Key arithmetic and supporting data behind the quantitative claims in the essay body. Inline citations map to the reference section below.

A. AI Capex Aggregate, 2025–2026

Technology platforms that once funded their own growth are borrowing to build physical infrastructure.

Source	Figure	Period	Notes
Hyperscaler AI/cloud capex guidance (Microsoft, Alphabet, Meta, Amazon, Oracle, CoreWeave)	>$700B aggregate	FY2026 plans	Industry sum from Reuters aggregation across hyperscaler guidance; Moody’s projects ~$820B in FY2027.[1]
Hyperscaler corporate bond issuance	$121B (2025) → $175B projected (2026)	Annual	Top-five U.S. hyperscalers; ~4× historical annual average. Amazon’s $54B March 2026 issuance was the company’s largest ever debt transaction. Bank of America projection for 2026.[2]
AI venture funding	~$255B	Single quarter (Q1 2026)	PitchBook tracked quarterly deployment, exceeding the full 2025 total. Dominated by OpenAI ($122B), Anthropic ($30B), and xAI ($20B) financings; excluding the top-five transactions, broader quarterly deal value contracts by 73%.[3]
U.S. datacentre investment	$45.7B PE	FY2025	S&P Global Market Intelligence; PE deployment represented 72% of $63.35B total sector investment.[18]

Capex, debt issuance, venture funding, and private-equity deployment measure different cash flows. Read separately, they show how much money the buildout is drawing and where it comes from.

Amazon’s trailing-twelve-month free cash flow fell roughly 95% to $1.2 billion in the same window it issued $54 billion in bonds. Meta issued $30 billion in October 2025. Alphabet issued $32 billion in February 2026, including a rare 100-year bond. Oracle planned $25 billion for 2026 against an upward-revised $50 billion FY2026 capex guide.

B. Energy Marginal Cost

Using the U.S. Energy Information Administration’s 2024 average operating heat rate for natural gas-fired generation (7,754 Btu/kWh)^[19] and front-month gas prices in May 2026:^[5]

Hub	Front-month gas	Fuel-only generation cost	Multiple vs. Henry Hub
Henry Hub (U.S.)	$2.83/MMBtu	~$22/MWh	1.0×
TTF (Europe)	~$17/MMBtu	~$132/MWh	~6.0×
JKM (Asia)	~$18.80/MMBtu	~$146/MWh	~6.6×

For a continuously loaded 100 MW campus (8,760 hours per year = 876,000 MWh), the fuel-cost difference between Henry Hub and TTF is approximately $96M/year. Delivered-power costs also include transmission, capacity charges, cooling, financing, and carbon costs.

C. Named-Project Facility-Load Estimates

Roughly a fifth of disclosed ordered load is plugged in.

Project	Reported scale	Facility load	Confirmed energised
xAI Colossus 1	220,000+ Nvidia processors[20]	~290 MW (consistent with reported facility power)	~290 MW; 300 MW expansion under build
xAI Colossus programme	~1,000,000 target	~1.3 GW target (heuristic)	Partially built
Oracle / OpenAI Abilene	400,000 GB200-class[21]	~1.2 GW (facility design, per Oracle / Crusoe disclosure)	2 of 8 buildings operational (~300 MW)
DOE Solstice + Equinox	110,000 Blackwell[22]	~140 MW (heuristic)	Equinox planned H1 2026
Meta roadmap 2025–2027	”Multi-million” Blackwell/Rubin[23]	Multi-GW	Undisclosed

Use disclosed facility design power where available; elsewhere assume roughly 1.3 kW per accelerator. This heuristic estimates scale. The comparison concerns announced versus energised load.^[24]

Summed across the three projects with disclosed status (Colossus ~1.3 GW + Abilene 1.2 GW + DOE 0.14 GW), the order book implies ~2.6 GW of facility load. The mid-2026 energised share, Colossus 1 (~290 MW) plus Abilene 2-of-8 (~300 MW), totals ~0.6 GW.

D. Energisation and Utilisation

Capacity must first receive power, then find useful work.

The energisation and utilisation figures concern different samples and stages. They cannot be multiplied:

• Energisation share describes ordered facility load that has been turned on. Source: public sample of named hyperscaler and sovereign projects in Appendix C. Roughly a fifth of disclosed ordered load in the sample is plugged in.
• Utilisation share describes the time-fraction of an active GPU spent on productive work. Source: Cast AI 2026 Kubernetes survey across AWS, Azure, and GCP enterprise clusters.^[7] Approximately 5% mean sustained.

Frontier training clusters likely run at much higher utilisation. The named projects cover only part of accelerator orders. Each sample exposes a different gap between purchased capacity and productive work.

The Cast AI 2026 State of Kubernetes Optimization Report shows the broader pattern: 8% average sustained CPU utilisation, 20% memory utilisation, 69% CPU overprovisioning headroom, 79% memory overprovisioning headroom, and fewer than 2% of GPU workloads on spot instances, primarily because spot capacity was effectively unavailable at hyperscaler scale. Specialised clusters using GPU time-slicing and MIG partitioning sustain up to 49% GPU utilisation.

E. Cost-Per-Token Decline

Training capex and inference opex have diverged by orders of magnitude.

Stanford HAI AI Index and Epoch AI compute-cost time series document a roughly 10× annual decline in inference cost at fixed capability level across the 2022–2026 window.^[15]

GPT-3.5-equivalent capability cost ~$20 per million tokens in late 2022 and ~$0.07 per million tokens in late 2024, a roughly 280× reduction over two years. GPT-4 launched at $30/$60 per million input/output tokens. GPT-4o offered the same capability tier at $3/$10 per million eighteen months later. DeepSeek V3 entered at $0.14 per million input tokens, undercutting GPT-4o by ~95%. Epoch AI reports a median annual decline of ~50× across the inference market.

Hardware, better algorithms, competition from open weights, and a maturing software stack all reduce inference cost. Quantisation, distillation, sparsity, speculative decoding, KV-cache optimisation, MoE routing, FlashAttention, and compilers contribute. Software and competition account for much more of the headline decline than hardware alone.

Large MoE training concentrates spending on synchronous power, interconnects, and coordination, with substantial instability risk. Memory-bound inference can use cheaper silicon, reading only the active experts per token. Appendix H examines where that advantage holds.

F. The Grid-Avoidance Tax

Grid queues have encouraged on-site gas generation. In 2026, Blackstone Tactical Opportunities and Halliburton invested $1 billion in VoltaGrid, supporting a 7.5 GW order book through 2030 and manufacturing capacity of up to 300 MW per month. Modular reciprocating engines and turbines promise power in months under simplified minor-source air permits.^{[4] [25]}

Earlier power carries a recurring fuel cost.

Thermal efficiency of a generator is the ratio of its electrical output to fuel input, conventionally expressed via the operating heat rate (Btu of fuel per kWh of electricity):

$$\eta = \frac{3{,}412.142}{\mathrm{HR}}$$

Using the U.S. EIA’s 2024 average tested heat rates by prime mover:^[19]

Generation class	Heat rate (Btu/kWh)	Thermal efficiency	Gas consumption (ft³/kWh)
Combined-cycle utility gas	7,548	~45.2%	~7.35
Average U.S. grid natural gas	7,754	~44.0%	~7.55
Reciprocating internal-combustion gas	8,924	~38.2%	~8.69
Simple-cycle gas turbine	10,999	~31.0%	~10.71

Gas consumption assumes approximately 1,027 Btu per cubic foot of pipeline natural gas.

The combined-cycle plant captures exhaust heat through a steam turbine. Modular simple-cycle and reciprocating units trade that recovery for speed and simpler installation. In the table above, the simple-cycle turbine consumes roughly 45.7% more natural gas per kWh than the combined-cycle plant, and emits CO₂ in direct proportion.

The 45.7% additional fuel is the grid-avoidance tax: a recurring premium for installing simple-cycle power sooner.

The tax compounds the regional cost spreads of Appendix B. Where imported gas already costs six or seven times the Henry Hub price, a simple-cycle installation can run at nine to ten times the fuel cost of a Henry Hub combined-cycle plant. Getting power sooner means burning more fuel for every computation.

G. Slot Density, the OpEx Paradox, and the Migration Path

If every powered slot were equally scarce, its operator would choose the newest, most efficient chip. Even free older hardware would carry the opportunity cost of displacing that chip. This is the OpEx versus slot paradox.

Contiguous power is scarce. Synchronous training needs hundreds of megawatts or gigawatts at one permitted site, with cooling, deep networking, and experienced operators. Transformer queues, substation permits, and grid connections restrict the supply.

Smaller powered sites are widespread. Colocation facilities, university and sovereign clusters, industrial sites, laboratories, offices, and homes have fragmented capacity. Some can host useful work without expanding the grid.

Slot tier	Power density	Site count (order of magnitude)	Workload type	Binding constraint
Frontier	100 MW – 1 GW contiguous	tens to low hundreds globally	Synchronous training	Substations, transformer queues, multi-year permitting
Plateau	1 MW – 100 MW distributed	thousands	Inference, fine-tuning, batch, simulation	Existing colocation, sovereign, and enterprise capacity
Edge	0.5 kW – 1 MW fragmented	millions	Local inference, agents, batch	Wall socket, room cooling, residential grid

At a $30,000–40,000 purchase price, an accelerator needs high utilisation and valuable work. Running a new chip at 15% utilisation in a 50 kW office slot can be uneconomic.

Depreciation lowers the utilisation needed to repay the purchase. A written-down chip can earn its place at 20% utilisation in a smaller slot that could not justify new hardware. Migration can open previously uneconomic sites as well as replace chips in existing ones.

A follow-the-sun virtual cloud could route latency-tolerant inference, fine-tuning, batch processing, and simulation to regions with cheap energy. Such work can use distributed capacity without a synchronous gigawatt-scale interconnect.

Verification in decentralised multi-tenant systems and WAN bandwidth remain constraints. Cheap hardware and available power create an incentive to solve them; the thesis does not assume they are already solved.

H. The Plateau Crossover

The crossover occurs when depreciation compensates for lower performance on the dimension a workload needs. Begin with purchase cost, then add software compatibility, memory fit, and operation.

Definitions. Four rates describe hardware growth, software improvement, depreciation, and workflow discovery; $\mu$ corrects for architectural incompatibility:

Symbol	Quantity	Value or assumption	Source
$g_q$	Annual frontier hardware capability growth per chip, on the binding dimension (FLOPS/chip, GB/s/chip, FLOPS/W, etc., not capability-per-dollar)	Per-chip values from H100 SXM5 → B200 SXM, 3-yr annualised: ~1.31× (FP16/BF16 throughput); ~1.34× (memory bandwidth); ~1.34× (memory capacity, 80→192 GB); ~1.47× (FLOPS/W); ~1.65× (FP8 throughput)	Nvidia published specs; per-chip values, not per-dollar
$r_s$	Annual software-efficiency multiplier at fixed capability	~50× (inference cost-per-token); ~3× (pre-training compute-per-capability)	Stanford HAI AI Index; Epoch AI
$\delta$	Annual secondary-market capital cost decay	~0.25 (used H100 SXM5: ~$35K new 2023 → ~$15K used mid-2026)	Industry pricing
$\mu(t,\tau) \geq 1$	Architectural divergence penalty (software gains hard-coded to new silicon that does not backport cleanly)	Modest for FP16/FP8 inference; significant for FP4 and the latest attention/decoding kernels	ISA differences
$\beta \geq 0$	Workflow-discovery recursion rate (AI-assisted engineering productivity feedback)	Empirically positive; bounded above by deployment friction	This essay

Use physical capability per chip for $g_q$; prices enter separately. Epoch’s roughly 1.37× annual FLOP/s-per-dollar improvement combines per-chip capability and launch prices. Substituting it for $g_q$ would count price twice. Each workload needs its own binding dimension.

The CapEx Crossover Identity. Let $\tilde{C}(t, \tau)$ denote useful capability per dollar of a chip purchased used at time $t$, with age $\tau$:

$$\tilde{C}(t, \tau) = \frac{Q_0 \cdot E(t)}{P_0 \cdot (1-\delta)^\tau}$$

where $Q_0$ is the chip’s physical capability on the binding dimension (fixed at manufacture) and $E(t)$ is the contemporary software multiplier. A new frontier chip at time $t$—assuming approximately stable launch prices across generations (the $p_F$ correction is treated below)—has capability $Q_0 \cdot g_q^\tau$ and price $P_0$:

$$\tilde{C}_F(t) = \frac{Q_0 \cdot g_q^\tau \cdot E(t)}{P_0}$$

The software multiplier $E(t)$ benefits both equally and cancels in the ratio. The structural factor—the CapEx Crossover Identity—is:

$$\boxed{\frac{\tilde{C}(t, \tau)}{\tilde{C}_F(t)} = \frac{1}{[g_q \cdot (1-\delta)]^\tau}}$$

When $g_q(1-\delta) < 1$, plateau hardware delivers more useful capability per dollar than frontier hardware on the binding dimension, with an exponentially growing advantage in $\tau$. This is a CapEx-only object. Operating costs enter in §H.5 below.

The threshold condition.

$$g_q \cdot (1-\delta) < 1 \iff \delta > \frac{g_q - 1}{g_q}$$

The secondary market must reprice faster than per-chip frontier capability advances on the dimension at hand. The condition is dimension-specific:

Binding capability dimension (per chip, H100→B200, 3-yr annualised)	$g_q$	Threshold $\delta^*$	Observed $\delta$	Identity fires?
FP16/BF16 throughput per chip	1.31	0.24	0.25	Yes (marginal)
Memory bandwidth per chip	1.34	0.25	0.25	Yes (marginal)
Memory capacity per chip	1.34	0.25	0.25	Yes (marginal)
FLOPS per Watt per chip	1.47	0.32	0.25	No
FP8 throughput per chip	1.65	0.39	0.25	No

At the rounded reference values, the purchase-cost advantage is marginal on bandwidth, capacity, and FP16/BF16 throughput. FP8 and FLOPS/W favour newer silicon. Rising launch prices change those thresholds.

The launch-price correction. Flagship retail prices have risen at $p_F \approx 1.05$/yr (H100 SXM5 ~$30–35K in 2023, B200 SXM ~$40K in 2025). Including this in the derivation: a plateau chip launched $\tau$ years ago had launch price $P_0/p_F^\tau$, and its used price today is $(P_0/p_F^\tau)(1-\delta)^\tau$. The full identity:

$$\frac{\tilde{C}(t, \tau)}{\tilde{C}_F(t)} = \left[\frac{p_F}{g_q(1-\delta)}\right]^\tau$$

with crossover condition:

$$\delta > \frac{g_q - p_F}{g_q}$$

Rising launch prices lower the depreciation threshold:

Binding capability dimension	$g_q$	Threshold $\delta^*$ ($p_F = 1.05$)	Observed $\delta$	Identity fires?
FP16/BF16 throughput per chip	1.31	0.20	0.25	Yes
Memory bandwidth per chip	1.34	0.22	0.25	Yes
Memory capacity per chip	1.34	0.22	0.25	Yes
FLOPS per Watt per chip	1.47	0.29	0.25	No (marginal)
FP8 throughput per chip	1.65	0.36	0.25	No

Rising launch prices strengthen the first three comparisons. FP8 and FLOPS/W still favour the frontier at these assumptions.

Numerical validation. For mid-2026 used H100 ($15K, 3.35 TB/s, age 3 yr) versus new B200 ($40K, 8.0 TB/s, retail) on memory bandwidth, the full identity predicts:

$$\frac{\tilde{C}_P}{\tilde{C}_F} = \left[\frac{p_F}{g_q(1-\delta)}\right]^\tau = \left[\frac{1.05}{1.34 \times 0.753}\right]^3 = \left[\frac{1.05}{1.009}\right]^3 = (1.041)^3 \approx 1.12$$

The annualised prediction is about 12%; Table H.2 gives 11.5% directly (0.223 versus 0.200 GB/s per dollar). These are two calculations of the same price-to-bandwidth comparison. The next question is how much capability the job needs.

H.1: Log capability scaling. Inference benchmark scores scale approximately as $\beta_K \log_{10}(\text{compute})$. A 4× FLOPS deficit corresponds to a benchmark gap of roughly 3–5 percentage points on MMLU, GPQA, or HumanEval at current scale. Many useful workloads have a capability threshold: a customer-support router designed around 85% intent-classification accuracy can buy the cheapest model that clears it. As $E(t)$ grows, older hardware clears more of those thresholds.

H.2: Memory-bound, not compute-bound. Autoregressive LLM inference is memory-bandwidth-bound, not FLOPS-bound. The relevant cap-per-dollar metric is bandwidth per dollar:

Hardware	State	Price	HBM bandwidth	GB/s per $
H100 SXM5	Retail (2023)	$35,000	3.35 TB/s	0.096
H100 SXM5	Used (mid-2026)	$15,000	3.35 TB/s	0.223
B200 SXM	Retail (mid-2026)	$40,000	8.0 TB/s	0.200

Used H100 delivers ~12% more memory bandwidth per dollar than a new B200 at retail. The B200’s structural advantages—192 GB VRAM, 4× FP8 throughput, NVLink 5 at 1.8 TB/s—bind on training workloads with synchronous gradient sync and on models exceeding 80 GB. Most production inference exhibits neither. Plateau hardware does not beat frontier globally. It beats frontier on the dimensions that govern specific deployment workloads.

Latent recursion. Research architectures such as COCONUT, looped transformers, and HRM/TRM reuse a small block across reasoning steps. Greater depth without proportionally more weights could help older chips clear H.1’s capability thresholds.

A fixed latent state with a flat KV cache shifts the constraint towards serial latency and compute efficiency. Sparse routing, proposed as a remedy for representation collapse in dense looped blocks, can instead require repeated HBM reads. These regimes favour different hardware.

The proposed frontier core keeps a compact working set near compute, following the Groq/Cerebras direction. A distilled, weight-streaming MoE descendant may remain limited by memory bandwidth per dollar. The first favours new hardware; the second can favour the plateau.

This is a predicted architecture split, not an established production trend. Test each design by the resource its workload makes scarce.

H.3: Architectural Divergence Penalty. Software gains need not transfer equally. Blackwell’s native FP4 and newer memory hierarchies can leave Hopper paying an emulation or compatibility cost. Let $\mu(t,\tau)\geq1$ represent that penalty:

$$\frac{\tilde{C}(t, \tau)}{\tilde{C}_F(t)} = \frac{1}{\mu(t, \tau) \cdot [g_q \cdot (1-\delta)]^\tau}$$

$\mu$ offsets the depreciation advantage. The penalty is expected to be modest for portable FP16/BF16/FP8 inference and larger for newer formats. Sustained lockout across useful workloads would defeat the migration thesis; distillation, quantisation, and portable kernels work against it.

H.4: VRAM step-function. A 70B-parameter model at FP8 occupies ~70 GB before KV cache. On H100 (80 GB), it fits with little headroom and typically requires two-GPU tensor-parallel configurations, introducing all-reduce communication overhead of 9–23% of end-to-end decoding latency. On H200 (141 GB) or B200 (192 GB), the same model fits single-GPU with concurrent KV caches and no TP penalty. The CapEx identity assumes capability decays smoothly. In reality, model-fit thresholds create non-linear cliffs. Plateau hardware wins inside the cliff and loses outside it. The migration thesis holds because most production workloads fit inside the cliff for previous-generation hardware, even as frontier training requires the next cliff.

H.5: The Economic Crossover (OpEx-extended). The first identity compares purchase costs. The full comparison also needs running costs. For a workload meeting a common quality threshold, define amortised capital cost per useful operation as $A(\tau, t, s, w)$ and operating cost per useful operation as $O(\tau, t, s, w)$, where $s$ indexes the slot and $w$ the workload. Economic capability per dollar:

$$\tilde{C}_{\text{econ}}(\tau, t, s, w) = \frac{1}{A(\tau, t, s, w) + O(\tau, t, s, w)}$$

Both costs are in dollars per useful operation, so their reciprocal gives useful operations per dollar. Throughput, software efficiency, architectural penalties and utilisation already enter the allocation of those costs over useful work. The comparison includes only configurations where the model clears the capability threshold, fits in memory, and has software support.

Electricity sets one operating floor. For a matched workload and accounting interval, revenue per token below the electricity cost per token cannot cover even that part of the running cost:

$$R_{\text{token}} \;<\; \frac{\mathrm{PUE} \cdot P_{\mathrm{IT,kW}} \cdot P_{\text{elec,site}}}{3600 \cdot \mathrm{TPS}}$$

Here $R_{\text{token}}$ is in $/token, $P_{\mathrm{IT,kW}}$ is average IT power allocated to the workload in kilowatts, $\mathrm{PUE}$ allocates facility overhead, and $P_{\text{elec,site}}$ is the site’s electricity cost in $/kWh. Power and token throughput $\mathrm{TPS}$ must cover the same interval, including any idle energy charged to the workload. Thermal design power can serve as a planning proxy; measured operating draw closes the comparison. The factor of 3600 converts hours to seconds.

Sanity check. A workload averaging 1 kW of IT power at $\mathrm{PUE} = 1.2$, $0.10/kWh electricity, and 100 TPS yields:

$$\frac{(1.2)(1)(0.10)}{3600 \cdot 100} \;\approx\; 3.3 \times 10^{-7} \; \text{\$/token}$$

—or roughly $0.33 per million tokens for electricity alone.

The heat-rate penalty in Appendix F raises the fuel needed per kilowatt-hour. It multiplies the fuel component of generation cost, not an entire utility tariff. Grid charges, fuel prices, plant costs and utilisation determine the actual site comparison. This is the link to plateau migration: written-down silicon still needs an affordable place to work. A university cluster, enterprise datacentre or regional cloud can offer one; its electricity bill and workload decide.

H.6: The workflow-discovery recursion. AI-assisted engineering can accelerate $E(t)$ by adapting software to available hardware. The prediction is $\beta>0$. The purchase-cost identity itself depends on shared software gains cancelling, rather than on their growth rate.

A stronger illustrative result follows if engineering throughput $L(t)$ scales with deployed plateau intelligence, $L(t) = L_0 \cdot E(t)$. Define the rate of software-efficiency improvement as:

$$\frac{d \log E}{dt} = \alpha_0 + \beta \log L(t)$$

Substituting the recursion and solving the resulting linear ODE in $u = \log E$:

$$E(t) = E_\infty \cdot \exp\!\big(A \, e^{\beta t}\big)$$

This illustrative double exponential suggests how software feedback might accelerate adoption. Verification, testing, integration, and organisational change constrain it. It is not a forecast; the weaker prediction is positive feedback.

H.7: Frontier slowdown. The $g_q$ values above assume the frontier continues to compound at recent rates. Three constraints suggest the rate may not hold indefinitely.

Moore’s law has stalled at the device level. The frontier still compounds, but through capital intensity, packaging innovation, and system-scale engineering rather than transistor scaling. Hyperscaler AI spending exceeds $700 billion in 2026.^[1] Frontier training compute has grown ~5×/year since 2020, but hardware FLOP/$ has improved only ~1.37×/year. Capability appears to scale approximately logarithmically with compute on standard benchmarks, so each additional order of magnitude buys a roughly fixed absolute increment, not a proportional one.

Slower frontier growth lowers the threshold. The illustrative $g_q=2.0$ per-release case and $g_q\approx1.15$ annual case use different intervals; comparisons must match them to depreciation. Continuing progress supplies new capabilities to distil. Slower progress can extend the useful life of existing capabilities.

H.8: Boom-driven input-cost inflation. Scarce HBM, packaging, substrates, networking, and power delivery raise new-hardware costs. The launch-price factor $p_F$ carries that pressure into the crossover.

Epoch AI’s teardown places B200 module production cost at roughly $5,700–7,300, with HBM and advanced packaging accounting for ~two-thirds of variable unit cost.^[26] SemiAnalysis estimates memory could rise to ~30% of hyperscaler AI data-centre capex in 2026, up from ~8% in 2023–2024, with HBM undersupplied through 2027.^[27] If boom-driven inflation pushes $p_F$ from 1.05 to 1.08 over the next two generations, the threshold on FLOPS/W (currently failing at 0.29 against observed $\delta = 0.25$) drops to ~0.26, on the cusp of firing. The wedge does not need the frontier to slow—only to keep getting more expensive.

Committed supply. Order books extend through Rubin and Rubin Ultra in 2027–2028 and Feynman in 2028–2029. The thesis expects a bust to reprice that committed pipeline faster than it changes delivery schedules. New and pre-ordered generations face the same slot economics.

Three depreciation regimes. The identity rests on distinguishing three depreciation rates that, in AI hardware, move at different speeds:

Regime	Driven by	Timescale	What it means
Frontier obsolescence	$g_q$, $\mu$	~2 years	Chip no longer optimal for the next-largest frontier training run
Capability obsolescence	Software stack abandonment, physical failure	5–10 years	Chip cannot run any useful workload
Economic obsolescence	OpEx vs revenue (slot $s$, workload $w$, site electricity cost)	Set by slot, workload, and electricity price	Output worth less than operating cost

Three empirical tests. Keep the dates and observations distinct:

Resale, end-2027. B200 SXM should clear at $22–28K, consistent with roughly 25–35% annual depreciation over two to three years. Prices above $35K into 2028 would contradict the predicted repricing mechanism and its empirical crossover.

Deployment, end-2028. Hopper and early Blackwell should carry the largest share of useful AI work, assessed through tokens, queries, decisions, and dollars saved. A majority remaining on contemporary frontier silicon would defeat the workload prediction.

Access pricing, end-2028. The hyperscaler/specialist spread on identical H100/H200 hardware should remain wide or widen from a current maximum near 10×. Convergence below approximately 2× would defeat the predicted separation of premium access from the cheaper substrate, and the associated revenue/workload-share claim. Check public prices monthly. Appendix J

Where repricing survives compatibility penalties and running costs, the second buyer inherits useful capacity at a price the first buyer could not obtain.

I. The Cascade in Progress

Older generations already serve different kinds of work at different prices. A bust makes that migration larger and more visible.

Generation-by-generation status, mid-2026.

Generation	Launch	Launch price (top SKU)	Secondary clearing (mid-2026)	Current production role
Turing T4	2018	~$2,500–10,000 depending on SKU	~$1,500–2,500	Default inference on AWS g4dn, GCP equivalent SKUs; embeddings, ranking, smaller-model serving across enterprise
Ampere A100	2020	~$15,000–20,000 (40 / 80 GB SXM4)	~$5,000–8,000	Primary production workload on AWS p4d, Azure ND A100, GCP a2; substantial share of enterprise AI inference globally
Hopper H100	2022/2023	~$30,000–35,000 (SXM5)	~$15,000 (per Appendix H)	Current workhorse; frontier-adjacent; per Table H.2 already trades at higher memory-bandwidth-per-dollar than new B200 retail
Hopper H200	2024	~$30,000–40,000 (limited disclosed)	Limited secondary supply	Current production tier; increasingly inference-priced as Blackwell capacity ramps
Blackwell B200	2024/2025	~$40,000	Negligible secondary supply	Current frontier tier; first observable depreciation cycle still 12–24 months ahead

Cloud SKU repricing as the dominant migration channel. Public AWS pricing data shows g4dn (T4) on-demand prices have fallen roughly 60–65% relative to 2019 launch; p4d (A100) by roughly 40–55% relative to 2021 launch. The chips never left Amazon. The migration ran entirely inside the hyperscaler’s balance sheet, visible only as falling unit prices on customer-facing SKUs—the quiet form of the cascade the body describes.

Appendix J records the second price surface: hyperscaler versus specialist access to the same hardware. Its spread measures the service bundle as well as chip economics.

Hyperscaler depreciation extensions. Microsoft (2022), Alphabet (2023), Meta (2023), Oracle (2024), and Amazon (2024) have each extended useful-life assumptions for server infrastructure from approximately four years to six years. Their audited accounts now assume a longer working life for the fleet. That expectation supports the distinction between leaving the frontier and ceasing to be useful.

Microsoft’s FY2023 extension alone added ~$3.7 billion to annual operating income. The aggregate impact across the top five hyperscalers is on the order of $15 billion per year in reduced depreciation expense. The same CFOs are simultaneously issuing record bond volumes to fund new silicon and extending useful-life on the existing fleet. Together they imply that the productive compute fleet is becoming both larger and longer-lived than the original capacity model assumed.

Falsification test on depreciation policy. The cleanest forward signal against the plateau thesis would be a hyperscaler shortening AI-GPU depreciation in audited filings. As of mid-2026, no major hyperscaler has done so. The keynote curve makes capability claims unaudited. The audited curve makes utility claims unmarketed. The thesis tracks the second.

Anthropic / Colossus as cascade in miniature. In May 2026, Anthropic took over the entire ~290 MW capacity of Colossus 1, the Memphis facility built around 220,000+ Nvidia processors. The SpaceX S-1 disclosed an agreement at $1.25 billion per month through May 2029, with reduced ramp fees and a 90-day termination right for either party.^[13] xAI shifted internal frontier training to Colossus 2 in Southaven, Mississippi, deploying next-generation GB200 Blackwell. The older facility acquired a new tenant while its builder moved on.

Abilene’s change of hands. The planned Oracle/OpenAI Stargate expansion from 1.2 GW to 2.0 GW collapsed in early 2026 amid financing complexity, revised OpenAI demand forecasts, and a winter weather event that took the liquid-cooling loop offline for several days. Microsoft took over the adjacent expansion with Crusoe, adding two buildings and an on-site 900 MW gas plant alongside Oracle’s existing 350 MW backup.^[14]

Cost-per-token compression as observable substrate-level repricing. Cost-per-token at GPT-3.5-equivalent capability fell ~280× over 2022–2024 (Appendix E). The decline is software-dominant. But the hardware-side decay rate is independently visible in SKU prices on ageing silicon, tracking the $\delta \approx 0.25$ band the framework predicts.

J. Hyperscaler Premium and the Plateau Substrate

The same chip sells at different rental prices depending on the services bundled with it.

Snapshot, May 2026. Public pricing for H100-class capacity, by access tier:

Access tier	Example	Per-H100 GPU-hour
AWS on-demand	p5.48xlarge (8× H100)	~$6.88 ($55.04/hr ÷ 8)
AWS reserved (Capacity Blocks)	p5.48xlarge reserved	~$4.33 ($34.61/hr ÷ 8)
Azure on-demand	ND H100 v5 class	~$12.29
Specialist cloud, mid-band	CoreWeave, Lambda, Civo, Denvr	~$2.25–$4.29
Marketplace / spot network	vast.ai-class supply, GPUPerHour	~$1.25–$2.75
Owned hardware, amortised	New H100 PCIe ~$25K; SXM ~$35–40K	under $1.50 at high utilisation

The maximum spread is roughly 10×. It prices availability, support, procurement, billing, and trust alongside hardware. Some customers need that bundle; others can trade it for lower cost.

What the spread implies. A chip can leave the frontier while its cloud rental stays expensive. The H100 remained priced as scarce on hyperscaler clouds after Blackwell launched, while marketplace rates approached amortised ownership cost. AWS raised H200 Capacity Block prices by 15% in January 2026.^[6] The premium buys a ready-to-use slot as much as a chip.

The bifurcation, mapped. The premium access layer captures: largest training runs, premium managed inference, latency-sensitive global APIs, regulated enterprise workloads, customers paying for procurement simplicity, organisations without internal MLOps depth. The economic substrate captures: batch inference, embeddings, fine-tuning, document processing, simulation, scientific computing, bounded agentic systems, retrieval and ranking pipelines, narrow-vertical Stockfish workflows. Both surfaces persist. They serve different customers.

Revenue and workload can separate. Hyperscalers could retain most revenue through the enterprise premium while lower-cost providers serve most tokens, queries, or useful inference-hours. The prediction concerns both shares, not revenue alone.

Multi-model routing volume as workload-share signal. OpenRouter—the multi-model routing platform serving 8M+ developers—disclosed token throughput growing from ~5 trillion to ~25 trillion tokens per week over the six months ending May 2026, a 5× expansion across a heterogeneous plateau-tier model surface (Anthropic, OpenAI, Google, Meta, DeepSeek, and open-weights).^[28] CapitalG led its $113M Series B with NVIDIA Ventures, Snowflake, Databricks, MongoDB, ServiceNow, a16z, and Menlo Ventures—the enterprise-data and silicon stacks underwriting the multi-model topology rather than a winner-take-all frontier outcome. The data confirms the topology the framework predicts. Whether the absolute level is durable depends on how the 2026–2027 procurement reset resolves.

Falsifier. The same-hardware spread falling below approximately 2× by end-2028 defeats this prediction; it is the third test in H. The May 2026 snapshot shows widening. See Appendix H.

Moving prices. Capacity Blocks, promotions, and spot markets reprice. The test concerns a durable spread on comparable hardware and service terms, not any single quoted rate.

Cycle prediction, recorded July 2026. The spread stays near 10× through the installation window, then compresses after repricing towards 2–3×. Falling below the 2× floor refutes the proposed bifurcation.

K. Macroeconomic Absorption and the Deployment-Phase Drag

The body argues that adoption takes years while hardware loses value in quarters. This appendix examines the gap between gains inside individual firms and gains across the economy. Brynjolfsson’s Productivity J-Curve explains the work needed between buying a technology and using it well. Carlota Perez places that work in the longer history of infrastructure booms. My reading is a 5–10 year primary deployment phase, with the economy still absorbing the change beyond 2035.

K.1. The Brynjolfsson J-Curve and intangible capital accumulation

Brynjolfsson, Rock, and Syverson (2021) formalised why a massive general-purpose technology boom can co-exist with stagnant aggregate productivity.^[29] Firms must accumulate substantial intangible capital—process redesign, retraining, verification scaffolding, organisational restructuring—before the technology’s productivity dividend appears in measured output. Because the intangible accumulation is expensed as opex (not capitalised), national accounts systematically under-measure both GDP and TFP during the absorption phase. Historical adjustments for computer-era intangibles found that true TFP was 15.9% higher than official measures by the end of 2017.

The J-Curve is now quantified in current AI data. PricewaterhouseCoopers’ 2026 AI Performance Study (1,217 senior executives across 25 industries) finds 74% of AI’s measurable economic value captured by the top 20% of firms, leaving 26% to the remaining 80%.^[30] The long tail of adopters has not yet completed the intangible-capital accumulation that would let them realise the gains, even where they have purchased the software.

The cash budget ratio is now visible. The SXSW 2026 CMO Survey of 400 organisations finds that positive-ROI AI deployments spent approximately $2.60 on training and change management per $1 spent on the AI software itself, with organisations failing to match this ratio experiencing tool abandonment rates of 60–70% within six months.^[31]

The implied market-value ratio is older but consistent. Brynjolfsson, Hitt, and Yang (2002) found $1 of physical computer hardware historically associated with approximately $9 of corporate market value—the market pricing the unmeasured complementary intangible capital long before the national accounts measured it.

K.2. Enterprise deployment friction is structural, not transient

The deployment friction is not theoretical. S&P Global Market Intelligence and the RAND Corporation (2025) report that 42% of corporate AI projects were scrapped in 2025 and 80.3% of AI initiatives failed to deliver business value—twice the failure rate of traditional, non-AI IT projects.^[32] Gartner projects that more than 40% of agentic AI projects will be cancelled by end-2027 due to escalating costs, lack of clear business value, and inadequate risk controls.^[32] Although 97% of surveyed enterprises have experimented with AI agents in some form, only 10–12% have successfully transitioned them into production environments.

At the workflow level, the downstream verification burden compounds the friction. The METR randomised controlled trial (2025) found experienced open-source developers were 19% slower using AI tools. The cognitive cost of reviewing non-deterministic AI output exceeded the speedup from generation.^[33]

Google’s DORA 2024 report associated a 25% increase in AI adoption with a 7.2% decrease in production delivery stability. Google Cloud (2026) reports 45% higher burnout among frequent AI users, with approval fatigue identified as the primary mechanism. GitClear’s analysis of 153 million changed code lines projected a doubling of code churn translating directly into delivery instability.

The proposed mechanism is that verification and integration absorb some isolated-task speedups, reported at 20–56%. These studies measure different tasks and outcomes; they do not supply one common estimate.

K.3. The labour-market signal

The Stanford AI Index 2026 reports roughly 20% lower early-career software-developer employment (ages 22–25) in exposed roles from 2024 to early 2026, alongside stable aggregate white-collar wages. This is consistent with human work shifting towards verification, integration, and architecture.^[34]

The employment pattern does not identify that mechanism. Hiring, demand, and organisational change also matter. Symbolic Superintelligence proposes testing what work remains and what makes the next task automatable.

K.4. Aggregate productivity dispersion and the Perez modification

U.S. aggregate labour productivity has grown at 1.6% annualised since Q4 2019—a modest acceleration from the 1.2% pre-pandemic decade. Forward projections are widely dispersed.

Goldman Sachs Global Macro Research projects AI-driven labour productivity acceleration to 1.7–1.9% through 2029, peaking at 1.9–2.3% in the early 2030s, with potential GDP growth elevated to 2.1–2.3% for the rest of the decade.^[35]

Penn Wharton Budget Model projects permanent annual potential-output gains of less than 0.04 percentage points, with aggregate GDP rising only 1.5% by 2035.^[36] Acemoglu (2024) similarly estimates aggregate GDP gains of 1.1–1.6% over a decade and a marginal annual TFP boost of approximately 0.05%.^[37]

The dispersion reflects genuine uncertainty about how fast the J-Curve resolves.

Perez’s installation → crash → deployment sequence separates speculative finance from production capital. After the telecom crash, consolidators such as Level 3 and Equinix acquired infrastructure cheaply and leased it into the next phase.

Hyperscalers combine those roles on one balance sheet. Cash flows and low-cost debt—$121B of bonds in 2025, per Note [2]—can fund new equipment while older chips move internally to inference. In this scenario, clearing concentrates among specialist clouds and developers; access changes hands through powered slots and leases as well as equipment sales.

K.5. Implications for the cascade

The adoption timetable changes the route to the plateau:

Timing. My estimate remains a 5–10 year primary deployment phase with further absorption beyond 2035. The evidence motivates a long adjustment; it does not measure that forecast.

Demand. In the optimistic Goldman case, rising workloads fill cheaper capacity sooner. In the Penn Wharton/Acemoglu case, slower adoption can extend the life of sufficient hardware. Both leave room for the plateau, provided useful demand covers its operating cost.

Access. Cheaper capacity lets firms experiment while they learn how to use the technology.

K.6. A note on method

The predictions are dated and testable. A failed prediction changes the argument.

M. The Q1 2026 Register

These figures preserve what was reported at the time. Revenue, losses, and adoption may change; the dated record lets later readers test the essay against what followed. Sources include Reuters, Bloomberg, and The Information.^[11]

• OpenAI: Q1 revenue ≈ $5.7 billion against an operating loss ≈ $6.95 billion (−122% margin), tracking toward a $36.6 billion annual loss. ChatGPT weekly active users stalled at 905 million against a 1 billion target. Sora shut down. The Disney partnership ended. The Walmart pilot shuttered.
• xAI: $2.47 billion Q1 loss on $818 million of revenue—roughly $1 billion a month of burn. Grok absent from the top-25 App Store downloads.
• Anthropic: $30 billion annualised revenue run rate by April—80× annualised Q1 growth against a 10× target, most plausibly a short-run anomaly of coding-adoption scaling rather than a rate that holds the full year. Claude Code past $1 billion ARR, on track for ≈ $11 billion of Q2 revenue at ≈ $600 million operating profit.

Adoption. Late-2025 surveys from Stack Overflow, JetBrains, and GitHub’s Octoverse put professional-developer adoption of AI tools above 70%.^[12] Growth increasingly depends on existing users doing more with the tools. Coding entered a field with compilers, tests, version control, and decades of deployment practice. Other fields are building their equivalents: NIST AI RMF, ISO/IEC 42001, sector-specific evaluation suites, and agent-testing tools.^[8] Medicine and finance show how capability can precede the institutions needed to use it.^[10]

Falsifier. The interpretation behind §II is that profitability follows the ability to put models to work. If, by 2028, the lab ranking by operating margin does not track the maturity of those deployment systems, the argument is wrong.

[1] Reuters / Breakingviews (2026). “How Big Tech’s USD 630 billion AI splurge will fall short.” See also Reuters market commentary: “Investors stay calm as AI capex boom eclipses dotcom mania.” (Industry capex aggregation across hyperscaler AI and cloud infrastructure guidance.)

[2] Bank of America Global Research / Moody’s Ratings / Reuters-linked market reporting (2026). “Amazon’s record USD 54 billion bond sale shows AI’s staggering costs.” See also Reuters commentary on AI infrastructure funding: “Investors stay calm as AI capex boom eclipses dotcom mania.” (Top-five U.S. hyperscaler bond issuance, 2026 issuance projections, and Amazon’s record debt transaction; a composite market-data reference standing in for the original Moody’s / BofA notes.)

[3] PitchBook (2026). “Q1 2026 AI VC Trends.” See also PitchBook coverage: “Q1 2026 AI funding blows past 2025 total with three deals accounting for 67% of capital.” (First-quarter AI venture funding of approximately USD 255 billion, dominated by mega-financings.)

[4] Reuters (2026). “Power grid delays challenge Amazon’s data center expansion in Europe.” (Amazon European infrastructure briefing on grid interconnection timelines, citing seven-year horizons against roughly two-year data-centre construction cycles.)

[5] CME Group / ECB / market quote pages (2026 snapshot). CME Henry Hub Natural Gas futures quotes. ECB euro foreign-exchange reference rates. ECB reference-rate archive for 15 May 2026. (Commodity and FX prices are dynamic; the essay’s arithmetic uses the May 2026 Henry Hub, TTF, and JKM observations recorded in Appendix B.)

[6] Amazon Web Services / Data Center Dynamics / InfoQ (2026). “Amazon EC2 Capacity Blocks for ML pricing.” “AWS quietly increases prices for H200 EC2 instances by 15%.” “AWS hikes EC2 Capacity Block rates by 15% in uniform ML pricing update.” (AWS H200 Capacity Block price increase in January 2026; interpreted as evidence of scarce energised, permitted capacity.)

[7] Cast AI (2026). “State of Kubernetes Optimization Report.” See also Cast AI summary: “2026 State of Kubernetes Resource Optimization: CPU at 8%, Memory at 20%, and getting worse.” (Average sustained GPU utilisation of approximately 5% across surveyed production clusters; not necessarily representative of frontier training clusters.)

[8] NIST / ISO (2023–2024). “NIST AI Risk Management Framework.” “Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1.” “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1.” “ISO/IEC 42001:2023—Artificial intelligence management system.” (Foundational AI governance frameworks; sector-specific extensions remain necessary for autonomous-agent deployment.)

[9] SAE International / Waymo / Reuters (2025–2026). “SAE J3016: Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles.” Waymo service areas. “Waymo launches robotaxi freeway service in San Francisco, LA, Phoenix.” (Level 4 autonomy as an analogue for the verification, regulation, and integration lag between impressive demonstrations and broad commercial deployment.)

[10] U.S. Federal Reserve / Office of the Comptroller of the Currency (2011–2026). “SR 11-7: Guidance on Model Risk Management.” “SR 26-2: Revised Interagency Guidance on Model Risk Management.” “OCC Bulletin 2026-13: Model Risk Management.” (Model-risk governance as a deployment gate for banking, capital-markets, and credit-decision systems.)

[11] The Information / Reuters / Axios / Wired / Anthropic (2026). The Information reporting on OpenAI and Anthropic Q1 2026 revenue. Axios: “Anthropic’s unprecedented revenue growth.” Wired: “OpenAI shuts down Sora.” Anthropic Series G funding announcement. (Composite Q1–Q2 2026 reference for private-company revenue, loss, ARR, and product-status claims; some source material is paywalled.)

[12] Stack Overflow / JetBrains / GitHub (2025). “Stack Overflow Developer Survey 2025: AI.” “JetBrains State of Developer Ecosystem 2025.” “GitHub Octoverse 2025.” (Professional-developer AI-tool adoption across independent industry surveys.)

[13] xAI / Reuters (2026). xAI: “New Compute Partnership with Anthropic.” Reuters reporting on Colossus 1 and Anthropic compute access. (Anthropic access to Colossus 1 and xAI / SpaceXAI compute reallocation; precise financial terms per the SpaceX S-1.)

[14] Data Center Dynamics / OpenAI / Reuters-linked reporting (2026). “Oracle/OpenAI drop plans to expand flagship Abilene Stargate site.” OpenAI: “Five new Stargate sites.” (Abilene expansion, tenant reallocation, financing complexity, and Microsoft / Crusoe / Nvidia-related reporting; composite reporting rather than a single transaction document.)

[15] Stanford HAI / Epoch AI / Artificial Analysis (2026). “Stanford HAI AI Index Report 2026.” “AI Index Report 2026: Economy Chapter.” “The Price of Progress: Price Performance and the Future of AI.” (Inference cost and fixed-capability cost declines over the 2022–2026 window.)

[16] Age of Wonders. “Create an Age of Wonders.” (Internal cross-reference for intelligence as infrastructure and the broader Age of Wonders thesis.)

[17] David, P. A. (1990). “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox.” American Economic Review, 80(2). RePEc record. (The canonical account of the multi-decade lag between electrification and its measured productivity payoff.)

[18] S&P Global Market Intelligence (2026). “Private equity investment surge sends US data center deals to 5-year high.” (Private equity deployment into U.S. datacentres during 2025.)

[19] U.S. Energy Information Administration (2024). “Average operating heat rate for selected energy sources.” (7,754 Btu/kWh used as the conversion basis for fuel-only generation cost.)

[20] Reuters (2025–2026). “Musk’s xAI buys third building to expand AI compute power.” See also Reuters reporting on Anthropic’s access to Colossus 1: “SpaceX signs cloud deal with Google.” (Colossus 1 reported holding more than 220,000 Nvidia processors; 300 MW expansion capacity and one-million-accelerator programme target referenced in xAI / SpaceXAI reporting.)

[21] Financial Times / Data Center Dynamics (2025–2026). “Oracle to buy USD 40 billion of Nvidia chips for OpenAI data centre.” “Oracle to spend USD 40 billion on Nvidia chips for OpenAI Texas data center.” (Abilene, Texas campus: 400,000 GB200-class chips, 1.2 GW design power, Oracle / OpenAI / Crusoe structure, and project financing.)

[22] NVIDIA / U.S. Department of Energy (2025). “NVIDIA and Oracle to Build U.S. Department of Energy’s Largest AI Supercomputer for Scientific Discovery.” “Energy Department Announces New Partnership with NVIDIA and Oracle to Build Largest DOE AI Supercomputer.” (Combined 110,000 Blackwell GPUs planned for Solstice and Equinox at Argonne National Laboratory.)

[23] NVIDIA / Meta Platforms (2026). “Meta Builds AI Infrastructure With NVIDIA.” See also Reuters: “Nvidia to sell Meta millions of chips in multiyear deal.” (Infrastructure roadmap referencing large-scale Blackwell- and Rubin-class deployments.)

[24] Reuters / industry power-density references. “Reuters reporting on Colossus 1 capacity and processor count.” (1.3 kW all-in facility load per top-end accelerator used as an order-of-magnitude heuristic, consistent with reported Colossus processor count and associated facility power. Real deployments vary with cooling architecture, networking density, and silicon mix.)

[25] Blackstone / VoltaGrid / Halliburton (2026). “VoltaGrid Announces USD 1 Billion Strategic Equity Investment from Blackstone and Halliburton.” VoltaGrid official announcement. (Modular natural-gas generation systems, Propell acquisition, 7.5 GW forward order book, and data-centre / industrial-site deployment.)

[26] Epoch AI (2026). “How much does it cost to produce an Nvidia B200 GPU?” “AI chip component cost shares.” (Component-level teardown of Nvidia B200 production cost; HBM and advanced packaging as dominant variable-cost drivers.)

[27] SemiAnalysis / Epoch AI / Tom’s Hardware (2025–2026). “Memory Mania—How a Once in Four Decades Memory Cycle Is Emerging.” SemiAnalysis datacenter industry model. “Memory will consume 30% of hyperscaler spending this year.” Epoch AI: “AI chip supply chain constraints.” (HBM, packaging, memory-share, and supply-chain constraints across the AI accelerator stack.)

[28] OpenRouter / BusinessWire (2026). “OpenRouter Raises USD 113 Million CapitalG-led Series B as Weekly Volume Explodes to 25T Tokens.” (OpenRouter as an inference-routing and workload-aggregation layer across heterogeneous model providers.)

[29] Brynjolfsson, E., Rock, D., & Syverson, C. (2021). “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies.” American Economic Journal: Macroeconomics. NBER Working Paper PDF. See also Atlanta Fed / Richmond Fed / Duke CFO Survey: “Artificial Intelligence, Productivity, and the Workforce.” Working paper PDF. (The structural lag between GPT arrival and measured productivity gains, mediated by complementary intangible capital.)

[30] PricewaterhouseCoopers (2026). “PwC 2026 AI Performance Study.” (Executive survey on AI value concentration, business-model reinvention, and autonomous decision-making.)

[31] SXSW CMO Survey / Brynjolfsson, Hitt & Yang (2002). “CMOs at SXSW 2026: For every AI dollar, invest USD 2–3 in training.” “Intangible Assets: Computers and Organizational Capital.” Brookings paper PDF. (On complementary intangible capital, training, change management, workflow redesign, and organisational adaptation.)

[32] S&P Global Market Intelligence / Gartner / Reuters (2025–2026). “Generative AI shows rapid growth but yields mixed results.” “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” Reuters coverage of Gartner’s agentic-AI forecast. (Enterprise AI implementation failure, project abandonment, and agentic-AI cancellation forecasts.)

[33] METR / Reuters / Google DORA / Google Cloud / GitClear (2024–2026). “Measuring the impact of early-2025 AI on experienced open-source developer productivity.” METR working paper on arXiv. Reuters coverage of the METR study. Google DORA Report 2024. “When AI writes the code, who reviews it?” GitClear AI Copilot Code Quality Report 2025. (Productivity-reliability paradox, review burden, approval fatigue, code churn, and delivery-stability concerns.)

[34] Stanford Human-Centered AI Institute / Stanford Digital Economy Lab (2026). “AI Index Report 2026.” “AI Index Report 2026: Economy Chapter.” “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” (Early-career software-developer employment effects and seniority-skewed labour-market impacts.)

[35] Goldman Sachs Global Macro Research (2023–2026). “Generative AI could raise global GDP by 7%.” (Goldman Sachs optimistic productivity and GDP uplift projections for generative AI.)

[36] Penn Wharton Budget Model (2025). “The Projected Impact of Generative AI on Future Productivity Growth.” (Conservative long-run estimates for generative AI’s effect on U.S. GDP and productivity.)

[37] Acemoglu, D. (2024). “The Simple Macroeconomics of AI.” NBER Working Paper No. 32487. RePEc record. (Conservative macroeconomic estimate of generative AI’s GDP and productivity effects.)

[S] Stockfish developers (2020). “Introducing NNUE Evaluation.” See also: Silver, D. et al. (2018), “AlphaZero: Shedding new light on chess, shogi, and Go.” (Both systems combine learned evaluation with search. Their comparison does not isolate engineering from learning.)

Background references

DeepMind / OpenAI / Anthropic technical references on distillation, quantisation, inference efficiency, and model-system optimisation over the 2022–2026 window. DeepMind AlphaFold. OpenAI model optimisation documentation. Anthropic research.

BlackRock (2026). BlackRock Investment Institute. (EMEA client survey and AI infrastructure / energy-investment commentary; use the specific survey PDF where available.)

Reuters / U.S. Department of Energy (2026). DOE Loan Programs Office. DOE Office of Nuclear Energy. (U.S. grid-side and nuclear-capacity responses to AI power demand; cite specific programme releases where available.)

Brynjolfsson, E. & McAfee, A. (2014). “The Second Machine Age.” W.W. Norton.

Smil, V. (2017). “Energy and Civilization: A History.” MIT Press.

Perez, C. (2002). “Technological Revolutions and Financial Capital.” Edward Elgar.

NIST AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework

International Energy Agency (2024). “Electricity 2024.” (Data-centre electricity demand projections through 2026.)