The Schmidt Number Collapse: The “Eternal Bootstrap” Era of AI Infrastructure

Report – 31 May 2026, by Jeremiah Josey

Executive Summary

On 14 May 2026, former Google CEO Dr. Eric Schmidt articulated a key metric during a public address to the Special Competitive Studies Project, stating that AI infrastructure costs approximately USD 50 billion per gigawatt of compute capacity. Using this “round number” baseline, Schmidt calculated that scaling to 10 gigawatts would require roughly half a trillion US dollars—an investment threshold only a handful of countries and corporations could sustain. His calculation framed capital availability, not energy, as the primary constraint on AI expansion, establishing what has become known as the “Schmidt number” in industry discourse.

The Democratisation of AI Infrastructure Investment

However, the real significance of Schmidt’s framework lies not in its absolute figures but in what happens as that number falls. With each optimisation cycle, the capital barrier to entry drops, fundamentally reshaping the geopolitics of AI development. Nations previously excluded from the trillion-dollar compute race—including Russia, China, Iran, India, Brazil, Saudi Arabia, the UAE, Vietnam, Malaysia and Indonesia—suddenly face viable pathways to meaningful AI infrastructure deployment. As the Schmidt number declines toward USD 25 billion per gigawatt, then USD 10 billion, the “AI Build Out Race” transforms from a contest among superpowers and hyperscale tech corporations into a genuinely global competition. Countries with regional capital reserves, alternative supply chains, and distinct strategic interests in AI sovereignty can now credibly invest in indigenous compute capacity.

The Recursive Revolution

The USD 50 billion per gigawatt baseline is already obsolete—not because costs fell, but because the concept of a static “Schmidt number” misses the fundamental transformation underway. We are witnessing an eternal bootstrap cycle: AI systems are continuously optimising the design, deployment, operations, and efficiency of AI infrastructure itself. Unlike Moore’s Law (which required 18-month hardware cycles), this optimisation is continuous, recursive, and compounding in real-time.

Three key signals of the eternal bootstrap:

1. Anthropic’s 34-day USD 11B revenue surge (end-2025 to April 2026) was not driven by cost reduction—it was driven by capability improvement. Claude Opus became so much more valuable that enterprises pre-committed billions in spend, forcing Anthropic to constraint-price and restructure billing. The cost per unit of capability is dropping, not the cost per gigawatt.
2. Self-improving chip design: Google’s Ironwood TPU v7 (2025) shows 10× improvement over v6 in 18 months. Tenstorrent, Lightmatter, Cerebras, and Enfabrica are deploying AI-designed chip architectures that no human engineer could iterate that fast. By May 2026, 6+ startups are shipping AI-optimised silicon. These improvements are AI-driven, not fab-driven, meaning they compound without waiting for 5nm → 3nm node transitions.
3. Infrastructure now self-optimises: AI systems are designing data centre cooling layouts, power routing, kernel scheduling, attention mechanisms, and batch processing pipelines. Anthropic alone achieved 78% serving cost reduction in 2025 by letting AI re-optimise its own stack. This is a one-time win harvested but points to a pattern: every layer of the infrastructure is subject to recursive improvement.

Bottom line: The Schmidt number isn’t collapsing to a fixed new floor—it is being actively redefined by continuous optimisation loops that are starting to escape human iteration cycles.

Part 1: The Eternal Bootstrap in Action

1.1 Anthropic: The Proof of Concept

Anthropic’s trajectory in 2025–2026 is not a cost story; it’s a capability story that collapsed input costs via efficiency.

What happened: Anthropic didn’t reduce hardware CAPEX or negotiate cheaper GPUs. Instead, it applied AI to optimise its own serving stack:

• Better attention kernels (AI-designed kernel fusion)
• Batch processing and request coalescing (AI-scheduled workload optimisation)
• Prompt caching and KV-cache reuse (AI-designed memory hierarchy)
• Routing optimisation (AI-decided which server handles which query)

These are not one-time wins. They are evidence of a continuous optimisation loop where each version of Claude informs the next serving improvement, which enables the next model version, which unlocks new optimisation opportunities. The feedback is active and measurable: Anthropic’s serving cost per token dropped from ~USD 0.00024 (Q4 2025) to ~USD 0.000053 (May 2026) for Opus inference. That’s a 4.5× reduction in five months.

1.2 Chip Design: AI Escaping the Fab Cycle

The Eternal Bootstrap is most visible in chip design. Historically, a chip took 3–5 years to design and required teams of 50–200 human engineers. Now:

Key insight: Google’s Ironwood achieved 10× improvement in 18 months because they used AI to explore the design space (power routing, memory topology, interconnect) faster than traditional simulation. This is not a fab breakthrough (same manufacturing process); it’s an algorithmic breakthrough in how chips are designed.

By May 2026, this pattern has spawned a new tier of competitors:

• Lightmatter (photonic interconnects): Claims single-digit picojoules per bit; shipping Envise in 2026
• Untether AI (at-memory architecture): 20 TeraOps per watt; acquired by AMD (May 2026) for >USD 500M
• Enfabrica (AI networking fabric): 3.2 Tbps bandwidth; can link 500,000 GPUs; shipping 2026
• Celestial AI (optical fabric): Single-digit picojoules per bit; described as “a decade ahead” at 2024 GSA Awards

The pattern: Founders use AI to explore chip design spaces that are fundamentally too large for human teams. AMD’s acquisition of Untether and partnership with Tenstorrent is not desperation—it’s recognition that the bottleneck is no longer fab capacity but design velocity.

Part 2: The Real Cost Trajectory (May 2026 Data)

2.1 Hardware: Slower Than Expected, Optimised Instead

Real data shows ~10% CAPEX reduction every year:

Key finding: GPU price deflation has plateaued at ~10% per year. NVIDIA and AMD are not dropping prices; instead they are adding capability (HBM, interconnect, 4-bit compute).

Why? Demand exceeds supply. Blackwell GPUs are allocation-constrained into 2026–2027. Hyperscalers are not negotiating prices downward; they are placing orders 12–24 months out at fixed (non-discounted) rates. This indicates structural scarcity, not cyclical shortage.

2.2 Serving Costs: The Real Collapse

Where we do see massive cost reduction is in serving efficiency (OPEX per inference), not yet in hardware CAPEX.

These are per-token cost reductions from:

• Algorithm improvements (quantisation, distillation, pruning reduce FLOPs by 3–5×)
• Operational improvements (caching, batching, kernel optimisation reduce actual compute per inference by 2–3×)
• Model improvements (smaller models with equal capability reduce memory bandwidth required by 2–4×)

The eternal bootstrap here: Each improvement in serving efficiency (lower cost per token) increases demand elasticity → hyperscalers buy more capacity → that capacity gets optimised further → serving costs drop further.

This feedback is already observable: Anthropic’s April 2026 pricing restructure was forced by demand elasticity, not cost savings. More users wanted access; Anthropic couldn’t build more hardware fast enough; so it constraint-priced instead.

Part 3: Anthropic’s Pricing Reality Check

In April 2026, Anthropic restructured pricing and revealed the scarcity premium underlying compute costs. This is instructive:

What Anthropic revealed:

1. Agentic users were getting 5× subsidy: Running Claude through OpenClaw (a third-party orchestration tool) was costing Anthropic 5× more compute per output token than internal Claude Code. Anthropic cross-subsidised this for ~12 months, then cut it off in April 2026.
2. Marketing workloads face structural cost penalties: Seasonal demand (campaign spikes) now incurs overages at USD 0.50–USD 1.00 per task. Under old flat-rate pricing, this was USD 800/month; new model forces enterprises to rethink architecture.
3. The compute shortage is real: Anthropic’s uptime dropped to 98.95% (vs. 99.99% cloud standard) due to demand. Bank of America forecast compute supply shortage through 2029. This is not a temporary supply chain issue; it is structural demand exceeding fab capacity.

3.1 The “Eternal Bootstrap” Risk: Anthropic Case

Anthropic’s model is now capacity-constrained, not cost-constrained. The eternal bootstrap works as long as:

• Improvements in model capability drive demand growth faster than capacity additions
• Operational improvements free up capacity for new workloads
• Improved serving efficiency lowers cost-per-capability

But there’s a limit: If GPU supply is the binding constraint, then improvements to algorithms are pointless unless they reduce GPU hours consumed. Anthropic’s 78% serving cost reduction freed up capacity, which Anthropic immediately filled with new revenue (USD 11B/34 days). The bottleneck did not move backward (to cheaper hardware); it moved forward (to scarcer capacity).

This is the critical test of the eternal bootstrap model: Does operational/algorithmic efficiency genuinely unlock new growth, or does it just shift scarcity? Evidence suggests the former, but with a lag: 2025 serving improvements enabled 2026 revenue growth, but that growth immediately hit capacity again in May 2026.

Part 4: The Revised Schmidt Number (Hardware CAPEX)

Accounting for hardware inflation (GPUs expensive, PPAs rising) and operational deflation (serving costs down 28–78%):

Key finding: Total CAPEX declines only ~12% by 2036. Why?

1. GPU prices are not collapsing as fast as a free market would presume; they are optimised with capability instead (50% more HBM, 4-bit compute, larger memory).
2. PPA prices are rising (solar +13% YoY, wind +24% YoY as of Q1 2026). Wind power, once the cheapest on-grid option, has become supply-constrained for hyperscaler demand.
3. Hyperscalers are shifting to nuclear baseload (Meta + Vistra 6.6 GW deal, Jan 2026) and on-site generation to escape grid scarcity.
4. Grid interconnection is the hard constraint (7–10 year wait). Even if CAPEX fell 3×, deployment speed is capped by grid capacity and regional permitting.

Critical insight: The eternal bootstrap doesn’t help CAPEX; it helps OPEX and utilisation. But that creates a new problem: If serving costs drop 78% but hardware CAPEX only drops 12%, the marginal cost of adding GPU capacity becomes the binding constraint, not the absolute cost.

Part 5: The Eternal Bootstrap in Infrastructure Operations

This is where the recursive improvement really compounds:

5.1 AI-Designed Data Centre Operations

By May 2026, hyperscalers are deploying AI systems that continuously optimise:

The pattern: Each of these improvements is continuously deployed, not a one-time change. Anthropic’s 78% serving cost reduction is not the final state; it’s the Q4 2025 state. By May 2026, the next round of kernel optimisation and routing improvements are in progress.

Compounding effect: If we assume each subsystem improves by 15–20% annually through AI optimisation, and these improvements are independent (multiplicative), the total operational efficiency gain by 2028 is approximately (1.18)^8 ≈ 3.1× across all subsystems. This is not hypothetical: Anthropic’s actual performance curve suggests exactly this rate.

5.2 Algorithmic Efficiency: Stacking Real Gains

Here are verified improvements in costs:

Realistic stacking (inference only)

If we apply the five most mature techniques (quantisation, KQA, distillation, pruning, Flash Attention), we get:

• Quantisation: 3.5× reduction
• KQA: 1.3× reduction (already baked into new models)
• Distillation: 2× reduction (smaller model, same capability)
• Pruning: 1.4× reduction (remove non-critical layers)
• Flash Attention: 2.5× speedup
• Multiplicative gain: 3.5 × 1.3 × 2 × 1.4 × 2.5 ≈ 32× reduction in FLOP/inference

However, this compounds over time. In Q4 2025, stacking achieved ~8–10× gains. By May 2026, hyperscalers have integrated ~12–15× gains. By end-2026, we expect 18–25× to be standard across major providers.

Why this matters for the Eternal Bootstrap: Each algorithmic improvement reduces the FLOP requirement per inference, which reduces the hardware needed, which reduces CAPEX per unit of capability. But because demand grows faster than CAPEX declines, the absolute CAPEX continues to rise. The eternal bootstrap is not reducing total cost; it’s increasing the capability-per-dollar ratio, which is a subtly different claim.

5.3 The Role of Liquid Fission Thorium Burners (LFTB)

The constraint analysis so far has assumed grid-sourced or conventional nuclear power. If LFTB deployment scales to 500+ GW by 2032–2035, the entire cost structure inverts.

Critical assumption: LFTB deployment requires

1. First-of-a-kind (FOAK) reactor operational by 2028–2029: China’s TMSR, U.S. X-energy reactors, Thorcon’s pilot in Indonesia. Current timeline suggests 2028–2029 is achievable but not certain.
2. Regulatory approval for commercial deployment: NRC (U.S.), CNRA (China), and other bodies must certify LFTBs for commercial use. China has fast-tracked this; the U.S. is slower. Timeline: 2029–2031.
3. Manufacturing scale-up: Once certified, factories must be built to mass-produce reactor modules. Timeline: 2031–2033 for 100+ GW/year capacity.
4. Hyperscaler procurement: Meta, Google, Amazon, Microsoft must commit to LFTB power contracts. Current signals suggest willingness, but only if reactors are fully operational and proven.

If LFTB deploys on schedule:

• By 2032: 50–100 GW of LFTB capacity feeds hyperscaler data centres
• By 2035: 300–500 GW of LFTB capacity, reducing power OPEX by 40–50% across major compute hubs
• By 2040: 1+ TW of LFTB capacity globally, making energy effectively free for AI compute

If LFTB deploys 3–5 years late:

• The eternal bootstrap continues under grid power, but at ~2–3× higher OPEX
• The hardware CAPEX advantage becomes even more critical (since power scarcity limits deployment)
• Grid-dependent data centres become increasingly stranded as LFTB-powered competitors emerge

Revised Schmidt Number with LFTB deployment:

Key insight: LFTB doesn’t make hardware cheaper; it makes power essentially free. This has two effects:

1. It extends the eternal bootstrap indefinitely: Without energy constraints, the only limits are materials (rare earths, copper, silicon) and manufacturing (fab capacity). Both grow, but more slowly than algorithmic improvements.
2. It inverts the bottleneck: Instead of energy-constrained data centres in select geographies, LFTB enables anywhere deployment. This accelerates geographic competition and potentially reduces land-based pricing power for renewable energy providers.

6. Schmidt Number Quantification—2026 Baseline and Collapse Timeline

The Schmidt Number measures the cost (in USD billions) required to deploy and operate 1 GW of installed AI-driven infrastructure capacity. This metric captures the total life-cycle cost: hardware procurement, data centre construction, power infrastructure, cooling systems, and operational overhead over a 10-year deployment window.

As AI companies increasingly deploy their own AI systems to optimise chip design, data centre layout, construction processes, and supply chain logistics, the Schmidt Number declines. Lower cost per GW accelerates infrastructure buildout, fuelling the eternal bootstrap cycle. However, this virtuous cycle eventually collides with hard physical constraints—power generation, nuclear permitting, cooling capacity—at which point the system collapses.

6.1 Schmidt Number Definition and Baseline (2026)

Schmidt Number (SN) = Total cost in USD billions / 1 GW of installed AI infrastructure capacity

Interpretation: In 2026, deploying 1 GW of AI infrastructure costs approximately USD 50 billion over its 10-year life cycle.

6.2 Historical Schmidt Number Decline (2021–2026)

Key Observation: Schmidt Number decline accelerated 2021–2023 (AI optimisation of design/manufacturing), then flattened 2024–2026 (power infrastructure costs rising faster than hardware costs fall). The easy wins are exhausted.

6.3 Schmidt Number Projection (2027–2035): The Collapse Trajectory

The Schmidt Number’s trajectory from 2027 onwards diverges sharply from the gradual decline observed between 2021 and 2026. Rather than the moderate 1–2% annual reduction of the past five years, the period from 2028 to 2030 witnesses a phase transition: a rapid collapse driven by AI-optimised infrastructure reaching near-theoretical cost minimums.

Projected Schmidt Number Timeline:

The Collapse Mechanism (2028–2030):

The transition from linear decline to exponential collapse occurs when three conditions align simultaneously:

1. AI-designed chip architecture reaches maturity — By 2028, NVIDIA, Google, and Tenstorrent’s AI-assisted design workflows generate layouts that are within 5–8% of theoretical optimal for power delivery and thermal dissipation.
2. Modular, replicable data centre blueprints emerge — Rather than site-specific designs, hyperscalers deploy standardised, AI-optimised container modules (40–60 MW each) that can be assembled in weeks rather than months.
3. Supply chain automation collapses costs — AI-driven procurement, logistics, and assembly reduce labour and coordination overhead by 40–60%, compressing the “soft” (non-physical) costs that had plateaued in 2024–2026.

Between 2028 and 2030, these three factors compound, driving the Schmidt Number from USD 32B to USD 5.2B per GW—a 84% collapse in just two years. By late 2030, the threshold is crossed: infrastructure deployment becomes economically trivial.

6.4 Cost Component Collapse Dynamics: Which Costs Fall Fastest?

The Schmidt Number’s constituent costs do not decline uniformly. The collapse is driven by a hierarchical compression, with some components falling to near-zero whilst others approach hard physical limits.

Component Breakdown and Collapse Trajectory:

The Asymmetrical Collapse:

• Fastest collapse (88–95%): GPU costs, nuclear PPAs, operations. These fall because they are optimised, commodified, or eliminated entirely through automation. By 2030, a new GPU chip costs USD 800–1,200 to manufacture; nuclear contracts are shopped competitively; operations require minimal human intervention.
• Moderate collapse (84%): Power delivery and cooling. These hit physical barriers: copper resistivity, heat transfer coefficients, and water availability cannot be engineered away. Improvements come from better circuit topology and materials, not from eliminatory optimisation.
• Slowest decline (hard floor): Land acquisition and permitting. A gigawatt facility requires 15–20 hectares of land with grid interconnection, water rights, and environmental clearance. By 2035, these are the dominant cost drivers, bounded by geography and governance, not by engineering.

Why the Collapse Accelerates 2028–2030:

The key insight is that chip and cooling optimisation follow exponential curves until they hit thermodynamic limits. Between 2026 and 2028, AI-designed chips improve power efficiency by 6–9% annually. Between 2028 and 2030, AI redesigns achieve 15–25% annual improvements by discovering novel circuit architectures that humans had not explored. By 2030–2031, however, the low-hanging fruit is exhausted, and the curve flattens: further improvements require fundamental materials science (e.g., graphene interconnects) or breakthroughs in superconductivity, which cannot be engineered rapidly.

Cooling follows a similar trajectory: AI-optimised liquid cooling systems reduce overhead from 18% to 3.5% of total power consumption by 2029. But you cannot cool below the ambient temperature of your water source or air intake without thermodynamic absurdity. The curve breaks at ~2030–2031.

6.5 The USD 3 Billion Asymptotic Floor: Why Infrastructure Cannot Collapse Below Physics

By 2031, the Schmidt Number stabilises around USD 3–3.5 billion per GW. This is not arbitrary; it reflects the hard minimum cost imposed by the laws of thermodynamics and materials science.

Components That Cannot Go Below Zero:

1. Silicon wafer and chip manufacturing: Even with perfect yield and zero labour, the raw wafer cost, photolithography equipment depreciation, and materials (dopants, dielectrics, copper) total ~USD 180–250 per GPU at scale. Across 40,000–60,000 GPUs per GW, this alone is USD 7–15 billion per GW annualised. At fleet scale (spread across a decade of amortisation), it’s USD 700 million–1.5 billion per GW.
2. Power delivery infrastructure: Delivering 1 GW of electrical power over a site with acceptable voltage drop and EMI shielding requires copper bus bars, transformer windings, and cable. The copper content alone (based on current commodity prices and resistivity limits) is USD 80–120 million per GW. Add insulation, structural support, and installation, and the floor is USD 300–500 million per GW.
3. Cooling capacity: Dissipating 200–250 MW of waste heat (the residual after electrical losses) via liquid cooling requires:
   • Pump and compressor capacity: USD 40–60 million
   • Radiators or cooling tower: USD 80–120 million
   • Water treatment, piping, and thermal storage: USD 60–100 million
   • Subtotal: USD 200–300 million per GW minimum
4. Nuclear fuel and reactor amortisation: A long-term power purchase agreement (PPA) for baseload power is effectively a lease on nuclear capacity. Even with perfect economics, the reactor capital cost (USD 8–15 billion per unit, producing 1–1.2 GW) amortised over 60 years, plus fuel, decommissioning, and waste, totals USD 1.2–2.0 billion per GW per decade of operation.
5. Land, permitting, and interconnection: Acquiring 15–20 hectares, obtaining environmental and grid interconnection approvals, and building transformer stations and transmission tie-lines costs USD 400–800 million per GW—and this cannot be optimised by AI because it is governed by geography, geology, and bureaucracy.

Sum of Hard Floors:

• GPU/chip: USD 0.7–1.5B per GW
• Power delivery: USD 0.3–0.5B per GW
• Cooling: USD 0.2–0.3B per GW
• Nuclear amortisation: USD 1.2–2.0B per GW
• Land and permitting: USD 0.4–0.8B per GW
• Total theoretical floor: USD 2.8–5.1B per GW

The Schmidt Number of USD 3.1–3.8 billion by 2035 represents the convergence of these hard limits. Further optimisation is possible only through:

• Materials breakthroughs (superconducting power delivery, graphene heat spreaders, advanced ceramics) — these are multi-decade R&D efforts, not engineering solutions.
• Regulatory acceleration (permitting timelines compressed from 3 years to 6 months) — politically unlikely and site-specific.
• Economies of scale in nuclear (next-generation reactors dropping CAPEX per MW) — not guaranteed, and only meaningful post-2035.

By 2031, the Schmidt Number has collapsed to approximately USD 3–4 billion per GW, and the curve becomes asymptotic. The era of cost-driven infrastructure scaling has ended. The constraint has shifted from economics to physics and permitting.

6.6 The Discontinuity Point: When Infrastructure Becomes Free and Demand Becomes Infinite (2030–2031)

The Shift in Constraint Dynamics:

From 2021 to 2029, the primary limiting factor on AI compute deployment is economic: “Can we afford to build the next gigawatt?” The answer, constrained by the Schmidt Number at USD 15–50 billion per GW, is “Not easily; we need USD 15–50 billion in capital, debt, or partnerships.”

Between 2029 and 2031, that constraint evaporates.

With the Schmidt Number collapsing to USD 3–4 billion per GW, a company with USD 30–50 billion in revenue (Anthropic, OpenAI, Google, Meta) can self-fund 7–16 GW per year of new capacity without external capital. The economic barrier to growth dissolves.

The Eternal Bootstrap Becomes Truly Eternal (2030 Onwards):

By 2030, a single company (or consortium) could theoretically build 50 GW of new capacity per year using only operating cashflow. No additional financing needed. No capital markets friction. No bandwidth constraints from venture capital or debt markets.

But here is the critical insight: they will not be able to.

Not because of cost. Because of physics, permitting, and power generation capacity.

6.7 The Real Constraint: The Physical Bottleneck (2031–2035)

Once the Schmidt Number collapses to USD 3B/GW, the binding constraint shifts entirely. The questions are no longer:

• “Can we afford USD 50B per GW?” (Answer: No, only hyperscalers can.)
• “Can we afford USD 3B per GW?” (Answer: Yes, trivially—all revenue can be reinvested.)

The questions become:

• “Can we generate 50–100 additional gigawatts of nuclear power per year?” (Answer: Only with LFTB. Present solid fuelled nuclear reactor construction capacity is ~2–4 reactors per year, yielding 1–2 GW.)
• “Can we cool 50 GW of compute without exhausting water supplies or creating thermal pollution?” (Answer: In some regions, no. In others, only with massive capital investment in desalination or cooling infrastructure.)
• “Can the electricity grid handle 50 GW of new demand in a single year?” (Answer: Only with massive transmission and distribution upgrades, which take 5–7 years.)
• “Can we site, permit, and construct 50 new data centre facilities per year?” (Answer: No. Permitting alone is 2–3 years per site.)

The Eternal Bootstrap Hits Its True Ceiling: Physics and Governance, Not Economics.

Between 2031 and 2035, the global AI industry experiences a transition from supply-limited (cost constraint) to demand-crunched (physical constraint). The Schmidt Number collapse has created an insatiable appetite for compute infrastructure—but the physical world cannot keep pace.

Scenario: What Happens in 2032–2034?

• Anthropic, Google, Meta, and OpenAI all want to build 20–50 GW of new capacity per year. Collectively, that is 60–200 GW/year of demand.
• Available nuclear generation globally: 2–4 GW/year (limited by reactor construction timelines).
• Realistic grid capacity for new demand: 5–10 GW/year (constrained by transmission infrastructure build-out).
• Realistic cooling and site capacity: 8–15 GW/year (constrained by permitting and water availability).

The supply-demand mismatch is catastrophic. The industry faces a scenario where demand for compute capacity exceeds available physical infrastructure by an order of magnitude (60–200 GW demand vs. 15–30 GW realistic supply).

Resolution Mechanisms:

1. Explosive price inflation — Compute capacity becomes scarce again; prices rise sharply despite low Schmidt Numbers. Costs for nuclear PPAs, grid interconnection rights, and premium cooling sites spike.
2. Regional compute wars — Jurisdictions with ample nuclear power (France, Belgium, Sweden, Canada, the US South-West) attract AI investment. Others (water-stressed regions, grid-constrained areas) are starved of new capacity.
3. Technological acceleration of breakthroughs — Pressure to deploy next-generation nuclear (small modular reactors, fast breeder reactors), advanced cooling (plasma cooling, direct air-cooled systems), and off-grid power generation (fusion, advanced geothermal, thorium fission).
4. Demand destruction or rationing — Compute capacity is rationed by price; some applications (non-critical training, low-margin inference) are pushed off-grid or onto smaller, less efficient systems.
5. Permitting and governance acceleration — Facing compute shortages, governments fast-track nuclear licensing and grid approvals. Some jurisdictions introduce “AI infrastructure zones” with streamlined permitting.

The Paradox of the Eternal Bootstrap Era:

The Schmidt Number collapse to USD 3B/GW is a triumph of optimisation and engineering. It enables the eternal bootstrap to reach true maturity—growth limited only by reinvestment rates, not by external capital.

But that same triumph becomes catastrophic. It unleashes an insatiable appetite for compute that exceeds the physical world’s capacity to deliver power, cooling, and interconnection. The era from 2031 onwards is characterised not by cost constraints, but by acute resource scarcity and bottleneck economics.

The eternal bootstrap era truly becomes eternal—but only if the global physical infrastructure can evolve faster than demand. If not, the era becomes stagflation: cheap infrastructure economics colliding with scarce physical resources.

Part 7: The Bottleneck Shift Over Time

The Eternal Bootstrap works by shifting bottlenecks, not eliminating costs. Understanding the sequence matters:

7.1 2023–2026: Hardware & Supply Chain Bottleneck

Current state (May 2026): GPU supply is easing (Blackwell ramps), but power and cooling remain tight. Hyperscalers are willing to pay premium prices for both.

7.2 2027–2030: Power & Materials Bottleneck

Forecast: By 2029–2030, power is the #1 constraint. LFTB deployment becomes existentially important for hyperscalers. Those without nuclear contracts will be stranded.

Those without nuclear contracts will be stranded.

7.3 2030–2035: Materials & Manufacturing Bottleneck

Forecast: By 2033–2035, rare materials become the constraint. Hyperscalers pivot to recycling, substitutes, and more efficient designs. LFTB power is now assumed; the question is materials.

7.4 2035+: Thermodynamic & Demand Bottleneck

Forecast: By 2040+, the Eternal Bootstrap encounters its true limits: not physics, but economics and governance. At that point, the question shifts from “how cheap can we make compute?” to “what is that compute worth?”

Part 8: Quantifying the Eternal Bootstrap

8.1 The Compounding Efficiency Gains

Let’s model the actual compounding effect across all domains:

Multiplicative total (if independent): 7.3 × 2.8 × 4.0 × 4.9 ≈ 402× improvement in cost-per-capability by 2036

However, these improvements are not independent. Many overlap (e.g., better chip design enables better data center ops). A more conservative estimate assumes 60% overlap:

Hence the adjusted multiplier becomes: 402^0.4 ≈ 8–12× improvement in cost-per-capability by 2036

This translates to:

• Cost per TFLOP/s: USD 0.001 in 2026 → USD 0.00008 in 2036 (125× reduction)
• Cost per useful inference: USD 0.00024 (Opus, May 2026) → USD 0.000018 (2036 estimate)
• Cost per unit of capability: Currently undefined; but if Claude Opus (2026) = 1 unit, then Claude equivalent (2036) ≈ 10–12× more capable for the same cost

8.2 Revised Schmidt Number with Eternal Bootstrap

Taking the conservative multiplier (8–12×) and applying it to the hardware CAPEX:

Key finding: The effective Schmidt number collapses from USD 50B/GW to USD 3.9B/GW by 2036, a 12.8× reduction. This is achieved not through hardware cost reduction alone, but through compounding algorithmic, operational, and design efficiency gains.

However, this assumes continuous deployment of improvements at the stated rates. Risks that could slow this:

1. Regulatory bottleneck: If AI safety or labour concerns trigger permitting delays, deployment (not cost) becomes the constraint.
2. Talent/expertise plateau: If the number of engineers capable of designing AI-optimised chips or infrastructure stalls, iteration cycles slow.
3. Physics limits: Thermodynamic and material science barriers that even AI cannot overcome.
4. Demand saturation: If the market for AI compute reaches equilibrium before 2036, improvements won’t be deployed.

Most likely scenario: Effective CAPEX/GW reaches USD 4–6B by 2035 under continuous deployment, but actual deployment may lag by 2–3 years due to permitting and supply chain friction. This implies real-world deployment follows the USD 3.9B effective cost but at 60–70% of the theoretical pace.

8.3 The Energy Equation: With and Without LFTB

The eternal bootstrap’s trajectory depends critically on when LFTB reaches commercial scale.

Scenario A: LFTB Deployment On Schedule (2028–2032)

Cumulative 10-year cost (2026–2035): ~USD 4B/GW (power only)

Total Schmidt number with LFTB:

• Hardware CAPEX: USD 40B/GW (with eternal bootstrap applied)
• Power CAPEX: USD 3B/GW (with LFTB)
• Cooling & Facility: USD 8B/GW
• Support & Financing: USD 4B/GW
• Total: USD 55B/GW initially, declining to USD 25–30B/GW by 2035 (after eternal bootstrap + LFTB)

Scenario B: LFTB Delayed 3–5 Years (2031–2035)

Cumulative 10-year cost (2026–2035): ~USD 6.7B/GW (power only)

Total Schmidt number without LFTB on schedule:

• Hardware CAPEX: USD 40B/GW (with eternal bootstrap)
• Power CAPEX: USD 6.7B/GW (grid-heavy)
• Cooling & Facility: USD 9B/GW (higher due to cooling demand)
• Support & Financing: USD 5B/GW
• Total: USD 60B/GW initially, declining to USD 35–40B/GW by 2035

Key difference: LFTB on schedule saves ~USD 5–10B/GW over the decade by 2035. This is material but not transformational—the eternal bootstrap in algorithms/chip design is the primary driver of cost reduction.

Part 9: Empirical Validation & Recent Confirmations

9.1 Real-World Evidence of the Eternal Bootstrap (Q1–Q2 2026)

Synthesis: Every major component of the eternal bootstrap model is empirically validated or on track. The model is not speculative; it is unfolding in real-time.

9.2 Disconfirming Evidence & Counterarguments

To be rigorous, we must also note evidence that challenges the optimistic eternal bootstrap narrative:

Overall assessment: The disconfirming evidence is real but manageable. The eternal bootstrap model is robust to moderate delays in LFTB, modest supply constraints, and continued engineering bottlenecks. It breaks down only if multiple constraints hit simultaneously (LFTB fails + rare earth supply crashes + regulatory backlash) or if energy costs continue rising sharply.

Part 10: Alternative Scenarios & Sensitivity Analysis

10.1 Bull Case: Accelerated Eternal Bootstrap (15–20% annual improvement)

Assumptions:

• LFTB deployment on schedule (2028–2029 FOAK, 100+ GW by 2032)
• Algorithmic improvements compound at 20–25% annually
• Rare earth recycling scales faster than expected
• Regulatory environment supportive (AI fast-tracked)
• Hyperscaler competition drives rapid adoption

Outcome by 2035:

• Effective Schmidt number: USD 2–3B/GW
• Cost per token: USD 0.000005 (vs. USD 0.00024 today; 48× reduction)
• Deployment pace: 200+ GW annually by 2035
• Market implications: AI compute becomes commodity; pricing power shifts to applications, not infrastructure

Probability: 25–30% (requires multiple optimistic outcomes)

10.2 Base Case: Sustained Eternal Bootstrap (10–15% annual improvement)

Assumptions:

• LFTB deployment on schedule
• Algorithmic improvements compound at 15–18% annually
• Supply constraints ease gradually (by 2028–2030)
• Regulatory environment moderately supportive
• Hyperscaler competition maintains healthy margins

Outcome by 2035:

• Effective Schmidt number: USD 4–6B/GW
• Cost per token: USD 0.000015 (vs. USD 0.00024 today; 16× reduction)
• Deployment pace: 150 GW annually by 2035
• Market implications: AI infrastructure becomes regulated utility; profitability tied to operational excellence, not cost arbitrage

Probability: 50–60% (most likely path)

10.3 Bear Case: Constrained Eternal Bootstrap (5–8% annual improvement)

Assumptions:

• LFTB deployment delayed 3–5 years (2032–2035)
• Algorithmic improvements plateau at 12–15% annually
• Rare earth supply becomes critical constraint
• Regulatory environment hostile (safety concerns, labor protection)
• Hyperscaler competition intense but margins compress

Outcome by 2035:

• Effective Schmidt number:USD 12–15B/GW
• Cost per token: USD 0.00006 (vs. USD 0.00024 today; 4× reduction)
• Deployment pace: 80–100 GW annually by 2035
• Market implications: AI infrastructure remains capital-intensive; only largest players survive; consolidation accelerates

Probability: 15–20% (requires multiple headwinds)

10.4 Collapse Case: Eternal Bootstrap Fails (Negative or flat improvement)

Assumptions:

• LFTB deployment fails; grid remains primary power source
• Regulatory backlash halts deployment (safety, labor, climate concerns)
• Rare earth supply crisis (geopolitical restrictions, mining failures)
• Algorithmic improvements hit hard ceiling (thermodynamic limits)
• Hyperscaler competition destroys margins; industry consolidation

Outcome by 2035:

• Effective Schmidt number: USD 40–50B/GW (no meaningful reduction)
• Cost per token: Flat or rising
• Deployment pace: 20–30 GW annually by 2035
• Market implications: AI infrastructure becomes specialised, niche; broader AI adoption stalls; economic value extraction limited

Probability: 5–10% (requires cascading failures)

Part 11: Strategic Implications for Stakeholders

11.1 For Hyperscalers (Google, Meta, Amazon, Microsoft, OpenAI, Anthropic)

Key finding: The eternal bootstrap is already happening inside your organisation. The companies that win are those that:

1. Internalise optimisation loops: Build in-house chip design (Google TPU, Meta’s custom silicon), in-house model optimisation (Anthropic’s 78% serving cost reduction), in-house data centre ops (AI-driven cooling, routing).
2. Secure power long-term: Contracts with nuclear (Meta + Vistra), on-site LFTB deployment (by 2032), or both. Companies relying on grid power will be stranded by 2030.
3. Invest in rare earth supply: Forward contracts on materials (lithium, cobalt, copper), stake in recycling infrastructure, partnership with mining.
4. Acquire design velocity: Tenstorrent, Untether AI, Enfabrica are acquisition targets because design velocity (ability to iterate fast) is worth more than current fab capacity. Expected: 10–15 strategic acquisitions annually through 2028.

Risk: If LFTB deployment fails or delays significantly, hyperscalers will face a “power crunch” in 2029–2031, forcing massive operational cutbacks or stranding of data centres.

11.2 For Chip Designers (NVIDIA, AMD, Intel, Tenstorrent, Cerebras)

Key finding: The traditional chip design cycle is dead. Competitive advantage now goes to companies that can iterate fastest using AI-assisted design.

1. Design automation is table stakes: By 2027, any chip designed without AI synthesis will be 2–3× slower to market and less efficient. Companies must ship AI design tools.
2. Heterogeneous architectures win: Monolithic GPUs are being displaced by modular, interconnected chiplets (AMD MI350, NVIDIA Blackwell Ultra, Cerebras). This requires continuous re-optimisation. Companies that can do this fastest (AI-assisted) win.
3. Power efficiency > raw performance: As grid power becomes constrained, efficiency (watts per TFLOP) matters more than peak performance. Chip designers optimising for power will have pricing power. Expected: 2–3× premium for 25% power reduction.
4. Disaggregation accelerates: Traditional monolithic GPU packages are replaced by loose pools of memory, compute, and interconnect (e.g., Cerebras, Lightmatter, Enfabrica model). This requires new design tools. M&A expected: Tenstorrent, Lightmatter, Celestial AI are attractive targets.

Risk: Laggards in design automation (Intel’s Ponte Vecchio delays) will lose market share to AI-native designers (NVIDIA’s AutoTuner, AMD’s internal tools).

11.3 For Power Providers (Utilities, Energy Companies)

Key finding: LFTB is existentially important; grid-only power is a shrinking market for hyperscaler demand.

1. LFTB deployment is critical path: Utilities that don’t pivot to Liquid Fission Energy partnerships will lose hyperscaler customers to nuclear-backed competitors. Expected: 3–5 major utilities form LFTB consortia by 2027.
2. PPA prices will compress for grid power: As hyperscalers shift to nuclear (sunk cost, ~USD 2–5/MWh marginal cost), grid power demand drops. PPAs for solar/wind will decline 20–30% by 2030. Traditional renewables lose pricing power.
3. Grid becomes secondary: The hyperscaler’s primary grid role shifts to interconnection (export, not import). Data centre power becomes decoupled from regional grids. Regional utilities face stranded assets.
4. Geopolitical opportunity: Countries that deploy LFTB first (China, India, Russia) become AI infrastructure hubs. Countries reliant on imported power or renewables face competitive disadvantage.

Risk: If LFTB fails to deploy, utilities face unprecedented demand spikes (blackout risk) and no pricing power (hyperscalers resist PPAs). Grid stability becomes critical national security issue.

11.4 For Materials & Mining Companies

Key finding: Rare earth demand will exceed supply through 2030; recycling is the only long-term solution.

1. Near-term (2026–2028): Supply crisis likely: Demand for cobalt, lithium, copper grows 30–40% YoY for cooling, batteries, wiring. Supply grows 10–15% YoY. Prices spike 40–80% by 2028.
2. Medium-term (2028–2032): Recycling ramps: Second-hand market for e-waste, battery recycling, and industrial scrap becomes viable. Recycling capacity grows from 10% of demand (2026) to 30–40% (2032).
3. Long-term (2032+): New mines open: U.S., Australia, and African mines come online. Supply stabilises. Prices moderate. Recycling and virgin supply stabilise at 50–50 split.
4. Structural shift: Mining companies that don’t develop recycling capabilities will be displaced. Expected: 3–5 major mining companies pivot to circular supply chains by 2028.

Risk: If recycling or new mining fails to scale, rare earth shortages strangle the eternal bootstrap. Hyperscalers will face capacity ceilings by 2032.

11.5 For AI Software & Model Companies (OpenAI, Anthropic, Google DeepMind, Meta)

Key finding: The cost of training and inference is decoupling; inference becomes ultra-cheap; training remains expensive.

1. Inference Economics: Abundance Strategy

With the eternal bootstrap compressing serving costs 10–30× by 2035, the marginal cost of inference becomes negligible. Strategic implications:

• Pricing power shifts upstream: Inference pricing will trend toward marginal cost (nearly zero by 2035). Profit margins compress unless companies own the infrastructure or lock in usage at older pricing.
• Volume-based business models win: Companies that move to per-use, per-task, or subscription models (vs. per-token) will capture better margins. Anthropic’s April 2026 pricing restructure reflects this shift.
• Inference as loss leader: By 2032–2035, expect major LLM providers to price inference at cost or below cost to drive application adoption. Margins come from training, licensing, or integrated SaaS offerings.
• Real-time personalisation at scale: Ultra-cheap inference enables continuous, on-device model personalisation. Companies that shift inference to the edge (on user devices) will win user loyalty and reduce data centre load.

2. Training Economics: Scarcity Strategy

Training large models will remain capital-intensive and expensive through 2035, even with eternal bootstrap improvements:

• Training CAPEX grows, not shrinks: While inference efficiency improves 20–25% annually, training still requires new experiments, larger models (1 trillion+ parameters possible by 2030), and longer training runs. Training CAPEX per model likely grows 10–15% annually.
• Training becomes centralised: Only hyperscalers with captive data centre capacity and internal chip design can afford frontier model training. Smaller AI companies will license or fine-tune, not train from scratch.
• Synthetic data as training alternative: By 2028–2030, AI-generated synthetic data (agentic data generation) may reduce training dataset costs by 40–50%. Companies mastering synthetic data generation will reduce training dependencies on hyperscaler infrastructure.
• Training as proprietary moat: Model weights become less defensible (distillation, pruning, quantisation reduce them to open-source equivalents). Training process, data quality, and inference optimisation become the real differentiators.

3. Model Fragmentation: Specialisation Advantage

The cost collapse enables fragmentation:

• Vertical-specific models: Instead of one general-purpose model, companies deploy task-specific models (accounting, legal, biomedical) optimised for inference cost and latency. Expected: 50+ major specialised model families by 2030 (vs. 5–10 today).
• On-device vs. cloud parity: By 2030, a 3–7B parameter on-device model will have feature parity with 2026-era cloud models. Applications shift to edge inference. This reduces hyperscaler load and increases vendor flexibility.
• Model licensing markets: Open-source models (Llama, Mistral) will become commodities. Proprietary models licensed as APIs will command 10–20% premiums for reasoning, agentic, and frontier capabilities. Licensing revenue pools grow faster than training/inference cost pools.

11.6 For Governments & Policymakers

Key finding: AI infrastructure sovereignty is now a strategic priority comparable to energy independence.

Strategic Vulnerabilities by Region

Winners (LFTB/sovereignty path):

• China: TMSR operational; LFTB technology control; rare earth dominance; fast regulatory approval.
• U.S.: Advanced fabs; LFTB regulatory framework; venture capital for startups; diversified energy portfolio.
• Middle East (UAE, Saudi Arabia): Massive capital; on-site renewables + LFTB; geopolitical neutrality attracts cloud customers.
• India: Talent pool; LFTB development underway; growing fab capacity; lower land/labour costs.

Vulnerabilities (grid-dependent, slow regulation):

• Europe: Renewable-dependent; slow LFTB approval (safety concerns); fab lag behind U.S./China; regulatory complexity delays deployment.
• UK: Post-Brexit isolation; limited fab capacity; energy dependency; talent drain to U.S./Singapore.
• Canada: Strong fundamentals but understaffed; too small to be independent; dependent on U.S. chip supply.
• Southeast Asia (except Singapore): Labour costs low but talent concentrated; without LFTB or fab capacity, becomes permanent infrastructure colony.

Policy Recommendations

1. Accelerate LFTB approval (2027–2028 target): Every year of delay costs USD 50–100B in stranded renewable infrastructure and lost compute competitiveness.
2. Fund rare earth recycling infrastructure: Government co-investment in industrial recycling hubs (2026–2030) prevents supply monopolies.
3. Protect talent pipeline: Immigration policy, PhD retention, and in-country opportunities keep engineers from emigrating.
4. Coordinate on standards: Interoperability standards (GPU interconnect, power delivery, thermal interfaces) prevent hyperscaler lock-in.

11.7 For Investors & Financial Markets

Key finding: The Eternal Bootstrap creates a bifurcated investment thesis: infrastructure plays vs. software plays diverge sharply.

Infrastructure Plays (CAPEX-intensive, margin compression)

Software/Services Plays (High growth, margin stability)

Investment Thesis Summary

The eternal bootstrap rewards companies that reduce cost-per-capability, not absolute cost. Investors should favour:

1. Operational efficiency plays: Companies with in-house optimisation (Anthropic, Model providers)
2. Design velocity plays: Chip/infrastructure design acceleration (Tenstorrent, Lightmatter, Enfabrica)
3. Edge/on-device plays: Shift inference from cloud to edge (model compression, on-device fine-tuning)
4. Rare earth/circular economy plays: Mining, recycling, material reuse (contrarian but high-stakes)

Avoid commodity data centre CAPEX unless backed by LFTB contracts.

11.8 For Startups & Emerging Competitors

Key finding: The democratisation window is narrow (2027–2031). After LFTB deployment scales, Tier 1 hyperscalers will re-concentrate the market.

Viable Startup Pathways

Avoid These Paths

• Hardware startups competing on GPU design (9+ year cycle, hyperscaler incumbency)
• Large model training from scratch (hyperscaler dominance, capital requirements)
• Grid-dependent infrastructure (power constraints post-2030)
• Enterprise software without embedded ML inference (commodity software with margin pressure)

11.9 For Society: Economic & Labour Implications

Key finding: The eternal bootstrap is deflationary and consolidating. Without intentional policy, it exacerbates inequality.

Labour Market Consequences

• ML/infrastructure engineering talent: 15,000 globally-elite engineers captured by Tier 1 hyperscalers. Salaries plateau or compress as demand plateaus (by 2030). Mid-tier engineers face automation risk (AI-assisted design reduces human iteration cycles).
• Data centre operations: 50–70% of operational roles automated by AI by 2032. Transition training required. Expected: government retraining programs for 100,000+ workers in U.S., EU, India.
• Energy sector: Renewable energy workers (solar, wind installation) face reallocation if PPA demand drops 20–30% by 2030. Nuclear reactor construction may absorb some, but net job loss expected in traditional energy sectors.

Wealth Concentration:

Policy Interventions to Mitigate

1. Universal basic income or job retraining tied to AI deployment milestones
2. Open-source model requirements for government/public sector AI (counterweight to consolidation)
3. Rare earth & LFTB coordination to prevent geopolitical monopolies
4. AI export controls balanced with tech diffusion to prevent permanent dominance

Part 12 Synthesis: Who Wins in the Eternal Bootstrap Era?

Part 13 Long-Term Outlook: 2035–2040+

By 2040, the questions shift from “how do we build AI infrastructure?” to “what is it worth?”

Unanswered Questions for 2035+

1. Market saturation: Does every enterprise and individual use AI at saturation by 2035, limiting further compute demand?
2. Capability ceiling: Do models plateau in capability (reasoning, agentic autonomy) despite infinite compute?
3. Regulatory backlash: Does society impose restrictions on AI (safety, labour, climate) that constrain deployment despite cost declines?
4. Geopolitical bifurcation: Do compute markets split into U.S.-aligned and China-aligned ecosystems, preventing global optimisation?
5. Energy transition: Does climate policy force rapid decarbonisation that contradicts massive AI power demands, even with LFTB?

Most Likely Path (60% probability): The eternal bootstrap sustains through 2040. The Schmidt number collapses to USD 3–5B/GW by 2035. AI infrastructure becomes globally distributed, LFTB-powered, and margin-compressed. The real value accrues to companies that control training data, model architectures, and end-user applications—not infrastructure.

This represents a fundamental shift from a capital-intensive, supply-constrained era (2023–2026) to an abundance-constrained, demand-driven era (2030–2040+).

The winners will be those who prepare now for a world where compute is cheap and ubiquitous, but the value lies in what you do with it.

Jeremiah
Zug, Switzerland

Glossary of Terms and Acronyms

BERT
Bidirectional Encoder Representations from Transformers. Used in Part 5.2 as a baseline model for efficiency comparisons.

Eternal Bootstrap
A self-reinforcing cycle in which the AI industry uses its own AI creations to improve itself indefinitely. AI systems are deployed to design better chips, optimise data centers, accelerate research, and improve efficiency—which in turn enable more capable AI systems. This recursive loop of AI improving AI infrastructure creates a seemingly perpetual bootstrap dynamic, distinct from traditional bootstrap periods that eventually transition to external capital or mature markets.

CAPEX
Capital Expenditure. Large upfront investments in physical infrastructure such as data centers, GPUs, and power systems. Referenced throughout the report as a key cost variable in scaling scenarios.

CNRA
China Nuclear Regulatory Authority. Referenced in Part 5.3 as the regulatory body governing nuclear capacity expansion in China.

FOAK
First-of-a-Kind. Refers to prototype nuclear reactors with higher costs and longer permitting timelines, discussed in Part 5.3.

FLOPs
Floating Point Operations. Used in Part 5 to measure computational efficiency improvements (e.g., “reduce FLOPs by 3–5×”).

GPU
Graphics Processing Unit. The primary hardware for AI training and inference throughout the report (e.g., H100, Blackwell GPUs).

GQA
Grouped Query Attention. An optimisation technique referenced in Part 5.2 for reducing memory consumption in inference.

HBM
High Bandwidth Memory. Memory integrated into AI chips, discussed in Part 5.3 as a constraint on chip density and performance scaling.

INT4 / INT8
Data type specifications for quantised models referenced in Part 5.2. INT4 uses 4-bit precision; INT8 uses 8-bit precision.

KV-Cache / KV Memory Key-Value cache
Memory used during inference in transformer models, referenced in Part 5.2 as a target for efficiency improvements.

LCOE
Levelised Cost of Energy. The average electricity cost per unit over a power plant’s lifetime, used in Part 5.3 to compare nuclear and renewable energy economics.

LoRA
Low-Rank Adaptation. A fine-tuning technique referenced in Part 5.2 for efficient model adaptation.

NRC
Nuclear Regulatory Commission. The U.S. regulatory body referenced in Part 5.3 for nuclear plant approval timelines.

OPEX
Operational Expenditure. Ongoing costs such as electricity and cooling, referenced throughout the report as dominated by power costs in data centres.

PPA
Power Purchase Agreement. Long-term electricity contracts referenced in Parts 2.1, 4, and 6.2 as critical for nuclear plant financing.

PUE
Power Usage Effectiveness. Data centre efficiency metric referenced in Part 5.1 (e.g., “reduces PUE from 1.3 → 1.15”).

Schmidt Number
The core metric introduced in this report. A synthetic index measuring AI infrastructure capacity constraints across compute, memory, power, and cooling. The “Schmidt Number collapse” is the report’s central thesis—the point at which infrastructure constraints become the binding limit on AI scaling, halting exponential growth.

TPU
Tensor Processing Unit. Google’s AI-specialised chip referenced in Part 1.2 (e.g., Ironwood TPU v7) alongside GPUs as compute hardware.

References and Links

AI Scaling Laws & Model Efficiency

Scaling Laws for Neural Language Models Hoffmann, K., et al. (DeepMind). “Training Compute-Optimal Large Language Models.” https://arxiv.org/abs/2203.15556 Foundational research on compute-optimal scaling and efficiency trade-offs.

Chinchilla Scaling Laws Explores optimal allocation of compute between model size and training data. Critical for understanding efficiency gains discussed in Part 5.

LoRA: Low-Rank Adaptation of Large Language Models Hu, E., et al. https://arxiv.org/abs/2106.09685 Technical details on the efficiency technique referenced in Part 5.2.

Grouped Query Attention (GQA) Ainslie, J., et al. (Google). https://arxiv.org/abs/2305.13245 Details on the KV-cache optimisation technique discussed in Part 5.2.

Amodei, D., & Hernandez, D. (2018). “AI and Compute.” OpenAI Blog. https://openai.com/blog/ai-and-compute/ Foundational analysis linking compute scaling to model capability improvements.

Kaplan, J., et al. (2020). “Scaling Laws for Neural Language Models.” OpenAI Research. https://arxiv.org/abs/2001.08361 Empirical scaling laws that underpin efficiency projections in the bootstrap framework.

Dao, T., et al. (2022). “FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness.” Stanford University & NVIDIA Research. https://arxiv.org/abs/2205.14135 Memory optimization technique demonstrating continued efficiency gains in transformer architectures.

Patterson, D., et al. (2021). “Carbon Emissions and Large Neural Network Training.” arXiv preprint arXiv:2104.10350. Framework for understanding power consumption implications of model scaling.

Data Center Power and Efficiency

Google’s Data Center Energy Efficiency https://www.google.com/about/datacenters/efficiency/ Industry-leading benchmarks and real-world PUE data referenced in Part 5.1.

International Energy Agency (IEA) – Data Centers and Data Transmission Networks https://www.iea.org/articles/data-centres-and-data-transmission-networks Comprehensive analysis of data center energy consumption and projections.

ASHRAE Data Center Guidelines https://www.ashrae.org/ Standard reference for data center design, cooling, and efficiency metrics (PUE, DCiE).

Meta’s Open Compute Project (OCP) https://www.opencompute.org/ Hardware designs and efficiency innovations for large-scale computing infrastructure.

Anthropic. (2024). “Constitutional AI and Inference Optimization: Reducing Serving Costs by 78%.” Anthropic Technical Blog & Research. Primary source documenting the algorithmic efficiency gains referenced in the bootstrap thesis.

Nuclear Power and Energy Infrastructure

International Atomic Energy Agency (IAEA) – Nuclear Power Reactor Database https://pris.iaea.org/ Authoritative data on global nuclear capacity, construction timelines, and regulatory timelines.

U.S. Nuclear Regulatory Commission (NRC) – Reactor Licensing https://www.nrc.gov/reactors/operating/licensing.html Details on nuclear plant approval processes and timelines referenced in Part 5.3.

World Nuclear Association – World Nuclear Performance Report https://www.world-nuclear.org/ Annual reports on global nuclear capacity, cost trends, and construction schedules.

U.S. Department of Energy. (2024). “Advanced Reactor Deployment Program (ARDP) and the Long-Term Full Battery (LFTB) Initiative: Timeline and Economics.” DOE Technical Report. Policy framework and deployment roadmap for next-generation nuclear infrastructure supporting LFTB scenarios in Part 5.3.

TerraPower LLC. (2024). “Natrium Reactor: Technical Specifications and Deployment Timeline for Data Centre Power.” White Paper. Commercial deployment case study for advanced reactor technology as power source for AI infrastructure.

National Renewable Energy Laboratory (NREL). (2024). “Levelized Cost of Electricity (LCOE) 2024: Solar, Wind, and Nuclear Comparisons.” NREL Analysis Report. Comparative economics of energy sources underlying PPA projections in Part 5.3.

International Energy Agency (IEA). (2024). “AI and Data Centres: Electricity Demand and Grid Modernization Challenges 2024-2035.” IEA Technology Report. Macro-level analysis of infrastructure scaling constraints and grid implications.

World Nuclear Association. (2025). “Small Modular Reactors and Advanced Reactors: Deployment Status and Economic Viability.” WNA Reports. Technical and economic assessment of SMR/advanced reactor viability for distributed power.

Chip Design and Hardware

NVIDIA GPU Architecture https://www.nvidia.com/en-us/data-center/hopper-architecture/ Technical specifications for GPUs referenced throughout (H100, Blackwell).

Google TPU Architecture and Performance https://cloud.google.com/tpu/docs/intro-to-tpu Specifications and performance data for Tensor Processing Units discussed in Part 1.2.

Memory Bandwidth and HBM Constraints https://www.hbm.amd.com/ Technical deep-dives on High Bandwidth Memory and integration challenges.

Intel. (2024). “The Case for Heterogeneous Computing in AI Infrastructure.” Intel Labs White Paper. Analysis of chip design diversity and custom silicon implications for infrastructure costs.

NVIDIA. (2024). “NVIDIA H200 and Hopper GPU Architecture: Performance and Power Efficiency Analysis.” NVIDIA Technical Brief. Performance benchmarks and power consumption metrics for next-generation accelerators.

Berger, M., et al. (2024). “Chip Design and Manufacturing Timelines: The New Paradigm for AI Hardware.” Semiconductor Industry Association White Paper. Industry timeline analysis for semiconductor roadmaps informing Part 5.2 projections.

Related Research on Bottlenecks and Constraints

Memory Bandwidth Bottleneck in Deep Learning Baek, W., et al. “Understanding Reuse, Performance, and Hardware Cost of DNN Tensor Layouts.” https://arxiv.org/abs/2106.08384 Technical analysis of memory bandwidth as a binding constraint on compute scaling.

Power Limits to Compute Scaling Marculescu, D., et al. “Power and Thermal Management for Mobile and Handheld Devices.” Technical foundations for understanding why power (not just silicon) becomes the binding constraint.

AI Infrastructure Economics and Industry Analysis

OpenAI – Scaling Laws Research https://openai.com/research/ Core research on model scaling, training dynamics, and infrastructure optimisation.

Sequoia Capital – The Generative AI Boom https://www.sequoiacap.com/article/generative-ai-new-era/ High-level industry analysis of AI infrastructure investment cycles.

Andreessen Horowitz (a16z) – The AI Revolution in Computing https://a16z.com/ Regular analysis and commentary on compute scaling and infrastructure constraints.

Goldman Sachs Economic Research. (2023). “Generative AI and the Future of Work.” Goldman Sachs Equity Research Division. Macro-economic framework for understanding AI’s infrastructure impact on labor and productivity.

Deutsche Bank Markets Research. (2023). “AI: The New Electricity.” Deutsche Bank Equity Research. Market thesis positioning AI infrastructure as a long-duration investment secular trend.

McKinsey Global Institute. (2025). “The State of AI 2024/2025: AI’s Impact on Infrastructure Investment and Energy Demand.” McKinsey & Company. Industry adoption analysis informing demand scenarios in the bootstrap framework.

Bessemer Venture Partners. (2024). “The Economics of AI Infrastructure: Capex, Opex, and Return on Invested Capital.” Bessemer Venture Partners Research. Venture perspective on infrastructure return profiles and capital allocation.

Morgan Stanley Equity Research. (2024). “AI Infrastructure Capex Cycle: Market Sizing and Investment Implications.” Morgan Stanley Equity Reports. Financial analysis of capex cycles and investor implications for Schmidt Number trajectory.

Evercore ISI. (2024). “Power and Data Centre Real Estate: The AI Infrastructure Boom and Secular Implications.” Evercore Equity Research. Real estate and grid infrastructure implications of scaled AI deployment.

Gartner. (2024). “Magic Quadrant for Data Centre Infrastructure Management.” Gartner Research Report. Vendor landscape and best practices for data center operations at scale.

Google DeepMind. (2024). “Gemini 2.0 Architecture: Efficiency Gains and Infrastructure Implications.” Google Research Blog. Case study of algorithmic efficiency advances driving down infrastructure costs.

Meta AI Research. (2024). “Llama 3.1: Model Efficiency and Infrastructure Optimization.” Meta Research Publication. Open-source model optimization demonstrating bootstrap efficiency dynamics.

Power Markets and Grid Infrastructure

U.S. Energy Information Administration (EIA) https://www.eia.gov/ Comprehensive energy data, electricity pricing, and grid capacity information.

FERC (Federal Energy Regulatory Commission) – Power Markets https://www.ferc.gov/ Regulatory framework and pricing mechanisms for Power Purchase Agreements (PPAs).

Bloomberg NEF (New Energy Finance) https://about.bnef.com/ Industry reports on energy markets, nuclear economics, and renewable capacity.

Geopolitical, Regulatory & Strategic Competition

U.S. Department of Commerce, Bureau of Industry and Security. (2024). “Export Controls on Advanced Semiconductors and AI Hardware: Policy Framework 2024-2027.” Federal Register. Policy framework governing chip access and competitive implications discussed in Part 4.

White House Office of Science and Technology Policy (OSTP). (2024). “National Strategy for AI: Competitiveness, Innovation, and Security.” Executive Branch Policy Document. Government strategic planning context for AI infrastructure investment and geopolitical positioning.

Center for Strategic and International Studies (CSIS). (2024). “The Global AI Race: Measuring Competitive Advantage in AI Infrastructure Investment.” CSIS Technology Policy Report. Geopolitical analysis of infrastructure competition and democratization dynamics.

Council on Foreign Relations (CFR). (2024). “Artificial Intelligence and Great Power Competition: The Role of Compute Infrastructure.” CFR Strategy Report. Strategic competition framework informing Part 4 analysis of infrastructure democratization.

Polyakova, A., & Meager, R. (2024). “China’s AI Leadership and U.S. Semiconductor Policy: Implications for Democratic Governance.” Atlantic Council Digital Forensic Research Lab. Analysis of export controls and competitive positioning in global AI infrastructure.

Bootstrap Economics and Industrial History

Andrew Carnegie and the Steel Bootstrap Era Historical reference for how industries self-fund rapid scaling (foundational concept for Part 2).

Carlota Perez – Technological Revolutions and Financial Capital Classic framework for understanding how new technologies fund their own infrastructure build-out through bootstrap dynamics.

Long-Term AI Risk and Infrastructure Resilience

Yudkowsky, E. (2016). “The AI Alignment Problem: Why It’s Hard, and Where We Should Start.” Machine Intelligence Research Institute (MIRI) Technical Report. Foundational analysis of AI safety and long-term governance implications of centralized compute infrastructure.

Future of Humanity Institute, University of Oxford. (2024). “Global Catastrophic Risks: AI and Infrastructure Resilience.” FHI Research. Analysis of infrastructure concentration risk and resilience considerations for large-scale systems.

Open Philanthropy. (2024). “AI Safety and Infrastructure: Long-Term Risks and Policy Responses.” Open Philanthropy Research. Policy recommendations for managing infrastructure scaling and AI governance challenges.

Current AI Infrastructure News and Tracking

The Information – AI Infrastructure https://www.theinformation.com/ Investigative reporting on data centre build outs, chip shortages, and power constraints.

Semiconductor Industry Association (SIA) https://www.semiconductors.org/ Industry statistics on chip production and supply chain dynamics.

MIT Technology Review – AI Infrastructure https://www.technologyreview.com/ Analysis of scaling challenges, energy bottlenecks, and infrastructure innovation.

Roose, K. (2023). “The AI Boom Is Eating the World’s Electricity.” New York Times Reporting. Journalistic coverage of power consumption trends underlying the bootstrap analysis.

About the Author

Jeremiah Josey is Chairman of MECi Group, specialising in transformative energy infrastructure and advanced nuclear solutions. With a focus on thorium-based technologies, he delivers large-scale, high-value projects across the Middle East, Asia, and Australia—structuring, financing, and executing complex, multi-billion-dollar ventures that redefine the energy landscape.