Second review: research answers and editorial decisions
28 October 2026. Supports v14. Transcript suggestions were treated as hypotheses where the speakers explicitly marked uncertainty. Existing evidence remains in the first review audit and financial dossier.
Answers to the open questions
- Cost distribution: no credible general 50% people / 20–30% inference / 20% evaluation split was found. Evaluation overlaps labor and inference; a pie would imply mutually exclusive measured categories we do not have. Start with portfolio spending categories, then allocate to workflows and accepted tasks. Keep the tree sparse and use accent color for the current branch. Sources S01 and the existing FinOps unit-economics record.
- Two-year amortization: not a general rule. Depreciation, economic useful life and investment payback are different. Use a workload-specific comparison with equivalent quality, response time, availability and utilization. Simple payback months = incremental upfront investment / (monthly avoided alternative cost minus incremental monthly operating cost), only if the denominator is positive and reasonably stable. Include people, energy/cooling, facilities/networking, support and redundancy; show demand and replacement sensitivity. S04 illustrates differing accounting lives, not a recommended GPU lifetime.
- Terminology: use Where should the model run? with three choices: managed API, self-hosted on rented GPUs, owned/on-premises. Renting can offer elastic capacity; it does not inherently limit product scalability. Ownership changes capital, utilization and replacement exposure. S02.
- Why own more of the stack? The more AI determines differentiation, the more deliberately the business should control evaluations, data, integration, model choice and exit options. This is our strategic inference, not an empirical law; buying hardware remains a separate decision. Specific air-gap/data requirements may determine hosting eligibility; GDPR alone is not a blanket on-premises mandate. S02–S03.
- Batching language: say Process non-urgent work together, later; mention “batch processing” once in speech. It is appropriate when deadlines allow. Combining cache reuse and asynchronous batch pricing is possible, but does not guarantee a workflow saving of the same percentage. S12.
- Founder story: keep mutual introductions. At Klaus, use one short naming callback only; repeat the longer story only if useful in Q&A. This is a presentation judgment, not a research claim.
- Software always creates value? More pull requests do not establish better accepted output, revenue or ROI. Existing DORA and updated METR evidence require quality and end-to-end measurement. Show positive and null/negative examples without calling all outcomes financial ROI. Use S05–S07 plus retained QJE support evidence.
- Future cost curves: historical prices for fixed capability fell sharply (S08–S09). We found no defensible general forecast that compute first rises then falls in 2027–28, or that all workflow costs will fall. Distinguish token price, cost at fixed benchmark quality, hardware price-performance, and total workflow cost. More calls, harder tasks, reviews and increased demand can offset unit-price declines.
- 2027–28 graph: include transparent planning scenarios, not a claimed forecast. Index 2026=100; show flat unit cost, a chosen 50%-per-year planning assumption, and mechanical extrapolation of the observed47%-per-quarter benchmark rate. These are neither confidence intervals nor probabilities. The historical rate is not guaranteed to persist. Scenario table is retained beside this note.
- Dataland / Smithsonian: public projects are verified as references (S10–S11). Dataland is in Los Angeles and powered by Google Cloud, not a museum hosted at Google DeepMind. These sources do not establish Halbach involvement or identify a downloadable demo. Do not conflate a prerecorded demonstration with a live installation. Final selected media remains needed.
- Caching title: keep one currency: Same context: $0.30 vs $3 per million input tokens for the verified Sonnet4.6 example. Do not mix dollar cents with euros. It is cache-read input pricing, not all tokens or the whole request.
Data and graph method
- Historical token-price CSV retains only two source-supported endpoints. Draw endpoint comparison or clearly label a connecting guide; do not invent intermediate observations or extrapolate the old series as measured2026 data.
- Scenario index formula: 100 × annual_factor^(year−2026). Annual factors 1,0.5 and0.53^4. The last applies the reported quarterly historical rate mechanically, not a source forecast.
- Small-model evidence belongs in notes/backup: the Stanford example concerns a specific MMLU threshold, not equivalence across every GPT task.
- Hardware efficiency improvements do not establish lower hardware sticker prices, cheaper rental in every region, or two-year payback. Avoid conflating these curves.
Implementation decisions
Keep measured financial outcomes distinct from productivity and experimental quality. Preserve v13 exports. Apply American “optimization” spelling to current authored prose; leave original source titles, historical filenames and URLs intact. Remove the demo recap, move Klarna to cost-per-task, replace normalized CareAccess charts with explicit86%cost and66%time reductions, and make scale/revise outcomes visible. Add a meaningful schematic improvement loop, not an unsupported performance curve.
Sources
S01 — FinOps for AI
No universal department or workflow cost percentages; define workload boundary and allocation.
Limit: No evidence found for the proposed 50/30/20 split. Quality labor overlaps people unless explicitly separated.
S02 — AWS inference stack
Managed APIs, managed hosting and self-managed rented infrastructure are distinct options.
Limit: Hardware ownership is not required for control over model serving.
S03 — EDPB international transfers guide
EDPB international transfers guide
GDPR permits transfers subject to applicable conditions and other processing requirements.
Limit: Does not imply an architecture is compliant. GDPR does not universally require on-premises AI.
S04 — Amazon 2025 audited property/equipment note
Amazon 2025 audited property/equipment note
Server/networking accounting useful lives are five to six years in this filing.
Limit: Not a GPU lifetime or payback benchmark. Two-year payback must be calculated for the actual project.
S05 — Generative AI and Firm Productivity: Field Experiments in Online Retail
Generative AI and Firm Productivity: Field Experiments in Online Retail
Pre-sale chatbot increased sales 16.3% against no pre-sale service in randomized 2023–24 experiments.
Limit: October2025 working paper v1. Not an improvement over human support or a net-ROI estimate. Table4/AppendixC1.
S06 — AI business advice experiment
640 Kenyan entrepreneurs: no statistically significant average revenue/profit gain; outcomes heterogeneous.
Limit: 2023 experiment, 2026 Management Science publication. Null average is not proof of zero effect in every firm.
S07 — Navigating the Jagged Technological Frontier
Navigating the Jagged Technological Frontier
758 consultants: on a selected outside-capability task, AI reduced correct recommendations by 19 percentage points.
Limit: June2023 experiment, March2026 publication. Real professionals, experimental task, not realized corporate ROI.
S08 — Stanford AI Index 2025: ten charts
Stanford AI Index 2025: ten charts
GPT3.5-level MMLU token price $20/M in Nov2022 to $0.07/M in Oct2024. A specific MMLU capability threshold was reached by PaLM540B then Phi3mini3.8B.
Limit: Price is blended input/output (3:1), not pure input. Two endpoints, not a sampled time series. Benchmark equivalence is not general model equivalence.
S09 — The plunging price of thought
September2026 analysis: average fixed-score inference cost decline about47%/quarter across five primary benchmarks since2023.
Limit: Historical estimate, not prediction or totalworkflowcost. Benchmark selection and frontier estimation have important limitations.
S10 — Google Dataland announcement
Dataland is a Los Angeles AI arts museum powered by Google Cloud.
Limit: Do not call this a museum at DeepMind or imply Halbach participation without separate verification.
S11 — Smithsonian Refik Anadol event
Smithsonian Refik Anadol event
Smithsonian Dreams is a public Refik Anadol Studio installation involving Smithsonian collection data.
Limit: A public reference, not proof of Halbach involvement or a live interactive web demo.
S12 — Claude Message Batches
Asynchronous batch API offers 50% token-price discount; prompt caching can also be used.
Limit: Batch processing trades immediate response for scheduling flexibility; cache-hit rates remain workload-dependent. Use matching currency in examples.