A Gated, Red-Teamed Valuation of the AI Compute-to-Power Capital Cycle
Abhinav Koppulu · Capstone project · August 2026
This capstone asks whether the AI infrastructure capital cycle can be underwritten as a financial problem rather than a technology narrative — and what breaks when you try. It connects semiconductor cohorts, workload economics, rack and facility power, regional time-to-power, monthly FP&A, three integrated statements, 13-week liquidity, capital allocation, and public-equity valuation into one system, built across 15 isolated workstreams and 6 sequential gates, against 65 registered point-in-time sources frozen at 2026-08-12.
Four results drive the conclusion. Time-to-power is a first-order value driver, not an operating detail: a 100-MW deployment carries $2.26bn (ERCOT) to $3.71bn (PJM) of modeled economic delay cost, a spread larger than most siting decisions contemplate. Accounting life is not economic life: the strongest hardware case returns +$127mm NPV in the base case and −$439mm probability-weighted, so the answer inverts with the scenario weights. Liquidity is the binding constraint: the constrained 13-week forecast draws $2.29bn and reaches the $4.0bn policy cash floor with zero headroom. The market is not obviously wrong: CoreWeave's probability-weighted DCF of $107.54 sits within $0.20 of the $107.73 event-date close, with terminal value at 104% of enterprise value — so the valuation is a duration bet, and the honest call is neutral/watch rather than a manufactured edge.
The result the project is actually built around is a disagreement it declined to resolve. Two independently constructed lease engines returned +$92mm and −$387mm on the same asset. Averaging them would have produced a clean, publishable, meaningless number. Instead the divergence was escalated to a binding release restriction: no lease authorization until one contract-level term sheet is run through both engines. An independent red team logged 12 issues, of which 2 remain critical and open under control. The contribution is a decision system that preserves material disagreement instead of laundering it into false consensus.
Keywords: capital allocation · DCF and scenario weighting · time-to-power · FP&A integration · 13-week liquidity · red-team review · point-in-time sourcing
Not investment advice. This is an analytical capstone and career portfolio. All prices and filings are frozen at a point in time and are stale by construction. Nothing here is a personalized recommendation, and no capital was allocated on any conclusion in this repository.
Should you invest in the giant build-out of AI data centers? This project treats that question the way a bank would — not as hype, but as a model that has to survive its own audit before anyone acts on it.
Three findings in plain terms:
- What decides whether a data center makes money isn't the price of electricity — it's how long you wait to switch the power on. That waiting can quietly cost billions ($2.3B–$3.7B on a single 100-megawatt site, depending on region), because you're paying for the building and the financing while it earns nothing.
- CoreWeave — the stock everyone points to — is priced just about right. Our from-scratch valuation lands within 20 cents of where the market actually traded it. No obvious bargain, so the honest call is "watch," not "buy."
- The plan runs out of cash before it runs out of good ideas. The short-term forecast hits the minimum-cash line with nothing to spare, so funding — not ambition — is the real constraint.
The most honest moment in the whole project: two of our own models disagreed about whether a lease made or lost money — by hundreds of millions of dollars. Instead of splitting the difference to get a tidy number, we made it a hard rule: nobody signs that lease until the two models are reconciled. Preserving the disagreement, rather than hiding it, is the point.
The cost of simply waiting for the power to be switched on, for one identical 100-megawatt data center, in three parts of the country. This is money lost to delay alone — before a single calculation about electricity prices. Every number here is frozen at August 12, 2026 and is deliberately out of date now: it's a snapshot of a method, not live advice.
The question is narrow and underwriting-shaped:
Under what conditions should capital be committed to AI compute infrastructure — and what evidence would have to be true first?
Most treatments of this subject argue a direction. This one builds the apparatus that would let a committee say no. Five design commitments distinguish it:
| Commitment | What it prevents |
|---|---|
| Workstream isolation — 15 workstreams built independently before reconciliation | Analysts converging on each other's assumptions before the numbers are in |
| Gates before modeling — QA designed at Gate 1, not retrofitted | Checks that are shaped to pass the model that already exists |
| Point-in-time freeze — one operating and one market timestamp | Silent restatement as data refreshes underneath a conclusion |
| Independent re-derivation — critical engines built twice | A single implementation error becoming the house view |
| Preserved disagreement — divergence escalated, not averaged | False precision, which is the failure mode this asset class rewards |
Figure 1. Gate progression. Each gate carries its own QA check count and had to clear before the next stage opened. Gate 2 passed with controlled exceptions rather than a clean pass, and that exception was carried forward rather than closed by assertion.
Six gates, each with a defined scope, artifact, and independent QA count:
| Gate | Scope | Status | Sheets | Formulas | QA |
|---|---|---|---|---|---|
| 1 | Charter and control design | PASS | 9 | 12 | 12 |
| 2 | Point-in-time data foundation | PASS with controlled exceptions | 21 | 303 | 20 |
| 3 | Technical engines and capital case | PASS | 24 | 19,619 | 37 |
| 4 | FP&A, statements, and liquidity | PASS | 30 | 4,890 | 20 |
| 5 | Public equity and investment committee | PASS | 20 | 431 | 15 |
| 6 | Final synthesis and red-team release | PASS | 15 | — | 25 |
Stages 2 through 5 carry 25,243 formulas and 4,071 comments; the comments are not decoration but the inline provenance trail tying cells back to registered sources.
Sixty-five sources are registered with organization, document, period, publication date, eligibility at freeze, URL, primary use, and a confidence grade:
| Source type | High | Medium-High | Medium |
|---|---|---|---|
| Technical primary source | 18 | — | — |
| Government / grid authority | 15 | — | — |
| Regulatory filing | 12 | — | — |
| Company primary source | 9 | 1 | 3 |
| Market / secondary source | 6 | — | 1 |
Every source carries an eligibility flag against the freeze timestamp, so a conclusion cannot quietly depend on information that postdates the decision it informs.
Fifteen tie-outs bind the stages to each other — Stage 4's FY2027 revenue must equal the figure entering Stage 5's DCF, Stage 3's lease NPVs must survive into Stage 4 unchanged, the balance sheet must close. Thirteen pass exactly; two are marked CONTROLLED rather than PASS, because they are diagnostics of a real tension rather than arithmetic that could be made to agree:
| Tie-out | Expected | Actual | Status |
|---|---|---|---|
| Stage 4 → Stage 5 FY2027 revenue | $26.26bn | $26.26bn | PASS |
| Balance-sheet check | 0 | 2.9 × 10⁻¹¹ | PASS |
| Market price vs weighted DCF spread | ≤ $1.00 | $0.19 | CONTROLLED |
| Terminal-value concentration | ≤ 1.0 | 1.043 | CONTROLLED |
Figure 2. Modeled economic cost of base energization delay for an identical 100-MW deployment across three regional archetypes. The $1.45bn spread between ERCOT and PJM reflects delayed operating contribution, financing carry, and interconnection capital — not a difference in power price.
Treating power as a $/MWh input understates it by an order of magnitude. The dominant term is when the facility energizes, because delay compounds against financing carry while contributing nothing. This reframes siting from a cost-minimization problem into a schedule-risk problem.
Figure 3. The capital case inverts under scenario weighting. The strongest base-case strategy — extend the H200 cohort, then refresh — returns +$127mm in the base case and −$439mm probability-weighted.
The decision rule that survives is workload-specific: retain and selectively refresh a hardware cohort while contribution margin at the workload level remains positive, and separate physical life, accounting life, frontier-eligibility life, secondary-market life, and economic life rather than collapsing them into a single depreciation schedule. A strategy that is positive only in its base case is not a strategy.
Figure 4. The constrained 13-week forecast. Cumulative draws of $2.29bn bring modeled cash to the $4.0bn policy floor with zero headroom — the constraint that gates every discretionary tranche in the final recommendation.
Three point-in-time forecast cycles were preserved and back-tested rather than overwritten. Revenue forecasting was conservative by $209mm; capex was understated by $3.6bn. The asymmetry is the finding: revenue error was noise, capital error was structural, and it traced to vendor delivery, construction-in-progress, powered-acceptance, and payment-timing controls that the original forecast did not carry.
Figure 5. CoreWeave at the frozen market timestamp. The probability-weighted DCF lands $0.19 below the event-date close — inside any honest error bar on a model whose terminal value is 104% of enterprise value.
| Measure | Value |
|---|---|
| Event-date close | $107.73 |
| Base DCF | $108.70 |
| Probability-weighted DCF | $107.54 |
| Terminal value / enterprise value | 1.043 |
Conclusion: neutral / watch. When PV of terminal value exceeds enterprise value, the valuation is a statement about duration and discount rate, not about next year's operations. Reporting a target price from that model would imply a precision it does not have.
The public-equity conclusion is a diversified bottleneck barbell across silicon, networking, electrical infrastructure, grid equipment, and generation, rather than a concentrated pure-compute bet. The top risk-adjusted screen result across the 24-company universe is AVGO, ETN, ASML, TSM, GOOGL — a screen output under stated constraints, not a recommendation.
Figure 6. Opportunity against crowding. High-quality AI beneficiaries can still deliver poor returns when entered at crowded valuations, which is the argument for breadth across the bottleneck rather than concentration at its most obvious point.
An independent red team reviewed the integrated model for model risk, internal consistency, and unresolved issues, logging 12 findings scored on impact × likelihood.
Figure 7. The 12 registered red-team issues by impact and likelihood. Two sit in the critical band and remain open under binding control.
| ID | Issue | Severity | Status |
|---|---|---|---|
| R01 | Lease-economics divergence — +$92mm vs −$387mm on the same asset | Critical | Open / controlled |
| R02 | Zero liquidity headroom — 13-week model reaches the cash floor | Critical | Open / controlled |
| R03 | Terminal value is 104% of enterprise value | High | Open / controlled |
| R05 | Capex understated $3.6bn across three backtests | High | Open / controlled |
| R06 | Hardware economic life not established by accounting life | High | Ongoing |
| R07 | Regional power cases are archetypes, not executed tariffs | High | Ongoing |
| R08 | Backlog is not guaranteed revenue; customers are concentrated | High | Ongoing |
| R04 | Peer prices observed at different July 2026 timestamps | High | Refresh required |
| R09 | Interest-rate and refinancing exposure | High | Ongoing |
| R12 | Crowding and factor exposure | Medium | Ongoing |
| R10 | Efficiency vs demand elasticity is genuinely two-sided | Medium | Scenario only |
| R11 | Some technical and regional inputs are time-sensitive estimates | Medium | Ongoing |
R01 is the project's defining result. Two independently constructed engines disagreed on the sign of the same lease. The reconciliation traced the gap to horizon, lease expense treatment, escalators, refresh capex, deposits, financing, residuals, and terminal treatment — and still could not close it without a contract-level term sheet that does not exist. The disagreement was therefore converted into a control, not a number.
Gate 6 passed with controlled open risks. Two restrictions bind any live use:
- Lease economics (R01). No lease authorization until one contract-level term sheet is run through both engines and the residual difference is under $50mm or fully explained by documented economic definitions.
- Liquidity (R02). No discretionary capital tranche without committed availability, customer funding or equivalent protection, staged vendor payments, and at least $500mm of modeled treasury headroom above the minimum cash requirement.
- Everything is frozen and therefore stale. Prices, filings, debt terms, and tariffs are pinned to 2026-08-12. The refresh runbook in
data/refresh_runbook.csvexists because the conclusions expire. - Peer prices are not synchronized (R04). Most comparable-company observations are row-specific July 2026 timestamps, not one market instant. This is disclosed rather than smoothed.
- Regional power cases are archetypes (R07). ERCOT, Georgia, and PJM cases are modeled representatives, not binding site-level utility commitments with executed tariffs and interconnection milestones.
- Terminal value dominates the DCF (R03). At 104% of enterprise value, the valuation inherits the full sensitivity of its discount rate and terminal growth assumption.
- Backlog is not revenue (R08). The $104bn figure is a demand signal subject to conversion, timing, credit, and termination rights.
- Single-company depth. The valuation work centers on CoreWeave as the pure-play expression; the 24-company universe is screened, not modeled to the same depth.
- No capital was deployed. Nothing here was tested against realized outcomes.
These deliverables are documents, so the verifiable claim is not "the tests pass" but "nothing has moved since the release was signed." Two independent controls run in CI on every push:
python scripts/verify_release_integrity.py # digests, sizes, and release membership
sha256sum -c SHA256SUMS.txt # independent coreutils cross-checkSHA256SUMS.txt pins a digest for each of the 162 released files. FILE_MANIFEST.json additionally records byte sizes and an authoritative file count, which catches a file being added — something a digest list alone cannot detect. CI runs both, plus a sha256sum cross-check so the result does not depend on the verifier script being correct.
The four primary deliverables carry these digests, unchanged since release:
| Deliverable | SHA-256 (first 16) |
|---|---|
| Control workbook | 2ac9a58f752f6102 |
| Handbook | 786265910b67b245 |
| Investment-committee deck | dc4e197eead29bb7 |
| Red-team report | 4c97dd0f42f8941f |
README.md is excluded from verification and reported on every run: it was rewritten after the release was signed, to present the project as a paper rather than a release note. Every deliverable, chart, register, QA record, and stage archive is verified.
What is not reproducible. The workbooks are Excel models, not a script that regenerates them from raw inputs. scripts/ contains the build and render tooling used to produce and QA the final artifacts, not an end-to-end pipeline. Re-running the analysis means following data/refresh_runbook.csv against fresh sources — 15 steps with defined owners, cadences, and acceptance criteria.
Primary deliverables:
| Artifact | What it is |
|---|---|
| Capstone handbook (DOCX) | 16-page methodology, results, and end-to-end walkthrough |
| Red-team report (DOCX) | 8-page independent model-risk and unresolved-issue review |
| Investment-committee deck (PPTX) | 18 slides, every one carrying a [Sources] block |
| Control workbook (XLSX) | 15-sheet cross-stage control, evidence, monitoring, and refresh workbook |
Evidence and registers:
| Path | Role |
|---|---|
data/source_coverage.csv |
All 65 sources with URL, period, eligibility, and confidence |
data/red_team_issues.csv |
The 12 findings with evidence, required control, and owner |
data/cross_stage_tieouts.csv |
15 reconciliations binding the stages together |
data/key_outputs.csv |
Headline metrics with units and interpretation caveats |
data/monitoring_triggers_final.csv |
Live monitors with trigger thresholds and actions |
data/refresh_runbook.csv |
15-step refresh with owners and acceptance criteria |
data/gate_summary.csv |
Gate-by-gate scope, status, and QA counts |
stage_packages/ |
Complete Stage 1–5 release archives |
qa/ |
Formula, render, accessibility, presentation, and archive checks |
charts/ |
All 19 figures |
Accessibility was treated as a release gate, not a courtesy: the handbook and red-team report each cleared with zero high-, medium-, and low-severity findings, and the deck passed overflow testing on all 18 slides.
Capstone project by Abhinav Koppulu (Finance and Supply Chain & Operations Management, Carlson School of Management, University of Minnesota). The aim was to find out what an AI-infrastructure investment case looks like when it is required to survive its own quality controls — and to publish the parts that did not resolve alongside the parts that did.
MIT License. Copyright © 2026 Abhinav Koppulu.
Third-party materials referenced in data/source_coverage.csv — SEC filings, grid-authority publications, and company disclosures — remain the property of their respective publishers and are cited, not redistributed.