Phase Space AI

Valuation

NVIDIA [NVDA]

NVIDIA [NVDA] — Valuation

Spot $190.01 (2026-07-29) · shares 24,221m · market cap $4,602bn · EV $4,560bn (lease-exclusive) / $4,565bn (lease-inclusive)

---

1. COMPANY STATE — declared first, before any instrument

STATE B — cyclical / capex-driven.

The A-versus-B choice is genuinely contestable and I am taking B. The consequence is not cosmetic: State A would let me build the terminal margin from trailing normalised economics with industry data as a sanity band only; State B requires the terminal margin to be mid-cycle-normalised on both revenue and margin, and requires me to DATE the cycle position. A also admits a wider vehicle set; B pushes toward equity and away from anything whose expiry has to survive a cycle turn.

Evidence that put it in B:

1. Operating-margin variance fails State A's test outright. GAAP operating margin by fiscal year: 27.18% (FY21), 37.31% (FY22), 15.66% (FY23), 54.12% (FY24), 62.42% (FY25), 60.38% (FY26). That is a 46.76pp range across six years. State A requires "operating margin low-variance across ≥5 years."
2. The downdraft is recent and demonstrated, not hypothetical. FY2023 revenue was flat — $26,974m against $26,914m, +0.2% — while operating income fell 57.9% ($10,041m → $4,224m) and gross margin fell from 64.93% to 56.93%. Three years ago this business had a channel-inventory correction that took 21.6pp out of its operating margin in a single year without revenue declining at all. That is the definition of operating leverage running in reverse against an exogenous demand cycle.
3. The demand driver is an exogenous capital-spending cycle in three budgets. Three direct customers are 54% of Q1 FY2027 revenue. Those are capital-expenditure decisions taken annually by three boards, not a diversified recurring revenue base. NVIDIA has no RPO, no backlog and no take-or-pay contract to bridge a budget pause (see Research §5).
4. The principal cost input sits in a supplier-power cycle with a contractual floor. MU discloses 16 take-or-pay supply agreements with a floor price through CY2030. HBM is NVIDIA's largest bought-in cost. NVIDIA's gross margin is therefore exposed to a memory cycle it does not control and, on the floor-price structure, cannot benefit from on the way down until CY2030.
5. A structural-basis break that State A explicitly excludes. State A requires "no business-model transition." NVIDIA's own 10-K attributes the FY2025→FY2026 gross-margin decline from 75.0% to 71.1% to "our business model transitioned from offering Hopper HGX systems to Blackwell full-scale datacenter solutions" — a change in what is being sold, and in how much third-party BOM passes through revenue. The reporting basis broke in the same window: revenue-by-market-platform was redefined in Q1 FY2027 and the Compute/Networking split retired.

The honest case for State A, which I am rejecting: NVIDIA has now held a 70–75% gross margin and a 54–65% operating margin for eleven consecutive quarters; the FY2023 downdraft was a gaming/crypto channel event in a business that is now 89.7% data centre and therefore arguably a different company; ROIC is ~171%; and it holds $41.9bn of net cash with a 47% FCF margin. A reader who weights "eleven quarters of stability in the current mix" above "46.8pp of six-year variance and a 54%-of-revenue capex dependency" lands on A. Eleven quarters is not five years, and the eleven quarters coincide exactly with the steepest capex expansion in the history of the industry — which is why I weight it the other way.

Falsifier for State B — what would move NVIDIA to State A

Both of the following, together:

1. Four consecutive quarters with GAAP operating margin inside a 62–68% band while the aggregate capital expenditure growth of the top three customers decelerates below 10% year-on-year. That is the specific observation that would show margin has decoupled from the customer capex cycle. Margin stability during capex acceleration is not evidence of it — it is what State B predicts.
2. Top-three direct-customer revenue concentration falling below 35% of total revenue (from 54%), i.e. the ACIE line genuinely broadening the demand base rather than re-labelling it.

If both occur, the terminal margin here should be revised up toward the demonstrated trailing level and the mid-cycle haircut removed. If instead operating margin breaks below 55% in any quarter in which customer capex growth is still positive, State B is confirmed and the terminal margin below is too high.

Dating the cycle position (State B requirement)

• Margin: at or on the peak. Q1 FY2027 GAAP operating margin 65.60% is an all-time high. Q2 FY2027 guidance implies 65.6% ($91.0bn × 74.9% − $8.5bn opex) — flat, i.e. guided to the peak, not through it.
• Revenue: not at the peak. Supply commitments +25% q/q ($95.2bn → $119bn); raw-material inventory +74.6% q/q; guided revenue +11.5% q/q. Nothing has rolled.
• Therefore the forecast begins at a margin peak with the volume cycle still expanding. That is the single most important framing in this memo, and it is why the terminal margin sits below trailing.

---

2. Verified inputs — every screen input checked, none silently adopted

Net cash, rebuilt from the 10-Q rather than taken from any API (the ORCL $133.9bn precedent):

Cash and cash equivalents                        13,237
Marketable debt securities                       37,098
                                              ---------
Liquid                                           50,335
Short-term debt                                  (1,000)
Long-term debt                                   (7,470)
                                              ---------
NET CASH                                         41,865      (0.9% of market cap)

Explicitly excluded: marketable equity securities $30,237m and non-marketable securities $43,364m — $73,601m of strategic stakes in AI-ecosystem companies, many of them customers. They are assets, not cash, their marks produced $15,929m of one-quarter non-operating income, and treating them as cash would understate EV by 1.6%. Including them, net cash is $72,102m. The conservative figure is used.

AV's shortTermDebt was 1,466 against a filed 1,000 — a $466m overstatement, apparently the current operating-lease liability folded in. Immaterial here (0.01% of EV) but recorded.

EV is stated on both bases (the CRWV lease-exclusive precedent that invalidated a seven-name ladder):

• Lease-exclusive: $4,560,367m — the headline, used for all multiples below.
• Lease-inclusive: $4,564,711m — adds $3,878m long-term + $466m current operating-lease liabilities.
• Not captured on either basis: $22.7bn of leases NVIDIA discloses it "expects to commence" between FY2027 and FY2030. If those were capitalised today, EV would be ~0.5% higher.

TTM operating figures (four quarters ended 2026-04-26)

EBITDA is computed from operating income plus the absolute value of cash-flow-statement D&A. AV's ebitda and ebit fields, and income-statement depreciationAndAmortization, were not used at all (documented defects: AV ebit = pretax + interest expense, 3.11x wrong on NOW; AV ebitda wrong on 12 of 20 names).

Current multiples: EV/TTM sales 17.99x · EV/TTM EBIT 28.10x · EV/TTM EBITDA 27.55x · P/E 28.83x (the P/E is flattered by the investment marks; the EBIT multiple is the clean one).

---

3. TERMINAL MARGIN — 58.0%, derived from NVIDIA's own demonstrated operating margin

This is the parameter that has failed five times in this project. It is not taken from a peer median, an industry percentile, max(own, sector median), or the universe-wide 14.4% constant that every stored scan_v2/scan_v3 terminal margin actually is. It is built from NVIDIA's own filed lines.

The opex bridge, reconciled from 10-K lines

NVIDIA's consolidated income statement carries exactly two operating-expense lines, so the bridge closes to zero residual — which is unusual and worth stating, because it removes the "other" bucket in which the NOW/INTU sellingGeneralAndAdministrative defect hid ~30pp of margin.

FY2026 10-K, exact: 153,463 − 18,497 − 4,579 = 130,387 = filed operating income. ✔ TTM, exact: 74.15% − 8.22% − 1.91% = 64.02% = TTM operating margin. ✔

Terminal (FY2031) bridge:

m_gross,T          70.0%
  less R&D          9.5%
  less SG&A         2.5%
                 -------
m_EBIT,T           58.0%

Hard constraint m_EBIT,T ≤ m_gross,T: 58.0% ≤ 70.0% — SATISFIED, with 12.0pp of headroom, and the full expense bridge (not merely the ceiling) is what sets it. The FN precedent — a 14.4% terminal margin against an 11.96% gross margin — cannot arise here.

Each line justified

Terminal gross margin 70.0% (against 74.15% TTM, 74.9% Q2 guide, 71.07% FY2026, 56.93% FY2023 trough):

• NVIDIA's own MD&A explains the FY2025→FY2026 fall from 75.0% to 71.1% as "our business model transitioned from offering Hopper HGX systems to Blackwell full-scale datacenter solutions." Rack-scale selling pulls third-party BOM (memory, CPUs, cabinets, power) through NVIDIA's revenue line. That mechanically dilutes gross margin while raising gross profit, and Rubin extends the pattern.
• HBM cost is contractually floored high. MU has 16 take-or-pay agreements with a floor price through CY2030. NVIDIA's largest bought-in input therefore has a four-year contractual floor under its cost, set in the supplier's favour, in a period when the supplier just printed an 84.6% gross-margin quarter.
• Hyperscaler ASICs cap merchant pricing at the margin (Research §6): custom silicon does not need to win to set a ceiling on what NVIDIA can charge its three largest customers.
• Offsetting, and why 70.0% and not lower: networking grew 141.5% and is 16.2% of Data Center revenue and rising; software/services attach is margin-accretive; and NVIDIA has demonstrated it can pass memory cost through — it held 71–75% gross margin straight through the entire Blackwell transition and a $4.5bn write-off. 70.0% sits 1.1pp below FY2026's actual and 4.2pp below TTM.

Terminal R&D 9.5% (against 8.22% TTM, 8.57% FY2026, 27.2% FY2021): rising, not falling. Q1 FY2027 R&D grew 58% y/y against 85% revenue growth — so R&D intensity fell this quarter — but the composition is turning capital-intensive: "a 112% increase in compute and infrastructure, a 31% increase in compensation and benefits... and a 204% increase in engineering development materials for new product introductions." Internal compute is a recurring, escalating cost of a one-year architecture cadence, and NVIDIA has committed $30bn of multi-year cloud-service agreements "primarily used to support our research and development efforts." Amortising $30bn of committed cloud across five years is ~$6bn/yr of R&D-line cost that does not scale down with revenue. 9.5% is 1.3pp above trailing.

Terminal SG&A 2.5% (against 1.91% TTM, 2.12% FY2026): modest de-leverage. NVIDIA's SG&A is extraordinarily thin because it sells to a handful of counterparties. The ACIE diversification strategy — the one that has to work for the concentration risk to fall — requires selling to "diverse AI purpose-built data centers and AI factories across industries and countries," i.e. a real enterprise and sovereign salesforce. That is a higher-cost distribution model. 2.5% is 0.6pp above trailing and still less than a quarter of what most enterprise-hardware businesses carry.

The terminal margin is BELOW trailing. Here is why that is not the documented error.

MEMO_BRIEF: "If your terminal margin is below the company's own trailing actual, that is almost certainly an error — say why it is not." The 84-name audit found 21 names below trailing (XZO −26.9pp, AVGO −24.3pp, MSFT −23.7pp). Those were mechanical: an industry-p75 cap or a universe median silently replacing the company's own economics. This is different in kind:

• Magnitude: −6.02pp against TTM 64.02%; −2.38pp against FY2026's filed 60.38%; and +18.6pp above NVIDIA's FY2017–FY2024 mean of ~31.5%. It is not a haircut to a peer median — it is above every single year NVIDIA printed before FY2025.
• State B requires it. The forecast starts at an all-time-high margin, guided flat, with the volume cycle still expanding. Carrying a peak margin to the terminal year while also carrying multi-year revenue growth is the exact internal inconsistency valuation.md warns about under State B: "a trailing peak margin combined with a cycle-average growth forecast is internally inconsistent and will always flatter."
• The causal bridge is named, not asserted (valuation.md Rule 2: a below-current terminal margin "requires an explicit causal bridge, not prohibition"): rack-scale BOM pass-through, a contractual HBM cost floor through CY2030, an ASIC price ceiling on the merchant accelerator, escalating internal-compute R&D, and a higher-cost ACIE distribution model. Every one of those is a filed disclosure, not a view.
• It is asserted, not clamped. The raw value (64.02% TTM), the transformation (−6.02pp), the reason (above) and the source (10-K/10-Q lines) are all stored in NVDA_analysis.json.

Sensitivity is reported over the exit multiple, not over the margin — but for completeness: at the base exit multiple, the required terminal margin at a 25% revenue CAGR is 53.65%, at 30% is 42.59%, and at 35% is 33.88%. All three sit below the underwritten 58.0%, so the margin assumption is not load-bearing for the verdict.

---

4. THE 5-YEAR IMPLIED-PATH TEST — the Valuation Criteria

Terminal value is 81.1% of EV at the base case, well above the 60% threshold, so the reverse DCF is mandatory as the primary long-horizon instrument.

The exit multiple is derived, not assumed

EV_T / EBIT_T = (1−t)(1−g/ROIC) / (WACC−g)

• t = 17% is the midpoint of NVIDIA's own guidance: "For the full year fiscal 2027, we expect GAAP and non-GAAP tax rates to be between 16.0% and 18.0%."
• Terminal ROIC 50% against a current ROIC of 171.4% (NOPAT $137,131m at 17% tax over invested capital of $80,008m = shareholders' equity less net cash less the $73.6bn strategic-investment book). A deliberate, large haircut: 171% is not a sustainable terminal input, and a high ROIC combined with a high exit multiple would double-count the same quality (Rule 4).
• WACC 9.5% — trailing 252-day realised volatility is 36.2%, so this is not a low-beta asset despite the size.
• Base exit 13.88x implies compression from today's 28.10x EV/EBIT of −50.6%. Stated as a number, per the requirement. A 24% grower becoming a 4% terminal grower should compress by roughly half; the multiple is not a free parameter.

The solve

Parameters held fixed: terminal margin 58.0%; exit multiple 13.88x; WACC 9.5%; horizon 5 years; FCF margin 46.97% (NVIDIA's own demonstrated TTM figure); starting revenue $253,491m; net cash $41,865m; shares 24,221m. Parameter solved for: the revenue CAGR required to justify EV of $4,560,367m.

Required 5-year revenue CAGR: 23.32%

Demonstrated 5-year revenue CAGR: 66.90% (FY2021 $16,675m → FY2026 $215,938m)

MARGIN = demonstrated − required = +43.58pp

Other demonstrated windows, for context: 3-year CAGR +100.05% (FY2023→FY2026), 2-year +88.27%, TTM year-on-year +70.68%. The 5-year figure is the headline because it spans the FY2023 downdraft — which is the correct window for a State B name and the least flattering of the four.

At the required 23.32%, FY2031 revenue is $723,104m and the PV of terminal value is $3,697,838m = 81.1% of EV.

Sensitivity — over the exit multiple, never over scenario probabilities

The margin is positive across the entire admissible exit-multiple range, and the narrowest reading — +38.50pp at a 61.0% multiple compression — still passes by a wide margin. The NOW precedent (verdict unchanged over 16pp of terminal margin because the exit multiple was the unexamined parameter) is directly addressed: here the exit multiple is swept from 10.97x to 28.10x and the sign does not change.

Verdict on the Valuation Criteria: PASS, +43.58pp.

One tooling defect measured here, because it matters

reverse_dcf.py's CLI has no --fcf-margin flag, even though DATA_DEFECTS records the terminal-only bias as "Fixed; pass fcf_margin". The function signature accepts it; argparse does not expose it, so any caller using the documented command line silently gets the terminal-only answer. Measured on NVDA:

At NVIDIA's 46.97% FCF margin the bias reaches +6.40pp, above the +6.04pp previously measured at a 35% FCF margin — the largest instance of this defect recorded in the project. The numbers above were produced by importing the module and calling project_ev directly with fcf_margin=0.4697.

---

5. THE 12-MONTH TARGET — anchored on NVIDIA's own multiple history

Consensus is unavailable; the base is built from guidance and disclosed commitments, and labelled as such

AV EARNINGS_ESTIMATES returned {"estimates": []} at HTTP 200 — the documented defect, INDETERMINATE, not FAIL, and it blocks nothing. data/pit/consensus_pit.jsonl carries no NVDA record. So the near-term base is the company's own guidance extended with decaying sequential growth, not consensus, and not a long-run house forecast:

Cross-check: $95bn of supply commitments payable in the remaining nine months of FY2027 against ~$74bn of implied COGS for those quarters at the guided 74.9% gross margin — the commitment book covers the path with ~28% to spare. The path is consistent with what NVIDIA has already contracted to buy.

The multiple anchor, and the regime-change problem stated honestly

Daily market-cap/TTM-revenue series, 2016-08-01 → 2026-07-29 (2,512 observations), TTM revenue stepped in at each 10-K/10-Q filing date so the series is not forward-looking, share count held at the current verified 24,221m so the series measures multiple movement. Current: 18.16x.

NVIDIA at 18.16x sales sits at the 0.6th percentile of its own AI-regime trading history — 796 observations, median 28.73x, minimum 17.50x. It has been cheaper on sales, in this regime, on five days out of 796. This is the BSX pattern again (a PASS at the 0th percentile of its own 8-year EV/Sales history), and it is now the second instance in this corpus.

The history spans a genuine regime change (pre-AI median 14.20x versus AI-regime median 28.73x), which valuation.md says must be declared rather than papered over. It does not make the anchor UNIDENTIFIED, because both sub-regimes are NVIDIA's own history and the growth-matching rule tells me which one to use:

In twelve months NVIDIA's trailing revenue will be ~$402bn and its forward growth rate will be decelerating from 85% toward roughly 25–35%. The NVIDIA observations with comparable forward growth are the pre-2023 ones, which traded at a median of 14.20x sales and a 10-year p25 of 12.80x — not the AI-regime observations at 28.73x, which were priced off 200%+ growth.

So the base multiple is drawn from the 10-year distribution, deliberately, and it is below where the stock trades today.

Target

12-month target: $266 — 16.00x NTM sales, the 43.3rd percentile of NVIDIA's own 10-year multiple history and the 0.0th percentile of its AI-regime history. +39.8% to spot $190.01, above spot.

The base multiple is below the current 18.16x and below every observation in the AI regime — but it is inside the 10-year distribution, which is the growth-matched window, so it does not breach the rule that the base may not sit below every stated anchor. The bear anchor (12.80x, the 10-year p25) still produces +11.8%, which is the load-bearing point: at $402bn of NTM revenue, the multiple can fall to the 25th percentile of a decade of history and the stock still rises. Revenue growth, not multiple expansion, carries the target.

No external professional target is on file for NVDA in this project, so no sanity-band gap is reported. Per calibration item B16 (16 of 16 house targets below spot, a median 46.1% below Street), the fact that this target is above spot is noted deliberately rather than treated as an error.

Named product-cycle events inside the 12 months

Each also appears in the catalyst calendar. Every one is drawn from a filing; none is invented.

1. Q2 FY2027 print against the $91.0bn ±2% guide, with the guided 74.9% GAAP gross margin and 65.6% implied operating margin, and zero China Data Center compute revenue assumed — so a China licence is pure upside to the guide.
2. Rubin ramp — the 10-K commits to a one-year architecture cadence "including our Rubin platform," and prepared-remark mentions peaked at 9 in FQ2026 Q4. The volume inflection is the single largest revenue variable in the window.
3. The $119bn supply-commitment balance re-disclosed each quarter. This is the only quantified forward series NVIDIA publishes; the direction of the next two prints is the cleanest available read on whether the cycle has turned.
4. Customer-concentration disclosure each quarter — 54% of revenue and 64% of receivables from three direct customers. Either number moving materially is thesis-relevant in both directions.
5. H200 China licence conversion — a February 2026 licence exists, has produced no revenue, requires US inspection and carries a 25% import tariff. Unpriced optionality on a business that was $25bn in FY2025.

---

6. Downside case — with a named cause (MEASURED; logged and scored, does not reject the name)

Named cause: one of the three direct customers defers or re-phases its 2027 accelerator budget in favour of its own ASIC programme.

The mechanism is fully specified by NVIDIA's own disclosures. The three customers are 21%, 17% and 16% of revenue. NVIDIA holds no RPO, no backlog, no take-or-pay and $160m of customer advances, so there is no contractual instrument to slow the transmission — the loss lands in the guided quarter. Simultaneously NVIDIA remains obliged on $119bn of supply commitments, $95bn of it inside nine months, which is what converted a demand shock into a $4.5bn inventory-and-purchase-obligation charge in Q1 FY2026.

Quantified: loss of the 21% customer's incremental spend, with margin compressing to the FY2026 level as fixed rack-scale supply is absorbed into a smaller revenue base. Revenue $402bn → ~$320bn NTM; operating margin 65.6% → 55%; a fresh inventory/purchase-obligation charge of $5–8bn on the $119bn commitment book (scaled from the H20 precedent); and the multiple to the 10-year p10 of 9.16x on a decelerating tape → roughly $121, −36%. The investment book adds a second-order hit: $73.6bn of stakes in the same ecosystem, marked at levels that assume the capex cycle continues.

Note the asymmetry this produces: the bear case is a −36% move driven by a counterparty decision NVIDIA cannot see coming and has no contract against; the base case is +39.8% driven by revenue NVIDIA has already contracted to buy the inputs for. That asymmetry — not the level of the multiple — is the actual risk in this name.

---

7. Criteria summary

No position verdict is issued. The memo scores criteria; the book decides.