AI data-centre semiconductors. AVGO / NVDA. Run 14 September 2026. Built from Broadcom 10-K FY2023 to FY2025 and 10-Q filings through fiscal Q3 2026, NVIDIA 10-K FY2024 to FY2026 and 10-Q filings through fiscal Q2 2027, both companies' most recent earnings calls, and cited third-party research. Events swept through 14 September 2026; most recent event checked: NVIDIA's acquisition of Hugging Face for $12.93bn, announced 3 September 2026, and Broadcom's fiscal Q3 2026 results and new fiscal 2028 AI revenue target, 2 September 2026. Figures in USD; both companies report in USD, so no FX conversion applies. Not a valuation and not a recommendation.
Broadcom's trailing rate is faster, and that is a base effect rather than a mechanism. AI semiconductor revenue grew 221% year on year to $16.7bn in fiscal Q3 2026 against NVIDIA data-centre growth of 117% to $89.0bn, but the forward guidance both managements have put on the record reverses the ranking. Broadcom guides AI revenue of $115bn in fiscal 2027 and $230bn in fiscal 2028; adding infrastructure software and non-AI silicon at current trajectories puts total fiscal 2027 revenue near $168bn against roughly $106bn in fiscal 2026. NVIDIA guides fiscal Q3 2027 revenue of $108bn and fiscal 2028 growth of approximately 70%, and describes that number as supply-constrained rather than demand-limited (Broadcom Q3 FY2026 call, 2 September 2026; NVIDIA Q2 FY2027 call, 26 August 2026).
The structural driver is buyer breadth. NVIDIA's three carrying cells are hyperscale ($49bn in the quarter, up 13% sequentially), the block management calls ACIE covering neoclouds, sovereigns, enterprises and industrial ($40bn, up 138% year on year, guided toward roughly half of data-centre revenue), and networking (record quarter, up 18% sequentially, Spectrum-X Ethernet up 2.6 times). Each is a separate demand pool with a separate failure mode. Broadcom's carrying cells are four named programmes: Google TPU, Anthropic, OpenAI's Jalapeño and Meta's MTIA, expressed by management as gigawatt commitments rather than as contracts. The one genuinely uncorrelated engine in this duel belongs to Broadcom, and it is not silicon: infrastructure software at $8.8bn a quarter, 94% gross margin and 15% ARR growth.
The as-reported three-year averages show a 21.9-point EBIT margin gap. Almost all of it is accounting, not operations. Broadcom's cost of revenue carries VMware purchase-accounting amortisation worth 9.1%, 18.0% and 12.6% of revenue in fiscal 2023 to 2025 against NVIDIA's 0.2%; strip it and Broadcom's gross margin averages 75.3% against NVIDIA's 72.9%, so Broadcom is 2.4 points ahead rather than 6.3 points behind, and the EBIT gap narrows to 8.7 points. Strip equity compensation as well, which runs at 9.7% of revenue at Broadcom against 4.1% at NVIDIA on the same three-year basis, and the two run at 60.0% and 63.1% operating margin. The cost engines are effectively tied. Neither company wins this duel on efficiency.
What separates them forward is pass-through, not cost. Both face the same memory-price shock. NVIDIA has announced executed price increases effective fiscal Q1 2028 and guides a 71% to 72% gross margin trough recovering to 72% to 73%. Broadcom is absorbing the same shock as mix: gross margin guided to approximately 73% for fiscal Q4 2026 against 78% a year earlier, attributed by its CFO to the increasing mix of XPUs with their increasing memory content. One company reprices and the other dilutes, because one sets list price to a queue of allocated buyers and the other negotiates inside a customer's co-development contract.
Broadcom's one exposure is design ownership. Fifty-six percent of revenue is AI silicon built to four customers' architectures, sold through a channel in which one distributor accounts for 42% of net revenue, up from 29% a year earlier, with the top five end customers at roughly 45% (AVGO fiscal 2026 Q2 10-Q). Broadcom owns none of those designs. The customer can re-tender the next generation, and press reporting from 20 August 2026 describes Marvell winning Google TPU design work and AMD reportedly engaged on a later generation, with MediaTek already holding TPU v8. Management did not address multi-sourcing on the 2 September 2026 call. This trigger is a single engineering decision at a single customer and can land in any quarter.
NVIDIA's one exposure is that it is financing its own demand. Non-marketable equity securities rose from $3.4bn to $22.3bn during fiscal 2026, management expects AI labs leveraging NVIDIA's balance sheet to contribute roughly a quarter of next year's business, and the balance sheet carries $95.2bn of inventory purchase and long-term supply and capacity obligations, flagged as a critical audit matter in the fiscal 2026 10-K, against $32bn of inventory at fiscal Q2 2027. If frontier-lab funding tightens, NVIDIA impairs the equity and strands the inventory in the same quarter. It has already demonstrated the mechanism at smaller scale with a $4.5bn H20 write-down in fiscal Q1 2026.
Broadcom breaks first because its failure mode is idiosyncratic, customer-specific and by external accounts already in motion, while NVIDIA's requires the entire AI capital-expenditure cycle to turn. NVIDIA breaks deeper because when that happens it absorbs the loss on three lines at once: revenue, inventory and investments.
These are the cells that carry roughly 80% of the combined economics. Every score is argued in the tabs below, and the scores are not summed, because adding a software cell to a China cell manufactures precision that does not exist.
| Cell (product and region, end-demand basis) | AVGO | NVDA | Why, one clause, sourced |
|---|---|---|---|
| AI accelerators, US hyperscalers and AI labs | 4 | 5 | AVGO: six XPU programmes and roughly 60% of custom-ASIC design-partner share (Counterpoint, 26 Jan 2026, estimate), but the customer owns the design and is re-tendering it. NVDA: $89.0bn data-centre quarter with purchase orders from every major hyperscaler, AI cloud and system OEM before Vera Rubin reached volume (Q2 FY2027 call). |
| AI and data-centre networking, US hyperscalers and AI labs | 4 | 4 | AVGO: merchant-silicon leader, first to market with the 100Tb Tomahawk 6 and Tomahawk 7 at 200Tbps taped out, plus its own EML, VCSEL and laser fabs (Q3 FY2026 call; FY2025 10-K). NVDA: number two in AI back-end Ethernet switching (Dell'Oro, 3 Sep 2026, estimate), Spectrum-X up 2.6 times, and NVLink uncontested in scale-up. A genuine split by layer. |
| AI data centre, sovereign, neocloud and enterprise | 1 | 5 | AVGO: these buyers do not design custom silicon, so Broadcom reaches them only through merchant networking components sold to someone else's system. NVDA: ACIE revenue of $40.0bn, up 138% year on year, guided toward roughly half of data-centre revenue (Q2 FY2027 call). |
| AI data centre, China | 3 | 1 | AVGO: $11.2bn of fiscal 2025 revenue on a ship-to basis, largely non-AI, with no export-control event found in the sweep window (FY2025 10-K). NVDA: China revenue fell from $25.0bn to $19.7bn between fiscal 2025 and 2026, and management assumes no China data-centre compute revenue in its outlook (FY2026 10-K; Q2 FY2027 call). |
| Legacy and consumer silicon, global | 2 | 4 | AVGO: non-AI semiconductors of $4.2bn, up 5% year on year and flat sequentially, held by legacy sockets in broadband, wireless and server storage (Q3 FY2026 call). NVDA: Graphics $7.9bn, up 46%, riding the same Blackwell architecture as the data-centre line (FY2027 Q2 10-Q). |
| Infrastructure software, Americas and EMEA enterprise | 5 | 0 | AVGO: $8.8bn a quarter at 94% gross margin and roughly 84% operating margin with ARR up 15%, and pricing power strong enough to attract an EU antitrust complaint from CISPE on 23 Mar 2026. NVDA: no reported software revenue line; CUDA is a stronger lock-in but is monetised inside the hardware price, so it produces nothing that survives a hardware pause. |
Scores are anchored to the exhibits in the three tabs and are per cell, never summed. Geographic basis caveat: Broadcom reports revenue by ship-to or delivery location, with title for most products transferring in Penang, Malaysia; NVIDIA reports by customer headquarters. Both state that a substantial share of Asian-billed revenue serves US and European end demand, so the cells above are scored on end demand rather than on the reported regional split.
Before any number: the two companies report geography on opposite bases, and the difference is not cosmetic. Broadcom discloses revenue by ship-to or delivery location, and title for most products transfers in Penang, Malaysia, which pushes revenue into Asia Pacific. NVIDIA discloses revenue by customer headquarters, recast in fiscal Q3 2026 from a prior basis, which pushes revenue into the United States. Both companies say so themselves: Broadcom writes that a substantial portion of product shipped to China ends up in devices its end customers sell in the United States and Europe, and NVIDIA estimates that 76% of data-centre revenue from Taiwan-headquartered customers was attributed to end customers in the United States and Europe. The reported regional splits below are therefore shown as reported with the basis labelled, and every judgement in this page is made on end demand, where the two look nearly identical: US hyperscalers and AI labs.
Segment labels are normalised as follows. Broadcom's AI semiconductor revenue, disclosed on calls inside the semiconductor solutions segment, maps to NVIDIA's Data Center inside Compute and Networking. Broadcom's non-AI semiconductors map to NVIDIA's Graphics plus Automotive plus OEM. Broadcom's infrastructure software segment has no NVIDIA counterpart and is marked ND on that side rather than forced into a comparison.
| Normalised cell | AVGO rev, % of total | AVGO y/y | NVDA rev, % of total | NVDA y/y | Leader and why |
|---|---|---|---|---|---|
| AI data-centre silicon and systems | $16.7bn, 56% | +221% | $89.0bn, 92% | +117% | NVDA on scale, 5.3 times larger; AVGO on rate of change |
| of which accelerators | approx. $12.2bn (inferred: 73% of AI revenue) | units up 3.5x | ND, not split | ND | NVDA, with AVGO closing |
| of which AI networking | approx. $4.5bn (inferred: 27% of AI revenue) | over +150% | record quarter, +18% q/q; $8.2bn in Q3 FY2026 | +162% in Q3 FY2026 | NVDA on systems revenue, AVGO on merchant switch silicon |
| Non-AI and consumer silicon | $4.2bn, 14% | +5% | $7.9bn, 8% | +46% | NVDA, whose legacy line still grows; AVGO's is flat |
| Infrastructure software | $8.8bn, 30%, ARR +15% | +29% | ND, no software segment | ND | AVGO, uncontested |
| Total | $29.6bn | +86% | $96.2bn | +106% | NVDA |
Broadcom figures: fiscal Q3 2026, quarter ended 2 August 2026 (Q3 FY2026 earnings call, 2 September 2026, and Form 10-Q, accession 0001730168-26-000080). NVIDIA figures: fiscal Q2 2027, quarter ended 26 July 2026 (Q2 FY2027 earnings call, 26 August 2026, and 10-Q). Broadcom's XPU and networking split is management commentary on the call only and appears in no filing, hence inferred.
Accelerators. Broadcom grows by adding named programmes, not by taking share in an open market: six XPU customers at Q3 FY2026, with Google's TPU Ironwood v7 and v8i, Anthropic, OpenAI's Jalapeño and Meta's MTIA named on the call. Management expressed the forward book in gigawatts rather than revenue: Anthropic at 1GW in 2026, 5GW in 2027 and line of sight to an incremental 10GW in 2028; OpenAI at 1.3GW in 2027 and over 5GW in 2028; Meta at roughly 3GW through 2028. NVIDIA grows by selling one platform to every buyer at once, with hyperscale at $49bn and ACIE at $40bn in the same quarter. The cost of growing differs by an order of magnitude: NVIDIA carries $95.2bn of purchase and capacity obligations to hold its position, while Broadcom spent $623m of capital expenditure in all of fiscal 2025.
Networking. Broadcom leads the merchant silicon layer: first to market with the 100Tb Tomahawk 6, Tomahawk 7 at 200Tbps taped out, leadership in every generation of PCI Express switching, and a vertically integrated optical component line of EMLs, VCSELs and CW lasers made in its own Fort Collins and Breinigsville fabs. NVIDIA sells systems rather than chips and is winning at that level: Dell'Oro ranks it number two in AI back-end Ethernet switching in both the first and second quarters of 2026, behind Celestica, with Spectrum-X up 2.6 times year on year. Note the layer confusion before drawing a conclusion: Celestica's number-one position is built on merchant silicon, much of it Broadcom's, so the two are not always competing for the same dollar.
Software and legacy silicon. Broadcom's software cell grew 29% to $8.8bn at 94% gross margin and roughly 84% operating margin on the VMware Cloud Foundation subscription conversion. ARR growth of 15% is the honest underlying rate; the gap to 29% comes from upfront licence recognition on non-terminable contracts, and the fiscal 2026 Q2 10-Q discloses $1,964m of upfront licence revenue reclassified into products in the quarter. Broadcom's legacy silicon is flat at 5% growth. NVIDIA's legacy cell, Graphics, grew 46% because gaming GPUs are cut from the same Blackwell silicon as the data-centre product.
| Broadcom | NVIDIA | |
|---|---|---|
| Channel | 56% of net revenue through distributors; one distributor is 42% of net revenue, up from 29%; top five end customers approximately 45%, up from 40% (FY2026 Q2 10-Q) | Direct to cloud providers, ODMs, OEMs and system integrators; largest direct customer 22% of FY2026 revenue, falling to 16% in FY2027 Q2 (FY2026 10-K; FY2027 Q2 10-Q) |
| Pricing mechanism | Programme-level co-development contracts with multi-year pricing schedules; remaining performance obligations of $33.3bn at fiscal 2025 year end (FY2025 10-K) | List-priced platform; management states executed price increases take effect in fiscal Q1 2028 to recover memory cost (Q2 FY2027 call) |
| Disclosed price and volume split of growth | ND | ND |
| Pricing power actually exercised | Semiconductor segment gross margin of approximately 76% in Q3 FY2026 despite rising XPU mix; separately, VMware price increases severe enough to draw an EU antitrust complaint | Raising prices into a memory shock while still guiding fiscal 2028 gross margin of 72% to 73% |
| Broadcom | NVIDIA | |
|---|---|---|
| Foundry | Approximately 95% of contract-manufactured wafers from TSMC in fiscal 2025, flagged by the company itself as a prioritisation risk (FY2025 10-K risk factors) | TSMC and Samsung (FY2026 10-K) |
| Memory | HBM sourced externally; Samsung memorandum of understanding, July 2026, covering HBM and 2nm foundry, not a firm order | SK hynix, Micron and Samsung all certified for HBM4 on Vera Rubin, June 2026 |
| Own fabrication | Yes: FBAR filters at Fort Collins, InP and GaAs lasers at Breinigsville and Singapore, both named as sole sources | None; CoWoS packaging via TSMC; assembly via Hon Hai, Wistron and Fabrinet |
| Inventory posture | $4.5bn at Q3 FY2026, roughly 3.6% of fiscal 2025 revenue | $21.4bn at fiscal 2026 year end and $32bn at Q2 FY2027, plus $95.2bn of purchase and capacity obligations |
| Disclosed incident | TSMC capacity named as the binding constraint, March 2026; substrate and HBM availability named on the Q3 FY2026 call | $4.5bn H20 write-down in fiscal Q1 2026 on loss of export licences |
Third-party estimates, each flagged as such. ASIC-based systems were 27.8% of 2026 AI server shipments against 69.7% GPU-based, the highest ASIC share since 2023 (TrendForce, 20 January 2026). Broadcom holds roughly 60% of custom-ASIC design-partner share for 2027 against Marvell at roughly 8%, while Google's TPU share of ASIC programme volumes falls from 64% in 2024 to 52% in 2027 as Meta's MTIA and Microsoft's Maia ramp (Counterpoint, 26 January 2026). In networking, Dell'Oro ranks AI back-end Ethernet switching in the second quarter of 2026 as Celestica first, NVIDIA second, Arista third and Cisco fourth and gaining (3 September 2026). Two competitive events in the window matter more than the share tables: AMD won a six-gigawatt multi-generation GPU commitment from OpenAI on 6 October 2025 with a warrant for up to 160m AMD shares, and press reporting on 20 August 2026 describes Marvell winning Google TPU design work with a warrant reported at up to $12.2bn, with AMD reportedly engaged on a later generation and MediaTek already holding TPU v8.
Read as share-taker or share-donor: in accelerators Broadcom is taking share from merchant GPU in aggregate while donating share inside its own largest programme. NVIDIA is donating aggregate accelerator share while taking networking share. One conflict is unresolved and worth naming: Broadcom management made no acknowledgement of Google multi-sourcing on the 2 September 2026 call, while press reporting from 20 August 2026 describes it in detail. Neither Broadcom nor Alphabet has confirmed it.
| Cell | The one risk that does the most damage | More exposed | Why |
|---|---|---|---|
| AI accelerators | A named customer redesigns to a rival partner or brings the next generation in house | Broadcom | Six customers carry 56% of total revenue, one distributor is 42% of net revenue, and Google multi-sourcing is already reported |
| AI networking | Rack-scale integration eliminates the discrete merchant switch socket | Broadcom | NVLink and Spectrum-X ship inside the rack; Broadcom must win a socket inside someone else's design |
| Legacy and consumer silicon | Cycle downturn | NVIDIA | Graphics rides the same Blackwell supply and the same AI cycle it is meant to diversify against |
| Software | Regulatory or renewal shock to VMware pricing | Broadcom | CISPE EU antitrust complaint of 23 March 2026 over VCF licensing terms |
| Whole book | Customer credit, because both are financing their own demand | Both, differently | Broadcom provides residual value guarantees inside an Apollo and Blackstone SPV; NVIDIA holds $22.3bn of non-marketable equity in the labs that buy its chips |
Design-partner incumbency. Durability: Medium. The moat is being the fastest and most reliable custom-accelerator partner, not owning the architecture. Management's proof is delivery: fastest time to market for TPUs from definition to production without respins, plus SerDes, chip-to-chip interconnect, HBM and SRAM integration and advanced packaging IP. External corroboration: roughly 60% of custom-ASIC design-partner share for 2027 against Marvell at 8% (Counterpoint, estimate). It is Medium rather than High because the customer owns the architecture and can re-tender it, and by press account has begun to.
Physical-layer integration. Durability: High. Broadcom owns fabs for FBAR filters and for the InP and GaAs lasers behind optical interconnect, and names Fort Collins and Breinigsville as sole sources. This is a years-not-money asset.
Software annuity. Durability: High operationally, Medium legally. $8.8bn a quarter at 94% gross margin with a mission-critical private-cloud installed base and high switching cost, but with a live EU antitrust complaint against its licensing terms.
Architecture plus CUDA. Durability: High. The moat is that customers write software against the platform and buy a rack that must be taken whole. Proof is behavioural rather than rhetorical: purchase orders for Vera Rubin from every major hyperscaler, AI cloud and system OEM before volume shipment, and revenue opportunity per gigawatt rising from $18bn on Hopper to $25bn on Grace Blackwell to $40bn on Vera Rubin as more of the rack becomes NVIDIA content. A well-funded rival needs years, not money: AMD has held a six-gigawatt OpenAI commitment since October 2025 without dislodging the platform.
System-level vertical integration. Durability: High. GPU, CPU, DPU, NVLink, Ethernet and InfiniBand and software sold as one product. Grace CPU trailing twelve-month revenue exceeded $5bn, and Vera is guided to more than double CPU revenue in fiscal 2028.
Software annuity: ND. No separately reported software revenue. CUDA is the stronger lock-in but produces no revenue line that survives a hardware pause.
The two concentrations are moving in opposite directions and that is the single most legible fact in this section. Broadcom's largest channel customer, a distributor, went from 29% to 42% of net revenue in a year, and its top five end customers from roughly 40% to 45%. NVIDIA's largest direct customer went from 22% of fiscal 2026 revenue to 16% in fiscal Q2 2027, with the top three at 16%, 15% and 13% across the half, as ACIE broadened the base toward what management expects will be roughly half the data-centre business.
Access quality differs in the same direction. Broadcom's own fiscal 2026 Q2 risk factors state that its top customers, including AI customers, have made greater demands on pricing and contractual terms, seeking to lease custom accelerators or requiring Broadcom to purchase and lease full AI racks, together with alternative financing arrangements, and that such arrangements have imposed and may impose financial obligations including backstops or guarantees. That is a company disclosing, in its own filing, that its buyers are writing the commercial structure. NVIDIA's equivalent disclosure runs the other way: it has been asked to offer financing arrangements to support customer build-outs and states it has not entered into any.
Broadcom sources approximately 95% of its contract-manufactured wafers from TSMC, a concentration it flags itself as a prioritisation risk, and it named TSMC capacity as the binding constraint in March 2026 and substrate and HBM availability on the September 2026 call. Its Samsung arrangement covering HBM and 2nm foundry is a memorandum of understanding, not a firm order. NVIDIA runs TSMC and Samsung at the foundry and has certified all three memory makers, SK hynix, Micron and Samsung, for HBM4 on Vera Rubin. Dual-sourcing coverage reads Low at the foundry and Medium on memory for Broadcom, against Medium and High for NVIDIA.
Pass-through is where the gap becomes financial. NVIDIA is Strong: it announced executed price increases effective fiscal Q1 2028 and still guides fiscal 2028 gross margin of 72% to 73% against what management called extreme pricing conditions in memory. Broadcom is Partial: it guided fiscal Q4 2026 gross margin down to approximately 73% from 78% a year earlier and attributed the decline to the increasing mix of XPUs with their increasing memory content, which is absorbing the shock rather than repricing it.
| Cell | Route control | Pocket price | Continuity | Outcome, share and margin | Confirming KPI |
|---|---|---|---|---|---|
| AI accelerators into US hyperscalers and AI labs | AVGO indirect, distributor at 42% of revenue, into four engineering organisations. NVDA direct into hundreds of buyers plus an ODM channel. | AVGO below: management's pitch is that Jalapeño runs OpenAI workloads at half the cost of a GPU. NVDA above, and raising. | AVGO constrained by TSMC. NVDA constrained but dual-sourced. | AVGO share up, margin down, because mix dilution is disclosed and guided. NVDA share down slightly in units, margin roughly flat because price recovers memory cost. | AVGO consolidated gross margin against the 73% Q4 guide |
| AI networking into the same buyers | AVGO sells merchant silicon to ODMs and switch vendors, including to Celestica, which Dell'Oro ranks first in AI back-end Ethernet. NVDA sells the network inside the rack. | NVDA above, captured inside a system price the customer cannot line-item. | Parity. | AVGO share flat in silicon, revenue per gigawatt down. NVDA share up and margin up. | Dell'Oro quarterly AI back-end switch vendor rank |
| Infrastructure software into global enterprise (AVGO) against consumer and workstation graphics (NVDA) | AVGO direct enterprise renewal motion at 94% gross margin. NVDA add-in-board partners and OEMs, Graphics operating margin 45.6% in H1 FY2027. | AVGO has raised effective prices hard enough to attract an EU antitrust complaint, which is the strongest available external evidence of pricing power exercised. | Parity. | AVGO share flat, margin up, with a regulatory tail. NVDA share up, margin up, but correlated to the same cycle as its data-centre business. | VMware ARR growth; European Commission action |
This cell has no true counterpart on the NVIDIA side, which is itself the finding, and it is the one place where the answer runs Broadcom's way.
Who owns the design. Impact: Major. Broadcom's customers own the architecture and specify the chip; Broadcom supplies IP and execution. NVIDIA owns the architecture and the software written against it. For Broadcom to close this it would have to become an architecture owner, which means a decade of ecosystem building it has never attempted. Not closable in eighteen months with money, which is what makes it a real moat difference rather than a lead.
Direction of customer concentration. Impact: Moderate. Broadcom's largest channel customer went from 29% to 42% of revenue in a year and its top five from 40% to 45%, while NVIDIA's largest direct customer fell from 22% to 16%. Broadcom closes this by bringing XPU customers five and six to material volume. It says it has six; revenue is visibly concentrated in four. Closable in two to three years if the programmes land.
Cost pass-through. Impact: Moderate. Broadcom guided gross margin from 78% to approximately 73% and attributed it to XPU memory content; NVIDIA announced price increases effective fiscal Q1 2028 and guides 72% to 73% after a trough. Broadcom closes this only by renegotiating programme pricing schedules, which realistically happens at the next generation rather than inside this one.
Both companies present costs by function and report in US dollars, so no by-nature reconstruction or currency conversion is required and all five ratios are defensible on both sides. Two mapping notes govern the tables. Broadcom's total cost of revenue includes amortisation of acquisition-related intangible assets of $1,853m, $6,023m and $6,031m in fiscal 2023 to 2025, and its reported selling, general and administrative line excludes the separately presented amortisation and restructuring lines. NVIDIA's acquisition-related intangible amortisation is immaterial and sits in unallocated acquisition-related and other costs, at $527m in fiscal 2026, so its cost of revenue is close to a pure product cost. Fiscal years are offset: Broadcom's ends in early November and NVIDIA's in late January, a gap of roughly twelve weeks, which is inside one quarter. Broadcom fiscal 2023 to 2025 is paired with NVIDIA fiscal 2024 to 2026.
| Percent of sales, 3-year average | Broadcom | NVIDIA | Gap | What drives it |
|---|---|---|---|---|
| COGS | 33.4 | 27.1 | 6.3 to NVDA | Broadcom's cost of revenue carries VMware purchase-accounting amortisation |
| R&D | 16.7 | 10.9 | 5.8 to NVDA | Broadcom funds many product lines and carries heavy equity comp; NVIDIA amortises one architecture over 3.5 times the revenue |
| SG&A | 6.9 | 3.1 | 3.8 to NVDA | Broadcom carries an enterprise-software sales and renewal organisation; NVIDIA sells direct to a handful of accounts |
| Gross margin | 66.6 | 72.9 | 6.3 to NVDA | Same purchase-accounting cause, inverted |
| EBIT margin | 37.1 | 59.0 | 21.9 to NVDA | The sum of the above; see the adjusted table below before reading anything into it |
| COGS % | R&D % | SG&A % | Gross margin % | EBIT margin % | |
|---|---|---|---|---|---|
| Broadcom FY2023 | 31.1 | 14.7 | 4.4 | 68.9 | 45.2 |
| Broadcom FY2024 | 37.0 | 18.1 | 9.6 | 63.0 | 26.1 |
| Broadcom FY2025 | 32.2 | 17.2 | 6.6 | 67.8 | 39.9 |
| NVIDIA FY2024 | 27.3 | 14.2 | 4.4 | 72.7 | 54.1 |
| NVIDIA FY2025 | 25.0 | 9.9 | 2.7 | 75.0 | 62.4 |
| NVIDIA FY2026 | 28.9 | 8.6 | 2.1 | 71.1 | 60.4 |
Broadcom fiscal 2024 is distorted by the first full year of VMware purchase accounting; NVIDIA fiscal 2026 carries a $4.5bn H20 export-control charge in cost of revenue, worth roughly 2.1 points of gross margin.
| Percent of sales, 3-year average | Broadcom | NVIDIA | Gap |
|---|---|---|---|
| Gross margin, as reported | 66.6 | 72.9 | 6.3 to NVDA |
| Gross margin, excluding acquisition-intangible amortisation | 75.3 | 72.9 | 2.4 to AVGO |
| EBIT margin, as reported | 37.1 | 59.0 | 21.9 to NVDA |
| EBIT margin, excluding acquisition-intangible amortisation | 50.3 | 58.8 | 8.5 to NVDA |
| EBIT margin, excluding amortisation and stock compensation | 60.0 | 63.1 | 3.1 to NVDA |
| Memo: stock compensation | 9.7 | 4.1 | 5.6 higher at AVGO |
| Memo: acquisition-intangible amortisation | 13.2 | 0.2 | 13.0 higher at AVGO |
Amortisation added back from both cost of revenue and operating expenses: Broadcom $3,247m, $9,267m and $8,062m for fiscal 2023 to 2025 per the 10-K segment note; NVIDIA's equivalent is immaterial. Stock compensation from the cash-flow statements: Broadcom $2,171m, $5,741m and $7,568m; NVIDIA $3,549m, $4,737m and $6,386m. Both are real costs to shareholders. The exhibit does not argue they should be ignored; it argues that they are differences in how the two businesses were assembled and paid for rather than differences in how efficiently they run.
| Broadcom Q3 FY2026, quarter to 2 Aug 2026 | NVIDIA Q2 FY2027, quarter to 26 Jul 2026 | |
|---|---|---|
| Revenue | $29,591m, up 86% | $96,221m, up 106% |
| Gross margin | 69.1% GAAP, 75% non-GAAP | 75.0%, GAAP equals non-GAAP |
| Operating margin | 53.9% GAAP, 67.9% non-GAAP | 66.2% GAAP |
| Next-quarter gross margin guide | approximately 73%, down from 78% a year earlier | 74% plus or minus 50bp, trough of 71% to 72% expected in Q4 |
| Broadcom FY2025 | NVIDIA FY2026 | |
|---|---|---|
| Property, plant and equipment, net | $2,530m, 4.0% of revenue | $10,383m, 4.8% of revenue |
| Capital expenditure | $623m, 1.0% of revenue | $6,042m including intangibles, 2.8% of revenue |
| Inventory | $2,270m, 3.6% of revenue | $21,403m, 9.9% of revenue |
| Operating cash flow | $27,537m, 43.1% of revenue | $102,718m, 47.6% of revenue |
| Committed forward purchases | Remaining performance obligations $33.3bn (revenue side) | Inventory purchase and long-term supply and capacity obligations $95.2bn (cost side) |
The most persistent difference across all three years is not gross margin and not EBIT margin, both of which move once you adjust them. It is operating expense intensity: Broadcom spends 16.7% of revenue on research and development and 6.9% on selling, general and administrative against NVIDIA's 10.9% and 3.1%, and the gap is present in every one of the three years in both lines. The mechanism comes straight from Step 1. Broadcom funds five semiconductor end markets plus an enterprise software portfolio, each with its own engineering roadmap and its own renewal sales organisation, and it pays a materially larger share of that cost in equity, at 9.7% of revenue against 4.1%. NVIDIA funds one architecture and sells it to a few hundred accounts, so every incremental dollar of revenue meets an expense base that barely moves.
Does this confirm or contradict the power map? It confirms it, but not in the way the reported numbers suggest. The company with the stronger pricing power should show it in gross margin, and once purchase accounting is removed Broadcom's gross margin is actually the higher of the two, at 75.3% against 72.9%. That is not a contradiction: Broadcom's product-level economics are genuinely excellent, helped by a software segment at 94% gross margin. What Broadcom does not have is NVIDIA's operating leverage, and what it is now losing is the direction of travel, guiding gross margin from 78% to 73% while NVIDIA guides a shallower trough and a recovery funded by price increases it can actually impose. The cost engines are tied; the pricing engines are not. That is why this duel is decided on the power lens rather than on the cost lens, and it is why a reader who anchors on the headline 37% against 59% operating margin has drawn the right conclusion from the wrong evidence.
| Metric | Threshold | By when | If it hits, it favours | Where published |
|---|---|---|---|---|
| Broadcom consolidated gross margin | At or above the guided 73% | Fiscal Q4 2026 result, around December 2026 | Broadcom: XPU pricing is absorbing memory inflation | AVGO earnings release and 10-K |
| NVIDIA gross margin at the guided trough | At or above 71% | Fiscal Q4 2027 result, around February 2027 | NVIDIA: pass-through works. Below 71% is flip condition three firing | NVDA earnings release and 10-K |
| Broadcom largest-distributor share of net revenue | At or below 45% | Each 10-Q through fiscal 2027 | Broadcom: concentration stabilising. Above 45% favours NVIDIA | AVGO 10-Q, MD&A net revenue discussion |
| Disclosed loss of a Broadcom XPU programme to a rival design partner | Zero | By end of fiscal 2027, November 2027 | Broadcom: design-partner moat intact. Any disclosed loss favours NVIDIA | AVGO earnings calls; customer disclosure |
| NVIDIA ACIE share of data-centre revenue | At or above 40% | Each quarter through fiscal 2028 | NVIDIA: diversification away from hyperscalers is real | NVDA earnings call CFO commentary |