KN Capital (back to top) Get in touch

KN Capital · Investing in the autonomous economy

The economy is learning to run itself.

KN Capital invests across the value chain of the autonomous economy: from the agents and robots at the surface, down to the silicon, materials, and energy that power them.

Scroll to descend How KN invests

The autonomous economy runs on two forces.

Artificial intelligence is automating cognitive work. Physical autonomy is automating physical work. Each rests on a stack of layers that make it possible, and both stacks stand on the same foundation of materials and energy. That is the map, and the layers are where KN Capital invests.

The autonomous economy as one mapA V shaped map. Down the left arm, the Bits stack: agents, models, AI factories, silicon, cooling. Down the right arm, the Atoms stack: the arenas; the robots; robot brains; data, simulation and safety; silicon for the machine; hands and senses; actuators. On both arms the interludes sit as faint marks between the bets. Both arms meet at a shared base of advanced materials and energy.BITSARTIFICIAL INTELLIGENCEATOMSPHYSICAL AUTONOMYTHE SURFACE01 AGENTS02 MODELS03 AI FACTORIES04 SILICON05 COOLING01 THE ARENAS02 THE ROBOTS03 ROBOT BRAINS04 DATA, SIMULATION AND SAFETY05 SILICON FOR THE MACHINE06 HANDS AND SENSES07 ACTUATORSB1 · ADVANCED MATERIALSB2 · ENERGY
The autonomous economy as one mapA V shaped map. Down the left arm, the Bits stack: agents, models, AI factories, silicon, cooling. Down the right arm, the Atoms stack: the arenas; the robots; robot brains; data, simulation and safety; silicon for the machine; hands and senses; actuators. On both arms the interludes sit as faint marks between the bets. Both arms meet at a shared base of advanced materials and energy.BITSAIATOMSPHYSICALSURFACE010203040501020304050607B1 · MATERIALSB2 · ENERGY
Bits · Artificial Intelligence Knowledge work is moving to AI. Descend the Bits stack Atoms · Physical Autonomy Physical work is moving to machines. Descend the Atoms stack

The Bits stackArtificial IntelligenceDepths 01 to 05

BitsThe surface

Knowledge work is moving to AI.

Intelligence gets an order of magnitude cheaper every year, and what gets that cheap gets used for everything. Agents that started by writing code now run whole workflows, and the work they can finish alone grows longer every few months. Each task they take over moves spend from salaries to software, and on to the tokens it runs on. At work, people move from doing the task to directing the agents that do it; at home, where most AI use already is, agents move from answering questions to running errands.

Who keeps the value when AI does the work?

Bits 01Agents

Put agents to work, build them, or enable them.

Agents cut the cost of work and open new revenue. The gain lands in three places, and very unevenly.

  1. Apply the agents.The largest share goes to whoever owns the work, and stays there while competitors are slow to move. Integrators earn a fee for the change. The most direct way to own it is to own the business.
  2. Build the agents.Labs and incumbents own the general purpose agent and set the pace. New entrants win in depth, one industry or one function at a time, and on distribution: their own engineers, partners who deploy for them, or users who pull them in.
  3. Serve the agents.The newest pool: the rails agents need. Inside companies: access, permissions, a runtime for long jobs, a record of every action. Between them: identity, payment and who pays when it goes wrong. Incumbents take the simple ones; neutral rails priced on work pull away.

Agents run on tokens. Where do tokens come from?

BitsWhere tokens come from

Intelligence has a supply chain.

Every agent runs on tokens, and burns through them. Reasoning models think before they answer, and an agent uses many times the tokens of a chat. Every token comes from a model running in a factory. The model is a fixed cost bet: built once at great expense and paid back across every user, so scale wins, and its moat is still moving. The factory is a capital business: every answer costs it compute, so it earns on how full it runs and how cheaply it borrows, and its rules are already clear.

First, the model.

The tokens behind one requestThree clusters of cyan dots, one dot for each token, growing left to right. A chat: a handful of dots in a soft glow, one answer. A reasoning model: about ten times the dots, because it thinks before it answers. An agent: the large cluster, because it works through the whole task. Demand compounds with every task handed over.THE TOKENS BEHIND ONE REQUESTDEMAND COMPOUNDS WITH EVERY TASK HANDED OVERA CHATONE ANSWERA REASONING MODELTHINKS BEFORE IT ANSWERSAN AGENTWORKS THROUGH THE WHOLE TASK
The tokens behind one requestThree clusters of cyan dots, one dot for each token, growing left to right. A chat: a handful of dots in a soft glow, one answer. A reasoning model: about ten times the dots, because it thinks before it answers. An agent: the large cluster, because it works through the whole task. Demand compounds with every task handed over.THE TOKENS BEHIND ONE REQUESTDEMAND COMPOUNDS WITH EVERYTASK HANDED OVERA CHATONE ANSWERA REASONING MODELTHINKS BEFORE IT ANSWERSAN AGENTWORKS THROUGH THE WHOLE TASK

Bits 02Models

The best model wins, until the next one.

Models are not a commodity yet, and may never be. In production, buyers still pay for the best. But the lead keeps changing hands, and holding it has a price.

Where value accrues

  1. The frontier labs.Ahead, and paid for it. Open weights close the gap on paper, not in production. Staying ahead is a capital race only a few can fund, and the clouds and chipmakers that fund it take their share twice: as owners, and as suppliers of the compute it buys.
  2. The neolabs.A different approach: a new way to reason, world models, a model built with its own hardware. Most will not hold a moat. One that works can reset the race.
  3. The data suppliers.Expert data for post training is in demand, but it is a service. Labs switch suppliers within days, and synthetic data and their own pipelines erode it. The moat does not compound.

Every model has to be run somewhere.

Buyers follow the best

And the best keeps changing

Buyers follow the bestA spotlight, what buyers run, falls on the best model, lit in ember on a stage line, with ember particles falling through the light. Past leaders sit to the left as dim outlines, fading out. Open weights, a cyan ring, sits just outside the light: close on paper. In a loop the next model slides in from the right, and every third time a new approach rises from below; the light moves to it, the old best dims, open weights catches up, and the row glides left. The best keeps changing.WHAT BUYERS RUNTHE BESTFOR NOWOPEN WEIGHTSCLOSE ON PAPERTHE NEXT ONE
Buyers follow the bestA spotlight, what buyers run, falls on the best model, lit in ember on a stage line, with ember particles falling through the light. Past leaders sit to the left as dim outlines, fading out. Open weights, a cyan ring, sits just outside the light: close on paper. In a loop the next model slides in from the right, and every third time a new approach rises from below; the light moves to it, the old best dims, open weights catches up, and the row glides left. The best keeps changing.WHAT BUYERS RUNTHE BESTFOR NOWOPEN WEIGHTSCLOSE ON PAPERTHE NEXT ONE

Bits 03AI factories

The token rush is short on factories.

An AI factory is power, a site, chips, and the software that serves the output. Chips are still scarce, but more and more they arrive before the site can run them: grid connections take years, and transformers and turbines are booked out.

Where value accrues

  1. Start from power.The strongest factories begin with the scarcest input: a grid connection secured early, or power made on site. Energy first builders, former miners and nations with power to spare hold today’s bottleneck.
  2. Run the chips.Neoclouds lease the site and borrow for the chips. Scarcity pays them today, but they rent the bottleneck, so their rents last as long as the shortage, however long that is.
  3. Sell the token.More tokens from the same chip is new supply. The serving layer routes each job to the chip that runs it best, keeps the fleet full, and sells the fastest token at a premium.

To see where the bottleneck moves next, look inside inference.

BitsInside inference

One request, two bottlenecks.

Training builds a model once. Inference runs it every time anyone, or any agent, asks, and it is now most of the compute in a factory. Every request has two phases with opposite appetites. Prefill reads the whole prompt at once and needs compute. Decode writes the answer one token at a time, rereading the model from memory at every step, and needs memory, the input in shortest supply today. Run both on one chip and much of it sits idle. Split them, and size each to its own bottleneck.

Opposite appetites mean different chips. Which silicon wins?

One request, two bottlenecksThree panels. Prefill reads every prompt token at once: compute saturated, bandwidth idle. Decode writes one token per step, rereading the model from HBM each time: bandwidth saturated, compute idle, and HBM is marked short. Then the split: eighteen chips in a prefill pool and a decode pool. For chat, four chips prefill and fourteen decode; for long context, ten prefill and eight decode, and on a switch only the six chips that change pool move across. Size each pool to its own bottleneck.PREFILLCOMPUTE · SATURATEDBANDWIDTH · IDLEREADS THE WHOLE PROMPT · COMPUTE BOUNDALL TOKENS AT ONCEDECODECOMPUTE · IDLEBANDWIDTH · SATURATEDONE TOKEN PER STEP · MEMORY BOUNDHBM · REREAD EVERY STEPSHORTSPLIT THEMWORKLOAD · CHATWORKLOAD · LONG CONTEXTPREFILL POOLCOMPUTE HEAVYDECODE POOLMEMORY HEAVYSIZE EACH POOL TO ITS OWN BOTTLENECK
One request, two bottlenecksThree panels. Prefill reads every prompt token at once: compute saturated, bandwidth idle. Decode writes one token per step, rereading the model from HBM each time: bandwidth saturated, compute idle, and HBM is marked short. Then the split: eighteen chips in a prefill pool and a decode pool. For chat, four chips prefill and fourteen decode; for long context, ten prefill and eight decode, and on a switch only the six chips that change pool move across. Size each pool to its own bottleneck.PREFILLCOMPUTE · SATURATEDBANDWIDTH · IDLEREADS THE WHOLE PROMPT · COMPUTE BOUNDALL TOKENS AT ONCEDECODECOMPUTE · IDLEBANDWIDTH · SATURATEDONE TOKEN PER STEP · MEMORY BOUNDHBM · REREAD EVERY STEPSHORTSPLIT THEMWORKLOAD · CHATWORKLOAD · LONG CONTEXTPREFILL POOLCOMPUTE HEAVYDECODE POOLMEMORY HEAVYSIZE EACH POOL TO ITS OWN BOTTLENECK

Bits 04Silicon

When the workload stabilizes, the chip specializes.

Every chip is a bet that its workload holds still. As AI splits into distinct workloads, a chip built for the job beats a general one on tokens per dollar per watt.

Where value accrues

  1. General purpose silicon.Still most of the money, as most work still moves, and its software moat is as deep as the hardware. The leader now builds chips for single phases of inference too.
  2. The hyperscalers’ own chips.The largest clouds design their own, and the few partners that turn those designs into silicon earn tens of billions a year.
  3. The specialists.Chips for one workload, video, voice, long context, or a model wired into the silicon. Each bets its workload holds still; the giants pay for the winners.
  4. New substrates.Computing with light, analog circuits or living neurons, for far less energy per operation. Long dated, but one that ships resets the cost of a token.

Silicon for the machine itself is on the Atoms stack.

Why not simply build faster chips?

One chip for every workload, or a chip for eachTop, a general purpose chip: a balance triangle equal on compute, memory and network, hatched ember inside. A dashed ember triangle traces the workload now running, and the hatched part outside it sits idle for that workload. Beside it: balanced for every job, so part of it idles on any one. Below, four specific chips, each a solid triangle shaped for its workload: video, compute heavy; voice, instant answers; long context, memory heavy; one model family, the model in silicon. In a loop the workload moves from video to voice to long context to one model family: the dashed triangle morphs to its shape, a dashed arrow runs from the general chip to the matching specific chip, and that chip lights up. Each is a bet that its workload holds still.GENERAL PURPOSE · ONE CHIP, EVERY WORKLOADCOMPUTENETWORKMEMORYHATCHED · IDLE FOR THIS WORKLOADBALANCED FOR EVERY JOB,SO PART OF IT IDLESON ANY ONENOW RUNNINGVIDEOVIDEOCOMPUTE HEAVYVOICEINSTANT ANSWERSLONG CONTEXTMEMORY HEAVYONE MODEL FAMILYMODEL IN SILICONSPECIFIC · A CHIP FOR EACH WORKLOADEACH IS A BET THAT ITS WORKLOAD HOLDS STILL
One chip for every workload, or a chip for eachTop, a general purpose chip: a balance triangle equal on compute, memory and network, hatched ember inside. A dashed ember triangle traces the workload now running, and the hatched part outside it sits idle for that workload. Beside it: balanced for every job, so part of it idles on any one. Below, four specific chips, each a solid triangle shaped for its workload: video, compute heavy; voice, instant answers; long context, memory heavy; one model family, the model in silicon. In a loop the workload moves from video to voice to long context to one model family: the dashed triangle morphs to its shape, a dashed arrow runs from the general chip to the matching specific chip, and that chip lights up. Each is a bet that its workload holds still.GENERAL PURPOSE · ONE CHIP, EVERY WORKLOADCOMPUTENETWORKMEMORYHATCHED · IDLE FOR THIS WORKLOADBALANCED FOR EVERY JOB,SO PART OF IT IDLESON ANY ONENOW RUNNINGVIDEOVIDEOCOMPUTE HEAVYVOICEINSTANT ANSWERSLONG CONTEXTMEMORY HEAVYONE MODEL FAMILYMODEL IN SILICONSPECIFIC · A CHIP FOR EACH WORKLOADEACH IS A BET THAT ITS WORKLOAD HOLDS STILL

BitsInside the chip

Inside the chip, the bottleneck moves too.

A chip is compute, a small fast SRAM on the die, a larger and slower HBM beside it, and the links to other chips. Compute has grown far faster than the memory that feeds it, so the wait moved from the math to the data.

The three walls

  1. Wall 1Wall 1 · Memory HBM cannot feed compute fast enough and holds too little, so decode waits and large models are split across chips. Three companies make it, all sold out, and it is now one of the largest costs in an accelerator.
  2. Wall 2Wall 2 · Packaging Memory only helps if it sits right beside the compute, joined in one package. That advanced packaging is a bottleneck of its own, held mostly by the leading foundry and booked ahead.
  3. Wall 3Wall 3 · Interconnect Splitting a model moves the wait to the links between chips. Copper reaches across a rack; beyond it, the answer is light, and the lasers and optics that carry it come from a few makers at volume.

The chip itself comes from one leading edge foundry, using machines from one lithography supplier, on wafers and gases from the base of this map.

Every way through these walls packs more into less space. What does that cost?

Inside the chip, the bottleneck movesAnatomy of a GPU package: a compute die whose cores flicker ember, with SRAM on the die, HBM stacks on both sides joined to it by links that cyan data pulses run along, links out of the package to other GPUs and the cluster, and cooling underneath, fed by wafers and gases from the base. Advanced packaging puts HBM beside compute; HBM has three makers and is sold out; the die comes from one foundry with one lithography supplier. A red glow marks where the chip waits, and it moves in a loop. Wall 1, memory: the pulses queue at the die and the cores dim, waiting. Wall 2, packaging: the stacks fill, and blocks of the model spill out to other GPUs, split across chips. Wall 3, interconnect: the pulses queue at the edge of the package. Next, wall 4, heat: the cooling box glows. Each fix exposes the next wall.SRAMWAITINGGPU PACKAGEADVANCED PACKAGING PUTS HBM BESIDE COMPUTECOMPUTE · FLOPSHBMHBMSTACKED DRAMTHREE MAKERSSOLD OUTONE FOUNDRYONE LITHOGRAPHY SUPPLIERINTERCONNECT+GPUSCLUSTERCOOLINGWAFERS · GASES → B1WALL 1 · MEMORYEACH FIX EXPOSES THE NEXT WALL
Inside the chip, the bottleneck movesAnatomy of a GPU package: a compute die whose cores flicker ember, with SRAM on the die, HBM stacks on both sides joined to it by links that cyan data pulses run along, links out of the package to other GPUs and the cluster, and cooling underneath, fed by wafers and gases from the base. Advanced packaging puts HBM beside compute; HBM has three makers and is sold out; the die comes from one foundry with one lithography supplier. A red glow marks where the chip waits, and it moves in a loop. Wall 1, memory: the pulses queue at the die and the cores dim, waiting. Wall 2, packaging: the stacks fill, and blocks of the model spill out to other GPUs, split across chips. Wall 3, interconnect: the pulses queue at the edge of the package. Next, wall 4, heat: the cooling box glows. Each fix exposes the next wall.SRAMWAITINGGPU PACKAGEADVANCED PACKAGING PUTSHBM BESIDE COMPUTECOMPUTE · FLOPSHBMHBMONE FOUNDRYONE LITHOGRAPHY SUPPLIERSTACKED DRAMTHREE MAKERSSOLD OUTINTERCONNECT+GPUSCLUSTERCOOLINGWAFERS · GASES → B1WALL 1 · MEMORYEACH FIX EXPOSES THE NEXT WALL

Bits 05Cooling

Cooling moves into the chip.

Every generation packs more power into less space, and a chip that runs hot slows itself down, wastes power and wears out sooner. On the same chips, liquid cooling alone gave up to a sixth more throughput than air.

Where value accrues

  1. Built for each chip.As chips near two kilowatts, standard cold plates struggle: coarse channels cool every spot the same. The next plates have channels a fraction of a millimetre wide, shaped to each chip’s hot spots, and the step after etches them into the silicon. Design plus precision manufacturing does not commoditize, and buyers pay billions for it.
  2. Materials that carry heat.Lab grown diamond spreads heat about five times faster than copper and is already fitted to GPUs; silicon purified to a single isotope conducts heat better than the natural mix. Early and unproven at scale, but each raises how hard a chip can run.

Every chip comes down to materials and energy. So does every robot.

Cooling moves into the chipAn exploded three dimensional view of an AI chip on its package, with memory alongside and heat glowing at three hot spots, looping through five steps that each bring cooling closer to the chip. A standard cooling plate high above with coarse channels, cooling every part the same: the chip is slowed. A plate built for this chip sits closer, its channels shaped to the hot spots. Lab grown diamond slides in, bonded on top of the chip, and spreads the heat out. A magnified view of the silicon shows heavier atoms scattering a wave of heat until it is purified to one isotope and the heat flows. Finally the plate comes down into the chip, and the channels are built into the chip itself. Heat rises across the gap while there is one, and a bar of work done rises at each step. A chip runs only as hard as it can be cooled.COOLING PLATESHAPED TO THE HOT SPOTSLAB GROWN DIAMONDON TOP OF THE CHIPTHE CHIPHOT SPOTSPACKAGEMEMORY ALONGSIDELAB GROWN DIAMONDBONDED TO THE CHIP · SPREADS THE HEAT OUTEARLYWORK DONEFASTERA CHIP RUNS ONLY AS HARD AS IT CAN BE COOLED
Cooling moves into the chipAn exploded three dimensional view of an AI chip on its package, with memory alongside and heat glowing at three hot spots, looping through five steps that each bring cooling closer to the chip. A standard cooling plate high above with coarse channels, cooling every part the same: the chip is slowed. A plate built for this chip sits closer, its channels shaped to the hot spots. Lab grown diamond slides in, bonded on top of the chip, and spreads the heat out. A magnified view of the silicon shows heavier atoms scattering a wave of heat until it is purified to one isotope and the heat flows. Finally the plate comes down into the chip, and the channels are built into the chip itself. Heat rises across the gap while there is one, and a bar of work done rises at each step. A chip runs only as hard as it can be cooled.COOLING PLATESHAPED TO THEHOT SPOTSLAB GROWN DIAMONDON TOP OF THE CHIPTHE CHIPHOT SPOTSPACKAGEMEMORY ALONGSIDELAB GROWN DIAMONDBONDED TO THE CHIP · SPREADS THE HEAT OUTEARLYWORK DONEFASTERA CHIP RUNS ONLY AS HARD AS IT CAN BE COOLED

The crossing

Both stacks bottom out in the same place. Cross to the other stack first, or continue down to the foundation.

Cross to Atoms Continue to the base

The baseB1 · Advanced materials

Every token and every joint needs critical materials.

Rare earth magnets move robots, lasers link AI racks, and chips are built from high purity metals and gases. The ore is found in many places, but the processing is concentrated, and access to it is now a lever in trade.

Where value accrues

  1. Process it elsewhere.The work was dirty and paid too little, so over decades most of it moved to one country, which supported it. Now buyers pay a premium for the scarcest materials from anywhere else, and governments back new plants with equity, price floors and long contracts.
  2. Recover before you mine.Much of what is needed already exists: chip metals left in the waste of working refineries and smelters, magnets inside retired motors and servers. Recovering them is usually faster and cheaper than opening a new mine.
  3. Make what comes next.The next chips and robots will call for materials few can make yet, like glass core substrates for the largest AI chips and magnets with little or no rare earth. They take years to qualify, so the first approved supplier often stays designed in.

The autonomous economy can only grow as fast as its energy.

The bedrockB2 · Energy

Energy is the slowest layer to build.

Electricity demand is rising fast, from AI data centers, new factories and fleets going electric. In the US and Europe, power plants and grid connections often take five years or more, longer than the data centers they serve, and turbines and transformers are booked years ahead.

Where value accrues

  1. Power at the load.Generation built beside the data center or the plant, so the load can switch on before the grid reaches it. Gas carries it today.
  2. Build reactors like products.Small reactors built on a factory line and added one unit at a time to match the load, from a single data center to a whole campus. The first test reactors of the new designs have gone critical. Designs that run on today’s fuel and reach series production first will likely set the cost for the rest.
  3. Make the fuel at home.Most advanced designs need fuel richer than today’s plants burn, and only two countries can make it at scale. Many reactors risk being ready before their fuel is, so whoever enriches it outside those two countries unlocks them.
  4. Store and firm.Solar and batteries take about a year to build once they clear the grid queue, and make up most new capacity. Batteries carry that power into the evening. Storage that lasts for days is only now reaching its first projects.

The autonomous economy turns energy into work, cognitive and physical.

How KN invests

KN Capital looks for where value will accrue next.

Every wave of technology moves value along its supply chain. KN Capital’s work is to understand that chain well enough to see where the value will settle, before the market agrees on it.

KN Capital starts with one question: if a technology works, who keeps the economics? From there it follows the constraints as they shift from one layer to the next, and looks for the new layers forming beneath the obvious ones. Much of what it backs is still proving its science or engineering, so it goes deep enough to judge it for itself.

The map that follows is how KN Capital does that work. It changes as the chain does.

The autonomous economy as one mapA V shaped map. Down the left arm, the Bits stack: agents, models, AI factories, silicon, cooling. Down the right arm, the Atoms stack: the arenas; the robots; robot brains; data, simulation and safety; silicon for the machine; hands and senses; actuators. On both arms the interludes sit as faint marks between the bets. Both arms meet at a shared base of advanced materials and energy.BITSARTIFICIAL INTELLIGENCEATOMSPHYSICAL AUTONOMYTHE SURFACE01 AGENTS02 MODELS03 AI FACTORIES04 SILICON05 COOLING01 THE ARENAS02 THE ROBOTS03 ROBOT BRAINS04 DATA, SIMULATION AND SAFETY05 SILICON FOR THE MACHINE06 HANDS AND SENSES07 ACTUATORSB1 · ADVANCED MATERIALSB2 · ENERGY
The autonomous economy as one mapA V shaped map. Down the left arm, the Bits stack: agents, models, AI factories, silicon, cooling. Down the right arm, the Atoms stack: the arenas; the robots; robot brains; data, simulation and safety; silicon for the machine; hands and senses; actuators. On both arms the interludes sit as faint marks between the bets. Both arms meet at a shared base of advanced materials and energy.BITSAIATOMSPHYSICALSURFACE010203040501020304050607B1 · MATERIALSB2 · ENERGY

Where KN invests

From the surface to the bedrock.

The autonomous economy runs on two stacks: AI doing cognitive work, machines doing physical work. Both stand on the same materials and energy.

KN Capital invests across all of it: in the agents and robots at the surface, and in the layers beneath the obvious that they run on.

Hover any layer to see what KN Capital backs there

Contact

Start a conversation.

Building, investing or thinking about the same problems?

I am