AI is advancing at an extraordinary pace.
Hundreds of billions of dollars are being invested across the AI ecosystem, and the demand for more investment continues to grow. AI is becoming pervasive, touching almost every aspect of our personal and professional lives.
Competition is intense. Development cycles are accelerating. At the same time, some industry leaders and policymakers are increasingly calling for appropriate safeguards and regulatory frameworks.
But there is another question that deserves more attention:
Are we getting proportional societal value from the enormous investment and effort going into AI?
I believe there may be a systemic reason why the answer is not always yes.
AI is a chain, not a single technology
Investment is happening across the entire AI value chain covering Data centers, storage and cloud infrastructure, Energy and communications, Chips and semiconductors, Foundation models, Data infrastructure, Engineering and coding, Cybersecurity, Robotics and industrial AI, Autonomous systems, Business-specific applications, AI skills and capability building, and Governance and regulatory frameworks
A significant proportion of investment is currently concentrated in infrastructure and foundational capabilities.
Yet, at the same time, we repeatedly encounter situations such as:
- AI is hungry, but the grid isn't ready.
- The pilot is successful, but the data isn't ready.
- The AI agent is ready, but cybersecurity and guardrails aren't ready.
- The capital is available, but the required compute or chips aren't available.
- The application is ready, but the regulatory framework isn't ready.
- The technology is ready, but the organization isn't ready to redesign its processes around it.
Each of these represents a mismatch somewhere in the AI value chain. And that mismatch can limit the value that the entire system can deliver.
Could this be a Theory of Constraints problem?
The Theory of Constraints (TOC) by Goldratt is based on a relatively simple principle: The throughput of a system is constrained by its most limiting constraint.
Once the constraint moves, we identify the next one.
I believe this thinking can be applied to the AI ecosystem.
But there is one important difference. A manufacturing system typically optimizes throughput for a business. For AI, the ultimate objective should be broader i.e. Maximize value AI creates for humanity
Applying TOC to AI
Classic TOC-- AI ecosystem interpretation
Identify the constraint-- Find what is currently preventing AI from creating more societal value
Exploit the constraint-- Get maximum utilization from the existing scarce capacity
Subordinate everything else-- Align other resources and activities around the constraint
Elevate the constraint-- Invest, innovate or change policy to increase the constraint's capacity
Repeat-- Once the constraint moves, identify the next constraint
What could the constraints be?
The constraints can exist at different points in the AI ecosystem.
- Capital: Is there sufficient financing for infrastructure, applications and deployment?
- Compute: Are sufficient GPUs, accelerators and computing resources available?
- Energy: Can electricity generation and availability keep pace with AI infrastructure?
- Physical infrastructure: Are data centers, grids, cooling systems and networks available?
- Data: Is the required data available, accessible, high-quality and usable?
- Talent: Do we have enough AI engineers and domain experts?
- Organizational capacity: Can organizations redesign processes quickly enough to take advantage of AI?
- Economics: Does the cost of AI deployment make economic sense relative to the benefits?
- Human adoption: Do employees, customers and citizens have the capability and willingness to use AI effectively?
- Problem selection: Are we applying enormous AI capabilities to the problems that can create the greatest value?
The constraint doesn't have to be technological.
In fact, the most important constraint may sometimes be organizational, economic, regulatory or human.
Let's consider two examples
1. Electricity availability
Imagine that the AI industry can build 100 GW of additional compute capacity, but the power system can support only 60 GW.
The additional 40 GW of compute cannot generate its expected value until the energy constraint is addressed.
In this situation:
Compute is not the binding constraint; Electricity availability is.
Investing more money in compute without addressing the energy constraint would therefore increase capacity without proportionally increasing societal throughput.
The same principle applies to data centers.
A data center may be constructed in 2–3 years, while energy generation, transmission and grid connection can require substantially longer planning and development cycles.
Therefore:
Building more data centers does not necessarily solve an AI capacity problem if the energy system cannot support them.
2. Capital allocation
Consider a hypothetical situation in which there are 10,000 potentially valuable AI applications, but only 1,000 have access to adequate capital.
Capital becomes a constraint.
But simply adding more capital isn't necessarily the answer. The more important question becomes:
Where will the next dollar of capital create the greatest increase in societal AI throughput?
Should it go toward: another foundation model or AI infrastructure or energy infrastructure or healthcare AI or education AI or cybersecurity or industrial AI or AI agents?
The answer should depend on which investment removes a binding constraint and enables the greatest additional societal value.
The multiplier effect of constraint removal
Some constraints may be more important than others because removing one constraint can unlock several others.
For example: Energy infrastructure enables Data-center capacity which enables Compute availability which enables AI applications which enables Enterprise adoption which creates economic and social value
In such a situation, investment in the original constraint can have a multiplying effect throughout the system.
This leads to an interesting question:
Should we measure investments not only by their direct return, but also by the additional AI capacity and societal value they unlock elsewhere in the system? I believe we should.
Constraint prioritization becomes critical
The AI ecosystem will almost certainly have multiple constraints simultaneously. Therefore, we need to prioritize them.
One possible approach is to evaluate each constraint based on:
- How much additional throughput will its removal create?
- How early is it in the AI value chain?
- How many downstream constraints will it unlock?
- How much investment is required to remove it?
- How quickly can it be removed?
And the constraint will keep moving
This may be one of the most important aspects of applying TOC to AI.
Today's constraint may be Compute, tomorrow it could be Energy then it could be Talent. Therefore, AI policy & action cannot be a one-time exercise. Constraint identification & removal needs to become a continuous process.
What role should governments and enterprises play?
I believe both have an important but different role.
Governments and Policy institutions
Governments can address constraints that affect entire ecosystems through:
- Infrastructure investment
- Energy and grid policy
- Education and workforce development
- Incentives
- Data policy
- Standards
- Regulatory frameworks
- Public-private partnerships
- Research and development
- International cooperation
The objective should not necessarily be for governments to pick individual AI winners. Instead, governments can help identify and remove systemic constraints that affect many participants simultaneously.
Enterprises
Enterprises should apply the same thinking internally.
Their constraints may be:
- Data availability
- Legacy systems
- Talent
- Capital
- Process design
- Integration
- Cybersecurity
- Change management
- AI governance
- Employee adoption
The question becomes:
What is the single most important constraint preventing our organization from obtaining more value from AI?
Rather than implementing AI everywhere simultaneously, organizations could focus their transformation efforts around that constraint.
The bigger question
AI capability may not ultimately be limited by the intelligence of the models.
It may be limited by our ability to build the infrastructure, energy systems, data ecosystems, organizations, skills, policies and human capabilities needed to convert that intelligence into value.
Perhaps the next AI challenge isn't simply building more AI.
Perhaps it is identifying and removing the constraints that prevent AI from delivering its full potential to society.
This is a concept I am currently developing, and I would genuinely value perspectives from people working across AI, infrastructure, investment, enterprise transformation, government policy and organizational change.
What do you think is the biggest constraint preventing AI from delivering greater value to society today?
No comments:
Post a Comment