag can fly on one side or the other — it's either "north" or "south." Now shrink that flagpole down to a single atom. The fluorine atom sticking out from a carbon is that flag. If it points up, that's a 1. If it points down, that's a 0. That's the entire memory bit — the physical position of one atom relative to a tiny carbon scaffold.
In this system, the fluorine atom bonded to each carbon can occupy one of two bistable orientations — protruding from the top face or the bottom face of the carbon sheet. These two positions are the binary states: 0 and 1.
What makes this remarkable is that the two states are energetically degenerate — neither position is favored over the other. In conventional memory, stored data gradually "wants" to decay toward a lower-energy state. Here, both states are equal. There's no thermodynamic push toward either value. Data doesn't degrade.
To change a bit, the fluorine atom doesn't detach and reattach. Instead, the central carbon and its fluorine undergo a pyramidal inversion — like an umbrella flipping inside-out in a storm.
The carbon atom sits atop a three-carbon "base" (C₁, C₂, C₃). During inversion, the carbon (C₀) and its fluorine pass through the plane of that base, emerging on the other side. The fluorine threads through one of the gaps between the base carbons.
The energy required to drive this inversion — the inversion barrier — is approximately 4.6 eV (confirmed at 4.8 eV by a higher-level calculation). This is the key number in the entire paper.
The barrier was computed using a small molecule called fluorophenalane (C₁₃F₂₂) — the smallest molecule that faithfully captures the local chemistry of a fluorographane memory bit. It has a bridgehead carbon bonded to three ring carbons and one fluorine, constrained by a tricyclic geometry mimicking the infinite sheet.
Crucially, fluorophenalane provides a rigorous lower bound on the real barrier. The ring strain in the small molecule destabilizes the equilibrium geometry relative to the infinite sheet, and the free edges allow the ring to widen during inversion. In the real material, the surrounding lattice is maximally constrained — the gap between base carbons can't open as easily, making inversion harder.
The transition state was verified as a first-order saddle point with exactly one imaginary vibrational frequency (-65.93 cm⁻¹), confirming this is the correct reaction pathway. At the saddle point, the C₀–F bond elongates to 1.73 Å (from ~1.38 Å equilibrium), and C₀ sits only 0.26 Å from the base plane — nearly flat.
DFT method: B3LYP-D3BJ/def2-TZVP. Confirmed by DLPNO-CCSD(T)/def2-TZVP at 4.8 eV. The bond dissociation energy (5.6 eV) exceeds the barrier at both levels of theory, so the C–F bond never breaks during inversion.
The stability case
Think of it this way
Imagine a ball sitting in a valley between two mountains. For the ball to spontaneously roll over the mountain to the other valley, it would need a massive, improbable burst of thermal energy. The 4.6 eV barrier is such a tall mountain that at room temperature, the probability of a spontaneous flip is essentially zero — not just for your lifetime, but for 10⁴⁷ lifetimes of the universe.
The paper analyzes three potential ways a bit could spontaneously change, and eliminates each one:
At room temperature (300 K), the thermal flip rate is ~10⁻⁶⁵ per second. Mean retention time: 10⁶⁵ seconds — about 10⁴⁷ times the age of the universe.
Fluorine's tunneling rate through the barrier: ~10⁻⁷⁶ per second. Even more negligible than thermal flipping.
Could fluorine hop to a neighboring carbon? Rate: ~10⁻⁸² per second. Requires breaking a 5.6 eV bond. Fluorine atoms do not wander.
Drag the slider to see how the retention time changes with the barrier height. Even at much lower barriers, the retention is astronomical.
Think of it this way
Hydrogen is like a ping-pong ball — light enough to quantum-tunnel through the barrier with alarming ease. Fluorine is like a bowling ball — 19× heavier — making tunneling essentially impossible. This mass difference enters the tunneling equation under a square root in the exponent, amplifying the suppression by tens of orders of magnitude.
The paper shows that graphane (hydrogenated graphene) would have a hydrogen tunneling rate of ~10⁻⁶ s⁻¹ — a spontaneous bit flip roughly once per fortnight. That's useless for memory. Worse, hydrogen atoms preferentially migrate to reduce sheet corrugation, systematically erasing stored data rather than just randomizing it.
Fluorine is the only element in the periodic table that simultaneously satisfies all four requirements: (1) forms a covalent, directional bond to sp³ carbon; (2) is small enough to pass through the C–C gap during inversion; (3) is heavy enough to suppress tunneling; and (4) creates a barrier below the bond dissociation energy, preserving the bond.
📐 Go deeper: Why bit drift is impossibleIn charge-based memory (DRAM, NAND flash), a stored bit gradually fades — charge leaks from a capacitor or floating gate, and the voltage decays toward an ambiguous middle state. This is bit drift.
In fluorographane, this concept is physically meaningless. The C–F inversion coordinate is a double-well potential with exactly two stable minima separated by a 4.6 eV barrier. There is no intermediate resting state. The fluorine atom is either in one well or the other. There is no continuum of positions where it can pause. A fluorographane bit does not degrade, weaken, fade, or erode. It is correct until deliberately overwritten.
The architecture
The paper proposes a tiered architecture, progressing from today's technology to future systems:
Think of it this way
Imagine running your fingertip across a surface to feel bumps and dips. An atomic force microscope (AFM) does this with an incredibly sharp needle tip — sharp enough to feel individual atoms. Where a fluorine sticks up, the tip feels a mound. Where the fluorine has been flipped to the other side, it feels a trough.
Reading: The topographic contrast between the two bit states ranges from ~0.06 nm (non-contact) to ~0.30 nm (contact mode). Commercial AFMs routinely resolve features of 0.01 nm — so the signal is huge by AFM standards.
Writing: A voltage pulse (3–5 V, 10–100 ns) through the probe tip deposits enough energy to flip the fluorine. Six independent write mechanisms exist: voltage pulse, electron beam, mid-IR vibrational pumping, mechanical tip force, heated tip, and combinations thereof.
Throughput: ~1–1000 bits/second. Slow — but already useful for write-once archival applications like space missions, nuclear records, or cultural heritage preservation.
Precedent: Kalff et al. (2016) built a 1 kilobyte rewritable atomic memory on chlorine/copper at 77 K. This proposal operates at room temperature with 46× higher density.
Think of it this way
Instead of one needle reading one atom at a time, imagine a grid of 100 million tiny flashlights, each shining infrared light on a small patch of the sheet. Each flashlight reads thousands of bits by detecting how the fluorine bonds absorb the light differently depending on their orientation.
The C–F stretching vibration at 1220 cm⁻¹ (~8.1 µm wavelength, mid-infrared) is the coupling channel. Sub-wavelength apertures achieve spot sizes of 8–80 nm, addressing blocks of ~10³–10⁵ bits each.
Writing: Resonant vibrational "ladder climbing" — pumping energy into the C–F bond through shaped mid-IR pulses, accumulating enough over multiple oscillation cycles to drive inversion. The write pulse must be shorter than the ~1–10 ps intramolecular energy redistribution timescale.
Throughput at full scale:
2 × 10⁸ apertures × 10⁹ bits/s per aperture, dual-face configuration.
Emerging techniques like plasmonic nanocavity confinement (demonstrated ~1 nm optical field localization), ultrafast shaped femtosecond IR pulses, and engineered THz antenna tips could enable true single-bit random access at terahertz speeds.
The physical speed ceiling? The time for fluorine to transit the inversion path at maximum kinetic energy: ~16 femtoseconds. The architecture targets 1 picosecond, leaving two orders of magnitude of headroom.
The system reads/writes from both sides simultaneously. A central controller:
Each face provides an independent, complementary view of the same bit: what Face A reads as "mound" (fluorine toward it), Face B reads as "trough" (fluorine away). This is a built-in consistency check no single-face technology offers.
Storage density
Think of it this way
A 1 cm² sheet — smaller than a postage stamp — holds 447 terabytes. That's roughly 450,000 copies of the entire English Wikipedia. Or about 90 million high-resolution photos. On a surface you could cover with your thumbnail.
The math is simple and exact. Each bit occupies the area of one carbon site in the hexagonal lattice:
\(A_{\text{bit}} = \frac{\sqrt{3}}{4} \, a^2 \approx 0.0279 \text{ nm}^2\), where \(a = 0.254\) nm is the lattice constant.
\(N = \frac{10^{14} \text{ nm}^2}{0.0279 \text{ nm}^2} \approx 3.58 \times 10^{15} \text{ bits} \approx 447 \text{ TB}\)
This isn't an estimate or a projection. It's an arithmetic consequence of atomic positions in a known crystal.
Roll the fluorographane into a tape and wind it on a spool — like old magnetic tape, but at atomic scale. The non-stick fluorocarbon surface means layers don't bond to each other. No spacers needed.
ZB = zettabyte = 10²¹ bytes = 1 billion terabytes
Write energy per bit: 4.6 eV ≈ 7.4 × 10⁻¹⁹ J. For comparison, NAND flash needs ~10 nJ per bit — 10 billion times more. At full dual-face throughput (25 PB/s), total write power is ≤148 mW. Passive air cooling suffices. Idle storage dissipates exactly zero power.
Context
Here's every number from the paper's comparison table, contextualized:
| Property | CF (this paper) | NAND Flash | DRAM | HBM3 | MRAM | PCM | ReRAM |
|---|---|---|---|---|---|---|---|
| Areal density (/cm²) | 447 TB | ~1 GB | ~0.1 GB | ~0.1 GB | ~0.01 GB | ~0.1 GB | ~0.1 GB |
| Throughput | 25 PB/s* | 0.5 GB/s | 50 GB/s | 1 TB/s | 10 GB/s | 1 GB/s | 1 GB/s |
| Retention power | 0 | ~mW/GB | ~mW/GB | ~mW/GB | 0 | 0 | 0 |
| Non-volatile? | Yes | Yes | No | No | Yes | Yes | Yes |
| Endurance (cycles) | >10⁶* | 10³–10⁵ | >10¹⁵ | >10¹⁵ | >10¹² | 10⁸–10¹² | 10⁶–10¹² |
| Radiation hardness | High | Low | Low | Low | Moderate | Low | Low |
| Bit drift | Impossible | N/A | Refresh | Refresh | N/A | Yes | Yes |
| Raw material cost | $0.20/TB | $50–100/TB | $3k/TB | $10k/TB | $500/TB | $100/TB | $80/TB |
*Tier 2 projection at full array scale. Tier 1 throughput ~1–1000 bits/s. Endurance is a conservative estimate; physical arguments support effectively infinite cycling.
DNA data storage — like fluorographane — encodes information in molecular geometry. But the density gap is stark: fluorographane's gravimetric density is ~7 × 10²¹ bits/gram, exceeding DNA's theoretical ceiling (~10¹⁸ bits/gram) by nearly four orders of magnitude and the best demonstrated DNA density by six.
DNA also suffers from throughput limits: writing requires chemical synthesis at ~1–10 nucleotides/second; reading requires PCR amplification and sequencing (hours to days); random access scales poorly. And DNA needs cool, dry, oxygen-free storage to prevent degradation — fluorographane needs nothing.
A 1 cm² fluorographane sheet has a mass of ~0.5 µg. At current lab prices (~$50–100 for CVD graphene on copper foil), this yields a cost of ~$0.10–0.20 per terabyte. NAND flash at current crisis pricing exceeds $50/TB for consumer products — a differential of four orders of magnitude.
Fabrication requires approximately four process steps (CVD growth, transfer, fluorination, quality check) and two materials (carbon and fluorine). Compare with modern 3D NAND: 700+ process steps, dozens of materials, and $20–30 billion fab facilities.
Robustness
Think of it this way
If one tile falls off a mosaic, only that tile is lost — the rest of the artwork is fine. Similarly, a missing fluorine atom (vacancy) only affects that one bit. It doesn't cascade. It doesn't propagate. And because fluorine atoms don't migrate (5.6 eV bond energy!), the vacancy stays exactly where it is.
The paper presents a complete degradation hierarchy, ordered from most to least likely at room temperature:
Rate: ~10⁻⁴⁸ s⁻¹ (~10³¹ years). Only at edge/grain-boundary sites — which are excluded from the bit map. The most vulnerable sites store no data.
Rate: ~10⁻⁶⁵ s⁻¹. This is the ceiling on data loss probability. Utterly negligible.
Requires 5.6 eV — even harder than inversion. The hardest thermal process for memory bits.
Localized trajectories. No cascading gate-oxide breakdown. Standard error-correcting codes restore data. Radiation-hard — suitable for space, medical, and nuclear environments.
The dual-face architecture also provides intrinsic defect detection: a valid bit always shows "present" on one face and "absent" on the other. Any deviation from this pattern flags a defect automatically, at read speed, with zero overhead.
🔁 Go deeper: Write enduranceEach write cycle inverts a bond orientation without breaking or forming any bond, depositing or removing any material, or altering any phase. The energy is absorbed by the lattice and dissipated as phonons within picoseconds — leaving no structural trace.
In NAND flash, endurance is limited by progressive oxide degradation (10³–10⁵ cycles). In fluorographane, three arguments support high endurance: (1) the inversion is geometrically reversible between crystallographically equivalent states; (2) the barrier energy is distributed across ~10 atoms, keeping per-bond strain far below dissociation; (3) the carbon scaffold is not modified by the switching event. The paper conservatively estimates >10⁶ cycles, but notes the physical arguments support effectively infinite cycling.
What it means
The entire stack of registers, cache, RAM, SSD, and tape archive exists because memory has been volatile, expensive, and slow. When a single medium is dense enough to hold everything, fast enough for direct access, and non-volatile by physical law — everything is RAM.
An entire frontier AI model stored on a USB-sized fluorographane module makes inference a local operation. The cloud becomes optional.
All data humanity has ever produced (~120 zettabytes) fits in a single nanotape cartridge. Stored permanently. With zero energy.
Data centers consume tens of gigawatts. A huge fraction goes to memory refresh, idle leakage, and data movement. Fluorographane eliminates retention power entirely and reduces write energy by 10 billion times per bit.
Transferring a 0.4 ZB cartridge over 1 TB/s fiber takes ~13 years. Physically carrying the cartridge is faster than any electronic link. Physical logistics beats telecommunications.
Space missions, nuclear facilities, cultural preservation — anywhere data must survive without power, without maintenance, without protection from radiation.
Scanning-probe prototype — constructible today from commercially available equipment.
Multi-bit array demonstration.
First commercial product: radiation-hard write-once archival storage.
The bottom line
Here's what you could confidently explain to someone else: