In three years NVIDIA went from a $27B chip company to a ~$5.4 trillion colossus — the most valuable company on earth — by selling the picks and shovels of the AI gold rush[15][4].

The genuinely open question is not whether NVIDIA is dominant — fiscal-2026 revenue of $215.9 billion (+65%) and a quarterly run-rate above $80 billion say it is[1][2] — but whether that dominance is durable. Its moat, its ~90% share, its premium margins, its customer base and its biggest growth market are each contested by serious people with real evidence. This study lays out both cases on every question; the verdict is yours.

The decisive questions

Each links to the section that lays out the evidence on both sides.

The climb that frames the debate

Annual revenue ($B, fiscal years ending late January). The flat FY2023 crypto trough, then the AI inflection — the speed is the bull case and the bear case at once. Hover a point for detail.

NVIDIA designs the GPUs, networking and software that train and run most of the world’s AI. It has been a 33-year overnight success: founded in 1993, near-bankrupt more than once, and the first company ever to reach a $5 trillion market cap in 2025[8]. The ascent is real — and exactly why every assumption behind it is now scrutinized.

What NVIDIA actually sells

NVIDIA designs accelerated-computing hardware and the software stack around it. Today that is overwhelmingly data-center AI systems — GPUs (Hopper, Blackwell, the new Rubin), CPUs (Grace/Vera), and the NVLink, InfiniBand and Spectrum-X networking that lashes thousands of chips into one machine — plus the CUDA software platform developers build on[19][16]. Gaming (GeForce), professional visualization, and automotive/robotics remain real businesses but are now a small minority of revenue (see Business Model).

From near-death to most-valuable

The arc is unusually dramatic. NVIDIA was reportedly down to a month of payroll in the late 1990s, survived the dot-com and crypto cycles, and saw revenue go flat as recently as fiscal 2023 ($27.0B, with gross margin collapsing to ~57% in an inventory writedown)[4]. Then ChatGPT arrived: data-center demand exploded, revenue went from $60.9B (FY24) to $130.5B (FY25) to $215.9B (FY26), and the stock rode it to the top of the world’s market-cap rankings[4][1][15].

The milestones

Both sides of the ledger

Even the company’s history reads two ways — weigh them yourself.

The bull and bear cases share one fact and split on its meaning: hyperscalers are spending an estimated ~$600 billion on infrastructure in 2026 and NVIDIA captures ~90% of AI-accelerator spend[18]. To NVIDIA, that is the early innings of a build-out where Huang sees $1 trillion+ of AI hardware sold through 2027[12]; to skeptics, it is a debt-funded peak that history says tends to overshoot[24].

The size of the prize

AI infrastructure spending is enormous and concentrated. The largest five hyperscalers are projected to spend ~$602B of capex in 2026 (about +36% YoY), of which roughly 75% (~$450B) targets AI; Goldman Sachs pegs 2025–2027 hyperscaler capex at ~$1.15 trillion[18]. NVIDIA estimates it captures ~90% of accelerator spend — an estimated ~$180B GPU/accelerator budget flowing largely through it[18]. Looking out, CEO Jensen Huang projects at least $1 trillionof cumulative AI-hardware sales through 2027 — and argues demand will exceed even that (“we are going to be short”)[12].

What is actually driving demand

The current wave is the shift from training ever-larger models to deployingthem — “agentic AI” doing productive work — which NVIDIA argues expands, rather than saturates, compute demand[12]. Inference is becoming the majority of AI compute, which is a double-edged fact: it grows the market, but it is also the workload where cheaper custom silicon is most competitive (see Competition)[24].

Both sides of the ledger

The product story is “the data center is the computer.” NVIDIA sells GPUs, Grace/Vera CPUs and — via the $6.9B Mellanox acquisition — the NVLink, InfiniBand and Spectrum-X networking that turns them into one machine; networking alone was ~$31B in FY2026[19]. The counter-argument is that custom ASICs now beat general-purpose GPUs on cost for specific workloads[7].

The full stack, on an annual clock

NVIDIA’s differentiator is integration. Rather than selling discrete chips, it ships rack-scale systems where the GPU, CPU and interconnect are co-designed — what CFO Colette Kress calls “extreme co-design across all chips of the supercomputer”[19]. Networking, built on Mellanox, has become a business in its own right: roughly $11B in a single quarter and $31B+ in FY2026, more than 10x its FY2021 level[19]. On top of that sits an annual product cadence — Hopper, then Blackwell, now Rubin — that compounds a performance lead competitors must chase[7].

Where custom silicon wins

The most credible technical challenge is not a better GPU but a different shape of chip. Hyperscaler custom ASICs— Google’s TPU, Amazon’s Trainium, Microsoft’s Maia, all increasingly co-designed with Broadcom or Marvell — trade general-purpose flexibility for efficiency on a fixed workload. Google claims its Ironwood TPU delivers ~44% lower total cost of ownershipthan NVIDIA’s GB200 with ~90% utilization; analysts cite custom-ASIC TCO advantages of up to 65% for inference at scale[7]. Custom-ASIC server shipments are projected to reach ~27.8% of the market in 2026[7].

Both sides of the ledger

NVIDIA is, financially, an AI data-center company with side businesses. Data Center was $193.7B of $215.9B (~90%) in FY2026, at a ~71% gross margin; gaming has been folded into a new “Edge Computing” line and is now under ~8% of revenue[1][20]. Pricing power is the whole story — and the whole risk.

Where the money comes from

FY2026 revenue by segment ($B). Data Center is the company; everything else rounds to a rounding error. Hover a slice.

The model is simple to state and hard to copy: design the most capable AI accelerators and systems, sell them at a premium into a market that cannot get enough, and reinvest in an annual product cadence. The result is a ~71% FY2026 gross margin (and ~75% in the most recent quarter, before any China charges), with $120B of FY2026 net income[1][2]. Capital returns are now large too — a new $80B buyback authorization and a dividend raised to $0.25 in Q1 FY2027[2].

The reclassification tells the story

In May 2026 NVIDIA stopped reporting Gaming as its own segment, folding GeForce into an “Edge Computing” platform alongside PCs, consoles, AI-RAN, robotics and automotive. Edge was $6.4Bin Q1 FY2027, and gaming is now under ~8% of revenue — NVIDIA hadn’t shipped a new flagship consumer GPU in ~18 months[20]. It is a clean signal of where management’s attention has gone, and a reminder that the consumer franchise that built the company is now a footnote to it.

Both sides of the ledger

NVIDIA holds an estimated ~80–92% of the AI-GPU market[23]— but the threats are real and well-funded: AMD’s MI450 won 6-gigawatt commitments from OpenAI and Meta[17], and Broadcom targets >$100B of custom-XPU AI revenue by FY2027[14]. The bull case is that NVIDIA is keeping ~90% share despite all of it.

Five Forces: a structurally tough market NVIDIA still rules

Click a force for the rated pressure and its basis. Four of five forces read high — yet NVIDIA dominates anyway, which is itself the point: the CUDA + full-stack moat is what overcomes the forces.

Where the players sit

A qualitative map (placements are judgments from the cited evidence, not scores): x-axis = how general-purpose the silicon is, y-axis = software-ecosystem depth. NVIDIA is the only player strong on both — custom ASICs cluster in the specialized, thinner-ecosystem corner. Hover a point for the basis.

The three fronts

AMD is the nearest merchant-GPU rival: the MI450 (CDNA 5, TSMC) won a 6GW OpenAI deal — with a warrant for OpenAI to buy up to 160 million AMD shares — plus a multi-gigawatt Meta commitment, taking AMD to an estimated ~13% of AI-accelerator share[17]. Custom silicon from Google, Amazon, Microsoft and Meta — largely co-designed with Broadcom and Marvell, who control ~95% of ASIC design — is the structural threat, with Broadcom alone guiding to >$100B AI revenue by FY2027[14]. And the buyersare the competitors: the hyperscalers funding all of the above are also NVIDIA’s biggest customers.

Both sides of the ledger

NVIDIA’s deepest moat is software, not silicon: ~6 million CUDA developers and 20 years of libraries built on the platform raise switching costs across the whole industry[16][13]. The clearest stress test of how durable that is comes from China, where export controls cut NVIDIA from ~95% share to roughly zero— and, by NVIDIA’s own admission, helped rivals build their own ecosystems[11].

The CUDA flywheel

CUDA, launched in 2006, turned NVIDIA’s GPUs into general-purpose computers and accumulated a developer base that is now the moat. NVIDIA reported ~6 million developersat “20 years of CUDA” in 2026, building on 60+ CUDA-X libraries and 33M+ downloads since 2008[16][13]. Because the major AI frameworks are CUDA-first, porting to alternatives (AMD’s ROCm) costs months of engineering — which is why competitive hardware repeatedly fails to dent NVIDIA’s share. Around that sits a second moat: full-stack systems and locked-up advanced-packaging supply (see Product & Technology).

The China stress test of the moat

Export controls offer a natural experiment in how durable the lock-in is when a whole market is removed. NVIDIA’s China data-center share fell from ~95% to ~zero; even after ~10 Chinese firms were approved to buy the H200 under stringent terms (a 25% fee to the U.S. Treasury, per-shipment licenses), no chips had shipped as Beijing blocked imports to push domestic alternatives like Huawei[11]:

The risk to the moat is not merely lost sales but a strengthened rival: a whole market learning to build on non-NVIDIA silicon erodes the very ecosystem advantage CUDA depends on.

Both sides of the ledger

NVIDIA earned $120B of net income on $215.9B of FY2026 revenue, and the latest quarter ran at $81.6B (+85% YoY) with ~75% gross margin and $48.6B of free cash flow[1][2][10]. The debate is about the second derivative: growth is decelerating from triple digits, and four customers were 61% of one quarter’s sales[5].

The disclosed picture

Fiscal years end late January. All figures are from NVIDIA’s reported results.

Sources: FY2026 results [1]; multi-year figures [4]. FY2023 was the trough — revenue flat at $27.0B and gross margin ~57% after an inventory writedown [4].

Cash, capital returns and valuation

Profitability funds large capital returns: a new $80B buyback authorization and a dividend raised to $0.25 in Q1 FY2027, on top of record free cash flow[2][10]. The market has rewarded it — NVIDIA is the world’s most valuable company at ~$5.4 trillion[15] — and the sell-side remains overwhelmingly positive: of 61 analysts, 48 rate Strong Buy and 10 Buy, with an average 12-month target near $297 (range $180–$500)[21]. Yet the stock has fallen after each of the last four beat-and-raise quarters, trading at a mid-40s P/E into the most recent print after an all-time high of $235.74[10][3] — a sign of how much success is already priced in.

Both sides of the ledger

NVIDIA Data Center [2]; AMD Data Center [9]; Broadcom AI semiconductors [14]; custom-ASIC shipment share and TCO [7]; margins/share corroborated [23].

Scale of the most recent quarter

Most-recent-quarter AI / data-center revenue ($B). NVIDIA’s data-center line alone is roughly an order of magnitude larger than its nearest public challengers. Hover a bar for the basis.

The margin gap

Gross margin (%) — the clearest measure of NVIDIA’s pricing power, and the thing competition is designed to compress. NVIDIA’s ~104% return on equity dwarfs AMD’s ~7% on the same comparison[23].

Detailed, sourced competitive evidence is in the Competitive Landscape section; the durability debate runs through Strategy & Moats.

Sell-side sentiment is near-unanimous — 48 of 61 analysts rate Strong Buy, average target ~$297[21]. But the stock has fallen after four straight beat-and-raise quarters[3], and short-seller Michael Burry calls NVIDIA the “Cisco” of an AI bubble[6]. Both camps cite real data; the cruxes are depreciation, concentration, and circular financing.

The bull sentiment

The consensus is that demand is real, visible and accelerating. NVIDIA itself frames it as “the largest infrastructure expansion in human history,” with networking up 199% and free cash flow at a record $48.6B[2][10]. Analysts have largely raised targets after recent results, and the muted stock reaction is read by bulls as a valuation issue (a mid-40s P/E into the print), not a business one[10][21].

The bear case: bubble, depreciation, circularity

The most prominent skeptic, Michael Burry, has disclosed shorts and argues the boom rests on faulty accounting and circular money. His sharpest claim is that hyperscalers depreciate AI GPUs over 4–6 years when the real useful life is closer to 2–3, given NVIDIA’s own annual cadence — which, if true, overstates AI profits[6][22]. Tellingly, Jensen Huang has reinforced the obsolescence point himself:

The counter, from operators like CoreWeave’s CEO, is that older GPUs hold value — 2020-era A100s remain booked and H100s reportedly fetch ~95% of original rental price — though no historical data set yet exists to settle a debate that only began with ChatGPT in 2022[22]. On circular financing, Burry points to NVIDIA’s up-to-$100B OpenAI investment and an $860M CoreWeave lease guarantee as vendor financing to cash-burning customers; NVIDIA’s rebuttal memo says strategic investments are only ~7% of revenue (~$3.7B of ~$53B)[6].

The concrete risks, ranked

• Customer concentration — four direct customers at 61% of a quarter’s sales[5].
• Custom silicon & AMD — structural margin/share pressure from buyers’ own chips (see Competition)[7].
• China — ~95%→~0 share, aiding domestic rivals worldwide[11].
• Demand/financing — depreciation, circular deals, possible over-capacity[22][24].
• Valuation — priced for sustained dominance; little margin for disappointment[3].

Both sides of the ledger

Method

Research proceeded by fan-out web search and direct fetching of primary and reputable secondary sources — NVIDIA’s own results releases and developer materials, peers’ investor disclosures (AMD, Broadcom), reputable trade and business press (Tom’s Hardware, Fortune, The Globe and Mail, CTech, Futurum), market-data aggregators, and named skeptics. Every URL cited here was opened and read during the run; each claim was then transcribed into a structured manifest that tags it with a tier (1 = primary/official, 2 = reputable secondary, 3 = aggregator/soft), a confidence level, and a stance (supporting / critical / neutral). The load-bearing figures for NVIDIA are its FY2026 revenue and segment mix, gross margin and net income, the ~$5.4T market capitalization, accelerator market-share estimates (~70–92%), and the forward AI-infrastructure and hyperscaler-capex projections that underpin the bubble debate. NVIDIA is a U.S.-based, English-language company, so no native-language research pass was required.

Frameworks used

The analysis applies the Pyramid Principle (an answer-first executive summary) to order the argument, Porter’s Five Forces to test competitive pressure, peer comparables and a 2×2 positioning map to locate NVIDIA against rivals, and a revenue-trajectory and segment-mix read alongside a SWOT to frame strengths against threats — each applied even-handedly, with high-pressure forces and risks given the same weight as strengths, since the frameworks organize the evidence rather than render a verdict. A formal discounted-cash-flow valuation was deliberately skipped because the forward AI-infrastructure inputs are too uncertain to support one, and the GPU-depreciation question that would drive it is itself unresolved.

Disclosed vs. estimated

Because NVIDIA is public, the core financials — revenue, segments, gross margin, net income, buyback, and dividend — are disclosed figures taken from its own results. Competitor figures for AMD and Broadcom are likewise their own reported numbers, but on different fiscal calendars, so cross-company comparisons are comparable-basis and directional rather than exact. The remaining headline numbers are estimates: the ~$5.4T market cap moves daily; forward AI-infrastructure figures ($1T of AI hardware through 2027) and 2026 hyperscaler-capex numbers are projections by NVIDIA and banks; and market-share numbers (~70–92%; AMD ~13%; custom-ASIC ~27.8% of shipments) are third-party estimates that vary by source and definition. Practitioner and analyst sentiment is labeled as sentiment, not fact.

Neutrality & independence

This is a compilation, not an argument: each section pairs the case for NVIDIA against the case against it, and positive and critical claims alike are attributed to their sources. The study is an independent research artifact, not affiliated with, sponsored by, or endorsed by NVIDIA or any company named here. It is point-in-time as of 2 June 2026, and corrections are welcome.

Cross-checked at build time by an automated link checker. Financial figures are from NVIDIA’s and peers’ public disclosures; market-share, TAM and valuation-multiple figures are reported estimates and labeled in Methodology & Limits.