CNC: input = G-code file · processing = firmware + stepper motor controller · output = machined metal part at 0.01mm precision. Electricity moves motors that move a tool that removes material.

3D Printer: input = STL file + slicing parameters · processing = Marlin/Klipper firmware · output = physical object layer by layer. Electricity heats a resistor and moves a nozzle.

Laser engraver: input = vector file + power/speed · processing = GRBL controller · output = engraved or cut material. Electricity excites a laser diode at a precise wavelength.

Inkjet printer: input = raster file · processing = print firmware · output = printed document. Electricity propels 2-picoliter ink droplets at 1200 DPI.

Language model (LLM): input = token sequence · processing = matrix multiplication across billions of parameters on GPUs · output = most probable token sequence. Electricity moves electrons through 3nm transistors performing floating-point additions.

In every case: controlled electricity → result according to parameters. The CNC works in the physical layer — produces matter. The LLM works in the logical layer — produces symbols. Not a difference of nature. A difference of output domain.

In 1956, John McCarthy chose the term "Artificial Intelligence" — not for descriptive precision, but for academic prestige. An LLM does not "understand." It predicts the probability distribution of the next tokens according to weights adjusted during training. Remarkable engineering. Not intelligence in the philosophical sense.

The Operation Dindon corpus documented that "Sovereign Cloud" sells a legal sovereignty that does not exist (physical ✓ / legal ✗). "Artificial Intelligence" is the same deception structure: it sells a philosophical intelligence that does not exist (matrix computation ✓ / intelligence ✗). Both use a strong word ("sovereign," "intelligence") to name the property the client desires without it actually being present.

Every generation experienced its own "incomprehensible revolution." In 1965, the reel-to-reel tape recorder recorded and replayed a human voice — considered almost magical by those who did not understand magnetic induction. In 1975, the VCR captured a television broadcast for replay at will — described as incomprehensible by the uninitiated. In 1980, Teletext transmitted text via the television signal — deemed futuristic and world-changing.

None were magic. All were controlled electricity applied to a new domain. Generative AI in 2024 is at its maturity stage what the VCR was in 1975: impressive for its era, already demystified by engineers who understand its low-level layers, destined to become commonplace in five to ten years. Not a reason to deify it now.

OpenAI, Azure OpenAI, AWS Bedrock, Google Vertex AI — identical discourse: AI is too complex, too resource-hungry, too specialised to deploy without a hyperscaler. Permanent subtext: "You cannot do this yourselves." Exactly the Serverless and NoOps discourse — applied to AI. Partially true for some use cases. Deliberately exaggerated for all others.

When a company uses Azure OpenAI to index internal documents via RAG, it indexes its intellectual property in proprietary vector databases — Azure AI Search, Amazon Kendra, Pinecone on AWS. Documents theoretically remain with the company. Embeddings reside in the hyperscaler's infrastructure. Invisible exfiltration of heritage memory. Using the hyperscaler's AI to index data means giving them the blueprints of the organisation's own applicative black box.

Context Caching API features store internal documentation, source code, business procedures sent with each request in the hyperscaler's clusters — under cover of "security telemetry."

The difference from RAG: RAG exfiltrates structured memory (documents). Context cache exfiltrates living memory — real-time prompts, source code sent with each request. The difference between stealing blueprints and installing a camera in the meeting room. Covered by "telemetry" clauses nobody reads — exactly like Terms §14.12.

Llama 3 (Meta) is under a restrictive commercial licence — prohibition on using weights to train a competing model, restrictions beyond 700 million users. Open-Weights, not Open-Source under OSI. Genuinely free licences: Mistral 7B (Apache 2.0) · Falcon 7B/40B (Apache 2.0). A unilateral licence change can block software infrastructure post-deployment — exactly like a cloud Terms modification.

Replacing OpenAI/Azure API with NVIDIA A100/H100 GPUs solves only one level. NVIDIA owns proprietary locks — CUDA, cuDNN, NCCL — and production depends entirely on TSMC in Taiwan (~90% of advanced chips worldwide). The answer: open abstraction runtimes. ROCm (AMD) · vLLM · Ollama · llama.cpp — run models on heterogeneous hardware. Owning your AI in bare-metal means the parameter file can run on any available silicon.

Cloud API: billed per token — pure variable OPEX. Bare-metal GPU server: CAPEX + fixed OPEX (electricity, rack, maintenance).

Cloud API cost: N requests × cost_per_token × avg_tokens
Bare-metal cost: GPU_CAPEX / amortisation + fixed_monthly_OPEX

From documented cases: for intensive use (>100,000 requests/day with long contexts), bare-metal becomes profitable in 6 to 18 months. Below this threshold, the cloud API remains competitive. Intellectual honesty requires naming this threshold — not pretending bare-metal always wins.

An AI compute chassis (HGX A100 8 GPUs) consumes 10 to 40 kW per rack. Classical datacentre: sized for 5 to 7 kW per rack. A civil engineering problem that TCO formulas without a datacentre engineer systematically ignore.

PUE factor: for high-density AI compute, PUE 1.5 means 40 kW GPU + 20 kW cooling = 60 kW billed at the floor. Without liquid cooling or hot-aisle containment, operating costs explode.

Complete TCO formula: GPU_CAPEX + cooling_CAPEX + (GPU_kW × PUE × kWh_cost × annual_hours) + maintenance. High-density PUE is what the application engineer forgets and the datacentre engineer cannot ignore.

Weights Sovereignty = Contractual Layer: do I legally own the model without future unilateral restriction? Runtime Sovereignty = Software Layer: do I execute on my infrastructure with an open framework without artificial complexity? Training Data Sovereignty = Material Layer: do I know the foundations of what I use?

Proprietary model · opaque runtime · CUDA locked to TSMC chips · unknown training data = total AI captivity. Same structure as cloud captivity documented in 64 corpus studies.