Case study: NVIDIA, the hypergrowth problem

Every era gets one company that breaks everybody's valuation models, and studying it teaches more than a decade of normal markets. For this era it is NVIDIA: a graphics-chip designer that became, for a stretch of the 2020s, the most valuable company on earth. This case study is not a recommendation in either direction; it is a walkthrough of how professionals actually analyzed a business whose revenue could double in a year, and what the toolkit's each piece did and failed to do.

The short version

Hypergrowth breaks trailing multiples (the P/E looks absurd precisely when the stock is cheapest on what is coming), makes the DCF entirely a terminal-value bet, and forces the real work into three questions: how big is the end market actually, what share and margin survive competition, and what does the current price already assume (the reverse DCF)? NVIDIA rewarded analysts who tracked the DEMAND LEDGER (hyperscaler capex guidance) and supply constraints instead of the P/E, and punished both the valuation-only bears (who were right about the multiple and wrong about the earnings) and the extrapolators who forgot that 2018 and 2022 style air pockets are part of the deal.

Why this case

NVIDIA is the cleanest modern specimen of the hypergrowth problem because the growth was real, gigantic, and repeatedly discontinuous. This was not a story stock without earnings: data center revenue went from roughly $3B in fiscal 2020 to a run-rate measured in the tens of billions per QUARTER by 2025, with gross margins in the 70s. Every classical tool gave a confident answer and most of those answers were wrong in both directions at different times, which is exactly what makes it worth studying.

The timeline that mattered

Period	Event	What an analyst had to decide
2015-2018	Gaming GPUs boom; CUDA quietly becomes the ML standard in research labs	Is this a cyclical component maker or a platform being born?
2018	Crypto-mining demand collapses; stock roughly halves	Was the growth real demand or a one-off distortion? (Both.)
2020-2021	Pandemic compute + gaming surge; data center passes gaming	Same question again, at higher stakes
Late 2022	Stock down ~60% amid the rate shock; ChatGPT launches in November	The single most expensive moment to be anchored on the past
May 2023	The guidance heard around the world: a ~50% above-consensus data center forecast	When a company beats by that much, the model was measuring the wrong thing
2023-2025	Hyperscaler capex race; supply-constrained quarters; export controls on China; Blackwell ramp	Demand ledger vs supply ceiling; concentration risk; geopolitics as a line item
2025-2026	Multi-trillion valuation; custom-silicon competition matures; the 'AI capex sustainability' debate becomes THE market debate	Is the terminal state a toll road or a commodity cycle at scale?

The hypergrowth valuation problem, precisely stated

Trailing multiples invert their meaning. In late 2022, NVIDIA traded near 60x trailing earnings, screaming "expensive," months before earnings tripled: on the earnings that actually arrived, the buyer at that moment was paying a teens multiple. In hypergrowth, the trailing P/E is highest exactly when the future earnings are about to make it cheap: the multiple measures the past, and the past is the one thing that does not matter.
The DCF becomes all terminal value. With near-term flows doubling, 80%+ of any honest DCF sits in the terminal assumptions: what is the STEADY state of AI compute demand, margin, and share? The model does not answer that question; it just prices whatever you assume about it, which is why serious analysts ran it in reverse instead.
Comps circularity. Against which peers? Semis at 20x? Platforms at 30x? Every "comp" was itself being repriced by the same AI question. Comparing AI beneficiaries to each other in 2023-24 measured relative enthusiasm, not value.

How professionals actually analyzed it

The reverse DCF, quarterly. Start from the market cap and extract the implication. At ~$1T (mid-2023): roughly $60-80B of steady-state data center revenue at sustained margins, plausible against announced capex. At ~$3T+ (2024-25): several hundred billion of annual AI-infrastructure revenue with durable 70%+ gross margins, meaning the price now asserted BOTH that AI capex keeps compounding AND that competition never compresses the take. The reverse DCF did not say sell; it said exactly which two claims a holder was underwriting, which is its entire job.
The demand ledger. NVIDIA's revenue is a handful of customers' capex, so the professionals' primary dataset became Microsoft, Alphabet, Amazon and Meta's capex guidance, disclosed quarterly, plus sovereign and startup GPU commitments. When your company's revenue is four companies' budget line, you model THEIR budgets, not your company's history. (The Power & Hyperscalers industry guide continues this thread.)
Supply as the near-term truth. For long stretches the binding constraint was advanced packaging (CoWoS) and HBM memory capacity, making TSMC's and SK Hynix's expansion schedules better near-term revenue predictors than any demand survey. Hypergrowth analysis often lives in the supplier chain, not the income statement.
The moat interrogation. The margin question was never "are the chips good" but "what defends 75% gross margin?" The working answer: CUDA's software lock-in, systems-level integration (networking, NVLink), and a yearly cadence competitors had to match. The bear version: hyperscalers' custom silicon (TPUs and successors) exists precisely to attack that margin, and its progress is measurable in their disclosures. Both sides of that ledger are checkable quarter by quarter, which is what made this a research problem rather than a faith problem.
Concentration and geopolitics as line items. Double-digit revenue percentages from single customers, and a China business repeatedly re-shaped by US export controls: professionals carried explicit haircut scenarios for both rather than a footnote.

The numbers that mattered (and the ones that did not)

The shape of the thing (approximate, fiscal years)

data center revenue:   FY2020 ~$3B -> FY2023 ~$15B -> FY2024 ~$47B
                       -> FY2025 ~$115B     (yes, that is the slope)
gross margin:          ~62% pre-boom  ->  mid-70s at peak pricing power
operating margin:      ~30%           ->  60%+ in the strongest quarters

what did NOT matter: the trailing P/E at every single decision point.
what DID: hyperscaler capex guidance, supply capacity, CUDA's hold,
custom-silicon progress, export-control headlines, and the reverse
DCF's two underwritten claims.

The 2018 and 2022 rehearsals

Twice before the AI era, NVIDIA halved: the 2018 crypto hangover and the 2022 rate shock plus inventory correction. Both taught the same pair of lessons. Hypergrowth companies carry cyclical skeletons (demand can be borrowed from the future, and inventory corrections are violent), AND a drawdown in a structural grower is not, by itself, evidence the structure broke. Distinguishing "the cycle turned" from "the thesis broke" is the entire job in a 50% drawdown, and the analysts who had written down their falsifiers in advance were the ones who could do it.

The bear cases, and what became of them

"The multiple is absurd" (2016, 2020, 2023, ...). Valuation-only bears were right about the multiple and wrong about the earnings, the signature failure mode of shorting hypergrowth: overvaluation is real but has no timing content whatsoever.
"Demand is a bubble" (crypto edition, 2018). CORRECT, that time: a chunk of demand was mining, it vanished, the stock halved. The bear case that identifies WHICH demand is fragile beats the one that just says "bubble."
"Custom silicon eats the margin." The serious long-run bear case, still live through 2025-26: hyperscalers deploying in-house accelerators for internal workloads compresses the pricing umbrella even while units grow. Trackable, falsifiable, and the reason the terminal-margin assumption is the entire debate.
"AI capex is unsustainable." The macro bear case: the capex boom must eventually justify itself with end-customer revenue, or it corrects the way telecom equipment did in 2001. Whether AI monetization fills the gap fast enough is THE empirical question the market argues about; an honest analyst holds it as a scenario with a probability, not a slogan in either direction.

The transferable lessons for any hypergrowth company

Run the machine backwards. The reverse DCF converts a scary price into a specific, checkable set of claims. Underwrite the claims or do not own it; "it feels expensive" and "it keeps going up" are both mood, not analysis.
Model the customers' budgets, not the company's history. Extrapolating a hypergrower's own trend is astrology; its customers' disclosed spending plans are data.
Find the binding constraint. When demand exceeds supply, the supply chain IS the forecast.
Interrogate the margin, not the growth. Growth gets priced quickly; the durability of extraordinary MARGINS is where hypergrowth theses actually win or die, and it hinges on the moat holding.
Expect the air pockets. 40-60% drawdowns occurred inside every great hypergrowth compounding, this one included. Size so you can hold through one, and write the falsifier down BEFORE it happens, because inside the drawdown, cycle-turn and thesis-break look identical.
Concentration cuts both ways. A handful of customers made the growth legible AND made the downside scenario sharp. Hypergrowth with concentration is a different risk object than hypergrowth without it.

Glossary for this guide

Reverse DCF: Running the machine backwards: instead of estimating value from assumptions, solve for the growth and margins the current price already implies, then judge whether those are beatable. Often more honest than the forward version, because it removes your favorite input.
Terminal value: The value of all cash flows beyond the explicit forecast, collapsed into a single number at the forecast's end. It routinely carries more than half of a DCF's total value, which is why its assumptions deserve the most scrutiny.
Trading multiple: Price standardized by a unit of performance so unlike-sized companies compare: EV/EBITDA, P/E, EV/revenue. Ten times EBITDA means the market pays ten dollars per dollar of annual EBITDA.
Forward multiple: A multiple computed on the NEXT twelve months' consensus estimates rather than the last twelve reported. The street's default, because prices look ahead, at the cost of inheriting the estimates' errors.
Moat: A structural barrier that stops a company's high returns from being competed away: network effects, switching costs, brands and patents, cost advantage, or efficient scale. The durable question is never whether a business is profitable but what protects the profits.
Switching costs: The money, time and risk a customer would incur to leave. High switching costs show up in the numbers as retention; the claim without the retention is just hope.
Gross margin: Gross profit divided by revenue: the share of each sales dollar left after the direct cost of what was sold. The first test of whether the product itself, before any overhead, makes money.
Operating leverage: How much profits amplify a change in revenue because costs are fixed. High operating leverage makes good years great and bad years terrible; it is a magnitude, not a virtue.
Total addressable market: The revenue available if a company served every possible customer. Useful as a ceiling and abused as a pitch; the discipline is asking what fraction is realistically serviceable and at what cost.
Falsifier: The observable fact, chosen in advance, that would prove the thesis wrong and trigger exit. Deciding it while calm is cheap; deciding it mid-drawdown is expensive, which is why professionals write it down first.
Catalyst: The event expected to force the market to reprice toward your view: earnings, a spin-off, a regulatory decision, a contract. Cheap without a catalyst can stay cheap for years; the catalyst is the thesis's clock.
Crowding: Too much capital in one trade. Crowded strategies unwind together, briefly correlating at the worst moment; measuring how much of a story is already positioned is half of risk management.