The physical systems behind abundant intelligence.

Artificial intelligence is making cognition cheaper faster than the physical world can adapt.

That changes where the bottlenecks live. When compute gets cheaper, memory bandwidth becomes scarce. When clusters get larger, interconnect becomes scarce. When data centers densify, power and cooling become scarce. When robot policies improve, deployment and reliability become scarce. When biological models generate more hypotheses, experiments become scarce.

AI progress does not remove bottlenecks. It moves them one layer closer to reality.

Atomic Analysis is a series about those moving constraints: the infrastructure that turns intelligence into useful work in compute, machines, and science.

The Bottleneck Moves

Every technology boom eventually moves its bottleneck.

For most of modern computing, arithmetic was expensive. The industry responded by making transistors smaller, processors faster, and specialized accelerators more efficient. As arithmetic became cheaper, memory hierarchy and data movement mattered more. GPUs made matrix multiplication abundant enough that HBM bandwidth, packaging, and interconnect became strategic technologies. Large models then made inference itself an industrial workload, shifting the constraint again toward memory capacity, utilization, power, and cooling.

AI accelerates this pattern because the capability curve is moving faster than the systems around it.

A model can be copied almost instantly. A data center cannot. Software can be deployed globally in an afternoon. A transformer can take years. A robot policy can improve overnight; the factory it operates in still has safety cages, worn actuators, maintenance windows, legacy PLCs, and humans who need the line to keep running. A biological model can propose a thousand experiments before a laboratory has finished the first ten.

Every time intelligence gets cheaper, scarcity moves one layer closer to the physical world.

This is the organizing idea behind the rest of the series.

Software scales differently from reality

Digital systems compound quickly because their replication cost is close to zero. Physical systems compound through factories, construction crews, qualification cycles, permits, field service, and accumulated operating experience. Their learning curves can be powerful, but they run on a different clock.

That difference was easy to ignore when AI mostly produced text on a screen. It becomes harder to ignore as AI begins to consume large fractions of electrical infrastructure, coordinate fleets of machines, and propose interventions in biological systems.

The hierarchy matters because progress at one layer increases pressure on the next. Better reasoning makes long-horizon inference more valuable, which increases memory and serving requirements. Larger clusters increase power density. Better robot policies increase the number of environments worth automating, which makes deployment labor and integration more important. Better biological models increase the number of plausible interventions, which raises the value of high-throughput experimentation.

The interface becomes the product

Many of the most interesting systems appear at boundaries between disciplines. An optical interconnect sits between semiconductor design and networking. A solid-state transformer sits between the grid and the data center. A workload orchestrator sits between a scheduler and an electricity market. A robot integrator sits between an AI model and a factory line. A scientific agent sits between computation and an instrument.

These interfaces are hard because they inherit constraints from both sides. They are also where old assumptions tend to fail first.

The Infrastructure of Intelligence

Intelligence is often discussed as if it lived inside the model. Economically useful intelligence is a much larger system.

A request enters a serving stack. Tokens are prefetched, weights are moved, KV state is retained, accelerators communicate, results cross networks, electricity is converted several times before it reaches the board, and the resulting heat has to leave the building. Long-horizon agents make the loop even more demanding because the system must preserve state, invoke tools, wait on external processes, and resume work over much longer periods.

The cost of thinking is therefore becoming a systems problem.

When FLOPs stop being the whole story

Training taught the industry to think in terms of accelerator count and peak FLOPs. Inference exposes a different set of constraints. Autoregressive decoding repeatedly reads model state while producing relatively little computation per token. Long contexts increase KV-cache requirements. Mixture-of-experts architectures reduce arithmetic per token but add communication and routing. Reasoning models spend more tokens per answer. Agentic workloads preserve state for longer and generate irregular patterns of tool calls and waiting.

The next unit of AI progress may be limited less by multiplying numbers than by moving state.

HBM solves part of the problem by placing enormous bandwidth close to the accelerator, but capacity remains expensive and finite. As models and contexts grow, systems increasingly spill state across devices or clusters. That turns memory architecture into communication architecture.

Scale-up fabrics try to make multiple accelerators behave like a larger coherent machine. Scale-out networks connect those machines into clusters. CXL and related approaches push toward more composable memory. Compression reduces the bytes that have to move at all.

Eventually the electrical links themselves become costly. SerDes power, reach, package-edge bandwidth, and switch density worsen as cluster bandwidth rises. This is where optical I/O, co-packaged optics, optical chiplets, and new laser architectures become strategically important.

Ayar Labs, Avicena, Lightmatter, Celestial AI, and the broader optical-interconnect ecosystem are responses to the same underlying fact: data movement is consuming a larger fraction of the system budget.

Once clusters scale far enough, the infrastructure constraint moves again. Every accelerator draws power. Every electrical link dissipates heat. Every conversion stage loses energy. Every watt ultimately has to be delivered through a transformer, busway, and power shelf before a cooling system removes it from the building.

The Megawatt Race

For most of the AI boom, the scarce resource was compute. GPUs were hard to get, advanced packaging was constrained, and networking bandwidth struggled to keep up with ever-larger clusters.

Those bottlenecks have not disappeared. But power moves on a very different clock.

A new accelerator can move from design to volume production in a few years. A large transformer can take three or four years to procure. A new grid connection can take five. A major transmission project can take a decade.

That mismatch is beginning to shape the physical geography of AI.

Microsoft has committed for twenty years to support the restart of an 835 MW nuclear plant. Amazon has contracted for up to 1.9 GW from Susquehanna. Meta has signed long-duration nuclear agreements covering several gigawatts. Oracle redesigned Project Jupiter around as much as 2.45 GW of on-site fuel cells.

The technologies are different. The objective is the same: secure power before the GPUs arrive.

The scarce product is not electricity in the abstract. It is 500 MW, at this location, by this date.

The grid wasn't built for AI factories

Data-center electricity consumption was roughly 415 TWh globally in 2024 and is projected to approach 945 TWh by 2030. In the U.S., projections imply tens of gigawatts of incremental peak demand over the same period.

The headline generation pipeline looks enormous. More than 2 TW of generation and storage sits in U.S. interconnection queues. But queue capacity is not the same as deliverable capacity. Historically only a fraction of projects reach operation, and even projects with signed agreements can wait years for transmission, substations, transformers, and switchgear.

Transformers have become a particularly visible bottleneck. Large units now carry multi-year lead times. Gas turbines are similarly constrained, with the major manufacturers carrying years of backlog. New manufacturing capacity takes years to commission.

Capital can accelerate the decision to build a factory. It cannot collapse a three-year physical lead time into a quarter.

AI scales in three directions

Scale-up packs more accelerators into a tightly coupled system, pushing rack power from tens of kilowatts toward hundreds of kilowatts.

Scale-out adds more racks until the building reaches the ceiling of its grid connection.

Scale-across connects multiple facilities into one distributed compute system. As networking improves, several smaller power positions connected by fiber can partially substitute for one extraordinarily difficult gigawatt-scale interconnection.

For the first time, network architecture can partially arbitrage power geography.

Firm power is becoming a strategic input

Different buyers value power differently. Hyperscalers can sign multi-decade contracts and underwrite first-of-a-kind projects. Neoclouds care acutely about both time-to-energization and cost per GPU-hour. Colocation providers effectively sell energized megawatts, so the date a building turns on is part of the product.

This diversity is pulling several generation technologies into the market at once.

Fuel cells and on-site generation

When the grid cannot deliver and turbines cannot be bought, developers turn to equipment that can be manufactured and installed sooner. Solid oxide fuel cells and reciprocating engines are the clearest examples.

Fuel cells convert fuel electrochemically rather than through combustion. They can reach electrical efficiencies around 60%, produce much lower local pollutant emissions than combustion systems, and output DC natively—an increasingly useful characteristic as data centers move toward higher-voltage DC distribution.

The trade-off is cost. Fuel cells remain expensive relative to large conventional generation, and both fuel cells and engines require access to constrained gas infrastructure.

The comparison is therefore not always cheap power versus expensive power. It can be expensive power next year versus cheap power four years from now.

Geothermal

Geothermal offers firm output without an ongoing fuel bill. The key technical shift is the use of oil-and-gas drilling techniques to access resources that traditional geothermal could not.

Enhanced geothermal systems create permeability in hot rock instead of relying on naturally occurring reservoirs. Closed-loop systems go further and circulate working fluid through sealed wells without depending on a reservoir at all.

Fervo's Cape Station is an important commercial proof point. The project secured non-recourse financing before reaching first power, suggesting lenders were beginning to underwrite the asset against contracted cash flows rather than treating it purely as experimental technology.

Sage Geosystems, XGS Energy, Eavor, Zanskar, and Quaise attack different parts of the same problem: how to find, access, and economically extract heat from a much larger portion of the subsurface.

Nuclear

Advanced nuclear fits the data-center load profile almost perfectly on paper: firm output, high capacity factor, small land footprint, and the possibility of colocation.

But demand is arriving before the supply chain is ready.

Kairos Power and TerraPower have projects under construction. Aalo Atomics, Antares, Valar Atomics, and Deployable Energy are among the companies progressing through newer demonstration pathways.

The reactor is only one bottleneck. Many designs require HALEU. Some add specialized TRISO or metallic fuel fabrication. Large forgings, qualified instrumentation, licensing support, and a specialized workforce can all set the schedule.

The reactor gets the attention. The deployment date may be determined by the fuel plant or forging supplier.

Long-duration storage

Storage can turn fast-to-build wind and solar into a firmer resource. But “storage” hides several distinct products.

A 100-hour iron-air battery solves an energy-duration problem. A four-hour lithium-ion system handles normal operating peaks. A rack-level buffer may need to react in milliseconds.

Those are not versions of the same product.

The existing grid may have more capacity than it appears

New transmission is necessary, but some capacity is trapped behind conservative operating assumptions and inefficient planning processes.

Dynamic line rating uses real-time weather and conductor conditions instead of static worst-case assumptions. LineVision, Heimdall Power, and Ampacimon are among the companies pursuing it. Advanced conductors can carry more electricity on existing towers. VEIR is exploring superconducting transmission for dense corridors.

The technology is often easier than the institutional problem. Utilities earn returns on deployed capital, so a technology whose primary benefit is deferring capital can be economically attractive to the system without being naturally attractive to the buyer.

Interconnection adds another layer of delay. Projects are studied under conservative assumptions, speculative requests occupy queue positions, and withdrawals can trigger restudies.

Some of the five-year wait is made of copper. Some of it is made of paperwork.

The data center itself is being rebuilt around power

For decades, medium-voltage AC entered a facility, passed through transformers, switchgear, UPS systems, and power distribution, and was eventually converted into low-voltage DC at the server.

At 5–15 kW per rack, the conversion losses were manageable. GPU density changes the physics.

The industry first moved toward 48-volt DC distribution inside the rack. That helped support 100+ kW systems, but it eventually runs into a current problem: at fixed voltage, more power means proportionally more current, more copper, and more heat.

At 10 kW per rack, conversion loss was an efficiency problem. At 1 MW, it becomes a geometry problem.

This is why 800 VDC is emerging. Higher voltage reduces current, conductor mass, and resistive loss. NVIDIA, Google, Microsoft, and the Open Compute Project ecosystem are developing a common architecture around it.

The shift creates new requirements in solid-state transformation, protection, arc interruption, insulation, connectors, and safety.

Heron Power, DG Matrix, Amperesand, and Ionate represent different approaches to medium-voltage and solid-state conversion. The technical topology matters, but the industrial questions may matter more: can the equipment last for decades, can it be qualified quickly enough, and can production scale without sacrificing reliability?

The GPU is becoming a grid event

Thousands of GPUs can change operating states together. A training job begins. A checkpoint completes. A fault occurs. At sufficiently large scale, those synchronized transitions can move tens of megawatts faster than conventional grid infrastructure was designed to respond.

This turns storage into a multi-timescale problem: milliseconds for GPU transients, minutes for generator bridging, and hours for grid services.

The sub-second layer is particularly new. Native DC distribution makes rack-level storage easier to integrate, and the requirement comes directly from compute behavior rather than today's equipment shortages.

Cooling is now compute capacity

Rack densities above 100 kW increasingly require liquid cooling. Single-phase direct-to-chip systems are currently the pragmatic default. Two-phase systems can support higher heat flux but add complexity in fluid management, materials, pressure, and maintenance.

Accelsius, ZutaCore, Corintis, JetCool and others represent different points in the emerging stack.

The economics go beyond electricity savings. A facility has a fixed power envelope. Every watt not consumed by cooling can be allocated to compute.

A watt saved on cooling is another watt available to a GPU.

What if the fastest power plant is software?

Every physical technology above asks how to supply a 500 MW data center. Load orchestration asks whether it actually needs 500 MW every second.

AI workloads are unusually controllable. Training can checkpoint. Fine-tuning can pause. Batch inference can defer. Some jobs can move geographically. Interactive inference cannot.

The interesting asset is not merely the scheduler. It is the empirical knowledge of how much each workload can flex, for how long, without violating an SLA—and the operating history required for a utility to trust that promise.

The code can be copied. Trust takes longer.

What survives when the shortage ends?

Transformer factories will expand. Turbine supply will improve. Interconnection reform will remove some speculative projects. Today's willingness to pay almost any premium for speed will not last forever.

That makes one question unusually useful: does the product still matter when lead times normalize?

Power density still rises. That keeps pressure on voltage, conversion, protection, and cooling. GPU transients remain. Better grid visibility remains useful. Geothermal drilling tools still improve project economics. Workload flexibility still increases asset utilization.

The durable layer is not simply whatever is scarce today. It is the infrastructure required by the architecture of increasingly dense compute.

The Deployment Gap

AI has made dramatic progress in perception, language, planning, and control. That has made robotics demos better very quickly.

Deployment has not improved at the same rate.

A robot that succeeds in a benchmark can fail in a factory because the floor is reflective, the bin changed shape, a cable loosened, a fixture moved two centimeters, or the process engineer cannot afford an hour of downtime to recalibrate it. The gap between a compelling demo and a reliable production system is where much of physical AI will be won or lost.

The difference between a robot demo and a robot business is everything around the model.

Physical autonomy depends on far more than inference quality. Perception has to survive changing lighting and clutter. Planning has to respect physical constraints. Control has to run at deterministic latency. Actuators, reducers, batteries, sensors, and wiring harnesses have to survive years of operation. Safety systems need predictable failure modes. Maintenance teams need diagnostics that make sense at 3 a.m.

Then comes integration with the environment: PLCs, conveyors, MES systems, safety cages, work instructions, quality systems, and human operators.

As foundation models make the “brain” more reusable, more of the deployment cost shifts into this systems layer.

Integration is the hidden labor market of robotics

Industrial automation is installed by a fragmented ecosystem of integrators that survey facilities, design cells, select hardware, program controls, certify safety, commission equipment, and maintain it afterward.

This work is expensive because every site is treated as a bespoke engineering project.

AI can change that workflow before it changes the robot itself. Site scans can become digital twins. Cell layouts can be generated automatically. PLC code can be synthesized and tested in simulation. Failure histories can become reusable templates. Commissioning can move from artisanal engineering toward repeatable software-assisted deployment.

If robots become cheaper while integration remains expensive, integration captures a larger share of system value.

Reliability compounds

A robot fleet creates a different kind of data advantage from a consumer software product. Each deployment produces intervention logs, failure modes, wear patterns, calibration history, and edge cases that rarely exist in public datasets.

The useful question is not only whether a company has a better policy model. It is whether every new installation makes the next installation easier.

The strongest physical-AI systems will likely combine models, simulation, teleoperation, diagnostics, maintenance, and deployment tooling into one feedback loop.

The Experiment Gap

Biology is increasingly becoming a problem of too many plausible hypotheses and too little experimental throughput.

Models can rank targets, propose molecules, predict structures, generate proteins, and search spaces that no human team could enumerate manually. That is valuable, but prediction is not evidence.

A hypothesis becomes science only after the physical world answers back.

Models are beginning to generate experiments faster than laboratories can run them.

The scientific loop has a physical clock

Biological discovery moves through a loop: form a hypothesis, design an intervention, synthesize or perturb, measure the result, interpret the data, and choose the next experiment.

AI can accelerate the first and last steps dramatically. The middle remains bounded by synthesis, cell culture, imaging, sequencing, assays, instrumentation, sample handling, and the time biology itself takes to respond.

That mismatch is the experiment gap.

Measurement becomes strategic

As models improve, the value of high-quality experimental data rises. Sequencing, single-cell methods, spatial biology, proteomics, imaging, organoids, and high-throughput perturbation are not simply data sources. They are interfaces between computational hypotheses and biological reality.

The more expressive the measurement, the more useful the feedback loop becomes. But rich measurements are often slower, more expensive, or harder to standardize.

Automation changes the learning rate of science

A conventional laboratory is organized around human operators and instruments that often behave like isolated islands. Samples move manually. Metadata is inconsistent. Protocol deviations are discovered after the fact. Instruments produce files rather than structured state.

An AI-native laboratory needs a different architecture: machine-readable protocols, automated sample handling, instrument APIs, real-time quality control, provenance, and software capable of choosing the next experiment based on the last result.

The goal is not to remove scientists. It is to increase the number of high-quality learning cycles each scientist can generate.

The Runtime for Science

The natural endpoint of scientific AI is not a chatbot that knows more papers.

It is a system that can move through the scientific loop itself.

It reads literature and internal data. It forms a mechanistic hypothesis. It runs a simulation. It chooses an experiment. It sends instructions to an instrument or laboratory workflow. It observes the result, checks provenance and uncertainty, updates its model, and decides what to do next.

Most of the components already exist in primitive form. What does not yet exist is the runtime that connects them.

Scientific work is fragmented across tools and worlds

A modern research program may involve papers, proprietary datasets, Python notebooks, molecular simulations, imaging systems, sequencing instruments, robotic liquid handlers, ELNs, LIMS, cloud compute, and conversations with domain experts.

Each system has its own representation of state. Most do not expose clean interfaces to agents. Provenance gets lost between steps. Experimental context lives in human memory. A result may be statistically significant but scientifically meaningless because the wrong assumption entered three tools earlier.

A runtime for science needs more than orchestration

It needs persistent identity and memory so long-running programs retain state. It needs permissions because experiments, instruments, and proprietary data cannot be exposed indiscriminately. It needs provenance so every conclusion can be traced back to measurements, transformations, and assumptions. It needs uncertainty because scientific decisions are rarely binary.

And it needs interfaces to the physical world: instruments, robots, simulation engines, and laboratories.

The long-term prize is closed-loop discovery

Once those pieces connect, the unit of scientific progress changes.

Instead of asking a model for a prediction and then handing the result to a human workflow, the entire research program can become iterative software controlling physical experiments.

That does not make science instantaneous. Biology still grows at biological speed. Instruments still have throughput limits. Experiments still fail for reasons no model predicted.

But it raises the learning rate of the system.

That is the same pattern seen throughout this series. Intelligence becomes cheaper; the bottleneck moves outward. In science, the scarce resources become trustworthy experiments, instruments, physical throughput, and the infrastructure that connects evidence back to models.

Parting thought

The next decade will contain extraordinary progress in models. But the economic and scientific consequences will be determined by what those models can actually touch.

Power systems. Memory. Networks. Factories. Robots. Laboratories. Instruments. Biology.

The frontier increasingly sits at the interface between intelligence and reality.