We are used to talking about AI progress as a race for bigger, smarter models. But the real “next level” of AI will not be won by another parameter count. It will be decided by infrastructure – and by the human choices baked into that infrastructure about power, access, safety and equity.
AI is moving from lab demos into places where it shapes hiring, credit, healthcare, education, public services and everyday work. That makes questions about data pipelines, monitoring, permissions, energy and compute fundamentally human questions, not just technical details.
1. Models are the engine, infrastructure is the road (and people are the drivers)
Modern AI systems only work in the real world when they sit on top of robust infrastructure: data platforms, deployment pipelines, monitoring, rollback mechanisms, access controls and governance. This is exactly what MLOps and AI platform engineering focus on – streamlining the machine-learning lifecycle from development to deployment and monitoring. IBM+2Medium+2
Without that stack, AI remains an experiment. With it, organisations can keep models updated, trace decisions, manage risk and deliver reliable services at scale. Google Cloud+1
For the humans who have to live with these systems – employees, customers, citizens – this is what determines whether an AI system feels like a helpful colleague or an unpredictable black box. Infrastructure is what allows people to say, “I understand what this system is doing, I can challenge it if needed, and I know someone is accountable.”
2. From one-off models to AI-native systems
The first wave of adoption often looked like bolt-on features: add a chatbot here, a summarise button there. The next wave is AI-native systems, where AI agents and models are embedded end-to-end in workflows: triaging cases, drafting responses, triggering follow-up actions and learning from outcomes.
That shift demands orchestration, shared context, tool-calling and clear policies for what agents can and cannot do. Open-source and cloud-native tools such as MLflow, Kubeflow and Ray have emerged specifically to manage workflows, experiment tracking and scalable deployment for this kind of system. osf.io+2SSRN+2
Designing this layer is not just architecture work. It is role design: deciding when humans must stay in the loop, when they are on the loop (monitoring) and when they can safely be out of the loop for low-risk tasks.
3. Governance and compliance live in the stack, not just in policies
Responsible-AI principles – fairness, transparency, accountability – only become real when they show up in the technical stack. In practice, that means:
- Logging which model, data and prompt produced a decision
- Versioning models and configurations and being able to roll back
- Monitoring for drift, failures and bias in real time
- Enforcing access control for sensitive data and high-impact actions
Regulation is moving in exactly this direction. The EU AI Act defines “high-risk” AI systems and requires human oversight, documentation, logging and risk-management processes from both providers and deployers, including explicit obligations to assign human oversight, monitor operation and retain logs.
For workers and affected communities, this matters because it creates the conditions for recourse: you can ask why something happened, request a human review and expect an audit trail rather than a shrug and “the system decided.”
4. Infrastructure decides who gets to participate in the AI economy
AI infrastructure is also a fairness and sovereignty issue. If only a few large firms and a handful of countries control the compute, proprietary tools and surrounding stack, everyone else is reduced to consuming AI as a service with limited control.
Recent work mapping the geography of public-cloud GPU compute shows a stark split: “Compute North” countries hosting most AI-training capacity, and “Compute South” countries hosting more limited infrastructure oriented toward deployment, reinforcing global inequalities in who can develop and govern advanced AI systems. AAAI Open Access Journal+3Oxford University Research Archive+3ResearchGate+3
At the same time, language technologies overwhelmingly favour a small set of dominant languages. Commentaries and studies on the “digital language divide” note that out of roughly 7,000 languages, only a tiny fraction have strong NLP support, leaving many communities without tools in their own languages.
Open and interoperable infrastructure – including open-source tooling, shared benchmarks and regional compute initiatives – helps smaller organisations, public bodies and communities build systems around their own languages, norms and needs, instead of passively importing systems built elsewhere. osf.io+1
Equity here is not just about who uses AI, but who gets to shape it:
- Who decides what is “good enough” performance for a minority language?
- Who sets the thresholds for risk in local credit, welfare or policing systems?
- Who controls the logs, metrics and dashboards that define “success”?
Infrastructure choices today quietly fix who has voice and leverage in these decisions.
5. The physical reality: energy, water, latency and sustainability
Behind the “cloud” are very material constraints. The International Energy Agency (IEA) estimates that data centres accounted for around 1.5% of global electricity consumption in 2024 (about 415 TWh), and projects that total data-centre electricity use could more than double by 2030 to around 945 TWh – close to 3% of global demand – with accelerated computing for AI as a key driver. Reuters+5IEA+5IEA+5 Earlier analyses found similar orders of magnitude (around 1% of global electricity use in 2018). datacenters.lbl.gov+1
Water is part of the story too. Reviews and policy reports find that AI-intensive data centres can consume millions of gallons of water per day for cooling, and AI’s water demand may reach billions of cubic metres annually in the coming years if growth continues unchecked. Investigative reporting has shown that some of this build-out is happening in already water-stressed regions, raising concerns about competition with local communities and agriculture.
From an equity lens, these physical constraints matter because the environmental and land-use costs of AI infrastructure often fall on specific local communities, while the economic benefits may accrue elsewhere. Latency and connectivity gaps also mean users in the Global South or rural areas may experience slower, less reliable AI services – or none at all – even as they are affected by decisions made in distant data centres.
6. Infrastructure is where human and equity questions crystallise
Once we see AI as an infrastructure question, the human element becomes impossible to ignore. Key design questions suddenly look like this:
- Who can override an AI agent, and how quickly in practice?
- Who decides what is logged, who can see it and how long it is kept?
- Whose expertise is encoded in workflows – only engineers, or also domain experts, frontline workers and affected communities?
- Who is accountable when something goes wrong, and how easy is it to pause or shut down a system in reality, not just on paper?
- Whose values and risk tolerance are embedded in policy engines and thresholds?
Infrastructure is not neutral. The way we design platforms, guardrails and workflows encodes assumptions about power, responsibility and risk. If only executives, vendors and engineers are in the room, the resulting systems are likely to reproduce existing inequalities; if unions, civil-society groups and communities are involved, different choices tend to emerge.
7. What “next-level AI” really looks like
If the next level of AI is infrastructure-led and human-centred, the key questions shift from “Which model is best?” to:
- What rails do we need so that humans can safely rely on, challenge and improve AI systems over time?
- How do we design roles where AI handles what it is good at – scale, pattern recognition, automation – and humans retain judgment, context and ethical responsibility?
- How do we ensure new models can be swapped in without rebuilding trust, governance and compliance from scratch?
- How do we distribute compute, tools and governance capacity so that more regions and communities can build AI on their own terms?
The organisations – and societies – that thrive in this next phase will not simply have the most powerful models. They will have the most thoughtful rails for intelligence: infrastructure that is reliable, transparent, energy-aware and explicitly built around human needs, human agency and global equity.
Selected references
- IBM, “What is MLOps?” – overview of MLOps practices across the ML lifecycle. IBM
- LakeFS / industry guides on MLOps pipelines and automation. lakeFS
- International Energy Agency (IEA), Energy and AI and Data Centres and Data Transmission Networks – data-centre electricity use and AI-driven growth projections. Energy+4IEA+4IEA+4
- Masanet et al., “Recalibrating global data center energy-use estimates,” Science (2020) – estimates of global data-centre energy use (~1% of electricity). datacenters.lbl.gov+1
- Lehdonvirta et al., “Compute North vs. Compute South: The Uneven Possibilities of Compute-based AI Governance Around the Globe” – mapping of global GPU-compute concentration. AAAI Open Access Journal+3Oxford University Research Archive+3ResearchGate+3
- Chinasa Okolo & Marie Tano, “Closing the gap: A call for more inclusive language technologies,” Brookings (2024), and related work on low-resource languages.
- EU AI Act
Noleen Mariappen is a purpose-driven impact strategist and tech-for-good advocate bridging innovation and equity across global communities. With a background in social and environmental impact and a passion for digital inclusion, Noleen leads transformative initiatives that leverage emerging technologies to tackle systemic inequality and empower underserved populations. Noleen is an active contributor to ethical AI dialogues and cross-sector collaborations focused on sustainability, education, and inclusive innovation. Connect with her on LinkedIn: https://www.linkedin.com/in/noleenm/
__
The views expressed in this article are those of the author and may not reflect the official stance of Consumer AI Protection Advocates (CAIPA).
CAIPA’s mission is to empower consumers by advocating for responsible AI practices that safeguard consumer rights and interests across various sectors, including electric vehicles (EVs), autonomous vehicles (AVs), and robotics.
#CAIPA #ArtificialIntelligence #ConsumerProtection #AutonomousVehicles #FutureofWork
