A Telegram item in Russian, which the newsroom flagged, says Penn State scientists built an energy-efficient memory for AI out of DNA. The underlying work is a hybrid memristor: synthetic DNA coated with silver nanoparticles, combined with a thin film of quasi-2D perovskite. A memristor is a resistive element that keeps its state without power, and memristor arrays are a long-standing candidate for in-memory and neuromorphic computing, where data is processed where it is stored. That is the reason the press release ties the work to AI and data-center energy use.
For builders, the useful question is whether this changes anything about the memory hierarchy behind AI systems: HBM, DRAM, NAND, or emerging resistive and phase-change RAM. The answer depends on how mature the device is, what the energy claim actually measures, and whether any commercialization path exists. The rest of this piece tests each of those.
What we checked
- Does the cluster reveal a broader market or architecture transition, and not just a single release? refuted. The evidence shows one university lab paper: a materials-science device result, with a patent application filed and no product, partner or roadmap. The researchers' stated next steps are only to refine the method and explore other bio-inspired electronics. Nothing in the sources points to a shift in market or architecture. [1][4]
- Is the story new (September 2026)? refuted. Penn State published its release on February 24, 2026 and Tech Xplore republished it the same day. Popular Science ran its piece on August 24 and Tech Briefs on September 1, so the Telegram item is a late echo of the same release. [1][2][4][3]
- Is the headline energy-saving figure well defined? refuted. Penn State's release says the device does the same memory function as similar technologies 'but only uses one-tenth of the power'. It also quotes Bed Poudel saying it 'consumes 100 times less power'. Neither statement names a baseline technology. Secondary outlets mostly repeat the 100x figure. [1][5][4]
What we found
- The work is a single peer-reviewed paper: 'Molecularly Engineered Highly Stable Memristors with Ultra-Low Operational Voltage: Integrating Synthetic DNA with Quasi-2D Perovskites' in Advanced Functional Materials (DOI 10.1002/adfm.202530539). Penn State announced it on February 24, 2026. The corresponding authors are Kavya S. Keremane, Bed Poudel and Luyao Zheng of Penn State and Rashmi Jha of the University of Cincinnati. Funding came from NSF, NIH, Penn State and the University of Cincinnati, and a patent application has been filed. [1][2]
- The device claims the release does support are reliable switching below 0.1 V, consistent operation up to almost 250°F, and stability at room temperature for more than six weeks. The authors use short, rigid synthetic DNA instead of tangled natural DNA to get controlled structure. Keremane is quoted as saying that neither DNA nor perovskite alone 'produce[d] near as robust a result as the combination.' [1][3]
- The energy claim contradicts itself inside the primary release. One passage says the device uses one-tenth of the power, while Poudel's quote says it uses 100 times less, and neither names what it is compared against. Popular Science, Bioengineer.org and others chose the 100x figure. Any coverage of the story should not repeat either number as a benchmark against production memory. [1][4][5]
- The '215 million gigabytes per gram' figure that runs through the coverage describes DNA as a molecular storage medium in general. It does not describe this device. The memristor uses DNA as a conductive, doped layer in an electronic stack, not as a store of sequence-encoded data, and no outlet gives a device-level density. [1][4]
- The paper's own metrics could not be verified on the publisher site, which returned HTTP 403. The abstract as indexed by search gives an Ag/perovskite/Ag-DNA/Pt stack, a power density of 0.01 W/cm², forming-free operation, 10^3 endurance cycles, 4×10^3 s retention, and an ON/OFF ratio above 10^5 held for more than six weeks. Treat these as unconfirmed. If they are accurate, the endurance and retention are modest for a memory technology: about an hour of retention, against years for non-volatile production memory. [6]
- None of the opened coverage quotes an independent expert. Popular Science's sources are all co-authors. Nothing reports array-scale integration, CMOS compatibility, switching speed, or a neuromorphic or AI workload demonstration. The release lists the next steps only as refining the approach and exploring other bio-inspired electronics. [4][1]
What it means
For anyone building or buying AI infrastructure, this result changes nothing in the planning horizon. It is a materials-science proof of concept: one lab, one paper, no workload demonstration, and no independent replication or commentary. Lead-halide perovskites are also known to have stability and manufacturing difficulties, and the reported stability window is measured in weeks. The 'AI memory' framing comes from the press release, not from an AI benchmark. Do not cite the 10x or 100x figures in cost or energy models: the source disagrees with itself and names no baseline.
The broader signal is weaker but real. Ultra-low-voltage resistive switching remains an active research area for analog in-memory computing, and hybrid organic/biomolecular layers are one more route to lower switching energy. Teams following neuromorphic or compute-in-memory hardware should watch for array-level results (variability, endurance above 10^6 cycles, retention in years, CMOS back-end compatibility, and matrix-vector multiply accuracy) before treating this material system as a candidate. Editorially, the story belongs in a 'lab result, resurfaced' slot, not a 'shift' update.
Still open
- What baseline does the 'one-tenth' and '100 times less power' claim refer to, and which figure appears in the paper?
- Can the abstract's 10^3-cycle endurance and 4×10^3 s retention be confirmed from the full text, and how do they compare at matched conditions?
- Has the device been shown in a crossbar array, or on any in-memory or neuromorphic compute task?
- How large are the devices, and is the process compatible with CMOS back-end fabrication? Are there lead-toxicity or manufacturing constraints?