본문 바로가기
AI,Tech,Sci/AI Infrastructure Dissection

[AI Infrastructure Dissection] Who Really Profits from the AI Boom? The Anatomy of Fabless Giants

by pragma 2026. 7. 3.
Latest Quarter Revenue
$81.6B $19.3B $10.3B
NVIDIA Q1 2026 · Broadcom Q4 2025 · AMD Q1 2026
Latest Quarter YoY Growth
+85% +29% +36%
NVIDIA Q1 2026 · Broadcom Q4 2025 · AMD Q1 2026
AI Infrastructure Strategy
NVDA — CUDA full-stack ecosystem AVGO — Custom ASIC + software AMD — x86 + Instinct GPU
cf.  This report groups NVIDIA, Broadcom, and AMD within the same S&P 500 GICS bucket: the Information Technology sector / Semiconductors & Semiconductor Equipment sub-industry. All three companies follow a fabless model — they design chips but outsource manufacturing to foundries like TSMC. A foundry is a contract chip manufacturer that owns the fabrication facilities. This is the opposite of an IDM (Integrated Device Manufacturer), such as Intel or Samsung, which designs and manufactures its own chips. The fabless model's core advantage is low capital expenditure: no factories to build or maintain. Beyond that shared model, all three face the same structural variables: US export control regulation administered by the Bureau of Industry and Security (BIS), and the capital expenditure cycle of hyperscale cloud providers (AWS, Microsoft, Google). That said, sharing a sector classification does not mean sharing a business model. This comparison exists to show where they diverge.
⚠ Fiscal Year Alignment Note — The three companies do not share the same fiscal year end. AMD's fiscal year ends in late December and aligns closely with the calendar year. Broadcom ends in late October. NVIDIA ends in late January. All data in this report has been converted to Calendar Quarter (CQ) labels. NVIDIA's internal fiscal quarter designations run one quarter ahead of the calendar (e.g., NVDA FQ1 2027 = Calendar Q1 2026). Broadcom lags by approximately one month. Calendar Q1 2026 data for Broadcom is not available — the corresponding AVGO FQ2 2026 filing had not been submitted at the time of writing.
Executive Summary

Three companies in the same S&P 500 semiconductor sub-industry, but they occupy fundamentally different positions in the AI infrastructure cycle. NVIDIA became the gravitational center of the industry with the Hopper architecture launch in 2023 — Compute & Networking segment revenue grew from $4.5B in Q1 2023 to $74.6B in Q1 2026, roughly 17x in 13 quarters. The engine behind that is CUDA (Compute Unified Device Architecture), NVIDIA's software platform for parallel computing. AI researchers and engineers have spent years building code that runs on NVIDIA GPUs via CUDA. Switching to a competitor's GPU would require rewriting that entire codebase. That switching cost is the moat — a structural barrier that makes the business something more than a chip company. Once an organization is deep in the CUDA ecosystem, exit is expensive.

Broadcom took a different path. Its $86.2B acquisition of VMware hedges semiconductor cyclicality with software subscription revenue — Infrastructure Software segment margins of 70–78% consistently exceed the Semiconductor Solutions segment's 55–60%. Broadcom's October fiscal year end means the latest available data here is Q4 2025 (AVGO FQ1 2026, ended February 1, 2026): $19.3B revenue, up 29% year-over-year. AMD survived between the two giants by taking server CPU share from Intel with its EPYC processors, while its AI GPU business (Instinct MI300 series) began contributing from 2024 onward — but the gap to NVIDIA on AI compute has not closed. What separated the three companies across these 13 quarters was not raw engineering capability. It was the depth of ecosystem lock-in.

On capital efficiency, Broadcom leads by a wide margin — free cash flow (FCF) margin has held at 40–50% consistently. NVIDIA is the stranger case: despite explosive revenue growth, it kept capex below 2–3% of revenue while returning $40.4B to shareholders in FY2026 alone through buybacks. AMD's FCF margin improved from single digits in early 2023 to 25% in Q1 2026, but the absolute scale remains incomparable to the other two.

Q2 2025 was the quarter where all three hit the same external shock simultaneously — US export controls forced NVIDIA to absorb $4.5B in inventory losses and AMD $800M. How fast each recovered from that shock is one of the cleaner tests of structural resilience in this dataset.

00

Fiscal Year Alignment

The three companies have different fiscal year end dates, so direct comparison of quarterly data requires a mapping step. All data in this report uses Calendar Quarter (CQ) labels. All section labels below follow CQ convention.

Fiscal Quarter → Calendar Quarter Mapping (Selected Quarters)
Calendar QuarterAMD Fiscal QuarterBroadcom Fiscal QuarterNVIDIA Fiscal QuarterNotes
Q1 2023FQ1 2023 (Apr 1)FQ2 2023 (Apr 30)FQ1 2024 (Apr 30)AVGO·NVDA lag by ~1 month
Q4 2023FQ4 2023 (Dec 30)FQ1 2024 (Feb 4) ⚠FQ4 2024 (Jan 28)AVGO FQ1 2024 = 14-week quarter (53-week fiscal year)
Q3 2024FQ3 2024 (Sep 28)FQ4 2024 (Nov 3) ⚠FQ3 2025 (Oct 27)AVGO FQ4 2024 = final quarter of 53-week fiscal year
Q4 2025FQ4 2025 (Dec 27)FQ1 2026 (Feb 1)FQ4 2026 (Jan 25)Latest available AVGO data
Q1 2026FQ1 2026 (Mar 28)N/A (FQ2 2026 not yet filed)FQ1 2027 (Apr 26)No Broadcom data available
01

Revenue & Growth

Revenue scale and growth velocity map directly onto where each company sits in the AI infrastructure cycle. NVIDIA's inflection came in Q2 2023 with the start of Hopper shipments — from $8.3B that quarter to $81.6B in Q1 2026, roughly 10x in under three years. Broadcom's revenue jumped in Q4 2023 following the VMware acquisition close, then continued expanding steadily. The latest available quarter is Q4 2025, where Broadcom reported $19.3B in revenue, up 29% year-over-year. Calendar Q1 2026 data is not available — the corresponding AVGO FQ2 2026 had not been filed at the time of writing. AMD struggled in 2023 as the PC market downturn dragged the Client segment sharply lower, but recovered from 2024 onward driven by Data Center growth.

Analysis Note — Same AI Cycle, Different Entry Points All three get labeled "AI beneficiaries," but the mechanism is different for each. NVIDIA is the de facto monopoly supplier of AI training and inference compute. Training is the process of building a large language model (LLM) — feeding it vast amounts of data until it learns patterns. Inference is when a finished model answers a question or generates output. Both processes require thousands of GPUs operating in parallel, and at the scale cloud providers operate, the H100 and Blackwell chips are the only practical option. That position is the direct cause of NVIDIA's revenue surge.

Broadcom's growth has two separate drivers. On the semiconductor side, demand grew for custom AI accelerators — chips called ASICs (Application-Specific Integrated Circuits), designed from scratch for one customer's specific workload rather than built as general-purpose hardware. Google's TPU (Tensor Processing Unit) and Meta's MTIA are prominent examples. Broadcom designs and supplies these. On the software side, the VMware customer base was converting from perpetual licenses to subscriptions, generating recurring revenue.

AMD's Data Center growth came primarily from taking x86 server CPU market share from Intel with its EPYC processor line. Its AI GPU business (Instinct MI300 series) started contributing meaningfully from 2024, but the gap to NVIDIA on AI compute workloads remains large.
Quarterly Revenue ($B)
YoY Revenue Growth (%)
Quarterly Revenue ($M)
QuarterNVIDIABroadcomAMD
02

Segment Decomposition

Consolidated margins do not show what each segment actually earns. NVIDIA has two reportable segments: Compute & Networking (data center GPUs for AI training and inference, plus networking chips like InfiniBand — a high-speed, low-latency interconnect standard used to link GPUs within large AI clusters — and Spectrum-X Ethernet, NVIDIA's enhanced Ethernet product optimized for AI workloads) and Graphics (gaming GPUs for PC and laptop, plus professional visualization cards).

Broadcom splits into Semiconductor Solutions (AI accelerator chips, networking ASICs, storage controllers, and other hardware) and Infrastructure Software (VMware cloud management platform, plus CA Technologies — enterprise IT management software — and Symantec — cybersecurity software — both acquired before VMware).

AMD operated four segments through most of this period: Data Center (EPYC server CPUs and Instinct AI GPUs), Client (Ryzen CPUs for laptops and desktops), Gaming (Radeon consumer GPUs and console chip supply), and Embedded (FPGAs and SoCs for industrial, communications, and aerospace use, obtained via the Xilinx acquisition in 2022 — Xilinx was the leading FPGA maker before AMD purchased it for $49B). From Q1 2025 onward, AMD merged Client and Gaming into a single segment. The Embedded segment anchored profitability above 40% operating margin; the Client segment was the drag, falling as low as -23% in early 2023 during the PC downturn.

Analysis Note — Segment Structure Changes to Watch AMD merged Client and Gaming into a single segment from Q1 2025 — direct comparison to pre-2025 data requires accounting for this. The Embedded segment deserves particular attention. AMD acquired it through the 2022 Xilinx purchase. Xilinx made FPGAs (Field-Programmable Gate Arrays — chips whose circuitry can be reprogrammed after manufacture, unlike standard CPUs or GPUs which are fixed at design time) and SoCs (System on Chip — integrating multiple functions onto a single die) for telecommunications, industrial equipment, and aerospace.

Because these chips serve specialized industrial customers rather than consumer markets, demand is less volatile quarter-to-quarter — but the segment runs on an inventory correction cycle that leads the broader economy. The sharp decline in Embedded revenue through 2023–2024 was exactly that: customers drawing down excess inventory before placing new orders. NVIDIA switched to a new segment reporting structure in Q1 2026 — this report retains the prior structure for consistency across all 13 quarters.

Broadcom's Infrastructure Software segment grew from 21% of total revenue before the VMware acquisition close (Q4 2023) to over 40% afterward. VMware is the dominant vendor of virtualization software for enterprise data centers — virtualization means running multiple virtual servers on a single physical machine, which is foundational to how modern data centers operate. Broadcom converted VMware's perpetual licensing model to subscriptions after the acquisition, shifting the revenue pattern from lumpy to recurring.
NVIDIA — Segment Operating Margin (%)
Broadcom — Segment Operating Margin (%)
AMD — Data Center Operating Margin (%)
NVIDIA — Segment Operating Margin
QuarterC&N Revenue ($M)C&N MarginGraphics Revenue ($M)Graphics Margin
Broadcom — Segment Operating Margin
QuarterSemi Solutions ($M)Semi MarginInfra SW ($M)SW Margin
AMD — Segment Operating Margin
QuarterData Center ($M)DC MarginClient/Gaming ($M)CG MarginEmbedded ($M)Emb Margin
03

Margin Dynamics & SBC

NVIDIA's margin trajectory has no precedent in the semiconductor industry. Gross margin expanded from the low-60% range in early 2023 to the mid-70% range by late 2025. Two things drove this. First, product mix shifted toward higher-margin Hopper and Blackwell systems. Second, NVIDIA increasingly sells in "AI Factory" configurations — not individual GPU chips, but complete rack-scale systems (a rack is a standard data center cabinet; rack-scale means the entire rack, containing hundreds of GPUs plus networking and cooling, designed and sold as a single integrated unit). The average selling price (ASP) for a rack-scale system is dramatically higher than for individual chips.

Broadcom's margins rose structurally as VMware's high-margin subscription revenue grew to over 40% of consolidated sales. AMD's margins have improved but the gap to NVIDIA on gross margin has not closed. Stock-based compensation (SBC) — the cost of paying employees in stock or options rather than cash, which appears as a GAAP expense but involves no cash outflow — has grown at all three companies, with NVIDIA's AI talent competition the primary driver of absolute dollar increases.

Analysis Note — Baseline Effects to Control For NVIDIA extended the useful life of server and networking equipment from 3 to 4–5 years in Q1 2023, reducing annual depreciation by $135M on a FY2024 basis (net income effect: +$114M). No cash changed hands — this was an accounting estimate revision.

Broadcom absorbed VMware integration restructuring charges of $712M in Q4 2023 and $345M and $361M in Q1 and Q2 2024. These pushed GAAP (Generally Accepted Accounting Principles) operating margins temporarily lower. Stripping them out under Non-GAAP (adjusted operating income, which excludes SBC, acquisition-related amortization, and restructuring) reveals the underlying business was far more profitable than the GAAP line suggested during that period.

For AMD, the single largest Non-GAAP adjustment is amortization of acquisition-related intangibles — assets from the Xilinx and other acquisitions written down over time at $551M–$823M per quarter. This is a GAAP expense with no cash impact.
SBC ($M) — Quarterly
SBC / Revenue (%)
SBC ($M)
QuarterNVIDIABroadcomAMD
04

Capital Intensity & Cash Flow

The core financial advantage of the fabless model — designing chips but outsourcing manufacturing to TSMC — is low capital expenditure. All three companies run capex-to-revenue ratios of 1–4%, compared to 20–30% for companies that own and operate fabrication facilities. Broadcom is the most capital-efficient of the three: capex below 1% of revenue while sustaining 40–50% FCF margins. NVIDIA managed to keep capex growth restrained even as revenue surged, achieving remarkable FCF margins relative to its revenue scale. AMD's cash flow became complicated for one quarter following the ZT Systems acquisition in Q1 2025, but FCF from continuing operations improved steadily throughout the period.

Why FCF Margin Matters Operating margin measures the fraction of revenue left after accounting charges. FCF margin measures the fraction of revenue that becomes actual usable cash. The two differ because some costs reduce accounting profit without requiring cash — depreciation is the clearest example, where assets are written down annually with no cash leaving the company. Others, like inventory build-up, generate accounting profit while cash gets tied up waiting to ship.

FCF = Operating Cash Flow minus Capital Expenditure. That is the cash available for dividends, buybacks, and acquisitions. A 40%+ FCF margin, as Broadcom sustains, means $40 of every $100 in revenue converts to free cash. AMD's single-digit FCF margins in early 2023 meant the business was generating very little freely deployable cash relative to its revenue — most of the operating cash was cycling back into the business.
Analysis Note — Reading the Anomalous FCF Quarters NVIDIA's Q1 2025 FCF margin compressed to 45.3% because the $4.5B H20 inventory charge directly hit operating cash flow (OCF). On the Q1 FY2026 earnings call, CFO Colette Kress laid out the numbers: NVIDIA recognized $4.6B in H20 revenue that quarter (shipped before April 9), was unable to ship an additional $2.5B, and wrote down $4.5B against existing inventory and purchase commitments. Writing down locked-up inventory and purchase obligations all at once pulled OCF lower. The charge came in below initial estimates because, as Colette noted, the company was able to reuse certain materials.

Q2 2025's FCF margin of 29% — the tracked-period low — was also not a structural problem. It reflected deliberate inventory build ahead of the Blackwell Ultra (GB300) product transition — GB300 is the next-generation Blackwell architecture, offering improved performance per watt over the original GB200. On the same quarter's earnings call, Colette disclosed that inventory increased sequentially from $11B to $15B to support Blackwell and Blackwell Ultra production ramp. Cash did not disappear — it converted into finished goods inventory. The Q3 revenue guidance of $54B, more than $7B above Q2, was the demand confirmation that validates that inventory build.

Data verification: Initial FCF figures contained errors corrected against SEC primary filings. NVDA Q1 2025 ($25,751M→$19,983M), Q3 2025 ($22,316M→$25,606M), Q4 2025 ($55,583M→$34,100M). AMD Q4 2025 also corrected ($2,378M→$2,082M) after separating continuing from discontinued operations following the ZT Manufacturing divestiture.
FCF Margin (%) — Quarterly
Capex / Revenue (%)
FCF ($M) & FCF Margin
QuarterNVDA FCFNVDA FCF MarginAVGO FCFAVGO FCF MarginAMD FCFAMD FCF Margin
05

Capital Allocation

Capital allocation decisions reveal management priorities more directly than almost any other metric. NVIDIA scaled shareholder returns alongside revenue — FY2026 buybacks of $40.4B were 19% above the prior year's $34B. Broadcom maintained a steady dividend program ($11.1B annually in FY2025) following the $86.2B VMware acquisition, managing the balance between debt service, dividends, and remaining buyback capacity. AMD concentrated on M&A — the ZT Systems acquisition at $4.4B — while keeping buybacks modest at $1.3B.

Annual Capital Allocation ($M)
CompanyYearRevenueCapexR&DR&D%BuybackDividend
NVIDIA FY202460,9221,0858,67514.2%9,746395
FY2025130,4972,83112,9149.9%34,015834
FY2026215,9384,51018,4978.6%40,388974
Broadcom FY202335,8195255,25714.7%5,8247,645
FY202451,5745169,31418.1%7,1769,814
FY202563,88771110,98117.2%2,45011,142
AMD FY202322,6805465,87225.9%9850
FY202425,7856366,45625.0%8620
FY202534,6399748,09123.4%1,3160
Analysis Note — M&A Events and R&D Intensity AMD's R&D intensity of 23–26% of revenue is the highest of the three. This is not a sign of inefficiency — it is structural. AMD has to simultaneously develop CPUs (central processing units, the general-purpose brain of a computer), GPUs (graphics processing units, optimized for parallel computation), and FPGAs, while competing against both NVIDIA and Intel. The x86 architecture, originally developed by Intel in 1978, is the instruction set that runs the vast majority of the world's servers and PCs. AMD is the only company other than Intel licensed to design x86 processors — that exclusivity is AMD's core competitive position in servers.

NVIDIA's R&D intensity has fallen sharply from 14.2% in FY2024 to 8.6% in FY2026 — R&D spending in absolute dollars increased, but revenue grew much faster. Key M&A events: Broadcom acquired VMware ($86.2B, FY2024); AMD acquired ZT Systems ($4.4B, FY2025, then divested the manufacturing unit for $2.4B); NVIDIA's attempted Arm acquisition failed ($1.35B termination fee, FY2023).
06

Loss Segment Efficiency

Of the three companies, only AMD reported a full-year operating loss in any segment — the Client segment posted -$46M in 2023 as the PC market downturn hit. NVIDIA and Broadcom maintained operating profitability across all segments throughout the tracked period. Note: AMD's Data Center segment also posted a quarterly loss of -$155M in Q2 2025, but this was caused by a single one-time event — the $800M inventory write-down tied to MI308 export controls, not a structural decline in the business.

AMD Client Segment Efficiency Over Time
YearSegmentRevenue ($M)Op. Income/Loss ($M)Loss/RevenueTrend
FY2022Client6,201+1,190 (profit)$0.81Baseline
FY2023Client4,651−46 (loss)$1.01Deteriorated
FY2024Client & Gaming*9,649+1,187 (profit)$0.88Recovered

*Client and Gaming merged from FY2024. AMD Q3 2023 returned to quarterly profitability (+$140M); full-year loss minimized.

07

Notable Anomalies

These are one-time items that distort direct quarter-to-quarter comparisons. Q2 2025 was particularly unusual — all three companies absorbed the same external shock simultaneously, when the US tightened semiconductor export controls.

NVIDIA · Q1 2023
Asset Useful Life Extension
Server and networking equipment useful lives extended from 3 to 4–5 years; assembly and test equipment from 5 to 7 years. Annual depreciation savings of $135M on a FY2024 basis, adding ~$114M to net income. Quarterly effects: Q1 +$33M, Q2 +$33M, Q3 +$41M, Q4 +$28M. No cash involved — accounting estimate revision only.
Annual depreciation savings: $135M
NVIDIA · Q1 2025
H20 Inventory Loss $4.5B — Export Control Direct Hit
H20 was a GPU NVIDIA designed with reduced performance to comply with earlier export rules — built specifically for the China market. On April 9, 2025, the US government applied new controls to H20 with no grace period. CFO Colette Kress disclosed the figures on the earnings call: $4.6B in H20 revenue recognized before April 9; $2.5B unable to ship; $4.5B written down against existing inventory and purchase obligations. The charge came in below initial estimates because the company was able to reuse certain materials. Jensen Huang declared this the "end of the road for Hopper" in China and estimated the China AI accelerator market at ~$50B — effectively lost. Q2 Data Center still grew despite absorbing an additional $8B H20 headwind, which itself demonstrates non-China Blackwell demand strength.
One-time loss: −$4,500M / China TAM (Total Addressable Market) lost: ~$50B
NVIDIA · Q1 2026
$16B Equity Gains — Net Income Distortion
NVIDIA holds strategic equity stakes in AI startups and other companies. Changes in the fair value of these stakes flow through the income statement. This quarter, unrealized gains on public equity securities of $13.4B plus non-marketable equity gains of $2.6B totaled $16.4B in "Other income." Added to operating income of $53.5B, this pushed pre-tax income to $69.9B and reported net income to $58.3B. Normalized net income (operating income × (1 − 16.6% effective tax rate, or ETR — the fraction of pre-tax income actually paid as tax this period)) = $44.7B — reported net income was $13.6B above the underlying business performance.
Reported NI: $58.3B → Normalized: ~$44.7B
BROADCOM · Q4 2023 – Q2 2024
VMware Integration Restructuring Charges
Following the VMware acquisition close (November 2023, $86.2B), integration restructuring charges were recognized across three quarters: Q4 2023 $712M, Q1 2024 $345M, Q2 2024 $361M. The charges covered redundant headcount severance and legacy VMware contract termination costs. These hit GAAP operating margins but are excluded from Non-GAAP — the underlying business was significantly more profitable than GAAP suggested during this period. Smaller integration-related costs continued into 2025.
Three-quarter total: −$1,418M
BROADCOM · Q2 2024
$4.5B IP Transfer Tax Charge — One Event, Two Reporting Stages
Broadcom transferred certain intellectual property rights from overseas entities to a US entity as part of a supply chain reorganization. US tax rules required recognizing the full future tax liability on profits that IP would eventually generate — a Deferred Tax Liability (DTL) of $4,522M recorded at the moment of transfer. No cash left the company; this recognized a future obligation at present value. The event had been disclosed in the prior quarter filing as expected to reach up to $4.9B. This is a single event reported across two quarterly filings — not two separate tax issues.
Tax charge: −$4,522M / GAAP net loss triggered
AMD · Full Year 2023
Client Segment PC Market Collapse
PC demand surged during COVID in 2020–2021, and manufacturers built large inventories. When demand normalized, distributors worked through existing stock before placing new orders — a process called inventory correction. AMD's Client segment average selling price fell -26% and unit shipments fell -55% year-over-year in Q1 2023. Annual Client revenue dropped to $4.7B from $6.2B the prior year, and operating income swung from profit to -$46M loss. The segment returned to quarterly profitability in Q3 2023 (+$140M) after inventory channels cleared.
Annual Client segment loss: −$46M
AMD · Q2 2025
MI308 Export Control Inventory Loss $800M
MI308 is AMD's AI GPU (Instinct series), designed for the China market in the same way NVIDIA designed H20. New US export controls in April 2025 triggered a $800M inventory write-down → Data Center segment operating loss of -$155M. On the Q3 2025 earnings call, CFO Jean Hu confirmed that Q3 results included no MI308 revenue from China shipments — the $4.34B Data Center quarter was achieved entirely without MI308, driven by MI350 series GPU ramp, EPYC Turin CPU shipments, and the OpenAI 6-gigawatt infrastructure contract. A partial export license was subsequently granted; $360M of the original charge was reversed in Q4 2025 → net annual impact -$440M.
Initial loss $800M → Reversal $360M → Net impact −$440M
AMD · Q4 2024 & Q2 2025
Tax Events — Charge and Refund in Opposite Directions
Q4 2024: A $373M deferred tax charge arose from integrating the legal entity structures of acquired companies. Q2 2025: An IRS examination resolved previously uncertain tax positions in AMD's favor, generating a $853M tax benefit. Both items flow through net income but not operating income — they will distort any net income-based comparison for those quarters and need to be separated.
Q4 2024: −$373M / Q2 2025: +$853M
08

Intra-Sector Positioning

The same semiconductor sub-industry classification covers three companies with fundamentally different competitive positions and revenue models. Without understanding the differences, the financial comparisons generate misleading conclusions.

Basis for the Common Classification All three share three structural exposures: the fabless manufacturing model (production outsourced to TSMC), US export control regulation, and the data center capital expenditure cycle of hyperscale cloud providers. Export controls are administered by the Bureau of Industry and Security (BIS), a division of the US Commerce Department. BIS requires licenses to export semiconductors above certain performance thresholds to D5 countries — a classification covering countries with weapons proliferation concerns, including China and Russia.

High-performance AI GPUs became a primary target of these controls from 2022 onward. The H20 (NVIDIA) and MI308 (AMD) export restrictions are the most consequential cases in this dataset. All three companies have China revenue exposure, so each tightening of these rules hits directly. Q2 2025, when all three recognized export control losses simultaneously, is the clearest demonstration of that shared vulnerability.
Structural Differentiators
DimensionNVIDIABroadcomAMD
Core MoatCUDA software ecosystem (5M+ developers)VMware customer base + custom ASIC design capabilityOnly x86 competitor to Intel
AI Benefit PathNear-monopoly GPU supply → direct beneficiaryCustom ASIC design + AI networking → indirect beneficiaryEPYC CPU share gains + Instinct GPU → partial beneficiary
Revenue ModelHardware + software stack bundleSemiconductor + software subscription (VMware) hybridSemiconductor-only (minimal software revenue)
Cycle HedgeLow (near-monopoly demand position)High (SW subscriptions buffer semiconductor cycle)Low (residual PC and gaming cycle exposure)
Key RisksCUDA ecosystem defection / export control expansionVMware customer attrition / debt loadAI GPU competition intensification / Intel recovery
09

Management Commentary — Context Behind the Numbers

Financial figures show what happened. Earnings call commentary shows why it happened and what management expects next. The three quarters below connect directly to the key anomalies in this report.

NVIDIA Q1 2025 — H20 Export Control Impact and Blackwell Transition (May 28, 2025) CFO Colette Kress laid out the H20 numbers precisely. NVIDIA recognized $4.6B in H20 revenue in Q1 (shipped before April 9, the control effective date), was unable to ship an additional $2.5B, and recognized a $4.5B charge against existing inventory and purchase obligations. The total H20 revenue that should have been recognized was approximately $7B. The charge came in below initial estimates because, as Colette noted, the company was able to reuse certain materials. Jensen Huang declared this the "end of the road for Hopper" for the China market and put the China AI accelerator opportunity at approximately $50B — a market now effectively inaccessible.

Jensen simultaneously emphasized Blackwell's acceleration. Blackwell contributed approximately 70% of Data Center compute revenue in Q1, with the transition from Hopper nearly complete. This quarter also produced the first concrete numbers on reasoning AI demand — Microsoft processed 100 trillion tokens in Q1, a five-fold increase year-over-year. Jensen explained that reasoning models (AI that works through problems step by step rather than giving a one-shot response) consume 1,000 times more tokens per task than older chatbot-style models. Grace Blackwell (GB200) — a rack-scale system combining NVIDIA's Grace CPU with Blackwell GPUs connected via NVLink (NVIDIA's proprietary high-bandwidth chip-to-chip interconnect, offering far higher bandwidth than PCIe) — was specifically designed for this compute profile. The Q2 revenue guidance of $45B — achievable despite absorbing $8B in H20 headwind — was itself evidence of non-China Blackwell demand strength.
NVIDIA Q2 2025 — FCF Trough and the ASIC Competition Question (August 27, 2025) The Q2 FCF trough was explained directly on the call. Colette Kress disclosed that inventory increased sequentially from $11B to $15B to support the Blackwell and Blackwell Ultra (GB300 — next-generation Blackwell with improved performance per watt) production ramp. Cash had not left the business — it had converted into finished product inventory. The Q3 revenue guidance of $54B, more than $7B above Q2, was the demand signal that validated the inventory build. Low FCF margin in this quarter was a symptom of strong near-term demand, not a structural problem.

The call also produced Jensen's clearest rebuttal of the ASIC-versus-GPU competition question. BofA analyst Vivek Arya asked whether Broadcom's growing ASIC business represented a threat. Jensen's response: "Many projects are started. Very few products go into production. Accelerated computing is a full-stack co-design problem." His argument is that AI model architectures are changing too quickly for fixed-function ASICs to keep pace — a chip designed for today's model architecture may be poorly suited to the next generation.

Jensen cited three structural NVIDIA advantages: the same programming model available across every cloud, edge device, and robotics platform; acceleration of the entire AI pipeline from data processing through pre-training, post-training, and inference; and performance-per-watt economics that translate directly into customer revenue. He also quantified NVIDIA's share of a 1-gigawatt AI factory for the first time: approximately $35B of the total $50–60B factory cost.
AMD Q3 2025 — Data Center Recovery and the ROCm Software Gap (Late October 2025) The central question for this call was whether the Data Center segment's recovery from -$155M operating loss in Q2 to +$1.1B in Q3 reflected genuine demand or simply the disappearance of the MI308 one-time charge. CFO Jean Hu answered directly: "Q3 results did not include any MI308 revenue from China shipments." The 34% sequential revenue increase came without MI308 — driven by MI350 series GPU ramp and accelerating EPYC Turin (AMD's 5th-generation EPYC server CPU, built on TSMC's 3nm process node) shipments.

On the Embedded segment, Lisa Su noted demand was strengthening in test and emulation, aerospace and defense, and industrial vision and healthcare. She cited over $14B in design wins — a "design win" is when a customer formally selects a chip for a future product, locking in multi-year revenue — on an annualized basis as evidence of structural recovery rather than a temporary bounce.

The most candid moment came on ROCm — AMD's software stack for AI workloads, the direct counterpart to NVIDIA's CUDA. Lisa Su acknowledged: "There is always more work to do, particularly for the new workloads where training and inference are converging with reinforcement learning." That is a clear admission that the software ecosystem gap to CUDA requires ongoing investment. The OpenAI 6-gigawatt infrastructure contract announced this quarter is the most significant lever AMD has to close that gap — at that scale, AMD's software has to work, which creates the deployment experience that builds ecosystem momentum.
Earnings Call Sources
QuarterCompanyDateKey Usage
Q1 2025NVIDIAMay 28, 2025H20 full loss breakdown, Blackwell ramp pace, reasoning AI token demand scale
Q2 2025NVIDIAAugust 27, 2025FCF trough cause (inventory $11B→$15B), Jensen ASIC rebuttal, AI factory economics ($35B/$50–60B)
Q3 2025AMDLate October 2025Data Center recovery confirmed (MI308-excluded), Embedded demand recovery, Lisa Su on ROCm gap
G

Glossary

TermFull NameWhat It Means
YoYYear-over-YearGrowth rate compared to the same quarter one year earlier
FCFFree Cash FlowOperating Cash Flow minus Capital Expenditure. Cash available for dividends, buybacks, and acquisitions after running the business and investing in infrastructure.
FCF MarginFCF / RevenueFCF divided by revenue. How many cents of freely usable cash are generated per dollar of revenue.
GAAPGenerally Accepted Accounting PrinciplesUS standard accounting rules. Includes all costs: stock-based compensation, depreciation, restructuring charges, acquisition amortization.
Non-GAAPAdjusted Operating IncomeAdjusted operating income. Excludes SBC, acquisition-related amortization, and restructuring charges. Presented by management as a cleaner view of recurring profitability.
SBCStock-Based CompensationThe cost of paying employees in stock or options. It shows up as a GAAP expense but involves no cash outflow, which is why GAAP margins are lower than Non-GAAP.
CapexCapital ExpenditureSpending on physical assets. For fabless companies, this is research infrastructure and office facilities rather than manufacturing equipment.
OCFOperating Cash FlowCash generated from the core business operations before capital spending. More reliable than net income as a cash measure because it excludes non-cash depreciation charges.
ASPAverage Selling PriceAverage price per unit sold. Rises when product mix shifts toward higher-end products; falls during inventory clearance periods.
FablessFab-lessA chip company that designs semiconductors but outsources manufacturing to a foundry like TSMC. Core advantage: no factory capital expenditure.
ASICApplication-Specific Integrated CircuitA chip designed from scratch for one specific application rather than general use. Google's TPU and Meta's MTIA are prominent examples. Broadcom designs and supplies these for hyperscale customers.
CUDACompute Unified Device ArchitectureNVIDIA's parallel computing software platform. The de facto standard for AI training and inference. The 5M+ developer ecosystem is NVIDIA's primary competitive moat — switching away requires rewriting years of accumulated code.
BISBureau of Industry and SecurityBureau of Industry and Security — the US Commerce Department division that administers export controls. The H20 and MI308 restrictions were implemented through BIS.
DTLDeferred Tax LiabilityDeferred Tax Liability. A future tax obligation recognized on the balance sheet today. In Broadcom's case, triggered by an intra-group IP transfer that created a taxable event.
CAMTCorporate Alternative Minimum TaxCorporate Alternative Minimum Tax — a 15% minimum tax on large corporations introduced by the US Inflation Reduction Act (2022). Affected Broadcom's Q3 2025 tax expense.
CQCalendar QuarterCalendar Quarter — January–March, April–June, July–September, October–December. Used as the unified reference frame in this report because all three companies have different fiscal year ends.
R

References

All figures in this analysis are sourced from SEC EDGAR filings. Key figures including FCF were cross-verified against primary filings; corrections are documented in the Section 04 analysis note.

NVIDIA (NVDA) — SEC EDGAR CIK 0001045810
Calendar QuarterNVIDIA Fiscal QuarterSEC EDGAR Filing
Q1–Q3 2023FQ1–FQ3 FY2024EDGAR 10-Q Index
Q4 2023FQ4 FY2024 (10-K)EDGAR 10-K Index
Q1 2026FQ1 2027nvda-20260427.htm
Broadcom (AVGO) — SEC EDGAR CIK 0001730168
Calendar QuarterAVGO Fiscal QuarterSEC EDGAR Filing
Q1 2023 – Q4 2025FQ2 FY2023 – FQ1 FY2026EDGAR 10-Q Index
FY2023, FY2024, FY2025 (annual)10-KEDGAR 10-K Index
AMD — SEC EDGAR CIK 0000002488
Calendar QuarterAMD Fiscal QuarterSEC EDGAR Filing
Q1 2023 – Q3 2025FQ1–FQ3 FY2023–2025EDGAR 10-Q Index
Q4 annual (FY2023–2025)10-KEDGAR 10-K Index
Q1 2026FQ1 2026amd-20260329.htm
Earnings Call Transcripts Used in Section 09
Calendar QuarterCompanyDateKey Usage
Q1 2025NVIDIA (FQ1 2026)May 28, 2025H20 loss structure ($4.6B recognized / $2.5B unshipped / $4.5B written down); Blackwell 70% of DC compute; reasoning demand 1,000x token scaling; China TAM $50B declared lost. Source: FactSet CallStreet
Q2 2025NVIDIA (FQ2 2026)August 27, 2025Inventory build $11B→$15B (FCF trough cause); Jensen ASIC rebuttal; NVIDIA $35B of 1GW AI factory. Source: FactSet CallStreet
Q3 2025AMD (FQ3 2025)Late October 2025Data Center $4.34B ex-MI308 confirmed (Jean Hu); Embedded design wins >$14B annual; ROCm software gap acknowledged (Lisa Su). Source: AMD IR