Business analytics case study
The strategic decisions, financial performance and industry shifts that turned a graphics-card manufacturer into the company the rest of the technology industry now builds on top of.
Behind: NVIDIA’s split-adjusted closing share price, monthly, January 2015 to June 2026, on a linear scale. Marked points are the first close above each trillion-dollar valuation. Source: Yahoo Finance; Reuters.
The argument
Analysis by Vedant Ambre
Between fiscal 2016 and fiscal 2026, NVIDIA’s annual revenue grew from $5.0B to $215.9B. The interesting question is not how fast that happened. It is why the company was in a position to capture it at all when it did.
The conventional account credits ChatGPT. It is the wrong causal story. By the time OpenAI released ChatGPT in November 2022, NVIDIA’s data-centre business had already overtaken gaming permanently, its accelerators were already the default hardware for training large models, and the software layer that made them difficult to replace was already sixteen years old.
What follows traces four decisions and the financial evidence for each. Three of them were made when the market they addressed did not yet exist, and they were expensive at the time. That is the pattern worth extracting — not the growth rate, which is not repeatable, but the sequence of choices that made the company the only available supplier at the moment demand arrived.
Revenue growth
Fiscal 2016 to fiscal 2026, from $5.0bn to $215.9bn.
From data centre
Share of revenue in the quarter ended April 2026. It was 6.8% in fiscal 2016.
Gross margin
Up from the mid-fifties for most of the previous decade.
Market value
Approximate, as of 24 July 2026. First company to pass $5tn, in October 2025.
Before the story
For most of the 2000s, the economics of computing were set by Intel. General-purpose processors improved predictably, and the sensible thing for any software problem was to wait for the next generation of CPUs rather than to rewrite the problem. Graphics processors were specialised parts for a specialised market, bought almost entirely by people who played games.
Two things changed that. The first was physical: single-thread processor performance stopped improving at its historical rate, so work that needed to go faster had to be made parallel. The second was that a particular class of problem — training neural networks — turned out to consist almost entirely of matrix multiplication, which is exactly what a GPU already did tens of thousands of times per frame.
The demonstration came in 2012, when AlexNet won the ImageNet competition using two consumer NVIDIA graphics cards. It was not a product NVIDIA had designed for that purpose. It was a product that happened to be programmable, because six years earlier the company had decided to make it so.
In 2006 NVIDIA began spending heavily to make its graphics chips programmable for work that had nothing to do with graphics. There was no market for it.
What happened
CUDA shipped in 2006 alongside the GeForce 8800. Technically it exposed the GPU’s processing cores to general-purpose C code, letting them communicate and share data instead of only running a fixed graphics pipeline. Commercially it was a bet that someone would want to do this, and a decision to carry the cost of the software layer until they did.
The cost was not trivial. Every subsequent GPU had to carry silicon and validation effort for compute features that gamers did not use, and NVIDIA had to fund compilers, libraries, documentation and developer support for an audience measured in thousands.
Exhibit 1
Annual research and development expense, US$ billions. NVIDIA fiscal years aligned to the calendar year they mostly cover.
ReadingIn 2015 Intel spent roughly nine times what NVIDIA did in absolute terms. NVIDIA was nonetheless committing over a quarter of its own revenue to R&D — $1.33bn against total revenue of $5.01bn. The absolute gap closed only in 2025, and as much because Intel cut its spending as because NVIDIA raised its own.
SourceSEC EDGAR XBRL, ResearchAndDevelopmentExpense (NVIDIA, AMD, Intel)
Why it mattered
The durable asset was never the chip. It was the accumulated body of code written against CUDA — libraries, frameworks, tuned kernels, and the training of the people who wrote them. A competitor can match a processor in a product cycle. Matching fifteen years of other people’s software is a different problem, and it is the reason AMD’s hardware has repeatedly been competitive on paper without displacing NVIDIA in practice.
Switching costs that accumulate in someone else’s codebase are the most durable kind, because your competitor cannot buy them back.
Exhibit 1b
NVIDIA's own disclosures. Four figures over six years, each measuring something different.
Developers from 200+ countries or regions who have used CUDA since 2006 (cumulative, not active)
SourceThe figures below are NOT a consistent metric. 'Registered developers in the NVIDIA Developer Program' (2020), 'CUDA developers' (2023, 2024) and 'developers who have used CUDA since 2006' (2025) are three different things. NVIDIA has never published a methodology. Treat the series as directional evidence of ecosystem scale, not as a precise time series.
ReadingIt is not possible to answer this precisely from public information, and that is worth stating rather than papering over. NVIDIA has never published a methodology, and the 2025 figure counts everyone who has ever used CUDA rather than current users.
SourceNVIDIA GTC and COMPUTEX keynotes; NVIDIA developer programme announcements
The lesson
A platform investment looks identical to waste right up until the market arrives. The distinguishing feature of a good one is not conviction about timing — NVIDIA did not predict the transformer architecture — but that the investment compounds in the hands of people outside the company. CUDA earned nothing for years. It was accruing switching costs the whole time.
NVIDIA became a data-centre company before anyone was watching — and six months before the event that is usually credited with causing it.
Exhibit 2
Quarterly revenue, US$ billions. 45 quarters, fiscal 2016 Q1 to fiscal 2027 Q1.
Chart showing fiscal 2016 to 2017. Gaming and the other non-data-centre markets are far larger than Data Center, which is close to the axis.
One continuous series across NVIDIA's fiscal 2027 reporting change. NVIDIA restated two quarters onto the new basis: Data Center was unchanged and Edge Computing equalled the exact sum of Gaming, Professional Visualization, Automotive and OEM & Other. 'Everything else' is that sum throughout.
SourcesNVIDIA CFO Commentary and quarterly results, fiscal 2016–2027; SEC EDGAR, NVIDIA Corporation filings
| Quarter | Data Center ($m) | Everything else ($m) | Data Center share |
|---|---|---|---|
| FY2016 Q1 | 88 | 1,063 | 7.6% |
| FY2021 Q2 | 1,752 | 2,114 | 45.3% |
| FY2023 Q1 | 3,750 | 4,538 | 45.2% |
| FY2024 Q2 | 10,323 | 3,184 | 76.4% |
| FY2027 Q1 | 75,246 | 6,369 | 92.2% |
Fiscal 2016
In the year to January 2016, NVIDIA’s data-centre business took $339M of $5.0B in revenue — under seven per cent. The company sold graphics cards to people who played games. Its data-centre line was small enough that a bad quarter in Gaming could erase it entirely.
What that figure conceals is the spending behind it. In the same quarter NVIDIA put $339M into research and development — 29.5% of revenue. Most of it went towards a market that did not yet exist.
2017 – 2019
Then cryptocurrency mining arrived and made the numbers unreadable. Gaming revenue climbed steeply through 2017 and 2018, and neither NVIDIA nor its investors could reliably separate demand from gamers, demand from miners, and demand from the early machine-learning researchers who had started buying the same silicon.
The answer arrived in the quarter ended 27 January 2019. Gaming revenue fell 46 per cent in three months, gross margin dropped to 54.7%, and the shares fell 18.8 per cent in a single session. NVIDIA later paid a $5.5m SEC penalty for inadequate disclosure of how much of its gaming growth had come from mining.
July 2020
In the quarter ended 26 July 2020, Data Center out-earned Gaming for the first time — $1.8B against $1.7B.
This is the moment usually cited as the turning point. It deserves an asterisk. NVIDIA completed its $7bn acquisition of Mellanox on 27 April 2020, three months into that quarter, and the networking revenue it brought is inside the figure. The first crossing was partly bought.
The reversal
It also did not hold. Gaming retook the lead the following quarter and kept it for six consecutive quarters, through the pandemic demand surge that pushed graphics cards into shortage.
Data Center led permanently only from the quarter ended 1 May 2022 — $3.8B against $3.6B. That date matters more than it looks.
The date that matters
NVIDIA became a data-centre company in May 2022. ChatGPT was released in November 2022.
The transformation was complete six months before the event that is usually credited with causing it. What ChatGPT changed was not NVIDIA’s direction but its scale — and the number of people who understood what the company had already become.
Fiscal 2027
Data Center is now 92.2% of revenue. In the first quarter of fiscal 2027, NVIDIA stopped reporting Gaming as a market at all, folding it into a new “Edge Computing” platform alongside robotics, automotive and workstations.
The company’s stated reason was that it was “transitioning to a new reporting framework that better reflects our current and future growth drivers.” Read plainly: the business that gave NVIDIA its name is no longer large enough to be one of its reportable markets.
ChatGPT did not change what NVIDIA was. It changed how many organisations concluded, simultaneously, that they needed to buy what NVIDIA already sold.
What happened
On 24 May 2023, NVIDIA guided to roughly $11bn of revenue for the following quarter. Analysts had expected around $7bn. The shares rose 24.4 per cent the next session. It remains the clearest example of a company telling the market that its own model of the business was wrong by half.
What followed was not a gradual acceleration. Data Center revenue went from $4.3B to $18.4B in four quarters. Crucially, gross margin rose at the same time — from 64.6% to 76.0%. Volume and price moved together, which happens only when the buyer has nowhere else to go.
One session
Share price move on 25 May 2023, after the guidance that repriced the AI trade.
Four quarters
Growth in quarterly Data Center revenue between April 2023 and January 2024.
Gross margin
Expansion over the same four quarters, from 64.6% to 76.0%.
Exhibit 3
GAAP gross and operating margin, quarterly, fiscal 2016 to fiscal 2027.
ReadingBoth, at once. Gross margin had sat in the mid-fifties to mid-sixties for seven years; it reached 78.4% by April 2024. Rising margin alongside rising volume is the financial signature of a supplier without a substitute — and the two collapses show what happens when that condition briefly fails.
The two ringed points are the $1.32bn gaming inventory charge in July 2022, and the $4.5bn H20 charge in April 2025 following the China licence requirement.
Why the response was so slow
AMD launched the Instinct MI300X in December 2023, roughly a year after the demand became obvious. The hardware was competitive on specification. It did not meaningfully change the allocation of orders, because the constraint buyers faced was not finding a chip — it was finding a chip their existing code, their existing networking and their existing operations staff could use at scale immediately.
Mellanox turned out to matter here as much as CUDA. Training a large model is a networking problem as much as a compute problem, and NVIDIA was the only vendor selling the whole rack.
When you are the bottleneck, you set the price. The strategic task is not to reach that position — it is to notice how temporary it usually is.
A company whose revenue is four other companies' capital budgets is no longer a supplier to an industry. It is a dependency of one — and it inherits that industry's risks.
Where the money comes from
NVIDIA’s revenue is, to a first approximation, a share of what a small number of companies choose to spend on data centres. That spending is disclosed, so the demand side of this business can be examined directly rather than inferred.
Exhibit 4
Annual purchases of property and equipment, US$ billions, from each company's consolidated cash flow statement.
ReadingCombined capital expenditure at the four largest hyperscalers went from $140bn in 2023 to $358bn in 2025. NVIDIA’s growth is not a mystery: it is a large and rising share of that number.
AMAZON: capex on this line includes very large NON-AI components — fulfilment centres, sort centres, delivery stations, transportation fleet, Whole Foods and physical stores, and (from 2025) Project Kuiper satellite infrastructure. Amazon's AI/AWS share of total capex is not separately disclosed on the cash flow statement. Do not treat Amazon's capex as an AI proxy. Microsoft's fiscal year ends in June, so its series is not calendar-aligned with the other three.
SourceSEC EDGAR XBRL, PaymentsToAcquirePropertyPlantAndEquipment
The concentration question
This is the obvious objection, and NVIDIA has been answering it with its disclosure. From fiscal 2027 the company splits Data Center into Hyperscale and everything else — AI clouds, industrial customers, enterprises and governments.
In the most recent quarter, hyperscalers were 50.3% of Data Center revenue — $37.9B of $75.2B. That NVIDIA chose to start reporting this split is itself informative: it is the metric the company expects to be judged on.
From hyperscalers
Share of Data Center revenue in the quarter ended April 2026. The remainder is AI clouds, industrial, enterprise and sovereign customers.
Supply commitments
Manufacturing, supply and capacity commitments at April 2026. NVIDIA has changed the basis of this disclosure repeatedly, so it is not a comparable series.
From China
Down from a fifth of revenue before the 2022 export controls, after successive restrictions and a $4.5bn H20 charge.
Operating margin
Quarter ended April 2026. Operating income of $53.5bn on revenue of $81.6bn.
What the market now expects
The most telling recent development is not in the revenue line. It is in how little the revenue line now moves the share price.
Exhibit 5
Share price move on the session after each quarterly report, against the year-on-year revenue growth reported.
ReadingNo longer. Through 2023 the two moved together. Since then they have separated: NVIDIA reported 85 per cent year-on-year growth in May 2026 and the shares fell. Growth is now the expectation rather than the surprise, which is the ordinary fate of a company that has become infrastructure.
The whole picture
NVIDIA passed a $5tn valuation in October 2025, the first company to do so. It peaked near $5.7tn in May 2026 and has since traded back to roughly $5.0tn. It is currently the most valuable company in the world, though the position has changed hands with Apple several times.
Exhibit 6
Total return index, January 2015 = 100, split-adjusted. Switch the vertical scale to compare readings.
ReadingOn a linear scale the last three years look like a discontinuity. On a logarithmic scale, where equal distances are equal percentage moves, the same data shows steady compounding from 2016 interrupted by two severe drawdowns. Both readings are true; the linear one is the one usually published. Note also that AMD’s return over the full period is of the same order as NVIDIA’s, from a 2015 base close to bankruptcy.
Prices are adjusted for the 4-for-1 split of July 2021 and the 10-for-1 split of June 2024, but not for dividends.
SourcesYahoo Finance daily closes, split-adjusted; Reuters, market capitalisation milestones
Becoming critical infrastructure is not the end of strategic risk. It is the point at which a company’s risks stop being competitive and start being macroeconomic and political.
The threats to NVIDIA’s position are no longer chiefly about chips. They are the possibility that hyperscaler capital budgets normalise; that custom accelerators — Google’s TPUs, Amazon’s Trainium, the designs Broadcom builds for others — absorb the predictable inference workloads while NVIDIA keeps the frontier training ones; and that export policy continues to redraw the map of who may buy what. China has already fallen to around nine per cent of revenue.
The record
Every event below is sourced to a filing, a company announcement, a government publication or a wire service. Select any point to read what happened and why it mattered.
Exhibit 7
68 sourced events, 2006 to July 2026, by category. Larger points are annotated elsewhere in this piece.
Product
AI ecosystem
Corporate
Competitive
Regulatory
Market shock
OpenAI released ChatGPT as a free research preview, fine-tuned from a model in the GPT-3.5 series. The conversational interface made large-language-model capability legible to non-technical users for the first time at scale.
ChatGPT converted large-model inference from a research cost into a consumer product with immediate compute requirements, and triggered the capital expenditure cycle that took NVIDIA data-centre revenue from $15 billion in fiscal 2023 to $193.7 billion in fiscal 2026. NVIDIA management subsequently used 'since the emergence of ChatGPT' as its reference point for the roughly 13x expansion in that line.
One million users within five days of launch.
Sourcesopenai.com; reuters.com
ReadingThe product lane starts in 2006 and runs almost alone for a decade. The ecosystem lane — the events that created the demand — does not become dense until 2022. The regulatory lane does not begin until 2022 and has not stopped since.
SourceNVIDIA newsroom, SEC filings, US Bureau of Industry and Security, Reuters and other wire services
What transfers
NVIDIA’s growth rate is not a template — it depended on a once-in-a-generation shift in what computers are asked to do. The decision-making underneath it is more portable.
CUDA earned almost nothing for the better part of a decade while consuming a meaningful share of a much smaller company's R&D. The test of such an investment is not whether it pays soon, but whether it compounds outside your own organisation. Switching costs accumulating in customers' codebases cannot be bought back by a competitor.
The cryptocurrency cycle of 2017–18 flattered NVIDIA's numbers and then removed 46 per cent of gaming revenue in a single quarter. The company was penalised by the SEC for not making the composition of that growth clear. Management teams that cannot decompose their own growth cannot forecast it — and neither can their investors.
NVIDIA's data-centre business passed gaming permanently in May 2022, six months before ChatGPT was released. The popular account inverts the sequence. A transformation that is visible in the financials before its supposed trigger had a different cause, and the difference matters when drawing lessons from it.
Gross margin expanding from 64.6 to 78.4 per cent while unit demand was also rising is not operating leverage; it is the absence of a substitute. That condition is valuable and inherently temporary, so the useful question is always what would end it — and in this case the answer was custom silicon and export policy, not a better GPU.
NVIDIA's principal exposures are now the capital budgets of four companies, the direction of US export policy, and the possibility that inference migrates to cheaper custom accelerators. None of these is a competitive threat in the conventional sense. Infrastructure businesses trade competitive risk for macroeconomic and political risk.
NVIDIA reported 85 per cent year-on-year revenue growth in May 2026 and the shares fell. A company can outperform every operational measure and still disappoint, because the valuation has already assumed the outperformance. This is the ordinary endpoint of a re-rating, and it arrives without warning.
In summary
NVIDIA reached $81.6B of quarterly revenue at a 74.9% gross margin because it spent the previous fifteen years making a general-purpose computing platform out of a graphics part, and because the workload that eventually needed that platform arrived with no viable alternative supplier.
2006 – 2016
CUDA makes the GPU programmable for general work. NVIDIA carries the cost of the software layer for a decade, spending over a quarter of revenue on R&D while data centre remains under seven per cent of the business.
2017 – 2022
Cryptocurrency demand obscures the underlying shift and then collapses. Mellanox is acquired in 2020, adding the networking half of the rack. Data Center passes Gaming permanently in May 2022 — six months before ChatGPT.
2022 – 2024
The May 2023 guidance repriced the sector in a session. Data Center revenue grows 4.3× in four quarters while gross margin expands 11.4 points — the signature of a supplier with no substitute.
2024 – 2026
Blackwell and Rubin ramp against hyperscaler capital budgets that reached $358bn in 2025. NVIDIA passes $5tn. China falls to around nine per cent of revenue, and strong results stop moving the share price.