# The economics of useful AI compute: why efficient silicon can coexist with a larger energy bill

> A source-led analysis of AI demand, model efficiency, HBM and networking constraints, useful utilization, electricity and capital recovery. Current primary evidence and reproducible scenarios show when efficiency substitutes for equipment and when it expands total consumption.

- **Category:** AI, Semiconductors & Compute Economics
- **Author:** Voltformer Research Desk
- **Publication date:** 2026-10-22
- **Reading time:** 19 min read
- **Key topics:** AI compute economics, semiconductors, HBM, inference efficiency, accelerator utilization, energy, capital expenditure
- **Body language:** en
- **URL:** https://voltformer.com/articles/ai-compute-economics-memory-utilisation-october-2026

By Voltformer Research and Editorial Team · Evidence retrieved 22 October 2026 · Analysis, with explicitly labelled conditional scenarios

AI infrastructure has entered a difficult economic phase: advances in chips and models are reducing the resources needed for particular computations while investors commit more capital to the machinery that performs them. The relevant question is how many economically useful tasks a system completes, at an acceptable level of quality and delay, over its operating life. Peak arithmetic, announced megawatts and raw token counts each measure only one part of that process.

Our thesis is conditional. Better efficiency raises the value of installed capacity when demand, software and supporting infrastructure can absorb the improvement. It reduces the need for additional capacity when demand saturates or smaller systems substitute for larger ones. The outcome depends on the response of demand, the distribution of workloads, effective utilization and the timing of supply commitments. There is no physical law requiring a cheaper token either to reduce total electricity consumption or to justify every planned accelerator purchase.

### 1. What the latest primary evidence establishes

TSMC's 15 October results put third-quarter US-dollar revenue at USD 46.72 billion; 2-nanometer technology supplied 8% of wafer revenue. Those figures describe manufacturing sales and product mix. They do not count installed AI systems. [TSMC, third-quarter 2026 results](https://pr.tsmc.com/english/news/3346).

NVIDIA reported USD 89.0 billion of Data Center revenue for the quarter ending 26 July, up 117% year over year. Strong supplier revenue establishes purchasing activity; customer profitability requires separate evidence. [NVIDIA, second-quarter fiscal 2027 results](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027).

| Dated evidence | Observation or claim | Classification | What remains unproved |
|---|---|---|---|
| 29 July: SK hynix | HBM4 mass shipments began in Q2 | Company-reported operational development | Qualified volumes available to each buyer [release](https://news.skhynix.com/en/q2-2026-business-results/) |
| 30 September: Micron | FY2026 net capital expenditure of USD 27.367bn | Reported non-GAAP measure | Output and yield of future capacity [release and reconciliation](https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Fiscal-Fourth-Quarter-and-Full-Year-2026-Results/default.aspx) |
| 17 September: MLCommons | Inference v6.1 adds end-to-end RAG and Agentic Edge tests | Benchmark release | An operator's production workload economics [working-group analysis](https://mlcommons.org/2026/09/chairs-mlperf-inference-v6-1/) |
| 12 October: AMD | Helion blueprint combines configurable compute with model routing | Product announcement | Installed fleet size and achieved efficiency [announcement](https://ir.amd.com/news-events/press-releases/detail/1301/introducing-amd-helion-agentic-ai-compute-at-rack-scale) |
| 19 October: Meta | Prometheus first phase completed and used for training | Operator-reported deployment | Industry-wide utilization [engineering account](https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/) |
| 21 October: Meta | FY2026 capex guidance narrowed to USD 135–145bn, including finance-lease principal | Management forecast | Realized returns on the spending [results](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx) |

The matrix is deliberately heterogeneous: financial results, deployed operations, research and announcements answer different questions. Treating them as interchangeable evidence creates a false impression that every stage between a wafer order and a profitable service has already been completed.

### 2. Demand must be measured in completed tasks

A token is an engineering accounting unit, not a constant unit of economic output. Different tokenizers divide the same material differently. A longer answer may contain additional value or additional waste. An agent can issue repeated model calls while waiting for tools, retrying failed execution or reviewing its own work. Revenue per token also depends on discounts, caching and the mixture of input and output traffic.

A useful demand model separates task arrivals, model calls per task, input and output lengths, and the probability of successful completion. For a fixed task population, more reasoning can increase computation while reducing human review. Alternatively, a smaller model can eliminate most of that computation without reducing task success. Both possibilities should be tested on the buyer's actual task distribution.

Define an accepted task through externally evaluated correctness, a latency limit and an explicit escalation policy. Record total accelerator time, CPU time and energy across retries. A supplier can grow token throughput while its customer sees no improvement in cost per successful task. Conversely, token throughput can fall while task economics improve because routing avoids unnecessary generation. This is why usage growth and productivity growth should be reported separately.

### 3. Silicon performance is a constrained flow

The Roofline model relates achievable arithmetic performance to compute capability, memory bandwidth and arithmetic intensity. In units, operations per second cannot exceed the smaller of peak operations per second and bytes per second multiplied by operations per byte. It is a bound for understanding bottlenecks, not a promise of application throughput. [Williams, Waterman and Patterson, Berkeley technical report, 2008](https://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/EECS-2008-134.html).

For AI procurement, extend that reasoning across the system. Prompt processing, token generation, expert routing and communication impose different resource demands. A fast arithmetic engine cannot compensate for missing memory capacity. An uncongested fabric does not remove a sequential dependency. Purchased accelerators waiting for compatible firmware, network qualification or commissioning contribute no delivered work.

| Constraint | Relevant unit | Economic consequence | Verification required |
|---|---|---|---|
| Matrix computation | Operations/s at stated precision | Determines one ceiling on execution | Same model, precision and accuracy |
| Memory capacity | Bytes available to weights and state | Limits residency and concurrency | Capacity after runtime overhead |
| Memory movement | Sustained bytes/s | Can govern generation speed | Application traces, not peak pins |
| Scale-up and scale-out fabric | Bytes/s and latency | Changes synchronization and idle time | Collectives under realistic contention |
| Reliability | Completed work per scheduled interval | Controls useful output from installed equipment | Restarts, checkpoints and spare policy |
| Commercial demand | Accepted tasks/hour | Determines whether technical capacity earns revenue | Paid usage and renewal cohorts |

These constraints interact. Increasing batch size can reuse weights more effectively but worsen response time. Spreading a model across more devices can solve capacity problems while adding communication. The economically preferred configuration lies on a quality, delay and cost frontier; the highest headline throughput may sit outside it.

### 4. HBM scarcity and model compression are simultaneous forces

HBM capacity and bandwidth are separate constraints. Compressing a model's state may allow more concurrent requests, but it does not automatically eliminate the cost of reading weights or exchanging routed tokens. Larger batches can convert saved memory into additional throughput. Longer contexts can consume that saving instead.

The 17 September DeepSeek-AI V4.1-Flash preprint reports an HBM-resident global KV cache of 890 bytes per token, approximately one quarter of V4-Flash's footprint; it also distinguishes persistent state held outside HBM. This is a model-specific research claim, with no independently established fleet electricity saving implied. [DeepSeek-AI, V4.1-Flash technical report, v1](https://arxiv.org/abs/2609.19969v1).

The economic implication is a change in the scarcity rent attached to memory for particular workloads. If existing demand fits into fewer devices, a buyer can defer purchases. If demand was constrained by context length or concurrency, the same improvement can make additional services viable. Memory suppliers can therefore experience strong demand during a period of rapid compression, especially when new applications consume more state per task or more tasks become affordable.

For a purchase decision, ask which memory tier is saved, whether additional compute reconstructs the compressed state, and whether an offload path shifts pressure into host memory, storage or networking. A fall in HBM bytes per request is a useful result; its value depends on the resource that becomes binding next.

### 5. Serving software changes the denominator

PagedAttention's foundational contribution was to manage growing and shrinking KV caches with less fragmentation and duplication. That is a system allocation improvement rather than a change in a GPU's nominal arithmetic capability. [Kwon and colleagues, PagedAttention/vLLM paper, 2023](https://arxiv.org/abs/2309.06180).

In May 2026, Google described a UCSD implementation of diffusion-style speculative decoding on TPUs and reported an average 3.13-fold tokens-per-second improvement in the stated comparison. It is a measured workload result, not a universal energy multiplier. [Google Developers, DFlash implementation and benchmark description](https://developers.googleblog.com/en/supercharging-llm-inference-on-google-tpus-achieving-3x-speedups-with-diffusion-style-speculative-decoding/).

Speculation adds drafting work to reduce expensive target-model iterations. Its benefit depends on acceptance, model compatibility and batching. Quantization reduces data volume but must preserve the quality required by the task. Routing avoids large-model calls but can make mistakes that trigger escalation. Prefill and decode separation can improve resource matching while creating state-transfer costs.

Do not multiply published speedups from separate experiments. Two optimizations may remove the same bottleneck, compete for memory or lose their advantage at a different batch size. Re-measure the combined system under the intended traffic mix, including tail latency and failure recovery. The relevant counterfactual is the best feasible alternative available to the buyer, not an unoptimized historical baseline.

The arithmetic of an unchanged workload is revealing. If half of end-to-end execution time is spent in a component that improves tenfold while the remainder is unchanged, total speedup is `1 / [0.5 + 0.5/10] = 1.818`, not tenfold. More generally, with unchanged time share `a` and optimized-component speedup `s`, the bound is `1 / [a + (1−a)/s]`. Tool execution and communication may form that unchanged share. The model assumes no overlap and no additional work; its purpose is to expose the denominator, not predict a particular system.

### 6. Utilization has several incompatible meanings

Device occupancy, allocation, scheduled time, training goodput and revenue-producing utilization are distinct. A GPU can be allocated while waiting for data. A completed kernel can belong to a job that later restarts. A successful request may consume free promotional credits.

Meta's October account reports more than 85% goodput for pretraining and more than 90% for reinforcement learning. It also describes 37,656 provisioned GPUs supporting a 32,768-active-GPU run. Arithmetic gives a 14.9% inventory uplift over active count, or a 13.0% reserve share of provisioned inventory: different denominators. Neither percentage is an industry utilization estimate. [Meta, Prometheus engineering account](https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/).

Reserves can be economically rational. They raise installed capital per active device but reduce the expected cost of interrupted work. Larger fault domains make reliability engineering more valuable; reserve inventory can improve useful output even while reducing a superficial utilization ratio. A comparison must count all owned equipment and the work lost to failures over the same period. Do not multiply published goodput and reserve percentages unless their denominators and overlap are known.

### 7. From tokens per joule to delivered cost

MLCommons specifies that validated power measurements cover average AC power for the full benchmark system at the wall; a stated TDP or power-supply rating is not such a measurement. Its result categories also separate available systems from previews and research configurations. [MLPerf Inference Datacenter methodology](https://mlcommons.org/benchmarks/inference-datacenter/).

NVIDIA's July architecture account presents an up-to-tenfold agentic-throughput-per-energy comparison using an internal two-trillion-parameter mixture-of-experts workload. That identifies the manufacturer's comparison boundary; it cannot establish a tenfold reduction in every customer's bill. [NVIDIA, Rubin architecture account](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/).

Facility energy adds cooling and electrical losses to IT consumption. Delivered cost adds capital recovery, maintenance, operations and the cost of capacity held ready. Carbon additionally depends on the timing and location of electricity supply. A purchase agreement or annual certificate does not itself measure the marginal emissions associated with a particular inference. These quantities belong in separate columns.

For an explicitly hypothetical compute block, let utilization be `u`, productive output retention be `g`, busy throughput be `R` tokens/s, idle and busy average IT power be `P0` and `P1` kW, and overhead multiplier be `h`. With 8,760 hours and 31,536,000 seconds per assumed year:

`T = R × u × g × 31,536,000` tokens/year

`E = h × [P0 + u(P1 − P0)] × 8,760` kWh/year

`C_million = 1,000,000 × [K/L + F + pE] / T` dollars per million retained tokens.

Here `K` is installed capital in dollars, `L` is straight-line recovery in years, `F` is fixed annual operating cost and `p` is dollars/kWh. This simplified accounting model excludes financing, tax and residual value. It allocates all idle energy to the block; `g` accounts for useful-output losses without reducing energy already spent.

### 8. A reproducible scenario: utilization can dominate the tariff

All following inputs are assumptions, not hardware quotations or benchmark observations: `K=USD 4m`, `L=4`, `F=USD 150,000/year`, `R=100,000 tokens/s`, `g=0.90`, `P0=20 kW`, `P1=100 kW`, `h=1.20`, `p=USD 0.08/kWh`. Their purpose is sensitivity analysis. They represent no specific vendor product.

| Utilization | Annual facility energy, MWh | Retained tokens/year, trillion | Fixed recovery and operations, USD /million | Electricity, USD /million | Total, USD /million |
|---|---:|---:|---:|---:|---:|
| 25% | 420.480 | 0.70956 | 1.621 | 0.047 | 1.668 |
| 50% | 630.720 | 1.41912 | 0.810 | 0.036 | 0.846 |
| 75% | 840.960 | 2.12868 | 0.540 | 0.032 | 0.572 |
| 100% | 1,051.200 | 2.83824 | 0.405 | 0.030 | 0.435 |

At 50%, the power average is 60 kW before overhead. Annual energy is `60 × 1.2 × 8,760 = 630,720 kWh`; retained output is `100,000 × 0.5 × 0.9 × 31,536,000 = 1,419,120,000,000` tokens. Annual recovery plus fixed operations is USD 1.15m; electricity is USD 50,457.60. Division reproduces USD 0.845917 per million tokens.

![Scenario only: fixed recovery and operations versus electricity cost per million retained tokens, varying utilization. Inputs: capital USD 4m, four-year recovery, fixed operations USD 150k/year, 100,000 tokens/s busy throughput, 90% retention, idle/busy IT power 20/100 kW, overhead 1.20, electricity USD 0.08/kWh. Not an observed market price or benchmark. Source: Voltformer calculations.](https://voltformer.com/content/media/5385eb00de14a5c379d38f19d2e39dc8752d68b5e82798236e8035f7ec47eaa0.png)

[Open figure at full resolution](https://voltformer.com/content/media/5385eb00de14a5c379d38f19d2e39dc8752d68b5e82798236e8035f7ec47eaa0.png)

Figure 2. Assumed utilization sensitivity, with fixed and electricity cost separated. This is a scenario curve, not an observed market price or a recommended operating point.

Doubling the tariff at 50% adds about USD 0.03556 per million; reducing utilization to 25% at the original tariff adds about USD 0.82221. Energy still rises as the block performs more work. Lower unit cost and higher aggregate consumption appear together without contradiction. Different capital prices, workloads or useful lives can reverse the relative importance of the components.

### 9. Capital commitments can outlive a bottleneck

NVIDIA's July-quarter 10-Q records USD 279bn of supply and capacity commitments, up from USD 119bn in the previous quarter. The maturity profile below comes from Note 10; some agreements can be adjusted, rescheduled or cancelled subject to terms and potential costs. It is a financial exposure, not a delivered-capacity schedule. [NVIDIA, fiscal 2027 Q2 10-Q, Note 10](https://investor.nvidia.com/files/doc_financials/2027/NVDA-2027-Q2-10Q-Final-including-exhibits.pdf).

![Reported supply and capacity commitments at 26 July 2026, USD bn. Fiscal buckets are the company’s. Rounded amounts sum to 279. Commitments are not delivered capacity. Source: NVIDIA Q2 FY2027 10-Q, Note 10.](https://voltformer.com/content/media/3960e3158f7ec88f13c5e3aefd29e537b1836cb2016a3e0c5b4d129a8c231a94.png)

[Open figure at full resolution](https://voltformer.com/content/media/3960e3158f7ec88f13c5e3aefd29e537b1836cb2016a3e0c5b4d129a8c231a94.png)

Figure 1. Reported supply and capacity commitments at 26 July 2026, USD bn: remainder FY2027 92; FY2028 87; FY2029 88; FY2030 6; FY2031 5; FY2032 and later 1. Rounded reported amounts sum to 279. Fiscal periods are the company's, not calendar years.

Capacity expands with delays. Buyers respond to a shortage by reserving components; suppliers respond by committing to equipment and facilities. If software subsequently reduces resources per task, contracted supply can arrive after the original shortage has eased. Conversely, underinvestment can prolong scarcity when several demand sources expand together. The cycle turns on timing and contractual flexibility as much as on long-run demand.

Microsoft's July call illustrates a further accounting issue: its datacenter and office-building useful-life revision from 15 to 25 years shifts some prospective leases between finance and operating classifications. Management adjusted calendar-2026 reported capex expectations to approximately USD 175bn while saying underlying investment expectations were unchanged outside that effect. Buildings and GPUs require separate treatment. [Microsoft, FY2026 Q4 earnings-call transcript](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4).

A longer depreciation schedule changes reported expense timing; it cannot improve obsolete equipment's physical output. Underwriting should test recoverable cash generation over an economic life, including resale, redeployment and migration cost. Supplier investment, customer capital expenditure and financing commitments also must not be added indiscriminately: overlapping transactions can describe the same equipment more than once.

The simplified scenario uses straight-line recovery for transparency. A financing model should instead discount cash flows and test terminal value. For level annual capital recovery, zero residual value, discount rate `r` per year and life `L`, the charge is `K × r / [1−(1+r)^(−L)]`, with the limit `K/L` at zero discount rate. Faster obsolescence raises the required annual charge even if the electricity tariff stays fixed. Delaying an equipment order can preserve information and avoid technological regret, but that option has a cost when customers would otherwise pay for scarce capacity.

### 10. Rebound is an empirical question

Let `f` be a genuine factor improvement in energy efficiency for a fixed-quality task. If the task price falls in proportion and task demand has elasticity `ε`, then the stylized relation is `Q_new/Q_old = f^ε`, giving `E_new/E_old = f^(ε−1)`. This assumes full price pass-through, unchanged task definition and no binding supply constraint.

For a fourfold improvement, elasticity 0.5 gives half the original energy; elasticity 1 leaves energy unchanged; elasticity 1.5 doubles it. These are algebraic cases, not estimated elasticities. Price pass-through may be partial because capital or scarcity rents dominate. Demand may respond to quality and convenience rather than price alone. Additional reasoning per task changes the task itself and requires another term.

The strongest rebound argument is that useful new applications become affordable. The strongest substitution argument is that routing, compact models and task saturation let operators deliver the same value with fewer expensive resources. Both can occur across different customers. Identify them through task-level consumption and cohort economics rather than treating one aggregate growth rate as a universal behavioral law.

### 11. Three conditional paths and their falsifiers

| Conditional path | Conditions required | Observable implication | Evidence that weakens it |
|---|---|---|---|
| Capacity absorbs efficiency | New paid tasks and more demanding work grow faster than resource savings | Revenue-producing throughput and power rise together | Weak renewals, idle inventory and discounting despite adequate supply |
| Efficiency substitutes for equipment | Stable task demand; routing and compression retain quality | Lower accelerator-hours per accepted task and delayed purchases | Savings disappear under real traffic or escalate human review |
| Supply arrives before monetization | Long commitments; commissioning succeeds; customer demand disappoints | Availability improves while realized returns deteriorate | Persistent qualified shortages with high paid utilization |

These paths have no assigned probabilities or target dates. Their value is to make decisions falsifiable. Aggregate demand can grow while one architecture, provider or region suffers overcapacity. Supply-chain exposure should therefore be tested at the component and customer level, with explicit assumptions about cancellation rights and useful life.

### 12. What operators and investors should measure next

An operator should procure against accepted tasks, quality and tail latency, then reconcile those results with metered IT and facility energy. Preserve the workload definition when comparing generations. Track paid utilization separately from technical occupancy, and inventory separately from active allocation. Include spare equipment, idle consumption and migration engineering in lifecycle cost.

An investor should test whether gross profit and cash collection catch up with installed capital. Follow component qualification, packaging yield, lead times, credit support and customer concentration alongside supplier revenue. Watch for a widening gap between commitments and actual deployment, and between deployed capacity and billable use. Price declines are ambiguous: they may reveal successful efficiency or an inability to earn capital recovery.

Our scenario's main limitation is its deliberate simplicity. It has one workload, linear power behavior, constant overhead, no queueing penalty and no financing. Public disclosures do not provide a representative, independently audited distribution of fleet utilization, realized accelerator purchase prices or successful-task economics. Those gaps prevent a credible universal cost-per-token estimate. The thesis would weaken if broadly comparable production measurements showed unit economics insensitive to utilization and supporting-system constraints. Until then, the defensible unit of analysis is useful work over an asset's life, with every denominator and measurement boundary visible.

### Primary bibliography and measurement notes

All sources were retrieved on 22 October 2026; none is treated as evidence of events after that date. Corporate operational and financial statements retain their reporting boundaries; the model paper is a preprint. The two figures combine one reported contractual dataset and one independently reproducible scenario, without merging their classifications.

1. [TSMC, *Third Quarter 2026 Results*, 15 October 2026](https://pr.tsmc.com/english/news/3346) — quarter ended 30 September; revenue and wafer technology mix.
2. [NVIDIA, *Q2 Fiscal 2027 Results*, 26 August 2026](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027) — quarter ended 26 July; segment revenue.
3. [NVIDIA, *Q2 Fiscal 2027 Form 10-Q*, signed 26 August 2026](https://investor.nvidia.com/files/doc_financials/2027/NVDA-2027-Q2-10Q-Final-including-exhibits.pdf) — Note 10, PDF page 18; commitments, not forecast shipments.
4. [SK hynix, *2Q26 Financial Results*, 29 July 2026](https://news.skhynix.com/en/q2-2026-business-results/) — company account of HBM4 shipment timing; preliminary financial release.
5. [Micron, *Fiscal Q4 and FY2026 Results*, 30 September 2026](https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Fiscal-Fourth-Quarter-and-Full-Year-2026-Results/default.aspx) — year ended 3 September; net capex reconciliation.
6. [MLCommons, *Inference v6.1 Working-Group Analysis*, 17 September 2026](https://mlcommons.org/2026/09/chairs-mlperf-inference-v6-1/) — benchmark definitions and new workload coverage.
7. [MLCommons, *Inference Datacenter Methodology*, live documentation](https://mlcommons.org/benchmarks/inference-datacenter/) — power boundaries and availability categories as retrieved.
8. [AMD, *Introducing AMD Helion*, 12 October 2026](https://ir.amd.com/news-events/press-releases/detail/1301/introducing-amd-helion-agentic-ai-compute-at-rack-scale) — architecture announcement, not fleet measurement.
9. [Meta Engineering, *Inside Prometheus*, 19 October 2026](https://engineering.fb.com/2026/10/19/data-center-engineering/inside-prometheus-meta-gigawatt-scale-supercluster/) — operator account of training, reliability and allocation.
10. [DeepSeek-AI, *DeepSeek-V4.1-Flash*, v1, 17 September 2026](https://arxiv.org/abs/2609.19969v1) — model-specific cache claims; preprint.
11. [Google Developers, *Diffusion-Style Speculative Decoding on TPUs*, 4 May 2026](https://developers.googleblog.com/en/supercharging-llm-inference-on-google-tpus-achieving-3x-speedups-with-diffusion-style-speculative-decoding/) — UCSD implementation and specified throughput comparison.
12. [NVIDIA, *Inside Rubin GPU Architecture*, 21 July 2026](https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/) — vendor performance comparison and its workload boundary.
13. [Microsoft, *FY2026 Q4 Earnings Call*, 29 July 2026](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4) — useful lives, lease classification and capex guidance.
14. [Kwon et al., *Efficient Memory Management for Large Language Model Serving with PagedAttention*, 12 September 2023](https://arxiv.org/abs/2309.06180) — original serving-system research.
15. [Williams, Waterman and Patterson, *Roofline*, UCB/EECS-2008-134, 17 October 2008](https://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/EECS-2008-134.html) — foundational performance model; historical research, not current hardware data.
16. [Meta, *Third-Quarter 2026 Results*, 21 October 2026](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx) — capex outlook is forward-looking, not realized deployment.

[← Back to Scope](https://voltformer.com/articles.md) · [View Related Equipment](https://voltformer.com/catalog) · [JSON](https://voltformer.com/articles/ai-compute-economics-memory-utilisation-october-2026.json)