Pernille AI - Advanced AI Solutions

Purpose of Disruptive Perspectives

When we analyze various themes, we spend a lot of time and use many tools (quarterly reports, analyses, dialogue with companies, company visits, Excel, calculators, and language models). We often create small notes, and sometimes large notes that we think of as perspectives. We are old enough to know that there are rarely truths, often just different perspectives.

Disruptive Perspectives has only one purpose: To share our perspectives on themes that shape our future. These are not academic papers, encyclopedia entries, or recommendations to do something, buy or sell something. Just good old-fashioned information sharing to make visible how we view various themes at the time of publication. Perspectives don't become less, perhaps rather more, when you share them. With that starting point; have a nice journey through our perspectives.

What is AI Infrastructure?

When we talk about artificial intelligence, it's easy to think about apps like ChatGPT, self-driving cars, or medical image diagnostics. But we rarely talk about what makes all this possible - the very foundation that Artificial Intelligence (AI) is built upon. This foundation is called AI infrastructure, and it consists of everything from hardware and software to data centers, networks, and the people who develop and operate the technology. Without this infrastructure, there is no AI.

What makes AI infrastructure unique is that it is extremely demanding. Where traditional IT infrastructure handles websites or databases, AI infrastructure requires massive computing power, enormous amounts of data, and continuous development. Modern language models are trained on several hundred billion words and run on thousands of GPUs (Graphics Processing Units) for weeks.

And where data infrastructure could previously be built once and used for years, AI infrastructure must be dynamic. Models become obsolete quickly, new datasets arrive, and algorithms must be updated. Therefore, modern AI infrastructure is often cloud-based and modular, with hardware like specialized GPUs and TPUs (Tensor Processing Units), and software that enables continuous development through MLOps (Machine Learning Operations).

Despite this, AI infrastructure resembles other infrastructure in some areas: it must be scalable, secure, and dependent on fast networks. The difference lies in the intensity and complexity. AI needs data. It is essential for further development and achieving AGI (Artificial General Intelligence) and ASI (Artificial Super Intelligence).

The Vision

When you build a highway, it's not just about asphalt and concrete. It's about what you enable. In this sense: movement, trade, contact between people. Similarly, AI infrastructure is not just about servers, data centers, and algorithms. It's about creating a platform for future innovation.

The grand goal of AI infrastructure is to make advanced artificial intelligence accessible to everyone. Not just for technology giants. This democratization makes it possible for startups, researchers, the public sector, and individuals to build solutions that were previously unthinkable. Think about climate models, personalized medicine, or intelligent learning tools. All of this depends on the underlying infrastructure being available, robust, and sustainable.

But to reach this goal, we must solve some fundamental challenges:

Scale: The infrastructure must handle both enormous datasets and extremely complex models
Environment: Today's AI training requires enormous amounts of energy, and without better energy efficiency, growth will become unsustainable
Accessibility: AI must be usable by small actors, not just those with billion-dollar budgets
Ethics, Security, and Reliability: We must be able to trust that the systems we build don't discriminate, misinterpret, or leak sensitive information

When we assess how good AI infrastructure is, we must therefore look broadly: How quickly are models trained? How much energy do we use per prediction? How accessible is the technology for new actors? How safe and fair are the systems in practice? The success of AI infrastructure cannot be measured in technical performance alone. It must also work for society.

The Six Enablers

When we lift our gaze, we see that AI infrastructure stands on the shoulders of 6 powerful "enablers":

1. Specialized Hardware

GPUs and neuromorphic chips provide the computational power for systems to learn. Here, Moore's Law still plays a role, but it is now supplemented by new principles like Amdahl's Law and innovations in quantum computing.

Key players: NVIDIA (H100, A100), Google (TPU), Intel (Loihi), Tesla (Dojo)

2. Data

Without data, there is no AI. But quantity alone isn't enough - data must be relevant, clean, and continuously updated. Edge devices and synthetic datasets are becoming increasingly important.

Key considerations: Quality over quantity, privacy preservation, synthetic generation

3. Software Frameworks

PyTorch, TensorFlow, and MLOps platforms enable models to be trained, deployed, monitored, and improved in real-time.

Essential tools: JAX, MLflow, Kubeflow, AWS SageMaker, Azure ML

4. Cloud & Edge Computing

Cloud platforms provide elastic capacity and global availability, while edge computing delivers low latency and local control - two complementary approaches.

Infrastructure: AWS, Google Cloud, Azure, Edge devices, 5G networks

5. Human Expertise

Technology alone isn't enough. AI requires interdisciplinary expertise from coding and statistics to ethics and domain knowledge.

Critical skills: ML engineering, data science, domain expertise, ethics

6. Regulatory Framework

Economic and regulatory conditions set the boundaries. Public support, tax incentives, and risk capital are crucial, as are regulations like GDPR that ensure responsible AI development.

Key frameworks: GDPR, AI Act, national AI strategies

These six factors don't work alone - they connect in a dynamic system. When one of them fails, it ripples through the entire ecosystem. This is systems theory in practice: everything is connected.

How Could AI-Generated Regulations Look?

A typical AI-generated regulation would likely be very detailed and technical. Using Natural Language Processing (NLP) and machine learning models, AI would propose regulations based on existing rules, technical specifications, and public input. An example could be:

"AI systems classified as high risk (defined as an error rate above 0.1% in critical applications) shall log all decisions in an auditable format and undergo annual bias testing."

Such proposals would be very precise but could easily lack necessary flexibility, cultural understanding, and local adaptation - elements that require human judgment.

The Optimal Solution: Humans with AI Support

Even though AI shouldn't get complete responsibility for regulatory work, that doesn't mean we should ignore its potential. The best approach is to use a hybrid model, where AI functions as a powerful support tool.

High Complexity + High Impact

Human-led with AI assistance

Low Complexity + High Impact

Human decisions with AI input

High Complexity + Low Impact

AI drafts with human review

Low Complexity + Low Impact

Automated with oversight

Infrastructure Categories

AI infrastructure consists of hardware, software, data storage, and networks that together enable development, training, operation, and maintenance of advanced AI models. This comprehensive system naturally divides into four main areas:

Computational Infrastructure = The Engine

This is the heart of AI: specially adapted processors like GPUs (e.g., NVIDIA H100), built for parallel processing of enormous amounts of data. Without sufficient computing power, modern AI models cannot be trained or run effectively.

GPU clusters for training
TPUs for specialized workloads
Neuromorphic chips for brain-like computing
Quantum processors for optimization

Data Infrastructure = The Fuel

Data is AI's fuel. This infrastructure includes everything from data lakes and warehouses to real-time streaming systems. The quality and quantity of data directly determine how good an AI model can become.

Data lakes and warehouses
Real-time streaming platforms
Data governance and quality tools
Privacy-preserving technologies

Software Infrastructure = The Control System

Frameworks and platforms that make it possible to develop, train, and deploy AI models. This includes everything from low-level libraries like CUDA to high-level frameworks like TensorFlow and PyTorch.

Deep learning frameworks
MLOps platforms
Model versioning systems
Experiment tracking tools

Deployment Infrastructure = The Highway

The systems that bring AI models from development to production. This includes containerization (Docker/Kubernetes), model serving platforms, and monitoring systems that ensure models perform as expected in the real world.

Container orchestration
Model serving platforms
A/B testing frameworks
Performance monitoring

AI Factories

These factories are the core of modern AI infrastructure. These are not factories in the traditional sense, but data centers built for one thing: to produce intelligence. The term was first popularized by NVIDIA, and describes facilities that combine enormous amounts of computing power with advanced software to train and operate AI models at scale.

What is an AI Factory?

An AI factory is a term that describes highly specialized, scalable computing facilities designed to develop, train, and operate artificial intelligence on an industrial scale. According to NVIDIA, AI factories are massive GPU clusters that function as "factories" to generate AI models, applications, and intelligent solutions, where data and energy are raw materials, and AI tokens (e.g., predictions, decisions) are the product.

These facilities are often cloud-based or hybrid-based data centers that combine powerful hardware, advanced software, and data management to support AI workflows like deep learning, natural language processing, and autonomy.

Components of AI Factories

Hardware Stack

• NVIDIA H100: Latest GPU for AI training
• Google TPU: Tensor processing units
• Tesla Dojo D1: Custom AI training chips
• BrainChip Akida: Neuromorphic processors

Software Stack

• TensorFlow & PyTorch: Model development
• MLflow & Kubeflow: Automation tools
• NVIDIA TensorRT: Performance optimization
• ONNX Runtime: Edge deployment

AI factories are crucial for enabling technologies like self-driving cars, humanoids, and AI agents, including those using federated learning in vehicles.

CUDA and Dojo: The Battle for AI Computing

CUDA: NVIDIA's Secret Weapon

CUDA (Compute Unified Device Architecture) is NVIDIA's proprietary programming platform that allows developers to use GPUs for general computing. Since its launch in 2007, CUDA has become the de facto standard for AI development.

CUDA's dominance creates a powerful network effect: the more developers who use CUDA, the more valuable it becomes to learn CUDA, which in turn leads to more CUDA-optimized software.

Market impact: 90%+ of AI training uses CUDA

Dojo: Tesla's Answer

Tesla's Dojo represents an ambitious attempt to break NVIDIA's dominance. Dojo is a custom-built supercomputer designed specifically for training Tesla's self-driving AI.

What makes Dojo unique: Custom D1 chips with unprecedented bandwidth, 3D chip packaging for minimal latency, integration with Tesla's massive fleet data, and optimization for specific use cases rather than general computing.

Potential: Could revolutionize autonomous vehicle training

Tokens - The New Currency in AI

In the AI world, tokens have become the fundamental unit of measurement. A token can be a word, part of a word, or even a character. When we talk about GPT-4 processing billions of tokens, we're talking about its ability to understand and generate text at massive scale.

Token Economics

The cost of AI operations is increasingly measured in tokens:

• Training costs: How many tokens were used to train the model?
• Inference costs: How many tokens does each query consume?
• Context windows: How many tokens can the model remember?
• Pricing models: Cost per million tokens processed

This has created a new economy where companies optimize for token efficiency, and where the price per million tokens becomes a key competitive metric. Understanding token economics is crucial for managing AI costs and performance.

Federated Learning

Federated learning represents a paradigm shift in how we train AI models. Instead of collecting all data in a central location, federated learning trains models on distributed data while keeping the data local.

Benefits

✓ Privacy: Sensitive data never leaves the source
✓ Efficiency: Reduces bandwidth requirements
✓ Compliance: Easier to comply with data regulations
✓ Scale: Can leverage data from millions of devices

Challenges

✗ Coordination: Synchronizing learning across devices
✗ Heterogeneity: Different data distributions
✗ Security: Protecting against malicious participants
✗ Communication: Minimizing network overhead

What is Edge AI?

Edge AI refers to running AI models directly on devices at the "edge" of the network - smartphones, IoT sensors, autonomous vehicles - rather than in centralized cloud servers.

Why Edge AI Matters

Latency

Decisions in milliseconds, not seconds - critical for autonomous vehicles and real-time applications

Privacy

Data stays on the device - personal information never leaves your phone

Reliability

Works without internet connection - functions in remote locations

Cost

Reduces cloud computing expenses - no bandwidth costs for data transfer

What are Neuromorphic Chips?

Neuromorphic chips represent a fundamental reimagining of computer architecture, inspired by the human brain. Unlike traditional chips that separate memory and processing, neuromorphic chips integrate both in a way that mimics biological neurons.

Key Characteristics

Event-Driven Processing

Only activates when there's input, like neurons firing

Parallel Processing

Thousands of simple processors working simultaneously

Adaptive Learning

Can modify their own connections based on experience

Ultra-Low Power

Uses orders of magnitude less energy than traditional chips

Leading Players

• Intel's Loihi: Pioneering research chip with 128 cores
• IBM's TrueNorth: Million-neuron chip for pattern recognition
• BrainChip's Akida: Commercial neuromorphic processor for edge AI

What are Spiking Neural Networks?

SNNs represent the next generation of neural networks, more closely mimicking how biological brains process information. Unlike traditional neural networks that use continuous values, SNNs communicate through discrete spikes, similar to neurons.

Advantages of SNNs

⚡ Energy Efficiency: 10-1000x more efficient than traditional neural networks
⏱️ Temporal Processing: Natural handling of time-series data
🔋 Sparse Computation: Only active neurons consume power
🧠 Biological Plausibility: Better model of actual brain function

Applications

Robotics

Real-time sensory processing and motor control

Brain-Computer Interfaces

Direct neural communication and prosthetics

IoT Devices

Ultra-low power smart sensors

Neuromorphic Vision

Event-based cameras and processing

Tesla and NVIDIA: A Complex Relationship

The relationship between Tesla and NVIDIA illustrates the dynamics of AI infrastructure development.

The Partnership Phase (2016-2019)

Tesla initially relied heavily on NVIDIA's Drive platform for Autopilot. NVIDIA GPUs powered the training of Tesla's neural networks, and Drive PX2 computers ran in Tesla vehicles.

The Breakup (2019)

Tesla decided to develop its own Full Self-Driving (FSD) chip, citing:

Need for specialized optimization
Cost considerations at scale
Vertical integration strategy
Desire for complete control

Current Dynamics

Despite developing custom inference chips, Tesla still uses NVIDIA GPUs for training. This illustrates a key principle: different parts of the AI pipeline may require different infrastructure solutions.

From Cloud to Edge: When Intelligence Moves Out to Products

The migration of AI from centralized cloud to distributed edge represents a fundamental shift in computing architecture.

Drivers of Edge Migration

🔒 Privacy Regulations

GDPR, CCPA, and other laws favor local processing

⚡ Latency Requirements

Real-time applications can't wait for cloud round-trips

💰 Bandwidth Costs

Video and sensor data too expensive to stream

🔧 Reliability Needs

Critical systems can't depend on internet connectivity

The New Feedback Loop

The combination of edge and cloud creates powerful feedback loops:

1. Data Collection at Edge: Devices collect real-world data with privacy-preserving aggregation
2. Cloud Training: Aggregate learning from millions of devices with model improvement
3. Edge Deployment: Updated models pushed to devices with gradual rollout
4. Continuous Improvement: Edge devices report performance, cloud identifies patterns, cycle repeats

Where Are the Greatest AI Opportunities in 2025?

Looking ahead, the most promising opportunities in AI infrastructure lie at several convergence points:

1. Specialized AI Chips

Vision processors, language accelerators, robotics controllers, brain-computer interfaces

2. Quantum-AI Hybrid

Optimization problems, drug discovery, financial modeling, cryptography

3. Photonic Computing

100x energy efficiency, massive parallelism, minimal heat, quantum integration

4. Distributed AI Networks

Decentralized training, model marketplaces, resource sharing, privacy-preserving AI

5. Neuromorphic Ecosystems

Brain-inspired hardware, spiking networks, real-time AI, brain interfaces

6. AI-Optimized Data Centers

Liquid cooling, renewable energy, edge caching, optical interconnects

Investment Implications

For investors, AI infrastructure presents both enormous opportunities and significant risks:

Opportunities

✓ Pick-and-shovel plays: Infrastructure over applications
✓ Network effects: Platforms becoming standards
✓ Vertical integration: Full stack control
✓ Emerging markets: Regional AI capabilities

Risks

⚠️ Technological disruption: Today's leaders, tomorrow's laggards
⚠️ Regulatory uncertainty: Government intervention
⚠️ Concentration risk: Few dominant players
⚠️ Valuation bubbles: Hype exceeding value

Key Metrics to Watch

• Performance per watt
• Cost per training run
• Model compression ratios
• Edge deployment numbers
• Developer ecosystem growth
• Market concentration metrics

Conclusion: The Infrastructure Revolution

AI infrastructure is undergoing a revolution comparable to the birth of the internet. We're moving from centralized, general-purpose computing to distributed, specialized, and intelligent systems. This transformation will:

🌍 Democratize AI: Making advanced capabilities accessible to all
🚀 Enable New Applications: From humanoid robots to brain interfaces
🏭 Transform Industries: Every sector rebuilt on AI foundations
⚡ Create New Challenges: Energy, privacy, security, and ethics

The companies and countries that master AI infrastructure will shape the next century. The race is not just about algorithms or data - it's about building the physical and digital foundations that will support humanity's cognitive augmentation.

As we stand at this inflection point, the choices we make about AI infrastructure will determine whether artificial intelligence becomes a tool for human flourishing or a source of division and control. The perspective we adopt today will shape the reality we inhabit tomorrow.

The journey from simple neural networks to artificial general intelligence will be built on infrastructure. Not just the visible applications, but the invisible foundations - the chips, networks, frameworks, and systems that make intelligence computable, scalable, and accessible. This is our perspective on AI infrastructure: it's not just technology, it's the platform for humanity's next chapter. And that chapter is being written now, one GPU, one model, one breakthrough at a time.

Disclaimer

The content in this article is not intended as investment advice or recommendations. If you have any questions about the funds referenced, you should contact a financial advisor who knows you and your situation. Remember also that historical returns in funds are never a guarantee of future returns. Future returns will depend on, among other things, market development, the manager's skill, the fund's risk, as well as costs for purchase, management, and redemption. Returns can also be negative as a result of price losses.