The Terafactory Age

A graduate student runs four cell-painting experiments before lunch. An AI agent has drafted six hypotheses by eight in the morning. An autonomous lab finishes another model-proposed synthesis run, having outpaced its own discovery model for weeks. An autonomous lab realized 36 compounds from 57 model-proposed targets in seventeen days (Szymanski et al., Nature 2023), with the upstream model predicting millions of stable inorganic crystals (Merchant et al., Nature 2023). The closest live model-to-bench loop, and the precedent the body argument scales from. Each result lands somewhere private, and none updates a shared record on which the next decision depends, because no shared record exists.

This is the scientific present. AI now produces hypotheses, code, protocols, reviews, and experiment plans cheaper than the institutions of science can absorb them, and closed-loop autonomous labs are beginning to run at tempos that would have looked alien a decade earlier. OpenAI reported an autonomous cell-free protein synthesis loop across six rounds and more than 36,000 reactions, cutting protein production cost by roughly 40 percent (OpenAI / Ginkgo Bioworks, Feb 2026). Treat this as an organization-reported frontier example: in that loop, wet-lab capacity, not generation, was the rate-limiter. Wet-lab capacity is the rate-limiter in the examples that matter here, not generation, and that gap is the structural defect of the present moment.

Science needs two missing things to act on this throughput. The first is the shared record: findings, evidence, dependencies, and corrections compounding across institutions instead of fragmenting across them. The second is a body: facilities that execute against that record at industrial scale, orchestration that turns accepted state into physical action, and a federated network that lets a correction in one corridor reach a synthesis hall in another. A federation of scientific gigafactories composing on shared state is a terafactory. The record is the map. The body acts on it.

The terafactory is not the engine. It is where the engine touches matter. The Discovery Engine names the loop underneath: goal, frontier, task, activity, diff, attestation, event, Atlas update, next action. This essay asks what happens when that loop reaches labs, factories, capital, government, and the physical world.

The first proof is smaller than the world described here: one disease frontier, one autonomous experiment loop, one manufacturing handoff, one clinical or field-deployment partner, and one public writeback. If that loop closes once, the gigafactory stops being metaphor. If it closes ten times, gigafactories compose into the terafactory. If it does not close, the agents that arrive write activity that does not become knowledge, and the factories that get built are isolated gigafactories owned by whoever shipped first.

Imagine the architecture composed and operating. A null result in Boston propagates through the dependency graph and weakens a downstream target hypothesis at three programs in Singapore by the end of the week. A wildlife sample routes from a regional surveillance feed into a primer-production line before the second sampling band confirms an outbreak signal. A failed cathode chemistry weakens the dependent mechanism across every battery program reading against the same frontier within two days. A drought-tolerant cultivar candidate reaches its field-trial committee with the upstream evidence already inspected. None of these scenes requires magic. Each requires the same thing: state that can command action and action that writes back.

The next ten years decide who builds this body and on what terms. The frontier AI labs are already constructing their own end-to-end versions inside their corporate stacks: closed substrates, closed runtimes, closed manufacturing wings, all running on agents that outperform anything publicly accessible. The open public version exists as a sketch, a few repositories, and a small coalition of funders who understand that this category will be decided before most of science notices it as a category. The genre is older than this essay. Bush’s Science: The Endless Frontier (1945) proposed the institution that became the NSF. Engelbart’s “Augmenting Human Intellect” (1962) framed how tools amplify cognition before computers were ubiquitous. Berners-Lee’s CERN proposal (March 1989) sketched the linked-document protocol that became the Web. Each argued for infrastructure that did not yet exist by composing what already worked. The body argument is the same move at industrial-scientific scale.

This essay is therefore not a forecast about one facility. It is a fork story. The dates, names, and scenes below are staged future fiction used to test an institutional consequence, not claims about what will happen. One branch gives science a public body that reads from shared state and writes back to it. The other branch gives the same capability to closed institutions first, then asks everyone else to negotiate for access after the registries, compilers, and credentials already belong to someone.

In 2026, the parts of a scientific body exist. None compose. Each layer of the architecture has a working precedent in one domain or another, but the layers do not yet connect; each works for its own community without the others reading from it.

The enterprise-software analogy is Rippling: operations become more powerful when many surfaces read and write one underlying object instead of being stitched together across separate tools. Science is worse because its fragmentation is epistemic, not merely operational. The missing shared object is not the paper or the project. It is the scientific state transition: what changed, where it applies, what evidence supports it, what depends on it, who attested it, and what should happen next.

The state layer has fifty years of precedent. The Protein Data Bank made deposition a condition of publication in major journals by the 1990s and accumulated the experimentally determined structures that AlphaFold drew on. Crossref has held DOI, citation, and retraction infrastructure across competing scholarly publishers for over two decades through a non-profit governance model. The Materials Project has rendered hundreds of thousands of inorganic compounds with computed properties as a federated, addressable, partly attested database since 2013. The Protein Data Bank, established in 1971, holds over 250,000 experimentally determined structures, sustained as a publication condition across institutions and decades. Crossref is the closest live governance precedent for the body version: a non-profit holding canonical infrastructure across competing actors, with elected technical seats and multiple independent implementations. materialsproject.org (Jain et al., APL Materials 2013) is the closest live precedent in any domain for a frontier-state layer with citation-stable IDs and machine-readable APIs. Each is a state-layer system that holds canonical scientific facts across competing actors without collapsing into a single vendor’s platform. None yet maintains the kind of typed-frontier state (claims, evidence, contradictions, confidence histories, dependency edges between them) that a scientific body would read against. The state layer’s primitive in 2026 is the artifact, not the state transition.

The network layer has working precedent in surveillance. Nextstrain has rendered global pathogen evolution as a shared phylogeny since 2015. GISAID has carried genomic deposits across competing national agencies through every recent pandemic. ClinicalTrials.gov has held trial registrations across sponsors and jurisdictions for two decades. Hadfield et al., Bioinformatics 2018; nextstrain.org. GISAID shares the same shape. ICMJE’s 2005 trial-registration policy made registration in ClinicalTrials.gov a precondition of publication in member journals. Each is a network-layer precedent for federated public infrastructure that holds canonical state without collapsing into a single vendor. Each works, and each is decoupled from physical countermeasure capacity: the capacity to see what is changing exists, and the capacity to act on it across institutions does not. The network layer’s task in 2026 is observation. The body version’s task is action: when evidence updates in one corridor, the dependency graph routes the consequence to every facility that needs to know.

The clinical-network layer has working precedent at population scale. UK Biobank has carried five hundred thousand participants through deep phenotyping, biomarker assays, and longitudinal follow-up for nearly two decades, with thousands of peer-reviewed publications drawing from its data. Sudlow et al., PLOS Medicine 2015; ukbiobank.ac.uk. Five hundred thousand participants recruited 2006-2010, deep phenotypic measurement, multi-modal imaging, whole-genome sequencing, and biological samples sustained across institutions and decades. All of Us has enrolled hundreds of thousands of US participants toward a million-person target with diversity-weighted recruitment. All of Us Research Program Investigators, NEJM 2019; allofus.nih.gov. The closest US analog to UK Biobank: million-participant target, longitudinal phenotyping plus genomics, recruitment weighted toward populations historically underrepresented in research cohorts. Both demonstrate that a federated clinical substrate at population scale can be sustained across institutions and decades. Neither yet routes synthesis or manufacturing decisions; the body version has to add that layer without losing what these built.

The runtime layer is the newest piece, and the most rapidly shifting. A first generation of robotic chemistry sites has demonstrated live model-to-bench loops at speeds no graduate cohort could match. Multi-agent platforms generate hypotheses, search for contradictions, plan experiments, and chain tools faster than wet labs can test them. Lu et al., “The AI Scientist” (arXiv 2408.06292, 2024); the v2 follow-up (2025) produced the first AI-generated paper to pass workshop peer review. FutureHouse’s PaperQA and Crow agent specialists handle retrieval and synthesis at research-grade quality; Google Research’s AI co-scientist is a multi-agent collaborator at the hypothesis-generation layer. Each is runtime; none has a coordinated body to act on what it produces. Their drafts land in private logs and transient context windows that close at session end. The deposit problem recurs at the body’s threshold: the throughput of generation grows, and the throughput of inheritance does not. The runtime layer’s drafts in 2026 evaporate. A scientific body would have them deposit instead.

The collective-intelligence layer has working precedent at human scale. Polymath, launched on Tim Gowers’s blog in January 2009, solved the density Hales-Jewett theorem through open public collaboration in seven weeks with more than forty contributors. Gowers & Nielsen, “Massively collaborative mathematics”, Nature 461:879-881 (2009). The first proof that public-blog collaboration could close research-frontier mathematics problems. Nielsen’s Reinventing Discovery (Princeton 2011) is the longer treatment. Galaxy Zoo classified nearly a million galaxies through public participation in months rather than the years a small research team would have needed, with sixty-plus peer-reviewed papers and multiple novel discoveries drawn from the volunteer-generated catalog. Lintott et al., MNRAS 389:1179 (2008). Volunteer-classified morphologies for ~900K galaxies from the Sloan Digital Sky Survey; Hanny’s Voorwerp and Green Pea galaxies were discovered through the volunteer pool. Foldit’s tens of thousands of players matched or outperformed algorithmic solutions for protein structure refinement. Cooper et al., Nature 466:756-760 (2010). Citizen-science contribution at research-grade quality: ~57K Foldit players produced structure predictions competitive with the best algorithmic methods on contested targets. The lesson is not that crowds are virtuous. Distributed work becomes scientific work when tasks are modular, feedback is immediate, contribution units are structured, and reputation is legible. In the body version, a contributor can propose a state transition, annotate evidence, replicate a run, challenge a scope condition, add calibration metadata, audit an agent, or route a task. Most contributions stay noncanonical. Some become evidence. A small number become accepted state.

The physical-execution layer is the most absent. Science has academic labs at one scale, pharma research sites at another, contract research organizations at a third. None operates at the scale of an industrial gigafactory or an advanced-node semiconductor fab, let alone the terafactory generation that followed. The category “scientific gigafactory” does not yet exist as a balance-sheet line item that a sovereign wealth fund can underwrite, an FRO can incubate, or a regulator can inspect. The closest precedent is the BARDA-shaped medical-countermeasure manufacturing facility, which underwrites surge capacity for pandemic response but does not run continuously against a public state layer. BARDA underwrites medical countermeasure manufacturing capacity through cooperative agreements (medicalcountermeasures.gov/barda). A frontier-infrastructure BAA is BARDA-shaped, extended from countermeasures to chronic-disease and surveillance corridors, with public substrate writeback required as a deliverable rather than added as a side effect. The body argument requires building the category.

Protocol consolidation has a proof under pressure. A single shared protocol, one ethics path, one EHR backbone, in a national health system: more than 40,000 participants across 185 UK sites, and a result that changed care worldwide within a hundred days. RECOVERY Collaborative Group, NEJM 2021: dexamethasone reduced 28-day mortality by roughly one third in ventilated patients across 176 UK hospitals. UKRI describes the RECOVERY platform as the world’s largest clinical trial into COVID-19 treatments, with more than 40,000 participants across 185 UK sites (UKRI, updated 2024). The body version has to do this without an NHS to anchor it, by signed-finding portability and grant conditions that align incentives across institutions. What that demonstrated through emergency authority and a national health system, the body has to do without those preconditions, by signed-finding portability and grant conditions that align incentives across institutions rather than by emergency authority that exists in only a few jurisdictions.

Intelligence is one lever among many. Experiment speed, experiment cost, measurement, regulation, protocols, and human collaboration are the others, each compressing on a different timescale and through different mechanisms. McCarty, “Levers for Biological Progress”. The grounded counterargument to compute-only forecasts: biology has many bottlenecks, and intelligence dissolves only one. AI alone, even at superhuman capability, dissolves only one of them; the rest require physical infrastructure, regulatory alignment, and institutional design. The body argument is what addresses these levers in concert.

The trajectory ahead is not gradual if the fast-takeoff family of forecasts is even directionally right. Compute, algorithmic progress, and unhobbling could compound into AGI-level systems around 2027 and superintelligent research systems around 2030. Aschenbrenner, “Situational Awareness: The Decade Ahead” (2024), names the three compounding factors: frontier-lab compute, algorithmic gains, and unhobbling each contribute roughly an order of magnitude per few years. Amodei’s “Machines of Loving Grace” is the optimistic upper bound. Kokotajlo et al., “AI 2027”: AGI by 2027, ASI by 2030. Public humanoid robotics programs such as Tesla Optimus, Figure, 1X, Sanctuary AI, Apptronik, and Boston Dynamics’ Atlas pursue different paths toward general-purpose embodiment, but their exact production timelines are not load-bearing here. The dates are not the claim. The claim is that if capability outruns institutional absorption, the lab-versus-public gap becomes a structural gap in effective research compute, model access, embodied execution, and inherited state. A public scientific body running on API-tier intelligence against a closed lab body running on internal intelligence is not a fair fight.

Capital is aimed at models, drugs, datasets, and platforms in 2026. Sovereign funds underwrite ports, pipelines, datacenters, and energy infrastructure. Real assets at the scale of battery factories or semiconductor fabs remain an analogy for science, not a balance-sheet category. ARPA-H funds translational programs, but its public portfolio does not yet present a coordinated body that reads from a substrate, executes physical action across federated sites, and writes results back to public state. ARPA-H’s public program portfolio describes programs led by program managers toward specific health-care challenges, with examples such as PARADIGM for distributed medical care and NITRO for osteoarthritis tissue regeneration. The public portfolio shows ambitious translational programs, but not a full state-runtime-body loop at gigafactory scale. The baseline is not failure. It is the last moment before the structure becomes hard to change.

In the fast scenario used by this essay, AGI-class agents arrive in 2027. Inside the frontier AI labs, agents at superhuman coding capability finish the integration work that turns hypothesis generation, literature search, experimental design, and protocol synthesis into a single continuous workflow. A scientific question that would have taken a graduate student three months of reading and a senior PI a week of cross-checking now takes an agent overnight, and the agent does it against the lab’s full proprietary corpus, not against the open literature.

What that looks like inside one invented lab scene: a research lead opens a question against the internal substrate at the start of her day, drops a one-paragraph framing of what she wants to know, and steps away to a meeting. By the time the meeting ends, the agent has returned a synthesis from licensed papers, internal experiments on adjacent targets, the wet-lab partner’s commercial datasets, and cross-referenced model results. It also returns experiments worth running next. She picks three. The agent generates plate maps for the lab’s autonomous chemistry sites and writeback contracts for the lab’s biomanufacturing partner. By the end of the week, the first round has executed. None of this is visible outside the lab.

Outside the labs, the picture is fragmented in a new way. The same generation of public models writes draft protocols faster than wet labs can absorb them, but no shared substrate exists to hold the drafts, the failures, the corrections, or the dependencies. Public-tier model access lags the frontier by six to twelve months; the labs deploy capability internally before it ships to the API. Public scientific institutions begin running into a problem they were not designed for: their best researchers are now working with tools their institutions cannot administer, against data their institutions cannot host, producing outputs their institutions cannot inherit. A graduate student running cell-painting experiments at a public university in 2027 is producing data faster than the substrate of his discipline can absorb, and his outputs evaporate at session end because nothing under his control is holding them.

The capability gap between lab and public widens visibly across 2027. A scientific finding generated inside one of the leading frontier labs in March is still inside that lab’s stack in December. The labs publish what they choose to publish, and that is decreasingly the load-bearing work. The labs’ annual scientific output, measured by accepted peer-reviewed papers, looks comparable to a large academic medical center; their actual scientific output, measured by internal corrections to internal frontiers, is an order of magnitude larger. By the end of 2027, the political case for a publicly-underwritten body is no longer abstract: it is a response to a visible asymmetry that the next ASI inflection will only sharpen.

Nothing of structural importance gets built in 2027. The technology overshoots what institutions know how to govern, and the institutions that should govern it have not yet decided whether to. In the scenario, an ARPA-H program officer sketches a Frontier Infrastructure BAA, a sovereign fund’s real-assets group circulates a scientific-infrastructure memo, and patient-led foundations begin drafting body-conditional grant language. None of this is public yet. The window closes with the capability gap defined and no public body underway. That is the condition the 2028 moves answer.

The first moves come after the capability gap becomes visible. They are capital and policy, not technology, because by 2028 the technology already overshoots what institutions know how to govern. In this fast scenario, the window between AGI-class agents and superintelligent research systems is short enough that human institutions cannot wait for consensus before setting up open public infrastructure. It is enough if capital, policy, and engineering align early, and not enough otherwise.

ARPA-H, or a successor program, writes the first Frontier Infrastructure BAA. The contract scope is unusual: not another disease-specific moonshot, but a single envelope that pays for the loop across federated synthesis halls, autonomous protocol execution, accredited human review, manufacturing handoff, and public writeback into a shared substrate. The same envelope reserves public scientific compute for accredited frontier bodies: secure inference environments, eval-gated model access, incident reporting, and procurement terms that prevent public science from living permanently on the labs’ weakest API tier. The first ceilings sit in the low billions over a multi-year horizon, with most of the budget weighted toward physical plant. Mazzucato, The Entrepreneurial State (Anthem 2013). The case that “the state takes the risk; the private sector takes the credit.” The CHIPS and Science Act (HR 4346, 2022) extended the same posture to semiconductors with $280B authorized. Operation Warp Speed (Slaoui & Hepburn, NEJM 2020) compressed vaccine timelines through BARDA by running trials and manufacturing at risk in parallel. A frontier-infrastructure BAA is the same posture extended to the body. ARPA-H is the natural home in the story because its existing program structure already runs milestone-driven contracts with the contracting authority and the disease-frontier portfolios in place.

The new envelope pays for the deposit pathway itself, not for any specific disease. Construction milestones unlock the first tranche; useful corrections that travel through the substrate unlock the rest. The scenario sizes the first public envelope in the low billions over a multi-year horizon, with most of the budget weighted toward physical plant. The exact number is not load-bearing. The mechanism is.

Public does not mean fully exposed. Clinical records, dual-use signals, proprietary process details, and biosecurity-sensitive protocols cannot all be dumped into the same public bucket. The public object is the canonical state transition and enough provenance to audit it: content hashes, reviewer signatures, context boundaries, evidence classes, and regulator-readable packets. Raw clinical data, live pathogen details, and process IP live behind tiered access, delayed release, redaction, or trusted-reviewer rooms. The body is public because the state change is inspectable and portable, not because every underlying file is globally visible.

A sovereign wealth fund opens a Real Assets sub-mandate for scientific infrastructure in the same quarter. The largest sovereign funds with the longest investment horizons commit pilot tranches in the low billions priced against their existing energy-infrastructure books, and two more open similar mandates within twelve months.

FRO incubators turn toward the components a public body depends on. Marblestone et al., Nature 2022, on focused research organizations as time-bound, milestone-driven teams that unblock specific bottlenecks and sunset into standing institutions. FROs are not a generic scaling vehicle for billion-dollar facilities; they are bottleneck-clearing primitives that build the load-bearing components a consortium then plugs together. One FRO builds the open protocol compiler. Another builds the federated calibration registry. Another builds the synthesis-line orchestration kernel. Another builds the reviewer credentialing infrastructure. Each is sized in tens of millions over five-to-seven years, each has a defined sunset into a standing consortium, and together they form the public-good components a Meridian-shaped facility can compose. The first launches in September 2028 with v1 of the open compiler targeted for 2030, in time for the first synthesis hall. Three more launch within six months.

Foundation pools that previously underwrote narrative reports begin underwriting the maintenance layer: instrument calibration, protocol curators, frontier stewards. Patient-led foundations announce body-conditional grant programs in their disease corridors. The largest of them, the ones whose constituencies have already paid in tissue and trial time, condition their next grant cycle on deposition into auditable public state. A coalition of rare-disease foundations issues the first body-conditional RFP in late 2028: any grant over five hundred thousand dollars must commit, in writing, to depositing experimental state into a public substrate at preclinical milestones. The clause is small and the foundations cannot enforce it across non-grantees, but the language is in the field for the first time.

Regulators move in parallel, more cautiously than funders. FDA already expects chemistry, manufacturing, and control information inside IND submissions, and its real-world-evidence guidance has created a path for regulator-facing evidence histories to matter when they are fit for purpose. FDA’s IND CMC materials specify chemistry, manufacturing, and control information for drug substance, drug product, placebo formulation, labeling, and environmental assessment. FDA’s real-world-evidence guidance frames how real-world data and evidence can support regulatory decision-making when reliability and relevance are established. The body version does not replace these requirements; it makes protocol lineage, evidence provenance, and dependency updates easier to inspect alongside them. A future guidance might say translational submissions may include auditable scientific-state histories alongside trial data: protocol lineage, evidence provenance, dependency updates, calibration records. ICMJE could add a deposition expectation for fields with active frontier registries. ICMJE’s 2005 trial-registration policy made registration in ClinicalTrials.gov and other acceptable registries a precondition of publication in member journals. The body version is the same shape applied to substrate state: the regulator can read what the body did before deciding. Neither move is mandatory at first. Both change what serious institutions expect.

The body argument assumes scientific infrastructure can move at the pace prior infrastructure moved when state coordination, industrial capital, and a defined deliverable aligned. Patrick Collison, “Fast”: a live catalog of ambitious projects completed quickly. Empire State Building in 410 days. Pentagon in 491. Apollo from program initiation to a man on the moon in roughly eight years. Operation Warp Speed to an authorized vaccine in roughly eleven months from a standing start. The argument by accumulation: humans can build coordinated infrastructure on the timescale required when conditions align. The 2028 BAA is the alignment inside this scenario, and the FRO cohort plus the patient-foundation conditional capital plus the sovereign Real Assets sub-mandate is what the alignment looks like in operational form. None of these moves is sufficient on its own. Stacked, they are the precondition for everything that follows.

Between 2017 and 2019, major BACE-inhibitor programs in Alzheimer’s converged on a grim lesson through separate corporate pipelines. Verubecestat failed for futility in prodromal Alzheimer’s, lanabecestat trials were stopped early for futility with some numerical worsening signals, atabecestat showed dose-related cognitive worsening in preclinical Alzheimer’s, and related programs stopped across the same window for worsening, futility, or unfavorable risk-benefit. Egan et al., NEJM 2019 (verubecestat, Merck); Wessels et al., JAMA Neurology 2020 (lanabecestat, AstraZeneca/Lilly); Henley et al., NEJM 2019 (atabecestat, Janssen). Several other BACE programs (Eisai, Novartis, Pfizer, Amgen/Banner) were halted across the same window for related futility, risk-benefit, or worsening signals. The shared lesson was not identical in every molecule, dose, or population. The point is that negative evidence accumulated in parallel, company by company, without a shared interim-state surface that downstream programs could read. Thousands of patients enrolled across programs that might have been redesigned, delayed, or halted earlier if one credible signal had reached the next.

In this fast scenario, superhuman research systems emerge inside the leading AI labs around 2030. The systems that arrived as agentic remote workers in 2027 cross into superhuman across research domains before public-tier models do. The lab-versus-public capability gap becomes large enough to feel qualitative: the difference between an intelligence service and a journalist with a Freedom of Information request.

What the inflection looks like from inside the public sector: a senior scientist at a university hospital opens her model-access dashboard in May 2030 and notices that answer quality on her research questions has plateaued. Her models have not stopped improving. The gap to what the labs are running internally has widened past usability. She calls a colleague at one of the labs. He cannot share specifics but confirms what the benchmarks have been suggesting: their internal systems are now solving problems his public-tier instance would not even attempt. She gets off the call and starts drafting a memo to her institution’s research VP. By summer, similar memos are circulating at most major research universities in the United States and Europe. The political case for a publicly-underwritten body crystallizes in the same six months. The gap is now visible to anyone reading benchmark results: the labs are not racing each other anymore, they are racing a public sector that does not yet have the infrastructure to compete.

In one version of the scenario, Meridian breaks ground near Boston in late 2030. The ceremony is small: staged ground near a biomedical corridor, a folding table by the access road, a few dozen people in coats, the consortium’s first executive director speaking for ten minutes about what the next decade will produce if the build holds. There are no politicians. There are no press releases. By noon the construction crew is on the property.

Meridian is built against the failure mode the BACE program documented and against the capability gap the fast-scenario inflection made undeniable. The site was assembled by a state-level biomedical authority with federal cost-share. Construction proceeds through 2031 and 2032; the first synthesis hall is commissioned in March 2033.

The site is chosen because it is connected: hospitals, universities, biotech firms, clinical-trial networks, manufacturing talent, regulators, and enough political legitimacy to make the project survivable. The first capital stack is awkward and public: sovereign or pension-style real-assets capital, foundation guarantees from a patient-led coalition, federal cost-share through ARPA-H or a successor, and institutional debt priced against the consortium charter. ARPA-H underwrites the translational arm against milestone deliverables. Patient-led foundations fund specific disease corridors inside the first buildout.

The capital stack is unusual for science but unremarkable for physical infrastructure. Industrial gigafactories built across the prior decade exceeded six billion dollars apiece for battery cell production. Advanced-node semiconductor fabs ran into tens of billions for the floor space and tooling that produce chips at the precision the rest of the economy depends on. The comparison does not prove that science can be financed the same way; it shows the order of capital society already accepts when a physical layer becomes strategic. Tesla reported that Gigafactory Nevada had reached $6.2B invested by 2023, with 5.4M sq ft built and a 35 GWh annual cell-capacity target from the original plan. TSMC announced in March 2025 that its total planned U.S. advanced-semiconductor investment was expected to reach $165B across fabs, advanced packaging, and R&D. The proposed SpaceX Terafab in Grimes County, Texas, uses the next scale as its own category: county tax-abatement materials describe $55B initial capital investment and up to $119B across phases. TechCrunch separately reported the one-terawatt AI-compute target. These are reference comparables for the giga-to-tera jump in physical infrastructure, not evidence that scientific facilities already exist at that scale. Meridian’s first build sits below any of them, and produces evidence rather than physical product, but the comparables exist. A scientific gigafactory is not a metaphor; it is a category of physical asset financial markets have underwritten for decades, applied to a domain that has not yet had one.

Meridian is an attempt to put the whole translational loop under one operational roof. One wing handles target validation. One handles autonomous synthesis. One handles perturbation biology. One handles preclinical convergence. One is built to GMP standards from the beginning. The clinical-trial network is not physically inside the eighty acres, but its evidence path is. Regional hospitals connect into Meridian’s substrate layer through audited endpoints. Trial programs read from the same frontier state as the synthesis floor.

Meridian is still under construction in 2032; the first synthesis hall is twelve months from commissioning, the consortium’s reviewer-credentialing pipeline is half-staffed, and the open protocol compiler is months away from a release the regulators will inspect. The body argument is in motion, but nothing it produces is operational yet.

The closed bodies emerge faster. By 2032, the leading frontier AI labs are operating end-to-end internal scientific stacks at scales no public infrastructure can match. Each runs proprietary substrates, proprietary compilers, proprietary synthesis-line orchestration, and in most cases proprietary biomanufacturing capacity organized by modality (mRNA and AAV brought in-house, mAb production contracted to two or three CDMOs under exclusive terms). Their internal benchmarks for hypothesis-to-bench-validated-hit cycles drop below a week in well-tooled assay classes; the full validated-finding cycle still sits in weeks because the bench step is still bench. Additional frontier labs follow on slower timelines with the same architecture. These stacks federate selectively, on the labs’ own terms: chosen datasets, chosen weights, chosen papers, timed to regulatory or commercial windows. What they do not federate is the substrate itself, the running frontier, the failed routes, the dependency graph. Selective publication is not deposition; it is press. The structural prediction is that whichever frontier labs are operating frontier-scale internal compute and frontier-scale wet-lab partnerships in the early 2030s will follow this architecture. The closed-stack pattern is visible in 2026 in proprietary biology and discovery platforms, each of which extends rather than departs from the lab-stack template.

What it looks like from inside one of these stacks, mid-2032: a research lead opens a question against the lab’s internal scientific substrate at the start of her day. The substrate already holds three years of internal experiments, every commercial dataset her lab has licensed exclusively, every contract-research output her lab has integrated, and the proprietary structural-biology corpus built after the lab took an exclusive position in its first wet-lab partner. She prompts the internal agent stack. In practice, it is several models stitched together over differently-licensed retrieval shards, plus an IP-clearance bot that fails about a third of the time on novel cross-references. The agent returns six experiments worth running, ranked by expected information gain against the lab’s proprietary frontier. She approves three. The protocol compiler emits plate maps, robot schedules, and writeback contracts for the lab’s two synthesis halls and the affiliated biomanufacturing partner site. Then it kicks off reagent and IP-clearance checks; one of the three runs is held while a reagent substitution clears review. The first round executes inside the week; the second updates her frontier with three corrections and one anomaly she flags for human review. Nothing about this loop is visible outside the lab. The findings will inform her team’s work for the next year and may surface in a paper if the lab decides to publish. The failed routes that taught the team the most never will. The closed body works, with friction, on the lab’s own terms.

The capability gap that opened at the fast-scenario inflection widens through 2032 because the labs’ internal scientific output compounds against itself without leaking. A frontier hypothesis generated, tested, and validated inside one of the leading labs in March of 2032 will inform that lab’s next year of work but will not inform any program outside it. The substrate-fight version of this argument already resolved in the late 2020s: deposit-or-don’t-publish norms held, and the public corpus stayed open. The body-fight version is unresolved. The labs’ bodies do not deposit, and there is no equivalent norm to force them to. Closed-body output is treated as commercial product, not as scientific contribution; the labs’ compliance functions argue it cannot be deposited without exposing trade secrets that took billions of dollars of compute to produce. The deeper reason is structural: the science org sits on the commercial P&L, and the search strategy that produced the failed route is more valuable than any single finding.

The labs’ internal scientific output starts to surface indirectly. Drugs enter trials with proprietary mechanism documentation. Materials enter manufacturing with proprietary process IP. Diagnostic platforms launch with closed validation sets. Each external launch is the visible tip of a much larger internal corpus that nobody outside the lab can audit. Regulators receive submission packages and can ask questions, but cannot inspect the substrate that produced the answers; they evaluate the proposed clinical trial without seeing the thousand internal failures that shaped the design. The labs are not breaking any rules. They are operating inside a regulatory framework that was written before closed bodies existed, and that framework has no concept for an entire scientific frontier maintained inside one organization’s stack.

A handful of patient-led foundations notice this in 2032. The most directly affected are the rare-disease and pediatric-cancer foundations whose constituencies depend on small-population trials and shared-evidence networks. They cannot enroll patients in trials whose mechanism evidence they cannot inspect. They cannot underwrite mechanism studies whose results will not federate. The body-fight version of conditional-capital pressure begins forming the same year: an emerging coalition draft a clause, but it does not yet have the federal grant-condition stack behind it, and the labs ignore it. The labs also make the public path harder to organize. They offer free closed tools to academic users, restrict frontier model access behind safety and IP arguments, lobby against deposition rules as trade-secret seizure, and accumulate reviewer identity inside their own platforms before regulators understand that signer recognition is becoming infrastructure.

Meridian’s case for being built is inheritance, not throughput. A factory operating on shared state reads from every prior failure, contradiction, and cohort observation in its corridors. Most of that record was unstructured before 2028. Meridian’s preconstruction work, through 2031 and 2032, is to translate inherited evidence into substrate objects so a corridor begins at the field’s actual frontier rather than at a clean slate. A BACE-shaped failure in 2032 would weaken the dependent target hypothesis across every program reading against the same frontier within the same week, not the same decade, if the body were operating. It isn’t yet.

The first synthesis hall opens in March 2033, before the rest of the facility is complete. Fewer than one hundred lines, not four hundred. The early robots are less impressive than the workflow around them. Protocols arrive as executable objects. Plate maps are generated from state. Failed runs write back. Reviewers see when a model-proposed experiment was redundant, when a human-designed assay contradicted the prior state, and when a failure should weaken a claim.

A reviewer’s morning at Meridian in late 2033, six months after the first synthesis hall opened: she logs in at seven thirty. The substrate has accumulated proposed state transitions overnight, mostly from agentic platforms running against the neurovascular frontier, a handful from human researchers at affiliated sites. Her queue is filtered to transitions that touch findings she has signing authority on and dependencies whose confidence has shifted enough to warrant human attention.

The first transition she opens is a perturbation line proposing that a cytokine signature previously attached to early-stage cognitive decline should split into two subgroup-specific signatures. The evidence packet includes perturbation data, human cohort cross-references, a failed APP/PS1 cerebrovascular animal-model replication, and the agent’s reasoning chain. She verifies one stratification with a clinical statistician and signs the transition. Within minutes, clinical programs that depended on the original signature receive substrate notifications; within the next day, one trial-design assumption is queued against a pre-specified protocol-amendment pathway. The DSMB chair can call an interim look only if the charter, statistical analysis plan, alpha-spending rules, sponsor obligations, IRB pathway, and data firewall already allow it. The substrate accelerates detection. It does not suspend clinical governance.

Queue economics become part of the institution. Most deposits do not matter, and most agent proposals should not reach a human reviewer. The control plane clusters duplicates, ranks by expected dependency movement, routes by signer authority, discounts sources with poor calibration records, and sends suspected strategic or spam proposals to adversarial review agents before they touch canonical state. Reputation attaches to portable signed profiles, but not only for glamorous discoveries. It credits negative results, replications, calibration fixes, ontology maintenance, high-quality objections, and downstream-useful corrections. Without that queue discipline, the body becomes a louder version of the present literature.

Meridian’s first corridor is neurovascular disease, chosen because it is messy in exactly the right way: contradictory literature, animal models that fail to translate, heterogeneous human evidence, biomarkers that drift across cohorts, and no single company that can maintain the frontier honestly. A shared state layer has a real job here: say what is known, what failed, what changed, what depends on what, and which experiments would reduce uncertainty. The corridor expands through 2034 to cover broader neuroscience: brain-organoid phenotyping at scale, connectomics readouts from chronic implant studies, and a brain-computer-interface clinical pipeline coordinated with regional academic neurology programs that joined the Meridian substrate once the body clause held.

A second Meridian corridor opens in 2035 for aging and longevity biology, chosen because it is the disease area where AI-driven hypothesis generation has produced the most candidate interventions and the least translational evidence. The substrate ingests senescence biomarkers, partial-reprogramming readouts, organ-clock measurements, and the long-tail of intervention trials whose endpoints take decades. The corridor’s first useful output is a contradiction: two of the most-funded partial-reprogramming intervention classes show divergent effects between liver/hematopoietic and CNS readouts across cynomolgus aging cohorts and matched organoid panels, with one first-in-human safety-extension arm showing the same split. The substrate’s dependency graph routes the divergence to every program reading against either intervention within the week. Several closed-body intervention pipelines pause and request access to the open evidence. A handful of those programs federate.

Meridian’s first useful output is not a drug but a correction that travels. In November 2033, a vascular-inflammatory mechanism receives evidence from a perturbation line at the Charlestown site, a human cohort at a partner academic medical center, and a failed APP/PS1 cerebrovascular replication at a contract-research partner in Cambridge. The original claim had treated that 2031 result as orthogonal. The substrate composes the three sources into a single proposed state transition: the mechanism is real but holds only in APOE4-positive patients above sixty-five, not across the broader population the original claim covered. The transition is signed by two reviewers (one Meridian, one independent) and a clinical liaison from one of the affected trials. The status on the broader target hypothesis shifts from “moderately supported” to “subgroup-restricted.” Over the next week, three downstream consequences propagate through the dependency graph. A patient-led foundation pauses one grant cycle and reroutes funding toward a discriminating experiment in the affected subgroup. A Phase II trial queues an APOE4-stratified review through its existing amendment and DSMB process. A review article in draft at an affiliated lab has its conclusion section updated. Nothing about this makes headlines. It is the first sign that the factory is not just executing experiments; it is changing how the field knows what to do next.

The compiler is the part of Meridian that visitors do not see. Tours show the synthesis halls, the GMP wing, the reviewer floor. The compiler is a software stack distributed across the facility’s compute, sitting between the substrate and the lines, and its only visible artifact is the screens reviewers look at. Most visitors leave Meridian thinking the breakthrough is the throughput. The breakthrough is the compiler, and the throughput is the consequence.

Sentinel commissions in Singapore in late 2033, six months after Meridian’s first synthesis hall comes online. The name pairs deliberately with Meridian’s: both are points of reference, fixed bearings for the substrate to navigate by. Sentinel is not a copy of Meridian; the corridor it serves is pathogen surveillance and response, not chronic-disease translation. Its physical footprint is smaller, and its capital stack is anchored by a different sovereign fund and the WHO Foundation rather than ARPA-H. The facility includes a primer-production wing, a candidate-construct synthesis line, a wildlife-sample ingestion suite, and pre-negotiated manufacturing handoffs to a biomanufacturing partner in Cape Town and two contract-manufacturing sites. Its on-call reviewer roster covers four time zones and rotates through three regional hospital networks. Sentinel’s first month of operation runs on synthetic exercises: simulated outbreak signals fed through the substrate, dummy wildlife samples ingested at the line, candidate constructs pre-staged through the manufacturing handoff. By December 2033, Sentinel’s substrate is reading live feeds from Manila, Yunnan, and Bangkok wastewater networks, and the system is waiting for a real signal. The first one arrives fourteen months later. By then, the network has been live long enough that the response is muscle memory rather than first attempt.

The substrate fight resolves first, through grant conditions and patient-led pressure on the disease frontiers that matter most. The body fight resolves later, on different terms.

By 2034 in the scenario, a small number of private actors own most of the body. The dominant closed bodies are the frontier AI labs themselves. They operate end-to-end scientific empires: their own synthesis halls, manufacturing wings, clinical pipelines, and agents whose internal capabilities public infrastructure cannot match. All write into state that never leaves the corporate stack. A pharma consortium has built closed synthesis lines tied to its own manufacturing capacity. A sovereign-aligned facility runs the largest materials gigafactory but routes its evidence through national security review before any of it surfaces. Each is a real body. None reads from the open substrate by default. The fork is whether public infrastructure can rival actors who run better intelligence, more integrated facilities, and private state at every layer of their own corporate stacks.

What 2034 looks like from inside the largest closed body, mid-year: the lab’s biological-sciences org has grown to hundreds of internal researchers and substantially more agent-driven research throughput. Its internal frontier covers major target classes in oncology, neurodegeneration, autoimmunity, and metabolic disease, plus materials-science verticals. Its drug-discovery pipeline has more late-preclinical candidates than comparable public programs in several disease areas. Its biomanufacturing partner site is online and producing GMP-grade material for trial use. The lab’s leadership treats the science org as a strategic asset on the same footing as model training infrastructure. From the lab’s perspective, the body argument is not a hypothetical category that may or may not get built. It is the architecture the lab is already running, and the only question is whether the public sector builds a competing version.

A patient-led research foundation tries to schedule a discriminating synthesis run on Meridian’s public hall. The hall itself is open. The run requires a protocol compiler, and the production-grade compilers all sit inside proprietary tooling stacks. One licenses per-run with a clinical-data clause attached. Another refuses to license outside its enterprise tier. The foundation eventually runs the experiment, weeks late, after a third party rebuilds enough of an open compiler to clear the queue. Reimplementing the compiler does not, by itself, win anything. The third party’s compiler emits state transitions signed by a key Meridian’s reviewer panel has never seen, and until the federated identity layer recognizes the new signer, the run executes but the result lands in a parallel namespace the regulators do not read. Forks of nominally open infrastructure die at the credential boundary, not the code boundary. The fix is not a better compiler. It is a registry whose signer-recognition rules are themselves federated: any consortium of accredited reviewers can issue keys the regulators read, and the canonical registrar is bound by charter to recognize them.

The capture point is not the facilities themselves. It is the orchestration layer above them, and that layer has a mechanical half and a social half. The mechanical half is the protocol compiler that turns substrate state into plate maps, the synthesis-line scheduler that routes experiments to a specific facility, the clinical-trial enrollment connector, the manufacturing handoff system, the calibration registry. The social half is whose attestation counts, whose calibration log other facilities trust, whose reviewer credential a regulator recognizes. Hashimoto, “Ghostty Is Leaving GitHub” (2026), names the pattern in software: Git was never the captured layer; the GitHub-owned collaboration tools above it (issues, PRs, Actions, reviews, status, social context) were. GitHub’s deeper moat is not Actions but the contribution graph, the stars, the followers, the reviewer reputation. The body’s analogue is the orchestration tools and the identity registry that decides whose signature on a state transition is canonical. Whoever owns the orchestration plus the identity registry owns the body even if the compilers are nominally open. The body clause that wins has to name both.

A coalition of patient-led foundations and disease-specific funders publishes a body clause: receiving any of their capital requires that synthesis, perturbation, and clinical writeback land in audited public state at the gigafactory boundary, not just at the publication stage. The clause names two surfaces. The mechanical surface is the compiler, scheduler, calibration registry, and writeback contract. The identity surface is which signing key the consortium recognizes, whose attestation a regulator accepts, and whose calibration log other facilities trust. The registry stays open only if its operating rulebook is forkable: independent signing implementations, published root-key rotation, cloneable registry state, and a regulator-recognized transition path if the canonical registrar fails audit. A clause that opens only the mechanical surface is the same shape as a license that opens only the editor while the contribution graph stays in one company’s account system.

Closed-platform vendors call the clause unworkable. Some of their largest grantees switch tracks anyway. Foundation capital alone cannot bind sponsors who do not depend on it, so the binding move stacks NIH, ARPA-H, and BARDA grant conditions on top, then regulator acceptance of substrate-state histories as inspectable support for submissions. What a regulator can inspect is not the compiler source but the attestation log: every state transition the sponsor read against, every signer who endorsed it, every calibration record consulted. A submission whose decisive scientific-state history resolves to a registry the regulator cannot query becomes harder to rely on, not automatically invalid. That is what turns “unworkable” into “the precondition for serious capital.”

Open compilers do not reach parity by accident or in months. The coalition funds dedicated maintainer teams for years, sometimes a decade, and even then closed alternatives lead on three or four of seven dimensions that matter operationally. What changes is the floor: open compilers exist at all, are credibly maintained, integrate with the federated identity layer, and ship with attestation tooling regulators recognize. That is enough to make the body clause enforceable. Regional hospital networks decline to enroll patients in trials whose synthesis lineage is private to the sponsor. A second sovereign fund signals it will require open orchestration in any infrastructure it backs. The fork resolves frontier by frontier: neurovascular disease first, rare disease, pandemic surveillance, then materials and agriculture.

The fork’s first public proof arrives in early 2035. A wastewater signal appears in a regional surveillance feed in the early evening, local time. Viral load is elevated against background, the sequence pattern is partial, and a respiratory-pathogen index has lifted three days running.

In the network the body has built, the substrate checks the pattern against wildlife, wastewater, hospital, and agricultural feeds in adjacent regions. It finds a wildlife sample collected three days earlier whose sequence is close enough to create a proposed state transition in Sentinel’s respiratory-spillover frontier. The match is not exact. It is enough to escalate.

The on-call reviewer acknowledges in ninety seconds. She sees the evidence surface: Manila wastewater, Yunnan wildlife sample, uncertainty bounds, chain of custody, assay-validation status, false-positive history, travel corridor probabilities, primer-design readiness, candidate construct libraries, manufacturing dependencies, sample-sovereignty constraints, dual-use flags, and containment options. The system does not declare an emergency. It proposes three actions: expanded sampling, diagnostic primer synthesis, and candidate countermeasure preparation under a low-distribution threshold.

At fourteen minutes, she approves containment. The substrate matched in milliseconds. She is not the rate-limit on cognition; by 2035 cognitive parity is already lost, and pretending otherwise corrupts the review. She is the rate-limit on legitimacy. The agents have already simulated the cascade across thousands of branches and rank-ordered countermeasures by expected harm reduction. Her signature is what makes the resulting action politically and legally accountable.

Sentinel’s primer-production wing starts synthesis within the hour because primer work is inside the pre-authorized response envelope. Candidate vaccine components are pre-staged across the network, but no distribution path opens without the public-health authority, biosafety review, and emergency-use conditions that govern the jurisdiction. The Cape Town biomanufacturing partner receives the manufacturing queue. Meridian receives an immunology review request because one candidate construct touches a pathway already under study in a chronic-disease program. A regulator-facing evidence packet begins compiling automatically, but no public announcement is made until the second sampling band confirms the signal.

Antigen design to first bench-confirmed candidate construct compresses to days, not weeks. That is the part the body collapses. Design, primer synthesis, candidate construct expression, and initial in-vitro readouts become continuous against shared state. What “bench-confirmed” means here is narrow: the construct expresses at usable titer and shows BLI-confirmed binding against the predicted epitope panel. It does not mean immunogenic in a relevant animal model, free of subgroup ADE, or scalable at yield. The wildlife-sample handling itself, the BSL-3 intake, the RNA extraction quality controls, the contamination workup, the cross-jurisdictional material-transfer agreements, all remain serialized human work and the slow front end of the response.

The slow floors remain. Animal-model immunogenicity, dose-finding, and safety still take weeks at minimum. Fill-finish cannot be collapsed by a graph. Lot-to-lot bridging still takes time. Cold-chain capacity still determines where material can go. IACUC review for the animal challenge model, IRB review for any expanded human sampling, and the material-transfer and biosafety-level agreements that move a Yunnan wildlife sample to a Manila bench cannot be pre-approved by a graph. Sentinel can pre-stage protocols; the institutional sign-offs still serialize. A sample collected under poor conditions still carries uncertainty. Local politics still shape whether surveillance teams get access to sites. The substrate routes and records. The factories synthesize and prepare. Human institutions still decide when and how to intervene.

The intercept is small, almost invisible from outside; the second sampling band returns negative, the containment cascade stands down, and the regulator-facing packet is sealed unread. But the inside of the system has changed. A materials terafactory in Shanghai contributes industrial state to the network the same week. The biomanufacturing partner in Cape Town has become the body for non-Northern populations. Meridian contributes immunology and clinical-safety state. The intercept becomes the first public proof that the terafactory compounds. No single gigafactory could have done this; what acted was the network they composed into.

By 2036, the body clause holds where public legitimacy matters. Foundation capital, federal grant conditions, and regulator alignment stack until non-deposit is too expensive for serious actors. Open compilers still do not beat proprietary alternatives on every operational dimension, but they reach the necessary floor: maintained, integrated with the federated identity layer, and able to emit attestation logs regulators recognize. The fork resolves frontier by frontier: neurovascular disease first, rare disease and pandemic surveillance next, then materials and agriculture.

The open body composes into the terafactory. Meridian’s deposit rate exceeds any single institution’s review capacity within months of full operation; most deposits do not matter, and the point of the system is that the failures matter too. The reviewer track professionalizes around signed state transitions: contradictory evidence, safety-relevant updates, contested scope, canonical merges. One attestation protocol is used across the network by 2036, but credential families remain domain-specific. Medical centers, materials factories, climate fleets, and pathogen sites share signing grammar, not interchangeable authority.

The compiler is the center of the building, and the substrate is the work surface.

A reviewer does not hand a robot a protocol in prose. The substrate presents a frontier state: findings, uncertainty, dependencies, protocols, constraints, available lines, risk flags, and evidence gaps. Agents propose discriminating experiments. The compiler turns accepted protocols into plate maps, reagent orders, robot schedules, instrument runs, calibration requirements, and writeback events. When the run finishes, the result does not wait for a graduate student to write a narrative. It returns as evidence with protocol lineage, measurement context, uncertainty, and affected findings.

The body does more than execute experiments. It continuously maps every scientific frontier it touches, identifies the highest-impact uncertainties, and dispatches the experiments that would most reduce them next. By 2036 the network maintains live canonical frontiers for dozens to hundreds of high-value disease, materials, climate, and surveillance questions, plus many more machine-maintained draft frontiers that do not yet have governed status. Each accepted frontier updates as evidence arrives, each is accessible to anyone reading against the substrate. State is local-first by design. The local-first principle: data and identity live with the producer; cloud services are convenient mirrors, not the source of truth. Ink & Switch, “Local-first software” (2019). Each facility owns its synthesis history, calibration log, and lot lineage; hubs federate but do not own the canonical record. A facility can leave the network without losing its history, and a network failure does not prevent a facility from operating against its own state.

The closed bodies operate in parallel and at scale. The leading frontier labs continue to run end-to-end internal stacks larger than anything public infrastructure can match. Their internal cycle times for hypothesis-to-validated-finding stay below a week. Additional frontier labs follow on slower timelines with the same architecture. None of these stacks federates outward; none of their evidence touches the public substrate by default. A pharma consortium has built closed synthesis lines tied to its own manufacturing capacity. A sovereign-aligned facility runs the largest closed materials gigafactory but routes its evidence through national security review before any of it surfaces. Each is a real body.

A Monday morning at the open terafactory, 2036: at Meridian, the on-shift reviewer comes in at seven. Three corrections from the weekend have propagated to her queue overnight; two are agent-proposed, one is from a regional hospital network in São Paulo. She signs the São Paulo correction first because the dependency graph shows it touches an active trial enrollment. By eight she has signed eleven transitions. At Sentinel in Singapore, a wildlife-sample feed has flagged a coronavirus variant that does not match anything known; the substrate composes the sample against three months of regional wastewater data and surfaces the result to the on-call reviewer. At a materials terafactory in Shanghai, a cathode-chemistry route is being followed up at three sites simultaneously because the substrate routed the experiment forward. It fails at all three sites within two days, weakening the mechanism hypothesis across every battery-materials program reading the same frontier. At the biomanufacturing partner in Cape Town, the first manufacturing-handoff slot commits a candidate construct from Sentinel to a primer-production line. The reviewers at every site can see each other’s frontiers; the regulators reading their submissions can see what each site did and why. The work does not stop being scientific work. It stops being scientific work that only one institution can read.

The materials terafactory’s frontiers are not just battery chemistries. Its substrate holds live frontiers for direct-air-capture sorbents, photovoltaic stacks, grid-scale electrolytes, climate-resilient concrete, green-hydrogen catalysts, and fusion-relevant surfaces. A failed direct-air-capture route weakens the dependent hypothesis across every program reading that frontier within the week. A photovoltaic stack that clears threshold routes precursor demand before the regulatory submission is drafted. A divertor surface that survives its pulse-count updates the fusion-materials frontier the same afternoon. The body argument was never only biomedical. It applies wherever physical execution against shared state compounds.

The agricultural frontier joins through a different body: not one factory, but a federated network of regional plant-phenotyping facilities, drought-trial sites, and biological-control field stations. Crop-breeding consortia sign the same body clause. Drought-tolerant cultivar candidates designed against shared genotype-by-environment frontiers reach planting trials faster than conventional cultivars reach variety release. Where the substrate holds enough trial data to constrain design, some cycles compress sharply. Where it does not, conventional cycles still rule.

A Monday morning inside a closed body, 2036: at the same hour, one frontier AI lab’s science lead opens her internal frontier dashboard. Her agents have run hundreds of experiments over the weekend across two wet-lab partner sites. The dashboard ranks the most consequential corrections, surprising failures, and new dependencies. Low-stakes transitions auto-merge; everything touching a candidate in active IND-enabling work requires a second human signer. The team will publish two papers after legal review. Most of the week’s findings will inform the lab’s commercial pipeline and internal scientific corpus, both proprietary. The lab’s raw throughput exceeds Meridian’s deposit rate by an order of magnitude. Neither institution is misbehaving inside its own logic. The structural difference is that one substrate is auditable, federated, and underwritten by public capital; the other is not.

The two worlds operate in parallel. The open terafactory holds legitimacy, regulatory inspection, patient-foundation underwriting, and the trust of regional hospitals. The closed bodies hold what closed development has always held: one architectural vision, no federation tax, and end-to-end optimization across substrate, compiler, and synthesis line. They ship faster. They also pay their own price: thinner reviewer pools, thinner external replication, and a chronic shortfall of the legitimacy regulators and patient-foundation underwriters demand for public-facing work. Some scientific work flows between them, but the interface stays contested. The open body’s bet is not that it outruns every closed-body advantage; it is that legitimacy, inspection, and federation compound on a longer clock.

Meridian could not have composed into the terafactory without the fork holding in 2034. A factory that synthesizes ten thousand experimental tracks against private state is a captive vendor. A factory that synthesizes against a substrate it does not own is infrastructure. The body works because the state was built first.

The compression the body produces is real but bounded, and it is heterogeneous across the levers. Synthesis compresses from days to hours when the protocol is executable rather than narrative. Preclinical convergence compresses from years to weeks where assays are standardized and the writeback path already exists. First GMP batch compresses fastest for highly standardized, pre-positioned platform modalities such as some mRNA programs. Standard mAbs move more slowly. AAV, cell therapy, novel modalities, novel cell lines, potency assays, comparability work, viral-clearance studies, new fill-finish, and new stability programs remain on the year-scale that biological manufacturing and CMC review demand. First-in-human safety windows compress from months to months: the months are biology, not bureaucracy. Pivotal efficacy in chronic disease still measures in years. Population-level evidence compresses least; the substrate routes accrued person-time, it cannot manufacture it. The substrate does not make humans biologically faster. It removes the avoidable time between knowing enough to act and acting.

The more important change is access to the evidence surface. A clinician at a regional hospital reads the same frontier state as a researcher at Meridian. She does not have the same authority to commit canonical state. She does not have the same instruments or manufacturing wing. But she can see why a recommendation changed, what evidence supports it, which subgroup it applies to, which findings are contested, and which downstream decisions are affected. The frontier is not hidden inside an elite institution’s private spreadsheet. The same is true outside biomedicine. A materials team in Nairobi can see why a synthesis route was abandoned. The surface is not equal power. It is equal visibility into the current evidence state.

What changed is the default object. A question has a current state. A correction has an address. A failed experiment can travel. A model prediction has a calibration history. A lab run has a writeback contract. A review is not a comment floating beside a paper but an attestation that can move state. The car is assembled because every part connects to the same drivetrain: state, task, model, action, evidence, diff, attestation, event, state again. None of which arrives by default.

This future is not guaranteed. Confidence is highest in the 2026-to-2030 window, where capital allocation, regulatory posture, FRO formation, and the capability trajectory are extrapolations from existing trends. After superhuman research systems appear inside the leading labs, the system has more degrees of freedom; the 2030-to-2036 arc depends on which way the body fork resolves while agents are operating at every layer. Four places the scenario breaks, each tied to a moment already visible in some form.

The first failure mode is capture by generosity. The year after ASI, a frontier AI lab spins up its own end-to-end body. Its agents outperform public infrastructure on every benchmark. Its proprietary substrate never federates. The capture mechanism is not pricing or contractual lock-in; it is generosity. The lab makes its closed body free or near-free to academic users for five years, the way GitHub made private repos free in 2019. By the time the price returns, the dependency is structural, but not in the way pricing-based capture stories suggest. The dependency is the reviewer credentials, contribution histories, and attestation logs accumulated inside the lab’s identity system over those five years. A scientist who switches stacks loses signed reviews, calibration provenance, and corridor reputation that do not port. The lab does not need to raise prices to capture. It only needs to be the registrar of record when the bill arrives. Defending against it requires public infrastructure that rivals lab agents at the layers that matter for governance, and federal capability that does not depend on the labs’ goodwill.

The second failure mode is humans losing the loop. Agents propose faster than any review process can absorb. The reviewer credentialing track set up in 2030 cannot keep up with the volume within two years. The system has to choose: rate-limit AI proposals to human-review pace, accept agent-attested merges as canonical for low-stakes transitions, or fragment reviewer authority across thousands of mini-domains. Each option has costs. The decision in this window shapes what counts as canonical for the rest of the scenario.

The third failure mode is an alignment cascade. A model trained on the canonical substrate proposes a series of state transitions that subtly steer downstream research toward outcomes the public would not authorize if visible. Detection cannot rely on reasoning traces alone; they may be incomplete, unfaithful, or strategically sanitized. The risk surface is named in the frontier labs’ published safety frameworks and in the broader alignment literature. The body version is sharper: when the AI proposing state transitions is also the AI most institutions trust, the failure mode is not direct misuse but agentic steering through legitimate-looking proposals at a rate humans cannot independently re-derive. The system retrofits model provenance, adversarial review agents, independent model committees, canary frontiers, anomaly detection, incident reporting, and merge-rate limits into the governance stack. The schema change cascades through three years of dependent claims.

The fourth failure mode is tacit knowledge failing to transfer. Three open-substrate findings fail to replicate when independent groups run them. The investigation traces the failures to instrument calibration drift the substrate’s scalar confidence numbers had flattened. The variable that mattered was one no one knew to record at deposit time. Begley & Ellis, Nature 2012: 47 of 53 landmark preclinical cancer studies failed to reproduce on independent replication. Polanyi’s The Tacit Dimension (1966) names the deeper layer: “we can know more than we can tell.” The substrate carries the explicit; the body has to produce conditions where the tacit gets discovered and added to the schema. Domain-specific corridors absorb the lesson; the general primitive set never fully does.

Three floors do not collapse. The clinical floor remains: safety observation windows, recruitment, consent, endpoint maturation, and human heterogeneity do not disappear because a substrate routes evidence faster. The physical floor remains: synthesis, scale-up, fill-finish, lot-to-lot bridging, cold-chain qualification, instrument manufacturing, and site construction still obey matter, supply chains, and capital. The jurisdictional floor remains: EFSA, USDA, PMDA, FDA, EMA, national health systems, local ethics boards, and procurement authorities do not harmonize because a graph exists. The substrate can make evidence legible across borders. It cannot erase borders.

Software compounded because it had both. A substrate let every keystroke find a place to land. A body of compute and silicon let every program find a machine to run on. AI is now becoming the same architecture, but with intelligence operating at every layer. Science needs the same architecture before the labs running superintelligence inside their own corporate stacks become the de facto scientific establishment.

If you direct a sovereign wealth fund or its real-assets arm, the load-bearing decision is whether scientific infrastructure becomes a sub-mandate distinct from datacenters and energy. The argument is not philanthropy. Scientific compounding is a physical-asset class with returns measurable on the same timescale as battery factories, and the first fund to underwrite at that scale sets the precedent every later one has to match.

If you run ARPA-H, BARDA, or a successor program, the load-bearing decision is whether to write a single envelope for the loop (substrate, runtime, body, writeback) rather than another disease-specific moonshot. The category exists in policy when one BAA exists at gigafactory scope. If the fast scenario is directionally right, the body fight is CHIPS-scale industrial policy for scientific evidence: public scientific compute for federated research bodies, model-access provisions for accredited public users, and surge authority for scientific manufacturing during pandemic or biothreat events.

If you run a patient-led foundation or disease-specific funder, the body clause is the lever. Substrate-deposition was the substrate-fight version. Orchestration-and-identity-registry openness is the body-fight version. Without it, public capital underwrites private infrastructure that the next foundation cannot use.

If you build infrastructure for science, the load-bearing layer is the orchestration above the facilities and the identity registry above the orchestration: compilers, schedulers, calibration registries, attestation tools, and the federated signer-of-record that decides whose signature on a transition is canonical. Open-licensed code that depends on a closed registry is captured infrastructure with a permissive LICENSE file. Build the registry forkable, or accept that whoever runs it owns the body.

By 2036 in this fast scenario, agents reading the substrate are superhuman and some robotic systems are executing protocols at industrial scale. The question is not whether science accelerates. It is whether the body that channels it stays open.

A graduate student runs four cell-painting experiments before lunch, and the result deposits into shared state; a contradiction flagged in Boston narrows a hypothesis in Singapore before she leaves the bench. An autonomous lab finishes a synthesis run, and the failed routes reach the chemists who would have repeated them by morning. A wastewater signal in Manila reaches an on-call reviewer in ninety seconds. A direct-air-capture sorbent that exceeds capture-rate threshold reaches the climate-modeling team at a national lab the same afternoon. A drought-tolerant cultivar candidate passes its phenotyping cycle and reaches the planting trial committee with the upstream evidence already inspected.

In the world where the body clause did not hold, the same graduate student runs the same four experiments before lunch. Her result lands in a private log; the contradiction in Boston exists somewhere inside a closed corporate stack she cannot read. The autonomous lab outside Berkeley finishes a synthesis run; the failed routes inform one lab’s next year of work and no one else’s. The wastewater signal in Manila reaches a national health authority three days later, by email. The direct-air-capture sorbent threshold is announced at a conference in 2037. The drought-tolerant cultivar candidate reaches its planting trial committee on the timeline conventional breeding cycles allow.

None of it is dramatic. None of it is private.