The Economics of AI: Big Tech is doing it wrong

The current AI development model—characterized by massive proprietary foundation models requiring exponential resource investment—faces fundamental economic and physical constraints that make it unsustainable.

In other words, the current "arms race" approach will eventually lead to the vast sums of capital, currently being spent on NVIDIA GPUs and data centres, drying up due to the lack of return on investment.

Similarly, given the land, energy and water requirements to run these AI powerhouses, there are literal physical constraints on how capacity can be built out, aggravated at least in the short-term by climate change policies (correctly in my view, but that's a different story).

The paradox is that I see an increasingly large population of people using AI, albeit with varying complexity - some quite superficially for simple search or idea generation, helping developers and analysts write code more efficiently, or more complex multi-agent orchestrations.

I think we have reached a critical mass meaning people would perceive a deterioration in the quality of life were none of this technology to exist. Pandora's box is open - there is no going back.

There is a gap here though. Market forces will inevitably fill this void, finding the path of least resistance. But what does that look like? That is what I discuss below.

In summary, a more viable future architecture emerges: open-source foundation models serving as commoditized infrastructure, coordinated by standardized agentic systems, with competition focused on specialized models and secure data integration tools.

Companies are spending vast sums of money on AI, without a clear economic model to get a return on the investment

The Unsustainable Status Quo

Resource Constraints Are Real and Binding

The AI industry's current trajectory assumes unlimited access to critical resources, but physical reality is imposing hard limits:

Land Scarcity: Prime data centre locations near fibre infrastructure, power grids, and cooling resources are increasingly scarce. Northern Virginia, a major hub, is approaching saturation. Zoning battles intensify as communities resist massive facilities that strain local infrastructure.

Energy Bottlenecks: Multiple major AI projects face delays due to grid capacity constraints. Data centres may soon account for 10-15% of local grid demand in some regions. Utilities cannot build new power plants fast enough to meet demand, creating competition between AI development, electrification efforts, and existing industrial needs.

Water Stress: Microsoft's water usage increased 34% in one year, largely attributed to AI operations. Facilities in drought-prone areas face restrictions, competing directly with agriculture and municipal water supplies as chip densities and cooling requirements intensify.

Infrastructure Cascades: The constraints compound through supply chains—semiconductor fabs hit limits on specialized materials, skilled technician shortages emerge for building and maintaining facilities, and bottlenecks develop for transformers, cooling systems, and backup generators.

Diseconomies of Scale

Unlike traditional software, AI exhibits diseconomies of scale due to resource competition:

• Exponential Resource Requirements: Each new model generation requires exponentially more compute, not just linear increases

• Resource Bidding Wars: As demand grows, companies bid up prices for prime locations, energy access, and specialized talent

• Geographic Dispersion: Companies may be forced into suboptimal locations simply to access available power and cooling

• Talent Inflation: AI engineers command large compensation packages as demand outstrips supply

This represents a fundamental departure from traditional tech economics, where marginal costs decreased with scale.

The Quality Measurement Problem

AI development faces a unique challenge: there is no objective measure of quality or accuracy. Unlike traditional software with clear performance metrics, AI outputs exist in subjective space where "better" depends entirely on context and use-case.

Swiss Army knife approach

Given the "Swiss army knife" approach for the big models, balancing complex domain and technical knowledge with creativity, you can easily test for some of that but not everything. I recognise that all sorts of stats do exist, which frankly I understand only at the level of "big number = good, small number = bad", but my point is that it will always be subjective based on the question and context.

"Temperature" settings allow for creativity versus consistency trade-offs, which makes it impossible to objectively justify the exponential cost increases of larger models. Hallucinations are a feature of creativity, not a bug.

You wait, just give it more time...

The standard argument is that "it's this clever now, just imagine how clever it will be in x years time". But we are not dealing with Moore's Law here. It is becoming more art than science, reflecting the complexity of human intelligence.

Anecdotally, people's experience is that "I use X for this type of work and Y for another" - they literally sense that one model is better at some things than others in ways they perceive but don't (and will never) understand. Or they may say "hmm, did something change?" as there is a subtle shift when algorithms are tweaked.

Why pursue perfection?

So if the pursuit of perfection is pointless, what happens to this current arms-race we have amongst the leading players? They all want to create the "best" model, to out-run the competition who then wither and die, leaving them a monopoly or oligopoly to exploit.

In this case, the big players all have deep pockets, and most have large independent income streams. I just don't see how you end up with a "winner", without enormous value destruction. I would point people to the work of Ed Zitron who has written extensively on the economics of AI for a fuller discussion. I don't endorse all the views and opinions but they are thought-provoking.

So...a recap of where are at right now

1. An arms race destroying capital in pursuit of unattainable model "perfection"

2. Physical constraints impacting the ability to scale, with tangible real world consequences

3. A growing user-base integrating AI into their lives

So how COULD the economic model be reorganised to be more sustainable and efficient.

A Sustainable Alternative Architecture

I want this article to be as accessible as possible, so in very simple terms:

1. A single open-source generally-capable AI model.

2. A competitive market place for small, specialized models, focused on very clear aspects like say, a single programming language, or creative output.

3. A standard open-source agent architecture - think browser extensions or Excel add-ins - so users can bring in the extra smarts in a standardised way. The generally capable AI model is that glue that binds the context and individual model activities.

4. A big Enterprise issue is proprietary data and data security, so the ability to allow local domain information to be analysed and queried using the AI smarts but without losing control of this locally (or say, private cloud), is critical.

When I started writing this article, I was taken back to my University days in the late 1990s. The internet was there but nascent. The Netscape web browser was how you interacted with the web. However, it raised immediate questions - how do I know what is on the web? Is there a glossary for this thing?

Of course, Google search became the answer to the problem, which spawned a whole new advertising income stream and to some extent a whole new industry.

It worked because it was free to use and made life easier for users to get what they wanted. Nobody questioned the implicit cost of doing business (our personal data).

Users want to ask questions or get AI to "do stuff" and get the best quality (I won't say "right") answers/output, preferably for free, but at least being able to evaluate cost vs benefit properly.

Companies have a commercial imperative to maximise profit. I have no problem with that, but the way they are going about it now is a massive misuse of resources.

Three-Layer Model

Layer 1: Open-Source Foundation Models (Commoditized Infrastructure)

• Models like Llama and Mistral already approach GPT-4 class performance

• Shared development costs across the community eliminate individual company resource burdens

• Can be fine-tuned locally on proprietary data without external data sharing

• Functions like infrastructure (similar to Linux) rather than competitive advantage

• Funding similar to the Linux Foundation, or PBS (notwithstanding some issues), which should also ensure that content creators receive compensation for their intellectual property (lawsuits pending).

Layer 2: Specialized Models (Competitive Marketplace)

• Small, targeted models for specific domains (coding, creative work, analysis)

• Measurable quality metrics within narrow scopes

• Lower development costs and barriers to entry

• Pricing reflects a charge back to the Foundation

• Subscription-based licensing model, probably tiered (free, standard user, enterprise)

• Competition drives innovation in specific applications.

• As above, the use of specific intellectual property will need to be licensed individually or centrally (where Representative Groups exist)

Layer 3: Agentic Orchestration Systems (Platform Value)

• Standardized, open-source coordination systems that route tasks between models

• Creates network effects and platform value while remaining accessible

• Enables quality control at the workflow level rather than individual model level

• Allows optimization of the entire system rather than individual components

Supporting Infrastructure: Secure Data Integration

A critical fourth component enables enterprises to leverage proprietary data:

• Local fine-tuning capabilities

• Retrieval-augmented generation systems

• Secure enclaves for sensitive data processing

• Tools that provide "good enough" results without compromising data security

I don't see this as far fetched. In fact, a version of this is the relatively defensive position that many large Companies have adopted. They have taken base models, hosted them locally, fed them proprietary data and given access to the new composite to the local user-base. Some have gone further deploying agents to undertake specific workloads or deployed 3rd party applications.

The only difference is that every company is doing it in their own way. It's chaotic and inefficient.

Economic Viability Analysis

Why This Model Works

Resource Efficiency: Eliminates duplicate foundation model development, dramatically reducing aggregate resource consumption while maintaining capability.

Clear Value Propositions: Each layer has measurable value and sustainable economics:

• Foundation models: Shared infrastructure costs

• Agentic systems: Platform network effects

• Specialized models: Clear quality metrics and competitive differentiation

• Data integration: Enterprise necessity with measurable ROI

Market Dynamics: Creates healthy competition where it matters (specialized applications) while commoditizing infrastructure that benefits everyone.

Scalability: Avoids the resource constraint ceiling by distributing rather than concentrating compute requirements.

Funding

The Hybrid Model:

• Government contribution: Treats the open-source foundation model as a public good, like funding basic research or maintaining the internet infrastructure

• Corporate contribution: Companies pay into the fund based on their commercial usage/revenue derived from the model

• Graduated commercial licensing: The more you commercialize, the more you contribute back to the commons

This actually mirrors how the internet developed - government funded the initial infrastructure (ARPANET, NSF), then commercial players built on top while contributing back through taxes and infrastructure investments.

Practical mechanics:

• Foundation model training costs get pooled across government grants + corporate consortium fees

• Companies pay based on revenue tiers from AI usage (like how they pay different rates for AWS based on usage)

• Content creators get compensated from this shared pool for training usage

• But inference usage (when someone actually queries the model) still triggers direct micropayments

The political appeal:

• Government can justify it as maintaining technological sovereignty/competitiveness

• Companies get predictable infrastructure costs instead of individual licensing negotiations

• Content creators get sustainable funding without depending on any single company

• Citizens benefit from having a "national AI utility" that's not controlled by Big Tech

I Never Said This Would Be Easy

Of course, I am not naive about this. I recognise that there are enormous vested interests, both internally within the US, for example, but also the wider geo-politics (US vs China) are perhaps insurmountable, certainly at this point in time.

Incumbent Resistance: Companies with massive sunk costs in proprietary foundation models will resist commoditization of their primary differentiator.

Coordination Problems: Standardizing agentic systems requires industry cooperation and technical consensus.

Quality Assurance: Ensuring reliable performance across heterogeneous specialized models requires sophisticated orchestration.

Business Model Evolution: Companies must shift from "model ownership" to "application excellence" business models.

Regulation/Ethics Frameworks: Supra-national policies may lead to different interpretations of the "truth" (China), or shared understanding of what is deemed ethical.

Market Implications

Investment Opportunities

High-Value Targets:

1. Agentic orchestration platforms (potential for platform dominance)

2. Specialized model development tools and marketplaces

3. Secure enterprise data integration solutions

4. Infrastructure optimization for distributed AI workloads

Declining Value:

1. Proprietary foundation model development

2. Massive centralized compute infrastructure

3. General-purpose AI platforms without specialization

Competitive Landscape Shift

The competitive advantage shifts from "biggest model" to:

• Best specialized applications

• Most effective agentic coordination

• Strongest enterprise data integration

• Most efficient resource utilization

Timeline and Catalysts

Only a brave man or a fool would talk in terms of timelines, but being practical, only so much can be achieved given the size of the global players, the politics and the enormous sums involved. A very loose timeline might be something like this.

Near-term (1-2 years): Resource constraints force some companies to reconsider pure scale strategies. Open-source models achieve "good enough" performance for most use cases.

Medium-term (2-5 years): Agentic systems mature and standardize. Specialized model marketplaces emerge. Enterprise data integration becomes a major market.

Long-term (5+ years): The three-layer architecture becomes dominant. Competition focuses on applications rather than infrastructure.

Conclusion

The current AI development model is economically and physically unsustainable. Resource constraints will force the industry toward a more distributed, specialized architecture that separates infrastructure (commoditized) from applications (competitive).

There will have to be large-scale re-evaluation of the fundamental economic model for AI. It may be completely different, but perhaps this article offers some food-for-thought.