A family in Maryland opens their monthly utility bill. It has jumped — not gradually, not with warning, but in a single billing cycle. The state grid is buckling under the weight of a new artificial intelligence (AI) data centre built three states away. The economic benefits flow out of state. The infrastructure cost stays.
A chief technology officer in London reviews a quarterly cloud invoice. The cost per token has collapsed by sixty percent. The total spend has doubled. Cheaper intelligence spawned more agents — reasoning, verifying, looping — most of them invisible to the humans who authorised the original workflow.
A policymaker in Geneva reads the latest United Nations (UN) environmental assessment. Four hundred and forty-eight trillion watt hours. Two hundred and eight million tonnes of carbon dioxide. More electricity than all but ten countries. More water than the annual needs of 1.3 billion people. The numbers are no longer projections. They are measurements.
Three people. Three cities. The same crisis — viewed from three angles that have yet to converge.
In human nutrition, there is a principle championed by Dr Michael Greger: what is good for our health is also good for the planet. The diet that reverses chronic disease — whole food, plant-based, minimal processing — is the same diet that minimises agricultural emissions, water use, and land degradation. Not two interventions. One.
The same convergence applies to the machine.
The industry has fractured the AI conversation into silos. Environmentalists talk about sustainability — the carbon, the water, the grid strain. Enterprise leaders talk about sovereignty — the risk of a provider disappearing overnight, the dependency on infrastructure they do not control, the cost of infrastructure they cannot predict. Regulators talk about national security.
Three conversations. Zero overlap. The architecture that makes an organisation sovereign is the same architecture that stops it burning a city’s worth of electricity to parse a mere spreadsheet.
The Jevons problem
When a resource becomes cheaper, consumption does not hold steady. It explodes. William Stanley Jevons documented this with coal in 1865. The pattern has a name — the Jevons Paradox — and it is now the defining dynamic of AI inference.
Token prices have fallen by orders of magnitude. The response has not been restraint. It has been proliferation. Agents spawn sub-agents. Systems loop, verify, and re-verify. Workflows that once made a single application programming interface (API) call now make hundreds — most of them invisible. The industry calls these “ghost tokens.” The user sees one answer. The system consumed a thousand tokens to produce that answer. Meanwhile, the unit itself is shifting — providers quietly changing how they count so the same workload costs more without a visible price increase. The denominator moves. Nobody announces it.
Cheaper tokens do not reduce total energy. They increase it. The UN report projects that by 2030, data centre electricity consumption will exceed the combined usage of Pakistan, Bangladesh, and Nigeria — six hundred and fifty million people. Water consumption will equal the needs of 1.3 billion.
Waiting for efficiency gains does not solve this. Efficiency gains are the accelerant. A recent paper on cognitive debt applies the economics of financial bubbles to AI adoption — stability breeds overconfidence, overconfidence breeds fragility. The more capable the tool, the less visible the dependency. Until it breaks.
The only intervention that actually reduces total compute is architectural. Not faster chips. Not better cooling. Architecture — the decision about which model handles which task, and whether the task needs a model at all.
Tiered intelligence
Most organisations running AI today are making the same structural mistake. Every task — classification, extraction, summarisation, formatting, genuine reasoning — routes to the same frontier model. A four-hundred-billion-parameter system sitting in a hyperscale data centre, drawing megawatts, processing a request that a model one-hundredth its size could handle better.
This is the private jet problem. Not because frontier models are wasteful by nature — they are not. They are wasteful when misapplied. Sending a routine document classification to a frontier reasoning model is like chartering a 747 to deliver a letter. The letter arrives. The cost is absurd. And nobody notices because the invoice is denominated in tokens, not fuel.
The alternative is tiered intelligence — what amounts to a transit network for AI operations.
Tier one: automation. Rules, templates, deterministic logic. No model required. Most organisations skip this entirely — reaching for AI when a conditional statement would suffice.
Tier two: domain models. Small, fine-tuned, running on local infrastructure. A twenty-seven-billion-parameter model on a single graphics processing unit (GPU), trained on the organisation’s own data. These models do not need to reason broadly. They need to execute precisely within a bounded context. For eighty percent of operational work — classification, extraction, structured generation — they outperform the frontier. Not because they are more capable. Because they are more specific. One-hundredth the compute. Better results.
Tier three: frontier reasoning. The genuinely hard problems. Novel inference across broad knowledge. Multi-step reasoning that requires the full weight of a model trained on the breadth of human output. This is what frontier models are for — and it is a narrow slice of what most organisations actually send them.
The control plane sits between the tiers. It intercepts the request, evaluates the complexity, and routes to the smallest capable model. The default path is not the frontier. The default path is local. The frontier is the exception — reserved, expensive, and replaceable.
Fewer than one in four companies can currently track what their AI costs at a task level. The rest are sending tier-one work to tier-three models because nobody built the instrumentation to know the difference. The routing is useless without a map.
The providers themselves are validating the architecture. Every major provider now ships a tiered family — Anthropic from Haiku to Fable, OpenAI from Luna to Sol, Google from Flash to Ultra. The models exist. What does not exist — for most organisations — is the routing layer that matches the task to the tier. The control plane is still the missing piece. Alibaba’s SkillWeaver framework demonstrated what becomes possible when that piece exists — compositional skill routing that cuts agent token consumption by ninety-nine percent. Not a marginal gain. Two orders of magnitude.
The convergence
In June, Anthropic’s Fable 5 was shut down by a US Commerce Department directive. Ninety minutes from letter to blackout. Every customer, every country, every workflow that called that model — dark. Restored weeks later. But the restoration made the point sharper than the shutdown did. Access returned at the provider’s discretion, not the customer’s.
An organisation that had routed eighty percent of its work through local domain models would barely have noticed. Most have not. That is the gap. The domain layer kept working. The orchestration layer kept routing. Only the narrow frontier slice degraded — and it degraded to a fallback, not to nothing.
That is sovereignty. And the architecture that produces it is the same architecture that reduces compute by two orders of magnitude.
Not two problems. Not two solutions. One design decision.
The small model that gives an organisation operational independence is also the model that draws one-hundredth the power. The routing logic that ensures resilience is also the logic that prevents wasteful frontier calls. Consider Datagrid Southland, in my adopted country of New Zealand — a two-hundred-and-eighty-megawatt campus underpinned by a fifteen-year renewable power agreement. Data sovereignty and renewable baseload, solved simultaneously.
Sovereignty and sustainability are not competing priorities. They are the same design decision, viewed from different angles.
The horizon
The near-term architecture is deployable now. Small models on local GPUs. Edge inference on phones — a 2.6-billion-parameter model running on Apple’s Neural Engine with no data centre, no API dependency, no sovereignty risk. The architecture does not require anyone’s permission to build.
The medium term is already under construction. Microsoft is restarting Three Mile Island — 835 megawatts by 2028. Google is contracting for small modular reactors. Amazon has committed 20 billion dollars to Susquehanna. Nuclear solves baseload for the frontier slice. Further out, startups are putting GPUs into orbit — where solar is unfiltered, cooling is passive, and land constraints do not exist.
But here is the discipline the efficient frontier demands. None of these developments — nuclear, orbital, renewable data centres — matter if the routing does not exist. Without tiered intelligence, cheaper energy simply means more wasteful computation at larger scale. The Jevons Paradox does not care whether the electricity comes from coal, from a reactor, or from orbit.
The efficient frontier — in the financial sense, the computational sense, and the environmental sense — is the curve where every unit of intelligence is right-sized to its task. Maximum capability per watt. Maximum sovereignty per dependency.
What is good for operations is also good for the planet. The question is whether organisations have built the routing to prove it.
References
- United Nations University, “AI’s Growing Environmental Footprint,” 2026.
- Tom’s Hardware, “Maryland citizens slapped with $2 billion grid upgrade bill for out-of-state AI data centres,” 2026.
- Exponential View, “Why AI bills rise as costs fall,” 2026.
- Quartz, “Orbital data centre startups competitive landscape,” June 2026.
- Datagrid, “Southland Data Centre — 280MW, 100% renewable,” 2026.
- Anthropic, “Statement on US government directive to suspend access to Fable 5,” June 2026.
- “Cognitive Debt: AI as Intellectual Leverage and the Dynamics of Systemic Fragility,” arXiv, June 2026.
- VentureBeat, “New Alibaba AI framework skips loading every tool, cutting agent token use 99%,” 2026.
- OpenAI, “GPT-5.6: Sol, Terra, Luna,” July 2026.