Evidence report · Technology analysis

The Quad-Core Economy: How AI Orchestrates Robotics, Space and Bitcoin

Evidence report: how AI, robotics, satellite networks and Bitcoin work, who competes, and how an AI agent layer can run all four. 33 sources.

Industry: Technology
Region: Global
Focus: AI orchestration across robotics, space and payments

$15.7T

Upper estimate of what AI adds to the world economy by 2030 (PwC, 2017)

Industrial robots in operation at the end of 2025 (IFR)

20×

Fall in launch cost per kg, Space Shuttle to Falcon 9 (NASA)

21M

Hard cap on Bitcoin supply, set by the protocol

Industry evidence report. Independent analysis of public sources, accessed 11 October 2026. Algorythmos has not been engaged by any organisation named here and builds none of the robotics, space or Bitcoin systems described. The delivery job is an illustrative scenario. Nothing here is financial or investment advice.

Challenge

Artificial intelligence, robotics, the space economy and Bitcoin are usually written about as four separate stories, each with its own headline number. Read apart, they look like four bets. Read together, they describe one stack: software that can reason, machines that can act, a network that reaches anywhere, and a way for machines to settle what they owe. The difficulty is that the popular figures are repeated more often than they are checked. Of fifteen widely repeated figures checked below, three are out of date or wrong, five have no source that supports them, and most of the rest need a caveat.

Solution

This report treats the four as layers of one system and explains each at the level of mechanism: the algorithms, the organisations building them, and the limits that still apply. AI is the control plane. A language-model agent plans work and calls tools [5]; robot fleets carry it out under path-finding and learned control [14, 15]; low-Earth-orbit constellations carry the telemetry where no fibre runs [21]; and a payment network with programmable settlement lets software pay for what it uses [29, 30]. To keep it concrete, one illustrative delivery job is followed through all four layers. It is a teaching device, not a deployment, and no figures are attached to it.

Approach

1
How an engagement would run: we work on the control plane only. Algorythmos does not build robots, satellites or payment networks; we build the agent, data and MLOps layer that directs them. Discovery maps one workflow, the systems it touches and the decisions a person must keep.
2
Agent design. The workflow becomes a planner with a small set of typed tools (a fleet API, a telemetry feed, a payments API), each with permissions, spending limits and an audit log.
3
Evaluation before autonomy. The agent runs against recorded cases and simulated failures (a lost link, a blocked aisle, a rejected payment) until its behaviour is measured, not assumed.
4
Production hardening. Monitoring, rollback and human approval gates go in before any action can move goods or money, and they stay in.

Results

AI adoption is broad and the frontier is contested: 78% of organisations reported using AI in 2024, up from 55%, and US institutions released 40 notable models against 15 from China and three from Europe. The cost of running a GPT-3.5-level system fell more than 280-fold in two years [2].
Robots are a deployed base, not a forecast: 5 million industrial robots were operating at the end of 2025, with more than 600,000 installed that year and 59% of them in China [11]. One operator alone reports over a million mobile robots across 300 sites [13].
Launch cost fell by a factor of 20 between the Space Shuttle ($54,500 per kg) and Falcon 9 ($2,720 per kg), and Falcon Heavy roughly halves it again [18]. By 9 October 2026, 12,988 Starlink satellites had been launched [22].
Bitcoin is best read as settlement infrastructure: a cap of 21 million units enforced by every node [27], blocks secured by proof-of-work [25, 26], and a second layer for small instant payments [29]. Mining drew an estimated 138 TWh a year, 52.4% of it from sustainable sources [32].
The convergence is real at the edges and unproven at the centre. Agents that call tools, fleets coordinated by learned models and laser-linked constellations all exist today [5, 13, 21]. An agent paying for services over Lightning exists as a protocol, L402, and not yet as common practice [30].

Planning an agent that has to act in the physical world or move money? Talk to us about a four-week design and evaluation sprint for the control layer.

Book a Consultation

1. Artificial intelligence: the planner

Where our job is now: A customer order arrives. The agent reads it, checks stock and delivery windows through its tools, and writes a plan: pick, load, track, release, settle.

How it works

Modern AI systems are built on the transformer, a neural network that uses attention to weigh every part of its input against every other part [3]. A large language model is a transformer trained to predict the next token of text. On its own it only writes. It becomes an agent when it is put in a loop: the model reasons about a goal, chooses a tool, reads what the tool returned, and reasons again until the goal is met [5]. Retrieval gives it facts it was never trained on, by fetching relevant documents and placing them in the prompt [4]. The same family of models now works on molecules: AlphaFold predicts a protein’s structure from its sequence [6], and diffusion models design new proteins to order [7].

Who competes

The frontier is concentrated. In 2024, US institutions produced 40 notable AI models, China 15 and Europe three, and US private AI investment reached $109.1 billion [2]. Competition runs on three fronts at once: model quality, the chips and data centres that train the models, and price, where the cost of a fixed level of capability fell more than 280-fold between November 2022 and October 2024 [2].

Limits and open problems

Models still state false things with confidence, so production systems need evaluation sets, monitoring and human approval on consequential actions. The economics are projections: the $15.7 trillion figure is an upper estimate made in 2017 [1]. In drug discovery the bottleneck is the clinic, not the design. A new drug costs about $2.6 billion once failures and the cost of capital are counted [8], only 7.9% of candidates entering Phase I reach approval [9], and the first drug with an AI-discovered target and an AI-designed molecule has so far reported a 71-patient Phase 2a trial [10].

The agent loop

How a language model turns a goal into finished work. Schematic.

1A goal arrives in plain language.
2The model writes a plan: a short list of steps.
3It calls a tool for the first step, such as a search or an API.
4The tool returns a result.
5The model reads it, ticks off the step and revises the plan.
6When every step is done, it returns the answer.

2. Advanced robotics: the hands

Where our job is now: The plan reaches the warehouse. Two mobile robots are assigned the pick, and both need the same single-lane aisle at the same moment.

How it works

A robot runs a sense-plan-act cycle many times a second. Cameras and lidar feed a map of the surroundings and of the robot’s place in it, a technique called simultaneous localisation and mapping. A planner chooses a path, and a controller turns it into motor commands while correcting for error. Two things changed recently. The first is fleets: when hundreds of robots share a floor, the problem is no longer one robot’s path but everyone’s at once. Multi-agent path finding searches for a set of routes in which no two robots claim the same cell at the same time [14]. The second is learning: vision-language-action models take a camera image and an instruction and output robot actions directly, carrying knowledge from web-scale training into physical tasks [15], and policies trained in randomised simulation transfer to real hardware [16].

Who competes

Factories installed more than 600,000 industrial robots in 2025. China accounted for 59% of them, and the United States moved into second place ahead of Japan [11]. In logistics the scale leader is Amazon, which reports over a million robots across more than 300 facilities and a foundation model, DeepFleet, that it says cuts fleet travel time by 10% [13]. Cheaper batteries widen the field: average lithium-ion pack prices reached a record low of $108 per kWh in 2025 [17].

Limits and open problems

Structured floors are well understood; unstructured ones are not. General-purpose humanoids are far newer than the industrial arms and mobile robots counted above, and this report found no independent deployment statistics for them. Vendor productivity claims are rarely measured against a stated baseline, which is why this report uses only figures that an operator or a statistics body has published under its own name.

Fleet path finding

Two robots, one single-lane aisle. Schematic of conflict-based search.

1Two robots receive picks; both must reach the dock.
2Each robot’s shortest route is planned on its own.
3The planner finds a conflict: both routes need the same cell at the same time.
4A constraint is added. One robot waits; the other goes through.
5The second robot follows once the cell is free.
6Both reach the dock with no collision and one short wait.

3. The space economy: the network

Where our job is now: The truck leaves mobile coverage. Its position and the condition of the load now travel by satellite.

How it works

Two engineering advances changed the economics of orbit. The first is the reusable booster. Landing one is an optimisation problem solved on board in real time: find a thrust profile that brings the vehicle to the pad at zero velocity without exceeding engine limits or running out of propellant. The problem is not convex as posed, but it can be transformed into one that is, so a solver is guaranteed to find the best answer quickly [19, 20]. The second is the mass-produced satellite with laser links. Light travels faster in a vacuum than in glass fibre, so a constellation that passes data from satellite to satellite can beat terrestrial fibre over long distances and can serve places no fibre reaches [21].

Who competes

Launch is led by SpaceX, whose Falcon 9 lists at $62 million for 22,800 kg to low Earth orbit, or $2,720 per kg, against about $54,500 per kg for the Space Shuttle [18]. In broadband, Starlink had launched 12,988 satellites by 9 October 2026, of which 11,137 were in orbit [22]. The World Economic Forum and McKinsey put the whole space economy at $630 billion in 2023 and project $1.8 trillion by 2035, with growth driven by communications, positioning and Earth-observation services [23].

Limits and open problems

A list price per kilogram assumes a full rocket; small payloads pay more [18]. Coverage is not the same as adoption: 2.2 billion people were still offline in 2025 [24]. Launch prices below $200 per kg are a design goal for vehicles still in development, not a published price.

Launch, land, link

A reusable booster delivers a satellite; a packet then crosses the constellation by laser. Schematic, not to scale.

1Lift-off: booster and upper stage climb together.
2Separation: the upper stage continues to orbit; the booster turns back.
3Guided landing: the booster solves its descent path in flight and lands on the pad.
4The new satellite joins the constellation.
5A packet goes up from a user terminal and hops between satellites by laser.
6It comes down at a ground station near its destination.

4. Bitcoin: the settlement layer

Where our job is now: The goods are delivered. The agent now owes small sums to several services: a fleet operator, a satellite link, a data provider.

How it works

Bitcoin is a ledger that thousands of independent computers keep in agreement without a central operator [25]. Transactions are grouped into blocks. Each block header contains a fingerprint of the previous block and a Merkle root summarising its transactions, so changing any past entry would change every fingerprint after it [26]. To add a block, a miner must find a number, the nonce, that makes the SHA-256 hash of the header fall below a target. There is no shortcut: miners try values until one works, which is what makes the work costly to produce and trivial to verify [25, 26]. The network retunes the target every 2,016 blocks to hold the average interval near ten minutes, and the reward for each block halves every 210,000 blocks, which is what caps supply at 21 million [26, 27]. The base layer is slow by design. The Lightning Network adds payment channels on top: two parties lock funds once, then exchange signed updates instantly, and settle on the chain only when they close [29].

Who competes

As settlement technology, Bitcoin competes with card networks, correspondent banking and other ledgers, and trades speed for independence from any single operator. Its position in regulated finance changed on 10 January 2024, when the US Securities and Exchange Commission approved the listing of spot bitcoin exchange-traded products while stating that the approval was not an endorsement [31]. Mining has become an energy business: a 2025 Cambridge survey covering 48% of network activity estimated annual use at 138 TWh, with 52.4% from sustainable sources and natural gas the largest single source at 38.2% [32]. In Texas, large mining sites are among the loads that agree to reduce consumption when the grid is short [33].

Limits and open problems

The base chain adds a block only about every ten minutes [26], so everyday machine payments depend on second layers that are far less proven. Flexible demand from miners is voluntary, not guaranteed [33]. This report describes the technology and gives no view on bitcoin as an investment.

Mining a block

Proof-of-work in six steps. Schematic: the meter stands for the hash value.

1Waiting transactions are summarised into a Merkle root.
2The miner picks a nonce and hashes the header. The hash is above the target: rejected.
3Another nonce, another hash. Still above the target.
4A nonce is found whose hash falls below the target.
5The block is chained to the previous one, and other nodes verify it with a single hash.
6Every 2,016 blocks the network moves the target to keep blocks about ten minutes apart.

5. Orchestration: how AI runs the other three

Put the four layers together and the pattern is a control loop. The agent holds the goal and the plan. Everything else is a tool it calls: a fleet manager, a telemetry feed, a payment channel. The engineering work is in the interfaces, the limits and the audit trail. Lightning already has a protocol for the last step: L402 lets a server ask for a small payment in its HTTP response and lets a program pay and retry without an account or a card [30]. That makes machine-to-machine payment technically possible. Whether it becomes common is an open question.

1
Plan. The agent breaks the job into tasks and decides which tool handles each [5].
2
Dispatch. Tasks go to the fleet manager, which solves the path-finding problem and reports progress [14].
3
Observe. Telemetry returns over whichever link is available, including satellite where there is no terrestrial network [21].
4
Decide. The agent compares what happened with the plan and replans when they differ.
5
Approve. A person signs off any step that moves goods or money above an agreed limit.
6
Settle. Payments are released against confirmed events, and every action is logged.

One job, four layers

The illustrative delivery job from start to finish.

1An order reaches the AI agent, which plans the job.
2The agent dispatches the pick to the robot fleet.
3Robots load the goods; telemetry goes up to the satellite network.
4The satellite link relays status back to the agent.
5A person approves the release.
6Payment settles and the whole run is logged.

Launch cost to low Earth orbit

Thousand US dollars per kilogram, in 2018 dollars, at full payload. Source [18].

Notable AI models released in 2024, by region

Number of notable models, by the home of the institution that produced them. Source [2].

The world online and offline, 2025

Billions of people. Source [24].

Electricity sources for Bitcoin mining

Share of electricity by source, from a 2025 survey covering 48% of network activity. Source [32].

Claims checked against sources

Figures that circulate widely in this field, each set against its primary source. Where no source supports a figure, this report does not use it.

The claim	What the source says	Verdict
“AI will add $15 trillion to the economy by 2030.”	PwC estimated up to $15.7 trillion, in 2016 prices, in a study published in 2017 [1].	Close, but an upper estimate from 2017
“A new drug costs $2.5 billion to develop.”	$2,558 million including failures and the cost of capital; $1,395 million out of pocket (2013 dollars) [8].	Right, if the cost of capital is counted
“More than 90% of drug candidates fail in clinical trials.”	7.9% of candidates entering Phase I were approved over 2011–2020 [9].	Supported
“Generative AI cuts four to five years of drug discovery to four to six months.”	No primary source gives this as a general result. The best-documented AI-designed drug has reported a 71-patient Phase 2a trial [10].	Not supported as a general rule
“3.5 million industrial robots are deployed.”	4,663,698 in operation in 2024 [12]; 5 million at the end of 2025 [11].	Out of date
“Battery costs fell 40% in five years.”	Average pack price of $108 per kWh in 2025, a record low, down 8% on 2024 [17].	Restated with the survey figure
“Humanoid robots will cost $20,000.”	A stated target for a future product, not a price paid.	Left out
“Robotic warehouses lift pick rates 300% and cut unit cost 40%.”	No attributable source found. Amazon reports a 10% cut in fleet travel time from its routing model [13].	Left out; a smaller, attributed figure used
“The Space Shuttle cost $10,000 per kg to orbit.”	About $54,500 per kg at the time, or $61,700 in 2018 dollars [18].	Understated about five-fold
“Reusable rockets launch for $1,500 per kg.”	Falcon 9 lists at $2,720 per kg; Falcon Heavy at about $1,400 [18].	True of Falcon Heavy only
“Next-generation rockets will launch for under $200 per kg.”	A design goal for vehicles in development; no published price.	Left out
“The space economy exceeds $1 trillion.”	$630 billion in 2023, projected to reach $1.8 trillion by 2035 [23].	A forecast, not a measurement
“One-third of the world is offline.”	2.2 billion people offline in 2025; 29% of the world in 2024 [24].	Out of date; now nearer a quarter
“Bitcoin supply is capped at 21 million.”	The block reward halves every 210,000 blocks, which bounds total supply just under 21 million [27].	Correct
“Companies holding Bitcoin have outperformed their peers.”	An investment claim. This report gives no investment view.	Left out

Frequently asked questions

What is the quad-core economy?

It is a way of describing four technologies as layers of one system: artificial intelligence plans, robotics acts, satellite networks connect, and Bitcoin settles. The term comes from strategy writing, not from a standard. This report uses it as a frame and checks each layer against cited sources.

How do AI agents work?

An AI agent is a language model placed in a loop. It reasons about a goal, calls a tool such as a search engine or an API, reads the result, and repeats until the goal is met. Retrieval supplies facts the model was not trained on, while permissions, evaluation and human approval keep its actions within limits.

How do warehouse robots avoid colliding?

A fleet manager plans every robot’s route together. Multi-agent path-finding algorithms such as conflict-based search plan each robot’s path, detect where two robots would occupy the same cell at the same time, and replan with a constraint that makes one of them wait or go around.

How much does it cost to launch a kilogram into orbit?

According to a NASA analysis, the Space Shuttle cost about $54,500 per kilogram to low Earth orbit, a Falcon 9 lists at about $2,720 per kilogram and a Falcon Heavy at about $1,400. Those figures assume a full payload, so small satellites pay more per kilogram.

How does low-Earth-orbit satellite internet work?

Thousands of satellites orbit a few hundred kilometres up, close enough for low delay. A user terminal steers a beam to a passing satellite, which forwards the data by laser to neighbouring satellites and down to a ground station near the destination. Routes are recalculated continuously because the satellites are always moving.

How does Bitcoin proof-of-work work?

Miners assemble transactions into a block and repeatedly hash the block header with a different nonce until the result falls below a target set by the network. Finding such a hash takes an enormous number of attempts, but checking it takes one. The target adjusts every 2,016 blocks to keep blocks about ten minutes apart.

Can AI agents make payments on their own?

Technically, yes. The L402 protocol lets a web service request a small Lightning payment and lets a program pay and continue without a card or an account. It is an emerging approach, not common practice, and any production system should cap what an agent can spend and require human approval above that cap.

Does Algorythmos build robots, satellites or Bitcoin infrastructure?

No. Algorythmos designs and operates the AI layer: agents, data pipelines, evaluation and MLOps. This report is an independent analysis of public sources and is general information, not financial or investment advice.

Sources

More Research & Evidence Reports

All case studies

The Quad-Core Economy: How AI Orchestrates Robotics, Space and Bitcoin

Challenge

Solution

Approach

1. Artificial intelligence: the planner

How it works

Who competes

Limits and open problems

The agent loop

2. Advanced robotics: the hands

How it works

Who competes

Limits and open problems

Fleet path finding

3. The space economy: the network

How it works

Who competes

Limits and open problems

Launch, land, link

4. Bitcoin: the settlement layer

How it works

Who competes

Limits and open problems

Mining a block

5. Orchestration: how AI runs the other three

One job, four layers

Launch cost to low Earth orbit

Notable AI models released in 2024, by region

The world online and offline, 2025

Electricity sources for Bitcoin mining

Claims checked against sources

Frequently asked questions

Sources

More Research & Evidence Reports

Keep exploring