One of the most famous unsolved questions in computer science is P versus NP.
At its core, the question asks whether every problem whose solution can be checked efficiently can also be solved efficiently.
Imagine a difficult Sudoku puzzle.
Finding the correct solution may take hours.
Checking a completed solution may take only a few minutes.
Or imagine a password.
Finding the correct password among billions of possible combinations may require enormous effort.
Checking whether a proposed password is correct takes almost no time.
This distinction between finding and checking sits at the centre of P versus NP.
P describes problems that can be solved efficiently.
NP describes problems for which a proposed solution can be verified efficiently.
The open question is whether these two categories are actually the same.
Does every problem that is easy to check also have a hidden method that makes it easy to solve?
No one knows.
Most computer scientists believe that P does not equal NP, but no proof has been accepted.
This article does not attempt to solve P versus NP.
Instead, it begins with a different question.
What if the practical difficulty of a computational problem is not determined by mathematics alone?
What if it also depends on the resources, infrastructure, capital, energy, incentives, and competitive behaviour of the people and organisations trying to solve it?
That leads to a broader idea.
Computational complexity may have more than one dimension.
The usual explanation of P versus NP depends heavily on time.
One problem takes a short time to verify.
Another takes a much longer time to solve.
The difference in required time is treated as evidence of computational difficulty.
But time is not experienced equally across history.
A task that required a room full of machines and trained operators in the 1960s may now be completed by a phone in milliseconds.
Searching through a large set of records once required expensive mainframes, punch cards, human operators, physical storage, and significant institutional resources.
Today, a person can search millions of records using a web browser.
The problem did not necessarily change.
The infrastructure changed.
The speed of processors changed.
The cost of storage changed.
The availability of software changed.
The cost of computation collapsed.
This raises an important distinction.
A problem can remain mathematically difficult while becoming economically easy.
It can also remain mathematically simple while becoming economically difficult because of limited access to hardware, energy, data, or capital.
Classical complexity theory measures how computational requirements grow as a problem becomes larger.
That remains essential.
A faster machine does not automatically turn an exponential algorithm into a polynomial one.
If a problem requires approximately 2n operations, then making a computer one billion times faster only changes where the problem becomes impossible.
It does not change the shape of the curve.
However, businesses, governments, and societies do not operate in asymptotic space.
They operate in budgets, deadlines, power constraints, supply chains, labour markets, and competitive environments.
For them, the relevant question is not only whether a problem is theoretically difficult.
The relevant question is whether it is affordable, accessible, and strategically solvable under current conditions.
This suggests that computational difficulty should be viewed through at least two lenses.
Mathematical complexity refers to the formal structure of a problem.
It asks how the number of required computational steps grows as the input size increases.
Sorting a list may require roughly n log n operations.
Searching every possible combination in a difficult optimisation problem may require exponential growth.
This form of complexity is largely independent of the price of electricity, the cost of GPUs, or the amount of capital available to a company.
It belongs to mathematics and theoretical computer science.
Economic complexity refers to the real cost of solving a problem under existing technological and economic conditions.
It asks questions such as:
How much compute is required?
How much energy is available?
What hardware exists?
How expensive is that hardware?
How much capital can be deployed?
Is the required data accessible?
Does the organisation possess the necessary expertise?
Can the solution be delivered within a useful timeframe?
Two organisations can face the same mathematical problem and experience completely different levels of economic complexity.
A frontier AI laboratory with access to hundreds of thousands of accelerators, large datasets, specialised researchers, and billions of dollars faces a different problem from a small company using rented cloud infrastructure.
The mathematics may be identical.
The effective difficulty is not.
The central proposal of this article is that effective computational complexity can be represented as a function of mathematical difficulty, economic constraints, and strategic behaviour.
This is not presented as a theorem.
It is a conceptual model for understanding how computational problems behave in the real world.
The model begins with:
Ce = f(M, E, S)
Where:
Ce represents effective computational complexity.
M represents mathematical complexity.
E represents economic complexity.
S represents strategic complexity.
Mathematical complexity describes the formal computational difficulty of the problem.
Economic complexity describes the cost and availability of the resources required to solve it.
Strategic complexity describes how competition between actors changes those costs over time.
A simplified multiplicative version can also be written as:
Ce = M × E × S
This form is useful because it shows that high difficulty in one dimension can amplify difficulty in the others.
A mathematically difficult problem combined with scarce hardware and intense competition may become practically unreachable.
A mathematically difficult problem combined with abundant compute, cheap power, strong algorithms, and shared knowledge may become commercially manageable.
Economic complexity can be decomposed further:
E = g(H, P, K, D, A, T)
Where:
H represents hardware capability and availability.
P represents power and energy availability.
K represents capital.
D represents data.
A represents algorithmic knowledge.
T represents time constraints.
Organisational capability can also be added:
E = g(H, P, K, D, A, T, O)
Where O represents organisational capability, including talent, coordination, management, execution, and institutional knowledge.
This addition matters because access to resources does not guarantee effective use.
A company may own hardware but lack the engineering talent to use it.
It may possess data but lack governance.
It may have capital but lack strategic clarity.
It may have algorithms but lack distribution.
Economic complexity is therefore not simply the market price of compute.
It is the total cost of converting resources into a useful solution.
The model becomes more powerful when game theory is introduced.
Computational problems are not solved in isolation.
They are solved by people, companies, governments, research laboratories, and alliances.
These actors have different incentives.
They compete.
They cooperate.
They conceal information.
They share research.
They subsidise infrastructure.
They restrict access.
They build standards.
They create monopolies.
They attack each other's cost structures.
This means the economic complexity of a problem is not fixed.
It is partly produced by the strategic behaviour of the players involved.
Consider two companies trying to solve the same computational problem.
Company A invests heavily in specialised chips.
Company B responds by building a more efficient algorithm.
Company A signs long-term energy contracts.
Company B partners with a cloud provider.
Company A acquires proprietary data.
Company B releases an open-source model to attract developers.
Each move changes the cost of solving the problem for both companies.
The problem becomes a repeated game.
The players do not simply solve the problem.
They reshape the environment in which the problem is solved.
This can be represented conceptually as:
St+1 = h(St, It, Rt, Ct, Xt)
Where:
St is the strategic state of the system at time t.
It represents investment decisions.
Rt represents rival responses.
Ct represents cooperation, alliances, and coordination.
Xt represents external shocks, including regulation, geopolitics, supply disruptions, and technological breakthroughs.
The next state of the system is produced by the current state and the actions of the players.
Complexity becomes endogenous.
The system changes itself.
The original P versus NP distinction can also be reinterpreted through game theory.
One player searches for a solution.
Another player verifies it.
But verification is not always neutral.
The verifier may have incentives.
The verifier may reject a valid solution.
The verifier may lack complete information.
The verifier may demand a costly proof.
The verifier may create standards that favour one type of solution over another.
The solver may decide how much evidence to reveal.
The solver may hide part of the method to protect a competitive advantage.
The solver may produce a solution that is correct but expensive to verify.
The verifier may use sampling rather than complete inspection.
Once real institutions are involved, finding and checking become strategic activities.
This appears in many areas.
A pharmaceutical company develops a drug.
A regulator verifies safety and effectiveness.
A software company develops a system.
An auditor verifies its security.
A startup develops an AI model.
An enterprise verifies whether it is accurate, compliant, and useful.
A government develops a policy.
Voters, courts, economists, and agencies assess whether it works.
In each case, the cost of finding a solution and the cost of checking it are shaped by incentives, institutions, trust, and information asymmetry.
Artificial intelligence is one of the clearest examples of the model.
Training a frontier model is not constrained by mathematics alone.
It requires accelerators.
It requires electricity.
It requires cooling.
It requires data centres.
It requires networking.
It requires large datasets.
It requires specialised talent.
It requires capital.
It requires access to semiconductor supply chains.
It requires time.
The effective complexity of building a frontier model is therefore much larger than its formal algorithmic complexity.
It is an infrastructure problem.
It is a capital allocation problem.
It is an energy problem.
It is a geopolitical problem.
It is a game between rival companies and states.
One company secures GPUs.
Another secures power.
Another develops a more efficient architecture.
Another releases an open-source model.
Another subsidises inference.
Another controls distribution.
Each player is trying to reduce its own economic complexity while increasing the complexity faced by competitors.
This is one reason AI leadership cannot be understood by model quality alone.
The winning actor may not possess the best algorithm.
It may possess the lowest effective complexity.
It may have cheaper power, better hardware access, stronger distribution, more capital, or a better ecosystem.
NVIDIA provides another useful example.
The company does not need to change the mathematics of every workload it serves.
It reduces the economic complexity of running those workloads.
Specialised hardware reduces processing time.
Software tools reduce engineering effort.
Networking reduces communication bottlenecks.
Integrated systems reduce deployment complexity.
A mature developer ecosystem reduces organisational cost.
NVIDIA therefore lowers effective complexity for customers while raising strategic complexity for competitors.
Its ecosystem creates dependency.
Its scale improves its economics.
Its software creates switching costs.
Its infrastructure shapes the choices available to other players.
The competitive advantage is not simply faster chips.
It is control over the economic conditions under which computation occurs.
Protein folding shows how effective complexity can fall rapidly.
The biological problem remained difficult.
The physical structure of proteins did not suddenly become simpler.
But better algorithms, larger datasets, specialised hardware, and sustained investment changed what could be achieved.
The mathematical and scientific difficulty remained significant.
The economic cost of producing useful predictions declined.
A problem that once required years of laboratory work could increasingly be approached through computation.
This does not mean the problem became easy in an absolute sense.
It means the frontier of practical solvability moved.
Cybersecurity demonstrates the strategic dimension even more clearly.
An attacker searches for a vulnerability.
A defender tries to verify that a system is secure.
The attacker may need to find only one weakness.
The defender must search across many possible weaknesses.
Attackers and defenders invest in different tools.
They respond to each other.
Automation reduces the cost of attack.
AI may reduce the cost of vulnerability discovery.
Improved verification tools may reduce the cost of defence.
Regulation can raise compliance costs.
Shared threat intelligence can reduce collective complexity.
The difficulty of the security problem changes as the players change their strategies.
This is not a static computational task.
It is an adversarial game.
Many logistics problems are mathematically difficult.
Scheduling vehicles, warehouses, workers, and inventory can require enormous search spaces.
But the practical difficulty depends on the quality of data, sensors, software, compute, and coordination.
A company with poor data may face a much harder problem than a company with real-time visibility.
A company with fragmented systems may face a higher organisational complexity than a competitor using an integrated platform.
A company may not need to solve the mathematically perfect optimisation problem.
It may only need a good solution faster than competitors.
This reveals another important point.
Business advantage often comes from relative complexity, not absolute complexity.
A company does not need to make a hard problem universally easy.
It only needs to make the problem cheaper for itself than for others.
This leads to the idea of relative computational advantage.
A company gains relative computational advantage when its effective cost of solving an important problem is lower than that of its competitors.
This can be represented as:
RCAi = C̄e - Ce,i
Where:
RCAi represents the relative computational advantage of organisation i.
C̄e represents the average effective complexity faced by the market.
Ce,i represents the effective complexity faced by organisation i.
If an organisation can solve a problem more cheaply, more quickly, or more reliably than the market, it possesses a strategic advantage.
That advantage can come from many sources.
Better algorithms.
Cheaper energy.
Exclusive data.
Specialised hardware.
Strong partnerships.
Faster decision-making.
Better talent.
More efficient organisational design.
Greater access to capital.
Lower regulatory burden.
Stronger distribution.
The future competitive landscape may therefore be shaped by organisations that systematically reduce effective complexity faster than others.
This framework can help leaders make better decisions.
Most organisations treat technology problems too narrowly.
They ask whether the software works.
They ask whether an AI model is accurate.
They ask whether they should buy more compute.
But they often fail to identify which layer of complexity is actually constraining them.
The model forces a more disciplined diagnosis.
Is the main constraint mathematical?
Is the algorithm itself inadequate?
Is the problem economic?
Is compute too expensive?
Is power unavailable?
Is the data inaccessible?
Is the organisation too slow?
Is the real problem strategic?
Will competitors gain access to the same technology?
Could a supplier create dependency?
Could regulation change the payoff?
Could an open-source alternative collapse the market price?
Could a rival secure a critical resource first?
By separating these dimensions, leaders can avoid spending money on the wrong constraint.
Organisations could apply the model through a recurring complexity audit.
For every strategically important problem, leaders could assess:
1. Mathematical complexity
What is the formal structure of the problem?
What algorithms currently exist?
Where are the known limitations?
Is the target exact optimisation, approximation, prediction, or decision support?
2. Hardware complexity
What compute is required?
Is specialised hardware necessary?
Is capacity available?
What are the supply risks?
3. Energy complexity
How much power is required?
What is the energy cost?
Can the organisation secure stable access?
4. Capital complexity
How much investment is required?
How long before the investment generates value?
Can the organisation sustain the cost?
5. Data complexity
What data is needed?
Who owns it?
Is it reliable?
Can it legally be used?
6. Organisational complexity
Does the organisation possess the required talent?
Can teams coordinate?
Can decisions be made quickly?
Can the solution be integrated into existing operations?
7. Strategic complexity
Who else is solving the same problem?
What are their incentives?
Which resources are becoming scarce?
Which players can change the economics of the market?
Where are the chokepoints?
Where are the dependencies?
What happens if the cost of computation falls by 90 percent?
What happens if it rises by 500 percent?
This audit can help organisations determine whether they are facing a technology problem, an economic problem, or a strategic problem.
The model suggests several ways future companies can build durable advantage.
Companies should continuously reduce the cost of solving their most important problems.
This may involve better algorithms, automation, improved data, specialised hardware, stronger infrastructure, or simpler organisational processes.
The goal is not merely to adopt more technology.
The goal is to lower the total cost of converting computation into business value.
A company that solves a problem once gains limited value.
A company that turns the solution into a reusable platform lowers future complexity.
Reusable data pipelines, models, tools, APIs, and decision systems can spread the cost of solving a problem across many use cases.
This creates compounding advantage.
Companies can gain strategic power by controlling a resource that reduces complexity for themselves and increases it for others.
This may be data.
It may be distribution.
It may be energy.
It may be compute.
It may be software standards.
It may be a developer ecosystem.
It may be customer trust.
Control over a critical input can matter more than owning the best algorithm.
Leaders should identify problems whose economic complexity is falling rapidly.
A task that is too expensive today may become commercially viable within a few years.
Companies that recognise the shift early can build capabilities before competitors understand the opportunity.
This is especially relevant in AI, robotics, drug discovery, energy systems, simulation, autonomous systems, and advanced manufacturing.
Some business models depend on a problem remaining expensive.
That can be dangerous.
If the cost of computation collapses, the value of the intermediary may disappear.
Companies should ask whether their advantage survives cheaper compute, better algorithms, open-source alternatives, and new infrastructure.
The future cost of computation will not be determined by technology alone.
It will be shaped by the decisions of chipmakers, cloud providers, energy companies, governments, research laboratories, and capital markets.
Leaders should monitor the actors capable of changing the cost structure of an entire industry.
A new chip architecture can reduce compute costs.
A new export restriction can increase them.
A power shortage can delay deployment.
An open-source model can destroy pricing power.
A regulatory standard can raise verification costs.
A major infrastructure investment can create a new geographic advantage.
The next forty years are likely to be shaped by a continuous expansion of computational capability.
More problems will move from human labour to software.
More decisions will be delegated to models.
More scientific discovery will depend on simulation.
More economic power will concentrate around energy, chips, data, and infrastructure.
The central strategic question will not simply be who has the best technology.
It will be who can reduce effective complexity fastest.
In the coming decades, companies may compete on their ability to solve increasingly difficult problems at lower cost.
Governments may compete on access to energy, semiconductor capacity, and technical talent.
Research laboratories may compete on the cost of experimentation.
Healthcare organisations may compete on the cost of diagnosis and drug development.
Manufacturers may compete on the cost of simulation and automation.
Financial institutions may compete on the speed and quality of risk analysis.
Military power may increasingly depend on the cost of intelligence, coordination, verification, and autonomous decision-making.
As these systems evolve, the boundary between software and infrastructure will become less meaningful.
Intelligence will increasingly depend on energy systems, data centres, supply chains, capital markets, and industrial policy.
The economic model of computational complexity provides a way to interpret that shift.
It explains why theoretically difficult problems can become commercially routine.
It explains why seemingly advanced companies can fall behind despite having strong algorithms.
It explains why energy, hardware, data, and organisation may matter as much as software.
It explains why strategic competition can change the cost of computation for everyone.
It also explains why the winners of the next era may not be those who solve the hardest problems in an absolute sense.
They may be those who make hard problems cheaper for themselves and more expensive for everyone else.
P versus NP asks whether problems that are easy to verify are also easy to solve.
That question remains unresolved.
But the real world introduces another layer.
A problem's practical difficulty is not determined by mathematics alone.
It is shaped by hardware, energy, capital, data, time, organisations, incentives, and strategic competition.
This leads to a broader framework:
Effective computational complexity is the combined product of mathematical difficulty, economic constraints, and strategic behaviour.
The model does not replace classical complexity theory.
It extends the conversation into economics, management, infrastructure, and game theory.
It asks not only whether a problem can be solved efficiently in principle.
It asks who can afford to solve it.
Who can solve it first.
Who can verify it.
Who controls the resources required.
Who changes the cost for everyone else.
And who builds a lasting advantage from that change.
That may be one of the most important strategic questions of the next forty years.