Built for Heat — Interview III

Q 01 · The 1922 Problem

The rack was designed in 1922. Are we still using it?

You have said in other contexts that most data center infrastructure debates ignore the most fundamental constraint — the enclosure itself. Walk me through what you mean. What is the rack, where does it come from, and why does its origin matter in 2025?

The standard that governs more than 85 percent of all rack deployments worldwide today is EIA-310. The current version — EIA-310-E — was ratified in 2019. It references hole spacing dimensions that were first codified by Western Electric on the 17th of March, 1922. That drawing — Western Electric Drawing 35-B-1, "Relay Rack Mounting Dimensions" — fixed the 19-inch mounting width for telephone repeaters, echo suppressors, and carrier terminal bays in central offices. The dimension was not engineered. It was inherited from the steel relay panels that already existed in manual switchboards. The holes were put where they needed to be so that new equipment could bolt into existing ironwork without re-drilling.

The computing industry adopted the rack in the 1960s because the military had standardized it in the 1930s, and IBM needed mounting structures for the System/360. Not because anyone evaluated whether 19-inch rail spacing was an optimal geometry for thermal management of high-density compute. They used what existed — the same reason DEC, Sun, and Compaq used it. The 42-unit, 19-inch rack is a telephone equipment frame with a server bolted into it. That is the literal history.

What makes this more than a historical footnote is what EIA-310 does not specify. In no version — A through E, from 1957 to 2019 — has the standard ever defined a maximum power density, a minimum airflow requirement, an inlet temperature limit, or any thermal parameter of any kind. The standard that governs the enclosure in which you are running 120-kilowatt AI workloads was written to hold telephone relays, and it contains zero thermal requirements. All the cooling — raised-floor CRAC units, cold-aisle containment, in-rack CDUs, immersion tanks — exists entirely outside the standard that defines the enclosure. The cooling was bolted on. Every decade, a new bolt.

EIA-310-E references hole spacing from 1922. It contains zero thermal requirements. The enclosure for a 120 kW AI rack was never designed for thermal management at all.

Q 02 · The Physical Ceiling

Where does the 19-inch form factor actually break?

The industry has managed to keep pace with energy density increases inside the existing form factor — cold plates, in-rack CDUs, rear-door heat exchangers. Is there a physical limit, or is this just continuous adaptation?

The adaptation has been impressive. But there is a physical ceiling, and we are approaching it. The core constraint is geometry. A 19-inch rack has a rail spacing of 18.312 inches and a depth range of 610 to 914 millimetres. That defines the frontal area available for heat exchange. At air cooling, the physics are hard: at standard ASHRAE A1 conditions — 18 to 27 °C inlet — air cooling at 1,200 CFM with a 25 °C temperature rise reaches a practical ceiling of around 20 kW per rack before airflow velocity becomes the binding constraint, not the electronics. Beyond that, the required air velocity creates acoustic and back-pressure problems that no fan array can practically solve inside the existing cabinet geometry.

Liquid cooling pushes the ceiling higher — direct-to-chip cold plates demonstrated at 80 to 100 kilowatts, the NVIDIA GB200 NVL72 at 120 kilowatts per rack in 2025. That is the current state of the art. At 120 kilowatts the problem has changed character: the enclosure was never designed to route fluid at 65 °C across 72 simultaneous connection points. Every solution remains a retrofit onto a frame designed for telephone relays.

The floor projection changes this further. Blackwell Ultra — shipping 2025 to 2026 — reaches 150 to 200 kilowatts per rack. NVIDIA's Rubin Ultra NVL576, confirmed by Jensen Huang at GTC 2025, targets 600 kilowatts in a single rack by H2 2027. Ten racks today, one rack in 2027, the same compute output. The physical footprint shrinks. The thermal density per slot does not. The case for enclosure redesign is not an original position — Vertiv is already reworking its systems. The 19-inch geometry has not moved. The cooling integration inside that constraint has. The argument is confirmed. The geometry remains to be solved.

The deeper problem is not the current ceiling — it is the trajectory. The energy density progression over sixty years represents roughly a 100-fold increase: 30 watts per rack unit in 1964 to the equivalent of 3,000 watts per rack unit on a GB200 NVL72. That rate does not slow when you retrofit liquid cooling onto a 1922 frame. The next generation of compute — whatever follows Blackwell — will not fit inside a cooling architecture that was designed for the previous one. At some point the question stops being how to manage more heat inside the existing enclosure and becomes whether the enclosure was ever the right starting point. And every kilowatt past that ceiling has to leave the rack somewhere. The standardized point where it leaves — at a known temperature and flow rate — is the Thermal Plug. That is the part of this worth owning.

The 42U, 19-inch rack — identical in external geometry to what Western Electric specified in 1922, now carrying 120 kW of AI compute. The cooling is bolted on. The enclosure was never designed for this job.

Q 03 · The Market Landscape

Has anyone already tried to replace it?

If the problem is structural, it is not new. Has the industry produced any genuine alternatives — not just better cooling for the existing form factor, but a fundamentally different enclosure geometry?

Yes — and the picture sorts into three buckets: proprietary hyperscaler departures, alternative cooling geometries, and academic work the commercial world hasn't absorbed. The most significant commercial departure is the Open Compute Project's Open Rack — introduced by Facebook in 2011, widened to 21 inches with a 48-volt DC bus bar replacing per-server supplies; V3 targets 40 kilowatts on air and supports liquid manifolds above that. More than a million OCP racks run inside hyperscaler campuses — yet the 19-inch market is still more than 85 percent of global shipments. OCP coexists with the standard; it has not displaced it. The proprietary departures go further — Google's Googol rack (2020): 21 inches, 48 volts, 100 kilowatts liquid-cooled; Microsoft's Project Zipline (2022): 800 millimetres, 120 kilowatts on two-phase liquid — but these are franchise moves by operators large enough to absorb a proprietary supply chain, and they too coexist with EIA-310 rather than replace it.

The genuine geometry changes sit at the edges. Immersion cooling — boards slid horizontally into dielectric fluid, natural convection replacing fans, no cold plates — is the most radical change in production, but board-qualification friction keeps it below two percent of global deployments. The academic work comes closest to the structural insight: Georgia Tech ran a chimney-rack prototype from 2014 to 2018 — a five-kilowatt rack with a vertical central shaft hitting a 40 °C temperature differential without mechanical fans; ETH Zurich published a chimney-server concept at ACM SIGARCH in 2015. The research keeps arriving at vertical flow and passive convection. The commercialization keeps not following.

So the gap is precise. What exists is either a dimensional departure from 19 inches that still assumes forced airflow or pressurized liquid, or a radical change that demands new board qualification. What does not yet exist commercially is an enclosure that changes the structural thermal architecture without requiring the compute modules to change — one that works with existing 8-GPU boards, existing cold-plate standards, existing rack-unit dimensions, and routes heat on a fundamentally different geometry around them. And that missing enclosure has a missing twin: the standardized, bankable connection point at its edge where the heat leaves for the grid. The enclosure is the gap the academic work points toward. The connection point is the Thermal Plug. I am defining the second so the first has somewhere to terminate.

Energy Density Progression — 1964 to 2025

Era	Rack Load	W / U	Notes
DEC VAX-11/780 · 1980	~1.5 kW	~40 W/U	Front-to-rear air, 300–400 CFM. Approximate — not a rack system; figures represent cabinet-level equivalent.
Compaq ProLiant · 1993	6–10 kW	~200 W/U	Raised-floor CRAC, 42U rack era. First-generation 1U blade chassis not until early 2000s.
IBM BladeCenter · 2002	14–18 kW	~400 W/U	Introduced 2002; 14–18 kW is rack-level aggregate across multiple chassis. First reports of hot-aisle temperatures exceeding 40 °C.
Hyperscale containment · 2015	30–40 kW	~800 W/U	Cold-aisle containment, free-air cooling, OCP Open Rack. Representative peak — installations vary.
NVIDIA DGX H100 · 2023	40–50 kW	~1,100 W/U	4-system rack aggregate, air-cooled configuration. In-rack CDU variants push higher; figure reflects baseline air-cooled deployment.
NVIDIA GB200 NVL72 · 2025	120 kW	~2,500 W/U	Liquid-to-liquid CDU, 65 °C secondary loop, 48U system — at the physical ceiling of the 19-inch rail spacing.

Figures are representative of era peaks. Individual configurations vary. Sources: EIA-310 specification history · NVIDIA product documentation · Eric R026 technical review.

"The research arrives at the same place every time: branching and passive convection. The market stalls on supply-chain inertia. What has not been built is the enclosure that absorbs the existing compute modules and changes only the geometry around them."

Dimitri Wolf

Q 04 · What Biology Figured Out

What does four hundred million years of evolution know about heat management?

You mentioned biological systems as a reference. That framing gets used loosely in engineering discussions. Where is it rigorous, and what specifically does it offer here?

The rigorous entry point is Murray's Law. Formulated by Cecil Murray in 1926 — four years after Western Electric filed the relay rack drawing — it describes the branching ratios of vascular networks: the cube of the parent vessel diameter equals the sum of the cubes of the daughter diameters. That is not analogy. That is a derivable minimum-energy result. It can be shown from first principles that a branching network following Murray's ratios minimizes the total energy required to move fluid across the network while maintaining constant wall shear stress. Every mammalian vascular system, every plant, every tree follows this branching law — not because they chose it, but because any network that departs from it is outcompeted.

Applied to a thermal enclosure, Murray's Law defines a 3-scale hierarchy. At the smallest scale — die to board — you need a continuous conductive path along the full length of each server board: a cold rail that makes contact with any component hot enough to establish a gradient. No individual cold plates. No precision alignment. The rail is the capillary. The GPU packages continue on liquid cold plates — that path is unchanged. The rail handles the rest: SSDs, network controllers, voltage regulators — components that generate heat below the liquid-cooling threshold. The rack frame itself, as a thermally conductive aluminium extrusion, is the outermost layer: a passive radiator. None of this replaces liquid cooling. It completes the architecture around it. At the intermediate scale — board to rack — each board's rail feeds a board-level manifold that aggregates heat from all components into a single thermal stream before it reaches the rack backbone. At the largest scale — rack to building — the rack backbone collects from all board manifolds, increases cross-section at each bifurcation in accordance with Murray's ratios, and delivers to the building-level infrastructure. The artery's endpoint — where the rack hands its heat to the building, and the building to the grid — is the Thermal Plug. The branching law runs unbroken from capillary to that connection.

The key insight from biology is directionality. Vascular networks do not solve the heat transfer problem by adding pump pressure. They solve it by ensuring that at every scale transition — artery to arteriole to capillary — the aggregate cross-sectional area increases, velocity drops, and residence time extends exactly where exchange needs to happen. Every current cooling architecture does the opposite: it collects heat into a manifold smaller than the source area, increases velocity, and moves it away as fast as possible. That is thermodynamically wasteful at the point where waste costs the most. Biology's solution is to slow the coolant where the exchange happens, not accelerate it.

Murray's Law is not a biological curiosity. It is the only branching geometry that minimises total network resistance — and it applies equally to blood vessels, rivers, and fluid circuits in buildings.

Murray's Law: d³_parent = Σ d³_daughters. Minimum energy. Derivable from first principles. Every vascular system follows it — not by choice, but because those that deviate are outcompeted.

Nature's Build · The Termite Solution

The second biological reference is the termite mound. Macrotermes subhyalinus — the African termite — maintains internal nest temperature within 1 °C of 30 °C regardless of ambient conditions, without mechanical fans, using a system of tapered central shafts and peripheral channels. The physics is buoyancy-driven stack ventilation: delta-P equals the difference in air density between ambient and hot, times gravitational acceleration, times the height of the shaft. At rack scale, this translates to a central exhaust shaft that widens from bottom to top — a taper that converts thermal energy directly into airflow velocity without fan power. The Georgia Tech chimney rack prototype demonstrated a 40 °C temperature differential without fans at 5 kilowatts. The physics is straightforward. The engineering is a geometric choice, not a new technology.

Every biological thermal system runs vertically. Capillaries in bone, xylem in trees, veins in leaves — the direction of flow follows gravity and convection, not the constraints of a manufacturing process. The data center rack is the only thermal system in the modern built environment that deliberately routes its primary airflow path against the direction physics prefers: horizontal, forced, fighting buoyancy at every point. Rotating the boards ninety degrees — from horizontal shelves to vertical panels — does not change a single chip. It changes the geometry of the air column from a channel that fights convection to one that completes it.

The mound is not insulated from the outside. It is connected to it — deliberately. The termites solved active thermal regulation without electrical input, at scale, in equatorial heat. The data center running at 99% uptime is solving the same problem. The mound solved it fifty million years earlier.

The chimney is not a novel concept. It is the geometry that every biological thermal system converges on. The question is why the industry has not.

Q 05 · The Trash Can Was Right

Apple got there first — and failed. What does that teach us?

Has anyone in the known industry actually attempted this kind of integrated thermal architecture at scale — not as a concept, but in a shipped product?

The Mac Pro 2013 is the most important enclosure design in computing history that nobody talks about correctly. Apple took a cylindrical housing roughly 250 by 250 millimetres, placed a single centrifugal blower at the apex, and arranged three heat sources — the CPU and two GPUs — radially around a triangular aluminium thermal core. The core was a unified, machined heat exchanger: one component, three heat sources, one airflow path. The entire machine dissipated 450 watts in a 9.9-litre enclosure. By the thermal physics of 2013, that was extraordinary. The enclosure was the heat exchanger. The structure was the manifold. That is exactly the principle that Murray's Law and the biological literature point toward as the correct architecture.

The failure was not in the thermal concept. The failure was in the coupling. By making the thermal core a unified, precision-machined component that was geometrically specific to the GPU configuration of 2013, Apple created a machine that could not evolve. When AMD released new GPUs with different die positions, different TDP profiles, and different form factors eighteen months later, the unified core could not accommodate them. Apple found itself with a thermal architecture that was optimal for a specific hardware configuration that had become obsolete — and no upgrade path that did not require redesigning the entire core. The machine was discontinued in 2019, replaced by a conventional 19-inch-compatible tower that could accept off-the-shelf expansion cards.

The lesson is precise. The principle — enclosure as thermal manifold, structure as heat exchanger — is correct. The failure point was the tight coupling between the manifold geometry and the compute modules it served. The right architecture separates the two: a fixed manifold geometry that follows Murray's Law branching ratios, into which compute modules connect through a standardized thermal interface. The manifold does not change when the GPU changes. The GPU plugs in. The module is the commodity. The manifold is the infrastructure. Apple got the geometry right and the modularity wrong. The next generation gets both right, or it will fail for the same reason. That standardized thermal interface — the clean line between the commodity module and the durable manifold — is exactly what the Thermal Plug specifies: at the slot here, at the building header in the architecture that follows.

Apple made the enclosure the heat exchanger — which is correct. They coupled the geometry rigidly to the hardware — which is fatal. The principle survives. The coupling error does not have to.

The Apple Mac Pro 2013: triangular thermal core, three heat sources radially mounted, one centrifugal blower at the apex. The enclosure was the heat exchanger. Right instinct. Wrong coupling. The lesson is not to abandon the principle — it is to fix the interface.

Q 06 · The Architecture

What does the enclosure actually look like when you get it right?

You have laid out the history, the ceiling, the biology, and the precedent. What does the correct architecture look like? Not as a concept — as a buildable thing.

The correct architecture has three components. They compose — each can be implemented without the others, but together they form a system that has headroom for the next generation and costs no more to build than current approaches.

The first and most consequential change is geometric. In a standard rack, server boards are horizontal — stacked like shelves. Air enters at the front and must travel horizontally across each board to the rear exhaust. Natural convection wants to move vertically. The fan is fighting physics every cubic metre of the way. In the triangular chimney rack, the boards are rotated ninety degrees. The long axis of each board runs floor to ceiling. Three vertical panels arrange into an equilateral triangular prism. Between each panel and the central shaft, air rises freely — natural convection and chimney effect aligned with the channel geometry, not opposed to it. Mixed convection theory confirms this. When buoyancy and forced flow act in the same direction, the heat transfer coefficient at the board surface improves — meaningfully, compared to horizontal forced flow alone. The fan is no longer fighting buoyancy. It is adding to it.

The primary cooling path is unchanged from current practice: liquid cold plates on each GPU package, coolant supplied by a rack-level CDU. In the vertical arrangement, coolant enters at the top of the rack and flows downward through the cold plates by gravity plus pump pressure. In a 2-metre rack with water at 40 °C, the hydrostatic head is 19.46 kPa — 10 to 39 percent of a typical CDU pump pressure range. The pump energy recovered is approximately 70 watts at full rack flow: a secondary credit, not the primary argument. The primary argument for top-to-bottom routing is thermal. Coldest coolant enters at the top where the highest-density GPU packages run. Heat extraction is most efficient at the cold end — the temperature gradient stays steep for longer before any coolant saturation occurs along the board.

The central triangular shaft is the chimney. At rack height — two metres — the stack effect provides 2.85 pascals of driving pressure. That sounds modest until the channel geometry is calculated. Through a 40-millimetre gap between board panel and chimney shaft, 2.85 pascals drives 5.72 metres per second of air velocity. That is nearly double the 3 metres per second a standard server fan is typically sized to produce. In a clean, unobstructed channel, the chimney geometry is sufficient to handle secondary air cooling without a fan. In a real rack with heatsink fins, connectors, and cable harnesses adding resistance, the fan's role becomes pressure reserve against those obstructions — not the primary driver. The chimney supplies the base velocity. The fan compensates for obstruction losses. This is the correct framing: in a well-managed channel geometry, the chimney drives the flow. The fan is the supplement.

The chimney shaft at rack scale is the bottom segment of a building-level conduit. For the chimney to function, the shaft must connect to outside air — either through a building-integrated thermal spine that exits through the roof, or through a pre-cooled supply plenum at the base. The data centre building cannot be a sealed box. Hot air that enters the building and recirculates loses its temperature differential and the chimney stops driving flow. This is a building design requirement, not a rack feature. It does not change the rack geometry — it determines how the rack row is positioned within the building plan. At building chimney height of 10 to 20 metres, the stack effect delivers 14 to 29 pascals and handles 4 to 6 kilowatts per shaft passively — a meaningful base-load contribution across a full row of racks with no moving parts.

The rack frame itself — the triangular aluminium extrusion — is thermally conductive. In contact with surrounding air, it radiates and convects approximately 1.5 kilowatts from its outer surface: free, continuous, no components. If the building infrastructure includes a liquid-cooled wall or floor panel in contact with the rack frame, that contribution scales further. Where the rack exits to the building header is the Thermal Plug interface — the Murray's Law artery. Heat leaves the enclosure at a known temperature and flow rate. The building connects to the district heating network. The chain from GPU die to district heating pipe follows the same branching logic at every transition. The rack is not the end of the story. It is where the story begins.

Murray's Law — 3-Scale Thermal Hierarchy for the Manifold Chassis

The law: d³_parent = Σ d³_daughters · Minimum total energy for fluid transport across a branching network. First derived for vascular systems (Murray, 1926). Applicable to any branching flow network.

Scale 1 — Die to board rail (capillary)

Boards vertical — long axis floor-to-ceiling · air rises through vertical channels alongside board surface · natural convection direction matches channel geometry · GPU packages: liquid cold plates with TIM contact, coolant top-to-bottom gravity-assisted · SSDs / NICs / VRMs: cold rail (conductive contact, no liquid required) · rack frame: passive radiator (~1.5 kW from Al extrusion outer surface)

Scale 2 — Board rail to rack manifold (arteriole)

Quick-disconnect thermal coupling at each board slot · board-level manifold aggregates all component heat into one thermal stream · remove board → coupling breaks clean → manifold continues operating

Scale 3 — Rack manifold to building header (artery)

Rack backbone collects from all slot manifolds · cross-section widens at each bifurcation per Murray's ratios · standardized TP interface at rack base → building liquid header → district heating network

The same branching logic, unbroken, from GPU die to district heating connection.

Closed-loop cooling is not heat reuse

The industry is already moving to closed-loop, direct-to-chip cooling — driven by water scarcity, not by heat. But a closed loop that rejects its heat to the air is just a larger water-cooled PC: it stops consuming water, yet it stays hydraulically closed and thermally open. The heat leaves as warm air. Nothing is exported.

Turning that loop into something a district network can take requires primary-to-secondary loop separation (TCS/FWS) across a heat exchanger, a controlled output temperature, and — often, because single-phase cooling exits at 40–60 °C while most networks supply higher — a chiller or heat pump in between. That engineered handoff is the Thermal Plug. The difficulty is the point: it is why this has to be a standard with a defined interface, not a bespoke build negotiated at every site.

"The enclosure stops being a box and becomes a vascular system. The frame is the heat exchanger. The branches follow Murray's Law. The passive geometry does the work that mechanical systems currently do at significant cost and maintenance burden. This is not a new technology. It is the right geometry — finally applied."

Dimitri Wolf