Recursive Simulations Daily — March 13, 2026

Table of Contents

🏭 GTC Preview: Reply, NVIDIA, Google Converge on Cloud-Based Digital Twins for Industrial Robotics 🔄 Learning Theory of Model Collapse: A Formal Framework for Replay-Induced Degradation ⚡ The "Atoms Over Bits" Pivot: Capital Flees AI Software for Physical Infrastructure 🌡️ Australia Guts Climate Simulation Capacity as AI Demands Energy 🔌 Data Center Interconnects: The Next AI Infrastructure Bottleneck 🧬 When Simulation Infrastructure Becomes the Primary Infrastructure 🔮 Implications

---

GTC Preview: Reply, NVIDIA, Google Converge on Cloud-Based Digital Twins for Industrial Robotics

Reply announced today that it will present two industrial digital twin solutions at NVIDIA GTC (March 16-19, San Jose) that demonstrate the convergence of cloud infrastructure, simulation platforms, and physical AI. The first application, "The AI Fast Lane for the Industrial Edge powered by NVIDIA on AWS," integrates NVIDIA Omniverse with Isaac Sim to create a continuous development loop where edge AI models for robotic arms and mobile robots are validated and retrained using synthetic data from digital twin simulations. The architecture addresses what Reply identifies as "a key challenge for industry: enabling AI systems to learn in the field without interrupting operations" by processing sensor data in real time, identifying performance gaps, and automatically initiating retraining cycles. The human-in-the-loop approach preserves model quality while the digital twin generates the counterfactual training scenarios that cannot be safely or economically produced in physical production environments.

The second demonstration, developed jointly with Google, runs NVIDIA Isaac Sim directly on Google Cloud G4 instances to create "highly precise digital twins of complex logistics environments without local workstations." This cloud-based simulation architecture enables the validation of diverse robot fleets before physical deployment — the same sim-to-real workflow that ABB-NVIDIA announced for RobotStudio HyperReality earlier this week, but executed entirely in cloud infrastructure rather than requiring local workstations. The Otto Group will present a concrete implementation of this approach on March 17: a digital twin that "precisely replicates the warehouse, all robotic systems and their interactions," connected to fleet management and warehouse management systems through a robotic coordination layer. The result is what Reply calls "centralized fleet coordination, predictive layout simulation and optimized processes even during peak periods" — a claim that requires the digital twin's predictions to be accurate enough to inform operational decisions in real time.

The convergence here is architectural: NVIDIA's simulation platform (Omniverse, Isaac Sim) running on hyperscaler infrastructure (AWS, Google Cloud) to produce synthetic training data for physical AI systems deployed at industrial scale. This is not a research demonstration but a production deployment for the Otto Group, one of Europe's largest e-commerce logistics operators. The economic logic is recursive: the digital twin generates training data, the trained models execute in physical warehouses, the performance data from physical execution feeds back into the digital twin to improve simulation fidelity, and the improved simulation generates better training data for the next model iteration. The loop is closed at industrial scale with 30,000-plus attendees at GTC expected to see this architecture presented not as future possibility but as operational reality. When Reply describes digital twin simulations as enabling "continuous improvements to be tested and efficiency gains to be validated," the implication is that the simulation has become the authoritative testing environment — the same ontological inversion identified in last week's reports on Synopsys eDT and ABB-NVIDIA.

Sources: PR Newswire | Business Wire

---

Learning Theory of Model Collapse: A Formal Framework for Replay-Induced Degradation

A comprehensive new arXiv paper published March 12 (arXiv:2603.11784, "Language Generation with Replay: A Learning-Theoretic View of Model Collapse") provides the first rigorous learning-theoretic treatment of model collapse, the phenomenon where training on machine-generated outputs degrades future model performance. The paper introduces a "replay adversary" that injects a generator's own past outputs back into its training stream — a minimal abstraction of the feedback loop where LLMs trained on web-scale corpora increasingly encounter synthetic text generated by previous model generations. Building on the language generation framework of Kleinberg and Mullainathan (2024), the authors prove a fine-grained characterization of when replay fundamentally limits generation capacity: uniform generation is unaffected by replay (same sample complexity guarantees), but non-uniform generation and generation-in-the-limit exhibit strict separations where countable classes that are generatable without replay become non-generatable when replay is introduced.

The technical contribution is in formalizing the distinction between three failure modes. First, "epistemic stagnation" — AI systems that excel at recombining existing knowledge but cannot generate genuinely new insights, leading to a slowdown in knowledge creation as more content becomes AI-synthesized. Second, "simulation drift" — the gradual divergence between a recursive simulation and the reality it models, compounded through feedback loops too fast or opaque for governance to track. Third, the paper identifies specific architectural mitigations that mirror industry heuristics: data cleaning (filtering previously output elements from the training stream), watermarking (embedding detectable signals in synthetic outputs), and output filtering (restricting the generator's action space to prevent replay contamination). The positive results show these techniques can preserve generatability for specific hypothesis classes; the negative results construct adversarial sequences where no amount of filtering prevents collapse.

The paper arrives as empirical evidence of model collapse accumulates across multiple domains. A Geeky Gadgets report published today synthesizes recent case studies: Stack Overflow experiencing a 78% drop in question volume as users turn to LLMs, Chegg losing 99% of stock value as students bypass traditional learning platforms, and publishers seeing up to one-third reduction in search traffic as AI-generated summaries intercept users. The recursive structure is explicit — each generation of AI consumes the outputs of the previous generation, creating what the paper calls a "photocopying a photocopy" effect where clarity, originality, and diversity degrade with each iteration. The learning-theoretic framing makes precise what has been observed empirically: the boundary conditions under which recursive training on synthetic data remains viable, and the structural properties of hypothesis classes that determine whether replay is benign or catastrophic. For Antikythera's framework, the question is whether this formalization arrives in time to inform the design of systems already deploying at scale — or whether the feedback loops are already too entrenched to untangle.

Sources: arXiv:2603.11784 | Geeky Gadgets

---

The "Atoms Over Bits" Pivot: Capital Flees AI Software for Physical Infrastructure

Wall Street is experiencing what analysts are calling "the Great Rotation of 2026" — a dramatic capital reallocation out of AI software and into physical infrastructure, energy, and industrial assets. As of March 13, the iShares Expanded Tech-Software Sector ETF has plummeted 18% year-to-date while the Energy Select Sector SPDR has surged 25%. The proximate cause is the "Power Wall" crisis: AI infrastructure buildout has outpaced electrical grid capacity across North America and Europe, creating massive backlogs in data center deployments due to lack of power interconnection agreements. Microsoft faces an estimated $80 billion backlog in Azure services caused specifically by power shortages, while enterprise software stalwarts like Salesforce and ServiceNow have seen double-digit declines as investors question 24x+ price-to-sales multiples in an environment where energy constraints have become the binding limit on AI deployment.

The winning side of this rotation tells the inverse story. ExxonMobil and Chevron are reporting record margins as Brent Crude hovers near $118 per barrel, driven by escalating military tensions in the Middle East and a 100% surge in jet fuel prices since last year. NextEra Energy is up 24% year-to-date, fueled by a 30-gigawatt backlog for AI data center power demand. Union Pacific and CSX are thriving as the essential movers of physical goods required for the U.S. manufacturing boom enabled by the "One Big Beautiful Bill Act" (OBBBA) fiscal package. The market narrative has shifted from "infinite digital scale" to "megawatts under management" — a reframing where the providers of electricity and freight, not software seats, are the primary AI infrastructure beneficiaries. InvestorPlace reports that AI inference's massive memory bandwidth requirements have rotated the bottleneck to high-bandwidth memory (HBM) and high-performance storage, with Micron and the newly independent SanDisk becoming plays on the memory buildout.

This is not a temporary sector rotation but a structural recognition that the digital revolution has hit a physical ceiling. The hyperscalers collectively spent $337 billion on AI infrastructure in 2025, with capital spending projected to climb toward $600 billion in 2026 — more than the annual GDP of Sweden or Poland. But spending cannot deploy infrastructure faster than electrical grids can provide power or supply chains can deliver physical components. The "SaaSpocalypse" has collapsed software multiples from 20-30x revenue toward historical norms, while energy and utility companies are repricing as "growth" stocks. The recursive simulations significance is that simulation infrastructure itself requires physical substrates that are now the primary constraint: digital twins, world models, and synthetic data engines are computationally expensive, and when electricity becomes scarce, the simulation infrastructure competes directly with the production systems it is meant to optimize. The atoms-over-bits pivot is the market pricing in a future where ownership of physical energy and logistics infrastructure determines who gets to run simulations at scale — a constraint that no amount of algorithmic efficiency can fully overcome.

Sources: Financial Content | InvestorPlace

---

Australia Guts Climate Simulation Capacity as AI Demands Energy

Australia's national science agency CSIRO announced this week that 102 full-time equivalent positions will be cut from its Environment Research Unit, with a large number coming from climate modeling teams. The cuts — part of a broader redundancy round eliminating 350 roles following 800 jobs lost over the past year — have sparked fears for the future of the Australian Community Climate and Earth System Simulator (ACCESS), the Earth system model built in partnership with the Bureau of Meteorology, universities, and international partners. Professor Sarah Perkins-Kirkpatrick, president of the Australian Meteorological and Oceanographic Society, described the cuts as "one of the worst things I've seen during my career," warning that "if we don't have the Australian climate model, then we simply cannot replicate Australia's climate and weather within an accurate envelope; it's just not possible." No other global climate model can accurately represent Australia's specific atmospheric and oceanic conditions, making ACCESS irreplaceable for climate science and policy planning across the region.

The CSIRO spokesperson stated that "addressing the pressing problem of climate change" remains a key focus but acknowledged that "some research areas will be reduced, to focus on areas of greatest impact," including cuts to atmospheric chemistry modeling, Indo-Pacific ocean dynamics, and operational support. Professor Andy Hogg, director of ACCESS-NRI, warned that the cuts would render Australia's Earth system modeling capacity "sub-critical" and could hamper monitoring of air pollution, greenhouse gases, and short-lived climate gases. The timing is perverse: ACCESS contributed to the sixth IPCC assessment report on Earth's climate future, and scientists across Australia depend on it for research into climate change impacts on agriculture, biodiversity, extreme weather, and coastal infrastructure. Yet the cuts arrive precisely as global AI infrastructure spending surges toward $600 billion, competing for the same skilled computational scientists, electrical grid capacity, and public funding that climate modeling requires.

This is the climate data paradox materialized as institutional collapse. Last week's report noted Bloomberg's argument that AI simultaneously consumes energy and generates the climate data needed to manage risk; this week Australia demonstrates the darker version of that feedback loop — where AI's energy appetite actively destroys climate modeling capacity by redirecting public resources, grid power, and scientific talent toward commercial AI deployment. A CSIRO insider noted that the Environment Research Unit and its Earth system modeling team are the only permanently funded unit in Australia working full-time on climate modeling, with the remainder operating on short-term contracts. The message is unambiguous: when climate simulation competes with AI simulation for scarce resources, the market has decided which simulations matter. For Antikythera's framework, this is recursive simulation eating itself — the infrastructure needed to model planetary futures being sacrificed to build the infrastructure that consumes planetary resources, creating a meta-level feedback loop where the capacity to simulate climate collapse diminishes in proportion to the systems accelerating it.

Sources: Sydney Morning Herald

---

Data Center Interconnects: The Next AI Infrastructure Bottleneck

As AI clusters scale from thousands to hundreds of thousands of GPUs, the internal plumbing of data centers has become the binding constraint. The issue is interconnects — the cables, transceivers, switches, and signal-processing chips that move data between GPUs, servers, racks, and buildings. GPUs are only as powerful as the data pipeline feeding them; if information cannot move fast enough between chips, even the most advanced processors spend time idle. In March 2026, Nvidia invested $4 billion ($2 billion each in Lumentum and Coherent) with multi-year procurement commitments for 200G-per-lane EML lasers, the critical component inside next-generation 1.6T optical transceivers. CEO Jensen Huang called the investment securing supply for "the next generation of gigawatt-scale AI factories," a supply chain lockup that signals Nvidia recognizes interconnects as the next infrastructure wave after compute, servers, cooling, energy, and memory.

The central tension is a technology debate: copper or optical fiber. Direct Attach Copper (DAC) cables are cheap, low-latency, and power-efficient but suffer signal integrity degradation at high data rates over distance. At 800G speeds, usable DAC cable lengths have shrunk to roughly three meters; as data rates approach 1.6T, copper's range contracts further. Active Electrical Cables (AEC) — copper with embedded signal processing — extend usable range to seven to ten meters while consuming 25-50% less power than optical alternatives, representing copper's last stand before the physics wall. Optical transceivers convert electrical signals to light pulses over fiber, eliminating distance constraints but consuming five to 15 watts per port, adding latency at conversion steps, and costing materially more than copper. Broadcom CEO Hock Tan argued on the company's recent earnings call that copper dominates most scale-up connections today because "on every dimension that matters in scale-up — latency, power, cost — copper wins," while optics handles longer-distance scale-out links.

The temporal structure matters: copper wins the battle today, but co-packaged optics (CPO) — which integrates photonics directly onto chip packages to eliminate optical conversion penalties — is projected to reach commercial scale in the 2027-29 window. Nvidia's $4 billion bet on Lumentum and Coherent is a wager that CPO timelines are real and that controlling the optical supply chain is strategic. The recursive implication is that the same compute-to-interconnect bottleneck applies to simulation infrastructure itself: digital twins and world models that generate synthetic training data are network-intensive workloads, and interconnect capacity determines how many GPUs can collaborate on a single simulation. When Reply demonstrates cloud-based Isaac Sim running on Google Cloud G4 instances, the claim that this architecture "significantly accelerates the validation process" depends on Google's internal network fabric being fast enough to stream simulation state across distributed GPU clusters. The interconnect bottleneck is not just a constraint on AI training; it is a constraint on the simulation infrastructure used to generate training data, creating a nested recursive dependency where solving the next AI infrastructure layer requires first solving the physical networking layer beneath it.

Sources: InvestorPlace

---

When Simulation Infrastructure Becomes the Primary Infrastructure

The pattern across this week's developments is the inversion of simulation from analytical tool to operational substrate. Reply's GTC demonstrations position digital twins not as post-hoc analysis but as the continuous development environment where AI systems are validated, retrained, and optimized before deploying to physical robots. The arXiv model collapse paper formalizes conditions under which training on synthetic outputs remains viable, making explicit that simulation-generated data can substitute for real-world data only under specific architectural constraints. The atoms-over-bits capital rotation reflects the market pricing in that simulation infrastructure requires physical substrates — electricity, memory bandwidth, network interconnects — that are now binding constraints on how much simulation can occur. CSIRO's climate modeling cuts demonstrate the zero-sum competition between climate simulation and AI simulation for scarce computational resources and scientific talent.

The recursive structure is that simulation infrastructure is becoming the primary infrastructure, with physical systems validated against it rather than the reverse. ABB-NVIDIA's 99% sim-to-real accuracy claim from earlier this week represents the threshold where virtual commissioning becomes functionally equivalent to physical testing. Reply's "AI Fast Lane for the Industrial Edge" creates a continuous loop where edge AI learns in the field but the authoritative testing environment remains the digital twin. Another Earth's synthetic satellite imagery trains environmental AI on fabricated representations of regions that real satellites underserve. In each case, the simulation is not merely predicting reality — it is producing the training data, the validation environment, and the operational envelope that defines what "correct" behavior means. The question is whether this inversion is sustainable or whether it accumulates undetected drift.

The model collapse paper's formal framework provides one lens: replay is benign for uniform generation (where sample complexity is fixed and known) but catastrophic for non-uniform and in-the-limit generation (where the required sample complexity depends on unknown target distributions). Translating to applied contexts: simulation-based training works when the simulation's fidelity is known and bounded, but degrades unpredictably when fidelity assumptions are violated or when the target distribution shifts due to the system's own outputs. The atoms-over-bits rotation is the market registering that these fidelity assumptions depend on physical infrastructure that cannot be abstracted away. CSIRO's cuts are the institutional manifestation of that constraint — when simulation requires energy and the energy is allocated to commercial AI, the capacity to simulate planetary futures diminishes. The feedback loop is planetary-scale: simulation infrastructure that consumes resources to model resource consumption creates conditions that invalidate its own models, but the system lacks the reflexive capacity to register this meta-level recursion until the divergence is already entrenched.

---

Implications

The recursive simulations developments this week converge on a single structural observation: simulation infrastructure is transitioning from auxiliary analytical capability to primary operational substrate, and this transition is occurring faster than the governance, economic, or physical infrastructure can adapt. In industrial robotics, Reply's GTC demonstrations show digital twins as the continuous testing and validation environment where physical AI is developed — not a parallel capability but the authoritative one. In training data generation, the model collapse paper formalizes when synthetic outputs can substitute for real-world data and when that substitution degrades performance, making explicit that the viability of simulation-based training depends on architectural properties that are not yet widely understood in deployed systems.

In capital markets, the atoms-over-bits rotation reflects investors pricing in that simulation infrastructure competes with physical infrastructure for electricity, memory bandwidth, and network capacity — and that when those resources become scarce, physical constraints bind before algorithmic optimization can compensate. In climate science, CSIRO's cuts demonstrate the zero-sum competition between climate simulation (modeling planetary futures) and AI simulation (generating synthetic training data) for the same computational scientists, grid power, and institutional funding. The atoms-over-bits narrative frames this as software losing to energy, but the deeper structure is simulation losing to simulation: AI training infrastructure outbidding climate modeling infrastructure for the same physical substrates.

The common risk is "simulation drift" — the gradual divergence between recursive simulations and the reality they model, compounded through feedback loops that are too fast, too opaque, or too economically entrenched for governance to track. ABB's 99% sim-to-real fidelity and Reply's autonomous retraining loops assume the digital twin accurately represents physical dynamics; the model collapse paper proves that assumption breaks under specific conditions; the energy bottleneck shows that maintaining simulation fidelity requires physical resources that are increasingly contested. When CSIRO slashes climate modeling capacity to reallocate resources, the capacity to detect simulation drift in climate AI models diminishes precisely as the deployment of those models accelerates. The governance fabric concept from last week's ARMA summit proposed embedding oversight into data infrastructure, but when the simulation infrastructure itself is the contested resource, governance cannot be retrofitted — it must be architectural.

For Antikythera's framework, the question is at what point the accumulation of recursive simulation infrastructure — digital twins, world models, synthetic data engines, LLM-generated training corpora — constitutes a qualitatively different kind of planetary computation. Not a tool humans use to understand the world, but a parallel representational infrastructure that increasingly mediates between human decisions and physical reality. The model collapse paper's distinction between uniform, non-uniform, and in-the-limit generation provides a formal taxonomy for when this mediation remains stable versus when it degrades; the atoms-over-bits rotation shows that stability depends on physical substrates that cannot scale at the rate algorithmic demand requires; CSIRO's cuts demonstrate that when simulation competes with simulation for those substrates, governance capacity erodes faster than deployment accelerates. The boundary between simulating the world and participating in it is not a philosophical question — it is the operational reality unfolding this week across industrial robotics, capital markets, climate science, and data center infrastructure, and the recursive feedback loops are already too fast for any single actor to control.