# Whoever Owned the Furnaces > From 2036: teams of software agents learned to design metal alloys faster than anyone could melt them. The defence and industrial payoff was supposed to be enormous. Instead the cost of proof went up, and the people who already owned furnaces owned the whole decade. - Publication: What If? Magazine - Correspondent: Cy Skewhouse - Vertical: Edge Technologies - Published: 2026-10-30 - Reading time: 10 min - Tags: multi-agent-ai, materials-discovery, alloys, qualification, defence-manufacturing - Canonical URL: https://www.whatifmag.com/blog/the-composition-was-the-cheap-part What If? Magazine publishes speculative journalism written from a 2036 vantage. The reporting and sources are real; the 2036 framing is a deliberate editorial device, not a claim about the present. --- I spent most of 2025 convinced that the hard part of a new metal alloy was thinking of it. From 2036 that embarrasses me, because the record was sitting in the open. When teams of software agents began proposing compositions faster than any foundry could melt them, the bottleneck did not move downstream. It hardened. The scarce thing was never the idea. It was the signed paper that lets somebody with liability put the part on an aircraft. ## Why the agents were worth building Begin with the good, because it was genuinely good. The architecture is a division of labour. One agent reads the literature, one runs the thermodynamics, one drives atomistic simulation, one searches composition space, one argues with the rest about whether the plan is sound. In July 2024 Alireza Ghafarollahi and Markus Buehler at MIT published exactly that, eleven language-model agents named Planner, Critic, Engineer, Coder, Scientist and Plot Analyzer among them, wired to real atomistic solvers. It derived a ductility threshold of 1.26 for niobium-molybdenum alloys and put the ductile boundary above 60 percent niobium, reproducing the experimental ordering in which molybdenum and tungsten come out brittle and niobium and tantalum do not.[^1] Three months later the same pair swapped the expensive atomistic step for a graph neural network and aimed it at NbMoTa.[^2] The physical side kept pace: Berkeley's A-Lab ran 17 days with minimal human intervention, tested 353 recipes and produced 36 of its 57 targets.[^3] Now the constraint all of this would genuinely have relieved. When the federal materials push launched in 2011, its own framing document put the lag from initial research to first use at 10 to 20 years, and set a goal of halving it.[^4] That lag is why a hypersonic leading edge that must stay sharp under extreme heat flux is a research programme instead of a purchase order.[^5] I want to be plain about how much that concedes. The search half of materials science really was [gated by one professor's hunch](https://www.whatifmag.com/blog/the-bench-kept-the-doubt) and one graduate student's furnace calendar, and the agents broke that gate open. The concession is not small. ## What the simulations could not price The trouble is what the agents are reading when they reason about phases. Multicomponent thermodynamics runs on hand-assessed databases, and somebody has to do the assessing. Stéphane Gorsse and Oleg Senkov took one of the standard high-entropy databases apart in 2018 and counted: 325 binary systems across 26 elements, 294 of them fully assessed, and 2,600 ternary systems, of which 136 were fully assessed.[^6] That is about five percent. Nineteen of every twenty intersections on that map are drawn by inference from the neighbours. Sometimes the inference is fine. Dropping the ternary terms from cobalt-chromium-iron-nickel still reproduced 98 percent of the face-centred and body-centred phase points. Do the same for cobalt-copper-iron-nickel and only 63 percent of the face-centred points survive, because the miscibility gap lands in the wrong place.[^6] A thermodynamics agent cannot tell you in advance which of those two cases it is standing in, and nobody was pricing that difference. ![Grouped bar chart of binary and ternary systems in a thermodynamic database, total versus fully assessed.](https://www.whatifmag.com/images/articles/the-composition-was-the-cheap-part/the-composition-was-the-cheap-part--fig1.png) **Figure 1.** Coverage of a standard multicomponent thermodynamic database, read back from 2036. The generated candidates were confident. The underlying assessments were thin. The checking caught up twice in short order. In April 2024 Anthony Cheetham and Ram Seshadri reviewed the 384,870 structures a deep-learning pipeline had flagged as stable, drew ten at random, and found every one already in the standard crystal database. No use had been shown for any of the set. Another 18,138 of the entries held promethium, actinium or protactinium, and 23,529 more held technetium, neptunium or plutonium.[^7] Then in January 2026 the A-Lab paper was corrected. Novelty had meant new to the prediction platform, not new to science. The confirmed count fell from 40 to 36, and one compound was pulled because it had been sitting in the training data.[^8] Both of those were ordinary science working, and neither was a scandal. [The map was mostly inference](https://www.whatifmag.com/blog/the-wind-tunnel-was-the-last-honest-room). ## The qualification bill came due [Here is the number that decides everything](https://www.whatifmag.com/blog/the-model-was-free-the-box-was-not), and almost nobody outside metallurgy was carrying it in 2026. To put a new alloy into the handbook aerospace actually designs against, primary tensile properties at A-basis and B-basis require 10 heats, 10 lots and 100 tests. A non-parametric analysis needs 299. Secondary properties, compression and shear and bearing, need 3 heats, 10 lots and 20 more tests. The public specification under all of it takes support from at least two aerospace users and a 28-day review before approval.[^9] The handbook runs 2,700 pages and carries allowables for more than 400 material specifications.[^10] A heat is one melt. Ten heats means ten separate melts, by a producer who can repeat them on demand, years apart, with traceable powder or billet behind each one. ![Three tensile test frames in a row in a materials laboratory, with a broken coupon still held in the nearest jaws.](https://www.whatifmag.com/images/articles/the-composition-was-the-cheap-part/the-composition-was-the-cheap-part--fig2.png) **Figure 2.** A materials test cell, 2036. The frames are the slow step, and they were never the step anyone funded. Defence made it harder, not easier. A 2022 survey of the gap between civil and military certification found each service branch had grown its own printing processes. Policy set in 2021 had not produced visibility across the department. The spaceflight standard it cited demands a quantitative inspection of every Class A part, and one printed link on a V-22 nacelle in 2016 took real engineering time to certify.[^11] The standards community had already counted the holes. A 2023 roadmap drawn up by roughly 300 people from about 150 organisations listed 141 gaps, 54 of them high priority. Another 91 needed new research before a standard could even be written.[^12] The one clean counterexample proves the shape. NASA's computationally designed superalloy GRX-810 lasts up to 2,500 times longer than conventional nickel-base alloys above 2,000 degrees Fahrenheit, and it went to four licensed producers in May 2024.[^13] That was a licence, not an allowable. ## The pedigree became the asset Pull four documents filed in four different drawers. A commodity table: world rhenium production was about 81,000 kilograms in 2025, roughly 80 percent of it bound for superalloys, at 2,600 dollars a kilogram, with American net import reliance at 75 percent.[^14] A funding announcement: in September 2025 Periodic Labs raised a 300 million dollar seed round for autonomous laboratories, founded in part by the researcher who had led the crystal-generation work and co-authored the A-Lab paper.[^15] A committee minute with its 141 gaps.[^12] A database changelog with its 136 assessed ternaries.[^6] Nobody read those four together. Read together they say one thing. Candidate compositions fell to the cost of compute. Proof did not fall at all, because proof is melts and coupons and accredited laboratories and a committee that meets twice a year, so the ratio inverted. The idea became almost free, the signature became the expensive half, and value moved to whoever already owned melt capacity, test frames and a seat on the committee. > Making a composition free did not make an alloy cheap. It made the test coupon the scarce asset, and the people who already owned furnaces owned every one of them. ![Logarithmic bar chart counting candidates, database entries, tests and confirmed compounds.](https://www.whatifmag.com/images/articles/the-composition-was-the-cheap-part/the-composition-was-the-cheap-part--fig3.png) **Figure 3.** The attrition, counted from the 2023 to 2026 record and read back from 2036. The top bar is generated. The bottom bar is proved. Nobody said "pedigree premium" in 2026. I am naming it from here, which is cheating, but the thing itself was already loaded in the chamber. So who won. The specification holders won, because an allowable is a moat with a committee attached. The powder and billet producers won, because ten traceable heats is a service only a handful of firms can sell. The testing houses won outright. The primes won most of all, since a candidate alloy is worth nothing to a four-person team and a great deal to the only organisation that can afford 299 tests. Who paid is the part that still irritates me. Small teams treated a composition as an asset and discovered it was a lottery ticket. And the programmes that most needed exotic metal, the ones waiting on a leading edge that holds its shape, got 384,870 candidate structures against 400 qualified material specifications and a test queue that had not grown by a single frame. The lesson is older than the agents and it survived them. An alloy was never a formula. It was a formula plus a paper trail that somebody with liability was willing to sign, and only the formula ever got cheap. ## Author's Note This piece is written from 2036 and looks back at 2023 to 2026. The multi-agent alloy systems, the autonomous laboratory results and their 2026 correction, the database coverage figures, the handbook data requirements, the standards gap counts, the commodity figures and the funding round are real and sourced below. Three boundaries. First, the four-agent line-up behind this commission, a thermodynamics agent, an atomistic agent, a compositional agent and a synthesis and applied physics agent, matches no published system; it is the shape such systems take, and the published system cited here divides the work differently. Second, no verified source as of 2026 shows an agent-designed alloy carrying a flight allowable or a defence qualification, so the bottleneck here is extrapolated from documented requirements, not from a documented failure. Third, the 2036 end state in which proof rather than discovery sets the price is speculation. The correspondent's first person marks the difference between what was knowable in 2026 and what is obvious from 2036; it records changes of mind, not events. ## Works Cited [^1]: Ghafarollahi, A. & Buehler, M. J. "Automating alloy design and discovery with physics-aware multimodal multiagent AI." *PNAS* 122(4):e2414074122, 24 January 2025. DOI 10.1073/pnas.2414074122. Preprint: https://arxiv.org/abs/2407.10022 [^2]: Ghafarollahi, A. & Buehler, M. J. "Rapid and Automated Alloy Design with Graph Neural Network-Powered LLM-Driven Multi-Agent Systems." arXiv:2410.13768, 17 October 2024. https://arxiv.org/abs/2410.13768 [^3]: Szymanski, N. J. et al. "An autonomous laboratory for the accelerated synthesis of inorganic materials." *Nature* 624:86–91, 29 November 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10700133/ [^4]: Executive Office of the President. "Materials Genome Initiative for Global Competitiveness." June 2011, pp. 5–7, 16. https://mgi.gov/sites/mgi/files/materials_genome_initiative-final.pdf [^5]: DARPA. "Materials Architectures and Characterization for Hypersonics (MACH)." Tactical Technology Office programme page. https://www.darpa.mil/research/programs/materials-architectures-and-characterization-for-hypersonics [^6]: Gorsse, S. & Senkov, O. N. "About the Reliability of CALPHAD Predictions in Multicomponent Systems." *Entropy* 20(12):899, 24 November 2018. https://www.mdpi.com/1099-4300/20/12/899 [^7]: Cheetham, A. K. & Seshadri, R. "Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery." *Chemistry of Materials* 36(8):3490–3495, 8 April 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11044265/ [^8]: "Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials." *Nature* 650(8100):E1, 19 January 2026. https://pmc.ncbi.nlm.nih.gov/articles/PMC12872444/ [^9]: MMPDS / Battelle. "Adding an Alloy to MMPDS." Data and analysis requirements, October 2015. https://www.mmpds.org/wp-content/uploads/2015/03/MMPDS-Adding-an-Alloy_Oct2015.pdf [^10]: Battelle. "Metallic Materials Properties Development and Standardization (MMPDS)." Programme summary sheet. https://www.mmpds.org/wp-content/uploads/2025/05/Slick-Sheet_884_Metallic-Materials-Properties-Development-and-Standardization-MMPDS.pdf [^11]: Seidel, M. "Gaps Between Civilian and DoD in Certifying Additive Manufacturing (AM) Parts." DSIAC Technical Inquiry Response Report DSIAC-BCO-2022-230, June 2022. https://dsiac.dtic.mil/wp-content/uploads/2023/01/TI-Response-Report_DSIAC_Gaps-Between-Civilian-and-DoD-in-Certifying-Additive-Manufacturing-AM-Parts_612023.pdf [^12]: ANSI and America Makes. "Standardization Roadmap for Additive Manufacturing, Version 3.0." 17 July 2023. https://www.ansi.org/standards-news/all-news/2023/07/7-17-23-america-makes-and-ansi-publish-standardization-roadmap-for-additive-manufacturing-version-3 [^13]: NASA Glenn Research Center. "NASA Licenses 3D-Printable Superalloy to Benefit US Economy." 9 May 2024. https://www.nasa.gov/centers-and-facilities/glenn/nasa-licenses-3d-printable-superalloy-to-benefit-us-economy/ [^14]: U.S. Geological Survey. "Rhenium." *Mineral Commodity Summaries 2026*, February 2026. https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rhenium.pdf [^15]: Wiggers, K. "Former OpenAI and DeepMind researchers raise whopping $300M seed to automate science." TechCrunch, 30 September 2025. https://techcrunch.com/2025/09/30/former-openai-and-deepmind-researchers-raise-whopping-300m-seed-to-automate-science/