A new battery chemistry takes roughly a decade to travel from a lab notebook to a shipping product. The world's largest vehicle markets have legislated the 2030s as the deadline for electrifying everything. Those two clocks do not fit together, and the entire battery industry now lives inside the gap.
The standard story says AI is closing the gap because models got smart. The real story is narrower and more useful. Batteries hit three walls in the 2020s, discovery, recovery, and provenance, and all three turned out to be the same problem wearing different clothes. Each is a place where the industry ran out of the ability to evaluate cheaply and to remember structurally. AI matters here for two operations, not one intelligence: it makes evaluation cheap (a learned model standing in for an expensive test) and it makes state persistent (a structured record that survives every handoff). Every system in this piece, the closed-loop labs, the materials models, the recycling lines, the plant that burned, the battery passport now written into EU law, is one of those two operations in a different material. Irys, the company behind this essay, sells those two operations for documents; watch where the battery industry keeps rediscovering them, because the ending is a regulation that made them mandatory.
The Three Walls
The demand side is arithmetic. More than seventeen million electric vehicles were sold in 2024, roughly one in five new cars worldwide by the IEA's count, up from about one in twenty-five in 2020. Every long-range EV pack holds thousands of cells, and the materials, lithium, nickel, cobalt, flow through a five-handoff chain (mine, refine, active materials, cell, pack) whose refining and cathode stages sit mostly in China. The chain loses information at every handoff: each step optimizes itself and passes on a spec sheet, not a memory.
Wall one is search. Chemical space is combinatorial; in early 2024, Microsoft and PNNL reported screening over 32 million candidate solid-electrolyte compositions in about 80 hours, where a strong lab fully characterizes perhaps a handful of new compounds a month. Wall two is entropy. A recycler does not receive a category; it receives a river, mixed chemistries (NMC 523/622/811, LFP), formats, damage states, and histories, which fixed automation cannot absorb and people disassemble by hand at real risk. Wall three is provenance. From August 2031, batteries sold in the EU must contain minimum recycled content (16% cobalt, 6% lithium, 6% nickel), and from February 18, 2027 most EV and industrial batteries must carry a digital battery passport. You cannot verify either rule from a warehouse of black powder. The state has to have been kept.
Prediction: Reading a Cell's Future
The cleanest demonstration of what AI actually is, industrially, came from a battery lab before it came from any of the famous model companies.
In 2019, a Stanford/SLAC/MIT/Toyota Research Institute team led by Kristen Severson published the dataset nobody had built: 124 commercial lithium-iron-phosphate cells from one production lot, each charged with its own randomly drawn fast-charging policy, cycled until death. Identical hardware lasted anywhere from about 150 cycles to about 2,300. Every fleet on the road is carrying that cloud inside it, unlabeled. The team's prediction result was almost unfairly simple: take the discharge-capacity curve at cycle 100, subtract the curve at cycle 10, voltage by voltage, and compute the variance of the difference. Quiet difference, long life; ragged difference, short life. One number per cell, from weeks of testing, predicted final lifetime with roughly nine percent error, and the team published the dataset open, which is why an entire prediction field now trains and competes on it.
A year later the same group closed the loop: the early predictor feeding a Bayesian optimizer that chose each next charging protocol to test, real cells doing short confirmations, every result retained. The loop found target-hitting fast-charging protocols in sixteen days; the team's estimate for exhaustive search without early prediction was over five hundred days. The mechanism matters more than the number: the experiment became a query against accumulated state. Read the expensive thing once, turn it into structured state, and every later question is cheap.
That accounting, pay for structure once, query forever, is the entire economic argument for stateful reasoning systems, and it is why the Irys Swarm API is built the way it is: a document corpus is processed once into typed, traceable state, and later questions cost a query instead of another full read. Irys reports $1.30 per task against $50.90 for comparable systems on the Harvey Legal Agent Benchmark, Irys's own number, on its own benchmark run, and the mechanism of the gap is the lab's mechanism: never pay twice for what you already learned. The parallel is at the economics layer, not the physics layer, and it is worth stating that plainly.
Discovery: Cheap Proposals, Expensive Verification
Upstream, the same pattern runs at the scale of matter itself. Density functional theory can price a crystal's stability from first principles, at hours to days of compute per structure. The Materials Project spent a decade turning that expensive evaluation into a public commons, over 150,000 computed compounds, queried by everyone, paid for once. Learned surrogates trained on that commons now evaluate candidates at inference cost; DeepMind's GNoME wrapped surrogates in an active-learning loop and flagged 2.2 million new crystal structures, about 380,000 of them predicted stable, DeepMind's framing compares the haul to roughly 800 years of the field's discovery pace. The Microsoft/PNNL funnel (32 million to 23 finalists, 18 new, per the company's own materials) ended with a synthesized electrolyte the companies say uses about 70% less lithium.
Then comes the part the press releases skip. A predicted crystal is not a made crystal. Only a small fraction of GNoME's 380,000 had been synthesized anywhere in the first years after release, and crystallographers have since documented duplicate and misidentified structures in the databases. The A-Lab at Berkeley, a genuinely autonomous lab that roboticized solid-state synthesis, announced 41 novel compounds in 17 days in Nature, and independent chemists re-reading the diffraction evidence concluded the count of unambiguously new materials was zero or a handful. In January 2026, Nature corrected the paper to 36 compounds from 57 targets. The lesson is architectural, not moral: in any closed loop, whatever is cheapest to measure becomes the objective, whether or not it is what you meant to optimize. The only structural defense is lineage, every claim carrying its evidence, every product its measurement file, so a broken claim fails locally and traceably instead of taking the paper's credibility with it. We learned the same lesson on our own research when a preregistered control killed our own headline result; the difference between a lab that corrects in a day and one that corrects in two years is whether the record was kept along the way.
Recovery: The Physics of Unmaking
Recycling runs three strategies. Pyrometallurgy melts everything at around 1,450°C, indifferent to what arrives, but lithium mostly reports to the slag. Hydrometallurgy dissolves the shredded "black mass" and separates metals through pH-staged chemistry and solvent extraction into battery-grade salts: high recovery, low temperature, and a grinding multivariable control problem whose optimal setpoints drift with every delivery of feed, the exact regime where learned models beat fixed schedules. Direct recycling, the least mature, regenerates the degraded cathode itself, preserving the engineered crystal value that the other two destroy, and it is only possible when the input is sorted, known, single-chemistry. You can only directly recycle a cathode you know. The thesis, disguised as a flowsheet.
On the identification front, spectroscopy (near-infrared and X-ray fluorescence) reads the returning stream at conveyor speed, robots adapt disassembly to whatever pack design arrives, and the systems that work keep correction-fed state, the operator's reclassification becomes the next labeled example. The plant side runs digital twins that simulate decisions before the physical line commits. The vendors' numbers are the vendors' numbers: Mercedes-Benz's opening announcement claims more than 96% recovery of valuable materials at its Kuppenheim plant built on Primobius technology, and Nth Cycle's own materials claim double-digit emission cuts for its electro-extraction. The mechanisms are real regardless, and the recycling plant of the 2030s is less a building with software in it than a state machine that eats batteries.
Failure: Fredericktown, October 2024
On October 30, 2024, the Critical Mineral Recovery plant outside Fredericktown, Missouri, 225,000 square feet of battery shredding, exploded and burned, and never reopened. Official investigators never announced a confirmed cause; the leading theory in industry analysis (not in any agency finding) is a hydrogen vapor cloud from battery off-gas during shredding. Thermal runaway vents hydrogen, ethane, and methane; hydrogen burns across roughly 4, 75% concentration in air, and a shredder is, by design, a hydrogen generator run under inert atmosphere with gas containment as the safety case.
The system-level failure was the control surface: periodic human attention watching continuous drift, and fires had already struck twice at facilities linked through shared ownership, per local reporting. The counterfactual is not speculative, sensors streaming against a digital twin's expected envelope, thresholds as interlocks that act at a fraction of the flammability limit, calibration records that keep the watching layer itself watched. We have written about this pattern before as the exception leak: a green dashboard while one reading drifts. In a plant, as in a legal workflow, the dangerous object has to be its own typed thing, a blocker with a threshold, an owner, and a memory, or the drift stays invisible until it is a fireball.
State: The Thesis, Now with the Force of Law
The European Union's Battery Regulation 2023/1542 is best read as a software document. Recycled-content minimums from 2031 (rising in 2036 to 26% cobalt, 12% lithium, 15% nickel); recovery targets (80% of lithium by 2031); and from February 2027, a per-battery digital passport carrying chemistry, provenance, carbon footprint, recycled content, and due-diligence data, role-scoped, append-mostly, outliving every transaction that touches it. Not a fleet average. Not a certificate. Typed, persistent, auditable state, attached to a physical object, at continental scale.
Every circular claim audits back through records the passport forces into existence, and the quieter layers matter as much: industry data spaces like Catena-X building the governance rails; national projects welding a semantic spine across a chain of rivals whose data has never joined. Second-life markets are the small proof: a retired pack with a biography and a health estimate carries its evidence and gets priced on it; a pallet of unknown packs gets priced at the worst case, and the honest sellers withdraw, Akerlof's lemons on lithium. State revives the market.
This is where the document-reasoning parallel stops being an analogy. The regulation just created an industrial document mountain: due-diligence dossiers, recycled-content verification chains from recycler to cathode plant to vehicle, conformity documentation, and patent landscaping in one of the most patent-dense electrochemical fields on earth. That is a many-long-documents reasoning problem with an auditability requirement, the exact shape of the Irys Swarm API, and the legal work itself is what Irys One is for. Irys is not a passport platform, and the passport does not need us to exist; two fields that never met converged on the same architecture, keep the state, trace the claim, one for atoms, one for documents.
The Document Mountain, Concretely
Strip the regulation down to work items and the battery passport creates four document problems that barely existed five years ago.
Conformity: every obligation in Regulation 2023/1542, recycled content, recovery rates, due diligence, footprint declarations, has to be mapped to evidence across supplier contracts, assay certificates, and process records, and re-mapped every time the regulation or a supplier changes. Patent landscaping: battery chemistry is among the most patent-dense fields in engineering, and every process choice in a recycling plant walks through a thicket someone else filed. Supply-chain due diligence: the cobalt-chain documentation a carmaker must assemble, keep, and defend. And the plant's own memory, permits, safety cases, quality records, the shadow plant from the Fredericktown section, now inspectable by regulators and insurers rather than reconstructable after the fact.
None of these are model problems. They are long-corpus reasoning problems with an auditability requirement: read thousands of documents once, extract the obligations, claims, and evidence into structured state, and answer every later question as a query that carries its lineage back to the source paragraph, the same cost structure the closed-loop lab proved out in atoms. That is what the Irys Swarm API is: read a document set once, query the accumulated understanding cheaply forever, with every finding traceable to the document, the step, and the evidence that produced it. The legal and compliance work on top of that state, the matters, the drafts, the review, is what Irys One is for. To keep the scope honest one last time: Irys is not a passport platform and does not run plants. It is the reasoning layer the passport's existence makes necessary, and the battery industry is only the first place where that layer stopped being optional.
The Loop Closes
China holds roughly seventy percent of global battery recycling capacity, the headline figure of a joint CAS, Deloitte industry analysis published in February 2025; industry-research figures, not census. That capacity is anchored by Brunp, the CATL subsidiary whose loop, cells to vehicles to scrap to cathode back to cells, is the national-scale version of everything above, compounded by standards, subsidies, and feedstock that comes home. The Western response runs on companies (Redwood's Bremerhaven line, ABTC with Siemens-built controls), regulation as guaranteed demand, and subsidies; its honest scoreboard is edges narrowing while the core loop compounds.
What replicates is the state layer, data standards, passports, measurement rules, born-digital plants. What does not easily replicate is a decade of integrated scale. The playbook question generalizes: any industry with expensive evaluation, entropic returns, and provenance pressure is next, and solar, 60 to 78 million tonnes of cumulative waste projected by mid-century, is already at the door.
The closing accounting is the one this essay opened with. Cheap evaluation, persistent state, and the third operation the retractions taught: verification, with lineage on every claim. The battery industry spent a century learning that the valuable thing was the lithium in the ground. The next economy's lesson is that the lithium was never the point, the state around it was. Whoever keeps the state owns the loop, and the loop, not the mine, is where the value sits.
Related: Two Random Vectors and Qwen3-4B · What AlphaEvolve Teaches Legal AI About Evaluated Work · Auditability Is the Product · Why Prompt-As-Memory Fails.
Irys Swarm API: irys.ai/irys-api · irys.ai · Partners · Book a demo · LinkedIn
