In Chapter 2 you learned that AI is industrial-scale pattern-matching with no underlying concepts. In Chapter 3 you learned that the model is the training…

In Chapter 2 you learned that AI is industrial-scale pattern-matching with no underlying concepts. In Chapter 3 you learned that the model is the training data, and what's in that data, who put it there, and who didn't consent to it being there. This chapter completes the foundations by going to the layer underneath: what AI is physically made of. The training data has to be processed by something. The processing happens on specific chips, sitting in specific servers, in racks, in buildings that draw real electricity, cooled by water or air, manufactured from materials extracted from specific places. None of this is abstract. All of it has costs that get paid by specific people in specific places. The Canadian AI conversation routinely treats "infrastructure" as a category of policy spending, when on the ground it is a category of physical things with physical origins and physical consequences. > EXAMPLE — "the cloud" is a building near you. The word sounds weightless, so picture the actual object: a windowless box the size of a few hockey rinks, drinking the power of a small town and enough water to keep itself from cooking. One is probably going up within an hour's drive of you. It has a street address and a hydro bill, whatever the brochure says. This chapter is shorter and more concrete than Chapters 2 and 3 because the work it does is mostly grounding the abstractions you already have. It is also the chapter where the through-line of the guide, Owen et al.'s 54% finding, gets its fullest physical treatment, because the materials AI is made of are the materials Owen et al. were counting. You will leave with: a working physical picture of what's inside an AI data centre, from rack to chip to substrate; an account of the supply chain that builds the chip; the Owen et al. extraction geography understood as physical fact rather than as statistical abstraction; Suzanne Kite's Lakota hardware-building protocol from the Indigenous Protocol and AI Position Paper engaged as a substantive alternative framework rather than as decoration; and the technical scaffold that Chapter 6 (Canadian infrastructure) and Chapter 8 (environment) will lean on. After this chapter, the guide's Part I Foundations is complete. The Canadian-arc chapters (5–9, already drafted) and the contested-questions chapters that follow can be read with all four foundations in place. ---

One: From the data centre, working inward

Start at the building scale and work down. Chapter 6 will name specific Canadian data centres; this chapter teaches the generic physical objects you'll find inside any of them.

The building. A data centre is a building optimized for one purpose: continuous high-density computing. The building is typically a low-rise warehouse-style structure (data centres rarely benefit from multi-story configurations because power and cooling infrastructure works better horizontally), sited where land is cheap, electricity is reliable, network connectivity is strong, and natural disasters are uncommon. The exterior is usually unremarkable (most large data centres don't advertise their presence), and the interior is highly engineered: redundant power feeds, redundant cooling systems, redundant network connections, physical security including biometric access, fire suppression systems designed for electrical equipment rather than for paper, and one to many large open rooms (called "data halls") where the actual computing equipment lives.

The rack. Inside the data hall, computing equipment is organized into standardized vertical metal frames called racks, each typically about two metres tall, with mounting holes spaced in standard "U" (rack unit) increments. A modern data centre might contain hundreds to thousands of racks per data hall, arranged in rows separated by aisles where technicians can access equipment. The rack standard (19-inch wide, 42U or 48U tall) is one of the most stable conventions in computing: equipment from the 1990s and equipment from 2026 use the same rack-mounting hardware.

The server. Within each rack are servers: flat horizontal computers, looking nothing like a desktop or laptop, designed to maximize density. A typical server is 1U to 4U tall (one to four rack units), contains one to several CPUs (general-purpose processors), substantial RAM (memory), local storage (SSDs), networking interfaces, and increasingly, for AI workloads, multiple GPUs (graphics processing units). A modern AI-optimized server, like an NVIDIA DGX H100 system, contains eight H100 GPUs alongside CPUs and supporting hardware, in a chassis 6-8U tall, drawing 10-12 kilowatts of electricity continuously.

The chip. Inside each server, the work happens on chips. The chips that matter most for AI are the GPUs, Graphics Processing Units originally designed for video games but turned out to be the optimal substrate for the matrix multiplications that neural networks require. NVIDIA's H100 and H200 chips are the current generation; the H100 was introduced in 2022, costs around US$25,000-40,000 per chip, contains approximately 80 billion transistors, draws about 700 watts of electricity continuously when working, and is manufactured exclusively by TSMC (Taiwan Semiconductor Manufacturing Company) on TSMC's most advanced process nodes. The H100 is what almost every major large language model is trained and operated on.

The substrate. The chip itself is built from a substrate of silicon: a precisely-grown crystal of nearly pure silicon, sliced into wafers, processed through hundreds of fabrication steps in fabs that cost billions of dollars to build and operate. The fab process requires extreme precision: features on the H100 chip are measured in single-digit nanometres, which is about the size of ten silicon atoms. The fabrication is environmentally intensive: a single advanced chip fab consumes tens of millions of gallons of ultrapure water daily, large quantities of specialty chemicals (some of which are toxic or have substantial production footprints), and significant electricity. Taiwan's water-supply shortages have repeatedly threatened TSMC's operations, with implications cascading through the entire AI supply chain.

The materials. And underneath the silicon, the materials. The H100 isn't only silicon. It contains substantial copper for interconnects, gold for some contacts, gallium for high-frequency components, germanium for certain sensing elements, tantalum (refined from coltan) for capacitors, tungsten for via interconnects, and rare earth elements (neodymium, dysprosium, terbium, others) for magnetic components and precision electronics. The packaging requires further materials. The server containing the chip requires substantial steel, aluminum, plastic, more copper for power cabling and cooling, and lithium-ion batteries for uninterruptible power supply backup. The rack contains additional steel and aluminum. The data centre building contains substantial concrete, steel, and copper.

This is what AI is physically made of. Claims about AI capability, benchmarks, commercial offerings, policy debates: all of it sits on top of millions of pieces of equipment, manufactured from physical materials, extracted from specific places, by specific people, with specific consequences. The abstraction "AI infrastructure" gets emptied of meaning if it isn't tracked back to the physical substrate. The rest of the chapter follows the substrate.

---

Two: The chip supply chain

The supply chain that produces the chips that train and operate AI is one of the most concentrated and geopolitically fraught supply chains on Earth. The Canadian AI conversation often treats "compute" as a generic resource. The actual compute supply chain is anything but generic.

Design. The H100, the H200, the upcoming Blackwell architecture: these are designed by NVIDIA, a US-headquartered company based in Santa Clara, California. NVIDIA's competitors (AMD, Intel for AI accelerators, and Chinese designers operating under export restrictions) are also US-headquartered or operate primarily from US-aligned jurisdictions. The intellectual property of modern AI chips is overwhelmingly American or American-aligned. This is a sovereignty question Chapter 9 engages: any Canadian AI strategy that depends on chips designed in the US is, at the IP layer, depending on US-controlled technology.

Fabrication. The actual manufacturing of advanced AI chips happens at TSMC in Taiwan, with some leading-edge capacity at Samsung in South Korea and Intel's recovering fabrication operations in the US (with substantial public subsidies). For the most advanced process nodes (3nm, 2nm), TSMC's Taiwan operations are effectively the only viable supplier. The geographic concentration is extreme: a single jurisdiction (Taiwan) sits at the centre of global AI hardware production, in a strategic environment where Chinese-Taiwanese-American tensions could disrupt that production with no clear backup plan. The 2022 CHIPS and Science Act in the US committed roughly US$52 billion to subsidize domestic semiconductor manufacturing, a recognition that current concentration is a strategic vulnerability. Canadian analogues exist at much smaller scale.

Equipment. The fabs themselves depend on extremely specialized equipment, most prominently the extreme ultraviolet (EUV) lithography machines from ASML (a Dutch company) that are used to print the finest features on advanced chips. There is one company globally that produces EUV systems. The machines cost approximately US$200 million each, weigh around 180 tonnes, and take months to install. The chip supply chain has, at its most critical bottleneck, a single supplier in the Netherlands.

Materials. The fabs require ultrapure silicon (mostly from a small number of suppliers in Germany, Japan, and the US), specialty chemicals (highly concentrated supply chains, primarily Japanese and German), and large quantities of water and electricity. The materials going into the chip (copper, gold, gallium, germanium, tantalum, tungsten, rare earths) have their own concentrated supply chains, which connects directly to Owen et al.'s 54% finding (next section).

Assembly. Once fabricated, the chips are assembled into servers, primarily by manufacturers in Taiwan, China, Mexico, and (recently, with substantial US incentives) the United States. The servers are shipped globally to data centres operated by hyperscalers and increasingly by sovereign AI projects.

The Canadian position in the chip supply chain. Canada has some semiconductor capacity, primarily through legacy operations and smaller specialty chip companies. Canada is not a player at the advanced-process-node level: the chips Canadian AI workloads run on are not designed, fabricated, or substantially assembled in Canada. The 2024 federal budget allocated funds toward semiconductor research and development, but the gap between "Canadian semiconductor research" and "Canadian frontier-AI-chip production" is large enough that closing it would require investments and timelines well beyond what AI for All commits to. Canadian AI sovereignty at the silicon layer is currently not on offer, regardless of which provider Canadian users choose.

This is what one of the structural findings the guide carries throughout means in practice: a "sovereign cloud" from Bell or Telus, hosting workloads for Canadian government departments, runs on US-designed chips, fabricated in Taiwan, on Dutch-built equipment, assembled in mostly East Asian factories, using materials extracted from largely non-Canadian jurisdictions. The corporate-jurisdiction layer is Canadian. The substrate-jurisdiction layer is not. Both are real. The guide asks readers to track both.

ALIGNMENT — which layer is the flag on? A "sovereign Canadian" AI claim can be perfectly true about the company holding the contract and silent about the chips, the lithography machines, and the minerals underneath it. The corporate layer wears a maple leaf while the substrate layer still answers to Taiwan, the Netherlands, and the Congo. Ask which layer of the stack the sovereignty claim actually touches, and which one it walks past.

---

Three: The materials, where they come from

The Owen et al. (2023) finding from Chapter 1 and Chapter 8 — that 54% of the world's energy-transition-mineral projects sit on or near Indigenous peoples' lands (69% counting peasant lands), with 62% of those in high-water-risk locations — does its full physical work in this chapter, because this is the chapter where the materials become specific objects rather than statistical aggregates.

GOBLIN FACTS — the chip is not weightless. The guide keeps returning to the 54% mineral-project figure because "cloud" infrastructure is physical infrastructure: extraction, water risk, power, cooling, land, and jurisdiction all travel with the hardware.

Walking through the major materials and their concentrated extraction sites:

Cobalt. Used in lithium-ion batteries (uninterruptible power supplies for data centres, the battery storage that pairs with renewable energy commitments) and in some semiconductor processes. Approximately 70% of global cobalt production comes from the Democratic Republic of the Congo, primarily from the Katanga and Lualaba regions. Industrial-scale extraction operates alongside extensive "artisanal" mining — the polite term for small-scale, often unregulated, often informal, often dangerous mining including substantial child labour. The cobalt supply chain has been the subject of extensive documentation by Amnesty International and other human rights organizations. A substantial share of the cobalt in the global electronics supply chain — the same chain data centres draw on — is mined in conditions, documented by Amnesty International among others, that would be unlawful in Canada.

Lithium. Used in the same battery systems. Major producers include Australia, Chile, Argentina, and China. The "lithium triangle" of Chile, Argentina, and Bolivia overlaps substantially with Indigenous territory: Quechua, Aymara, and Atacameño communities in the Andean salt flats where lithium brine is extracted. Bolivia's Salar de Uyuni is one of the world's largest lithium deposits and is on land with contested governance involving Indigenous communities, the Bolivian state, and Chinese, Russian, and other foreign extraction interests. Water consumption for lithium brine extraction in already water-stressed regions has produced documented harms to Indigenous agricultural communities.

Coltan (refined into tantalum). Used in capacitors throughout electronic equipment, including AI hardware. Concentrated extraction in the eastern Democratic Republic of the Congo, in regions of long-standing armed conflict. The "conflict minerals" designation under US Dodd-Frank Act Section 1502 was created substantially in response to coltan-supply-chain concerns, though the regulation's effectiveness remains contested.

Copper. Used throughout electronics, power infrastructure, and cooling systems. Major producers include Chile, Peru, the Democratic Republic of the Congo, the United States, and Canada. Copper mining in Canada — at sites including BC's Highland Valley, the Sudbury basin, and proposed projects in the Ring of Fire in northern Ontario — happens substantially on or adjacent to Indigenous territory, with First Nations' positions on specific projects varying from active opposition to negotiated participation. The Ring of Fire specifically has been the subject of long-running disputes among First Nations, the Ontario government, the federal government, and proposed mining companies.

Tungsten. Used in semiconductor manufacturing equipment and in some chip components. Major producers include China, Vietnam, Russia, and Bolivia. Chinese tungsten production has been the subject of US export restrictions and reciprocal Chinese restrictions on tungsten exports, making tungsten one of the materials at the centre of US-China AI-related trade conflict.

Rare earth elements (neodymium, dysprosium, terbium, others). Used in precision electronics and magnetic components. Approximately 60% of global rare earth mining occurs in China, primarily in Inner Mongolia and other regions, with significant environmental degradation including the Baotou tailings ponds. Australia is the second major producer, primarily from the Mount Weld mine. Smaller production is in Canada, the United States, and other jurisdictions. Chinese export restrictions on rare earths (imposed in late 2024 in response to US chip-export controls) have been a major factor in the broader AI-related trade tensions.

Silicon. The substrate of the chip itself. Sourced from ultra-pure silicon production primarily in Germany, Japan, the United States, and Norway. The most environmentally and geopolitically benign of the major chip materials, though the fabrication process that turns silicon into a chip is itself extremely water-, electricity-, and chemical-intensive.

The aggregate picture, in plain terms. The materials that make up the physical substrate of AI — every server, every chip, every battery, every cooling system — are extracted from a small number of geographically concentrated locations, often in jurisdictions with weaker labour and environmental protections than Canada, often on Indigenous land, often in zones of armed conflict, often in water-stressed regions, with environmental and human costs paid by communities that are largely invisible from the Canadian AI policy conversation. None of this is hidden. It is simply the way the physical substrate of AI gets produced, visible to anyone willing to follow the supply chain. And to be precise about what the Owen et al. 54% finding is: a count of where the extraction projects sit (on or near whose lands), not a measure of material volumes, and drawn from a study of energy-transition minerals broadly rather than computing hardware specifically. The count is the study's. The extension to AI is the guide's, and it rests on the overlap in mineral families, which is real and traceable.

A Canadian connection. Canada is itself a producer of several of these materials (copper, nickel, cobalt in smaller quantities, and some rare earths), primarily from mining operations that intersect with First Nations and Métis territory across the country. The Sudbury basin, the Ring of Fire, the Tahltan territory in BC, the James Bay lowlands. The 54% finding is not only about elsewhere. Canadian AI infrastructure, to whatever extent it is built on Canadian-sourced materials, is partly built on materials extracted from Canadian Indigenous lands. The sovereignty conversation in Chapter 9 connects directly to this. AI for All's "Sovereign Foundations" pillar does not engage the extraction layer at all.

---

Four: Cooling, water, and the heat problem

Every watt of electricity that enters a server emerges as either useful computation or waste heat, and at the physical level, almost all of it emerges as heat. A modern AI data centre is a large industrial heater. The continuous removal of that heat is what keeps the chips from damaging themselves within minutes.

The three approaches to heat removal, in their generic form:

Free-air cooling. Outside air, blown through the building, carrying heat away. Works well in cool climates for substantial portions of the year. Consumes electricity for fans but no water for cooling itself. Limited by air temperature: when outdoor temperatures exceed about 25°C, free-air cooling becomes insufficient and other methods must take over.

Evaporative cooling. Water is sprayed into the airflow or onto heat exchangers, absorbing heat as it evaporates. Very efficient at heat removal but consumes substantial water. A large data centre using evaporative cooling in summer months can consume hundreds of thousands to millions of litres of water daily. The water is typically lost to the atmosphere (the entire point of evaporative cooling is that water carries heat away by evaporating), not returned to the local watershed.

Mechanical refrigeration. Industrial-scale chillers, working on the same principles as residential air conditioning but at much larger scale. No water consumed for cooling itself, but substantial additional electricity required, often increasing total facility electricity consumption by 30-50% above what the computing equipment itself draws.

Most large data centres use a blend of these approaches, switching between them based on weather, water availability, and cost. Canadian data centres, as Chapter 6 will show, benefit from cold climates that enable substantial free-air cooling. This is a real Canadian advantage at the per-facility water-consumption level. It is also part of why Canadian per-facility water numbers are modest by global standards. It does not address the much larger water consumption upstream in the chip fabrication process, which happens overwhelmingly outside Canada and is not affected by Canadian climate.

A specific finding from the Luccioni paper introduced in Chapter 8: GPT-3 inference (the operational use, not training) consumes approximately half a litre of water per 10–50 medium-length responses (Li et al., 2023). At the scale of billions of queries per day across all major commercial AI services, the aggregate water cost of AI inference is now substantial. Most of this water consumption happens in data centres outside Canada, but Canadian AI workloads are part of the aggregate.

The heat itself: what gets done with it. In most current data centres, the heat removed from servers is simply released to the atmosphere through cooling towers, chiller exhaust, or warmed outflow air. Some facilities are beginning to capture and reuse waste heat: district heating systems, agricultural greenhouses, industrial processes. Canadian examples are limited but growing, including some Quebec projects that route data centre waste heat to local district heating systems. AI for All does not commit to waste-heat-recovery requirements; whether it should is a contested policy question with substantial European precedent (Stockholm and Helsinki both have major district heating networks supplied partly by data centre waste heat).

---

Five: Suzanne Kite's protocol — what hardware ethics looks like when it's specific

The Indigenous Protocol and AI Position Paper (2020), introduced in Chapter 1 and engaged in Chapter 3 (Whaanga) and Chapter 9 (OCAP/NISR/CARE), contains an essay by Suzanne Kite (Oglála Lakȟóta, multidisciplinary artist and academic) titled "How to Build Anything Ethically." The essay is one of the Position Paper's most concrete contributions, and it bears specific examination in the physical layer chapter because Kite is doing something the broader hardware ethics conversation rarely does: working through what a protocol for materially ethical hardware-building would actually involve, step by step.

A non-Indigenous-led guide cannot speak on behalf of Kite's framework. What it can do is present the essay's substance in its own terms and note what the practical implications would be if hardware were actually built this way.

Kite's protocol draws on Lakȟóta sweat lodge construction protocols (the relational, intergenerational, materially-specific process by which a sweat lodge is built and maintained) and applies the underlying logic to building "anything," including computing hardware. The essay's argument, distilled: building things ethically is not primarily a matter of intentions or principles. It is a matter of specific relationships maintained through the entire process of acquiring materials, working with materials, building, maintaining, and eventually unbuilding the resulting object.

The protocol involves, in Kite's framing:

Acquiring materials in good relations. The materials that go into the thing being built must be acquired in ways that honour the relationships with the beings (human, non-human, land) the materials come from. For a sweat lodge, this means specific protocols for harvesting saplings, gathering stones, sourcing fabrics. For computing hardware, applying the underlying logic would mean: cobalt extracted in conditions that honour relationships with the communities and lands of the Katanga region; lithium extracted in conditions that honour relationships with the Atacameño, Quechua, and Aymara communities of the Andean salt flats; rare earths extracted in conditions that honour relationships with the Inner Mongolian communities affected by Baotou tailings; copper extracted in conditions that honour relationships with the First Nations territories of British Columbia, the Sudbury basin, and the Ring of Fire.

Building with knowledge transmission. The thing being built must be built by people who know how to build it, with that knowledge passed forward to others through the building. For sweat lodges, this means elders and learners building together. For computing hardware, applying the underlying logic would mean: production processes structured to develop and transmit skilled labour, with workers having genuine agency and the work having genuine social value, rather than the current global structure of low-wage assembly labour concentrated in jurisdictions with weak protections.

Maintaining the thing across its life. The thing must be maintained in ways that honour what it is and what relationships built it. For sweat lodges, this means specific protocols for use, care, and repair. For computing hardware, applying the underlying logic would mean: hardware designed for repair, upgrade, and longevity rather than for planned obsolescence; supply chains for parts and skilled repair; reuse and refurbishment as the default pathway; and recycling that actually recovers materials rather than producing the global e-waste flows that Chapter 8 surfaced (1.2-5 million additional metric tons of AI-related e-waste projected by 2030, with only 22% of global e-waste formally collected and recycled).

Unbuilding when the time comes. Things that are no longer needed must be unmade in ways that return materials to good relations. For sweat lodges, this means specific protocols for disposing of materials respectfully. For computing hardware, applying the underlying logic would mean: end-of-life pathways that recover materials, return them to viable secondary uses, and avoid the dumping of e-waste in lower-income jurisdictions (Ghana, India, and others receive substantial illegal e-waste exports from wealthy countries, with documented health and environmental harm).

The framework is concrete and workable — not a vague invocation of "ethical AI" but a specific account of what ethical hardware-building would involve at the materials, labour, lifecycle, and disposal layers. It is also so far from how current AI hardware is actually built that the gap is itself the finding. Current AI hardware is built in conditions that violate Kite's framework at essentially every layer: materials acquired without good relations, assembly labour conducted under conditions that don't transmit knowledge or honour workers, hardware designed for short useful life, end-of-life pathways that mostly produce harm.

The guide notes this without claiming to speak for Kite or the framework. The practical implication is direct: Canadian AI policy that committed to something like this framework, even at modest scale, even in partial form, would commit to procurement requirements (no AI hardware produced under documented labour or environmental harms), to lifecycle requirements (refurbishment, repair, secondary use), and to recycling requirements (genuine material recovery, not e-waste export). None of these requirements appear in AI for All. Whether they could be added is contested. Whether they should be is a question the Position Paper has been asking, in published form, since 2020.

---

Six: The e-waste end of the cycle

What happens to AI hardware when it's no longer useful? This is the part of the supply chain most invisible from the policy conversation, and it bears specific attention because the answer determines whether the materials originally extracted return to use or compound the original extraction harm.

The global picture, in numbers from the UN's 2024 Global E-waste Monitor and from Wang et al. (2024) on AI-specific e-waste:

62 million tonnes of e-waste generated globally in 2022, an increase of 82% since 2010 and projected to reach 82 million tonnes by 2030.
Only about 22% of global e-waste is formally collected and recycled, meaning the materials are actually recovered and returned to use rather than landfilled or burned.
AI-specific hardware will add an estimated 1.2-5 million metric tons of e-waste annually by 2030, on top of the broader e-waste flow. This is from the Luccioni paper's citation of Wang et al. (2024).
Substantial illegal e-waste exports flow from wealthy countries to lower-income jurisdictions — Ghana (the Agbogbloshie site, until its 2021 demolition), India, Pakistan, China (before tightened restrictions), and others — where informal recycling produces documented health and environmental harms including heavy-metal contamination, dioxin exposure, and groundwater pollution.

Canada's specific position. Canada produces approximately 1 million tonnes of e-waste annually (a per-capita rate similar to other high-income countries). Provincial extended producer responsibility programs exist in several provinces: Ontario, Quebec, BC each have producer-fee-funded e-waste recycling programs. Federal e-waste regulation is minimal. The Canadian rate of formally-collected e-waste is around 20%, roughly the global average, meaning 80% of Canadian e-waste is unaccounted-for in formal recycling streams, mostly going to landfill or to informal export channels.

The structural problem the AI conversation creates. AI hardware turns over rapidly. NVIDIA's H100 was the standard in 2024; the H200 was the standard in 2025; the Blackwell generation is the standard for 2026. Each generation makes the previous generation less competitive economically, and operators of data centres face pressure to upgrade rapidly. The hardware being retired is often still functional, just not competitive at the leading edge. Whether retired hardware finds a secondary market in less-demanding workloads or is sent to e-waste depends on operator choices and on the economic terms of secondary markets.

The federal AI strategy does not address e-waste. Canadian data centres are not required to publish hardware-lifecycle information. The materials extracted to build the chips currently being deployed will, on the current trajectory, be largely lost to landfill or informal recycling within five to ten years. The closed-loop ideal Kite's framework points toward — where materials are recovered and returned to use — is operationally absent from the current Canadian AI infrastructure conversation.

🧌 GOBLIN CHECK — Canada formally recycles about one-fifth of its e-waste. The other four-fifths is officially "unaccounted for," which is auditor-speak for on an adventure. The goblin — a creature that hoards things on purpose, professionally — finds the industry's accidental hoarding-in-landfills frankly insulting to the craft. If these materials were worth displacing communities to dig up, they're worth a system to get back.

This is one of the specific silences the guide surfaces. It isn't the loudest, but it may be the most structurally consequential, because the e-waste end of the cycle determines whether the original extraction harms are bounded (the materials get reused and the original extraction was a one-time cost) or compounding (the materials are lost and each generation requires fresh extraction).

---

Adding up the physical layer

CHAPTER RECAP — you now have: - A working physical picture of AI infrastructure from data centre building down to substrate materials — buildings, racks, servers, chips, silicon, and the materials inside the chip. - The chip supply chain mapped — US-designed, primarily Taiwan-fabricated, Dutch-equipment-dependent, with concentrated material sources — and the Canadian position in that supply chain (limited at every layer, with no near-term path to silicon-layer sovereignty). - The materials underneath the chips placed in their geographic and political context — cobalt from the DRC, lithium from the Andean lithium triangle, copper from BC and Ontario First Nations territories, rare earths primarily from China, with Owen et al.'s 54% finding doing its physical-substance work here. - The cooling and water dimension at the chip-and-data-centre level, with Canadian climate as a real per-facility advantage that doesn't address upstream chip-fab water consumption. - Suzanne Kite's "How to Build Anything Ethically" framework engaged substantively — the Lakȟóta protocol for hardware building applied operationally to the materials, labour, lifecycle, and disposal layers, with the gap between Kite's framework and current AI hardware practice named honestly. - The e-waste end of the cycle as the structural silence in Canadian AI policy — current trajectory loses most materials to landfill or informal export, with closed-loop alternatives operationally absent from AI for All.

This completes the guide's Part I Foundations. Chapters 1-4 together establish the conceptual moves (Chapter 1), the technical mechanics of how AI actually learns (Chapter 2), the substantive content of training data with all its political weight (Chapter 3), and the physical substrate the whole system runs on (Chapter 4). The Canadian-arc chapters (5-9, already drafted) make sense more solidly with these foundations in place. The contested-questions chapters that follow (10-14) will deepen specific threads from these foundations.

You can now read any Canadian AI infrastructure claim with the equipment to ask the right questions: what physical hardware, designed where, fabricated where, from materials extracted where, by whom, with what labour conditions, with what end-of-life pathway, with what waste-heat recovery, with what water consumption upstream and downstream? Most current claims will fail most of those questions. The failures are the structural conditions of current AI hardware, not the exceptions, and reading them clearly is the precondition for any serious infrastructure-layer policy work.

The next chapter (Chapter 5, already drafted) takes the Canadian-policy story from the AI for All launch through the response cycle, with the foundations now in place to make sense of what the strategy commits to, what it doesn't, and what the political stakes of those choices actually are at the practical level the foundations chapters have surfaced.

---

Bias label for this chapter: technical and political-economic analysis of AI hardware. Author lean: skeptical of "infrastructure" framings that obscure the physical substrate; willing to track the supply chain from chip design through fabrication through materials through extraction through assembly through deployment through end-of-life; sympathetic to Indigenous-led framings (Kite, Owen et al. implications) that engage the materials and labour layers substantively; explicit about the limits of a non-Indigenous-led manual representing Indigenous-led frameworks. Industry sources (NVIDIA technical specifications, TSMC operational disclosures) labelled and read accordingly. Independent academic sources (Wang et al., Luccioni et al., Owen et al.) treated as primary on their respective domains. Indigenous-led primary sources (Kite essay in the Indigenous Protocol and AI Position Paper) presented in their own terms with operational implications named.

Primary sources cited or relied on in this chapter: NVIDIA technical documentation on H100 architecture; TSMC public disclosures on fabrication; ASML public disclosures on EUV lithography; Owen, Kemp, Lèbre, Svobodova & Pérez Murillo, Nature Sustainability 2023; Wang et al. (2024) on AI-related e-waste projections; Li et al. (2023) on AI inference water consumption; Luccioni, Strubell & Crawford (FAccT 2025); UN Global E-waste Monitor 2024; Kite, "How to Build Anything Ethically" in Indigenous Protocol and AI Position Paper (2020); Amnesty International documentation on cobalt supply chain; US CHIPS and Science Act (2022) text and analysis. Detailed citations in the Sources appendix.

---