Semiconductors: processors and data-center AI compute. INTC / NVDA. Run 23 September 2026. Built from the research folder: Intel 10-Ks FY2021 and FY2023-FY2025 and 10-Q Q2 2026 (quarter ended 27 Jun 2026); NVIDIA 10-Ks FY2022 and FY2024-FY2026, 10-Q Q2 FY2027 (quarter ended 26 Jul 2026) and the Q2 FY2027 earnings call (26 Aug 2026); web sources for market share and events as cited. Events swept through 23 Sep 2026; most recent checked: Intel, 16 Sep 2026 reported SK hynix talks over Intel's Ohio fab (not finalized); NVIDIA, 10 Sep 2026 reported DOJ antitrust inquiry into the Groq licence deal. US dollars as reported. Fiscal years are offset by about four weeks (Intel ends December, NVIDIA late January), so Intel FY2025 is paired with NVIDIA FY2026. NVIDIA's latest quarter is six times Intel's, so the duel compares ratios, not absolutes. Not a valuation and not a recommendation.
Three NVIDIA cells carry the next five years. AI compute sold as rack-scale systems grows with customer capacity and with content per gigawatt, which management puts at about $18B for Hopper, $25B for Blackwell and $40B for Vera Rubin; its top five customers alone expect to spend nearly $800B of capex in 2026 and $1.3T in 2027 (Q2 FY2027 call). Demand is also widening: AI clouds, industrial and enterprise (ACIE) revenue grew 138% year on year against 102% for hyperscale (Q2 FY2027 10-Q). Networking (Spectrum-X up 2.6x) and CPUs (Grace and Vera above $5B trailing, expected to more than double in FY2028) ride the same racks (Q2 FY2027 call). Intel's 2026 rebound is price in a shortage: client ASPs rose 27% while units fell 8%, and server ASPs rose 48% on units up 9% (Q2 2026 10-Q), while AMD took 7.3 points of x86 server unit share in a year (Mercury Research via HotHardware, Aug 2026). Intel's one structural new engine, external foundry, billed $293M in Q2 2026 (Q2 2026 10-Q).
NVIDIA prices a system against the value of the compute it delivers and buys wafer capacity from TSMC, so its costs flex with volume; the Compute & Networking segment earned 67.3% operating margin in FY2026 and 71.0% in Q2 FY2027 (FY2026 10-K; Q2 FY2027 10-Q). Intel's designer arm earned 25.9% in FY2025, and then its factory, which bills about 98% of its revenue to Intel itself, lost 57.9 cents per dollar (FY2025 10-K, Note 3). The same company wins growth and margin, so the tension in this duel is direction, not level: NVIDIA guides gross margin down to a 71-72% trough in Q4 FY2027 on memory costs before 72-73% in FY2028, while Intel guides Q3 2026 to 41.0% from 34.8% in FY2025 (Q2 FY2027 call; Intel Q2 2026 release, 23 Jul 2026). That narrows a 37-point gap by a few points; it does not close it.
Intel's exposure is cost rigidity under borrowed pricing. Its 2026 margin recovery rests on shortage-era ASPs and on substrate and memory constraints it expects to persist "into next year" (Q2 2026 10-Q); when supply normalizes, AMD and Arm designers buy more TSMC capacity, prices reset as they did in FY2025, and fixed fab cost turns every price cut into lost margin. It starts from a thin cushion: FY2025 operating cash flow of $9.7B against $14.6B of capex (FY2025 10-K), bridged by a $20B equity raise in August 2026 (Intel release, 11 Aug 2026). NVIDIA's exposure is a commitment asymmetry: supply and capacity commitments rose from $119B to $279B in one quarter, plus up to $105B of lease guarantees for OpenAI's Ohio campus, while customers can cancel purchase orders "without penalty" (Q2 FY2027 10-Q; FY2026 10-K). Three direct customers took 44% of first-half revenue and design competing chips. Intel breaks first because it has almost no margin to lose; NVIDIA, with a 60% operating margin, can absorb a slowdown but not a pause that strands $279B of commitments (inferred).
The first four cells carry 96% of the two companies' combined latest-year revenue; the fifth is Intel's $17.8B internal factory, which decides its margins. Neither company discloses a product x geography cross-tab, so the dominant region is named from each company's geographic note (inferred).
| Cell (product x region) | Intel | NVIDIA | Why (one clause, sourced) |
|---|---|---|---|
| AI accelerated compute, US hyperscalers and AI clouds | 1 | 5 | NVIDIA $162.4B, +104% a year, prices raised into a cost shock; Intel "unsuccessful to date in becoming a meaningful participant" (FY2026 10-K; FY2025 10-K). |
| PC client, OEMs and ODMs billed in China and Asia | 3 | 4 | Intel holds 69.7% of x86 client units at a 28.9% segment margin but ceded about 6 points in a year; NVIDIA Gaming grew 24% a year to $16.0B (Mercury via HotHardware; 10-Ks). |
| Data-center networking, global | 1 | 5 | NVIDIA $31.4B, +91% a year, number one in data-center Ethernet switching; Intel's networking no longer reported separately (IDC 1Q26; 10-Ks). |
| Server CPUs, US cloud and enterprise | 2 | 3 | Intel holds 65.5% of units on legacy installed base after FY2025 discounting; NVIDIA's Grace drives Arm's record 13.6% but standalone Vera is unproven (Mercury; 10-K; Q2 FY2027 call). |
| Leading-edge manufacturing, US | 2 | 0 | Intel is the only US leading-edge logic manufacturer but external revenue was $293M in Q2 2026; NVIDIA is fabless by design (FY2025 10-K; Q2 2026 10-Q). |
Scores are anchored to the exhibits in the three tabs and are not summed; the three lenses get their verdicts in the Three Answers above. NVIDIA's PC score of 4 is inferred from Gaming growth outpacing a shrinking PC market; no discrete-GPU share tracker was checked.
Neither filing crosses product with geography, and the geography bases differ: Intel reports revenue by customer billing location, NVIDIA by customer headquarters (changed in Q3 FY2026). The cells below are therefore product cells with a regional overlay. NVIDIA's revenue is 69% US-headquartered customers, 20% Taiwan (76% of whose data-center revenue NVIDIA attributes to US and European end customers) and 9% China (FY2026 10-K). Intel's is 30% US, 24% China, 18% Singapore and 15% Taiwan, largely where PC makers are invoiced (FY2025 10-K).
Normalized labels: Intel CCPG (formerly CCG, with networking folded in from Q1 2025) and NVIDIA Gaming are both "PC client", although Intel sells the processor and NVIDIA the graphics chip. Intel DCAI includes custom ASICs. Intel does not disclose AI accelerator or networking revenue, so those cells read ND on its side.
| Cell | Intel rev (% of total) | Intel growth | NVIDIA rev (% of total) | NVIDIA growth | Margin signal / leader |
|---|---|---|---|---|---|
| AI accelerated compute | ND | ND | $162.4B (75%) | +104%/yr | NVIDIA; Compute & Networking segment margin 67.3% |
| PC client | $32.2B (61%) | -0.1%/yr | $16.0B (7%) | +24%/yr | Split; Intel CCPG margin 28.9%, NVIDIA Graphics segment 40.8% |
| Data-center networking | ND | ND | $31.4B (15%) | +91%/yr | NVIDIA; inside the 67.3% segment |
| Server CPU | $16.9B (32%) | +2.9%/yr | >$5B TTM | ND | Intel by scale; DCAI margin 20.2% FY2025, 35.5% 1H 2026 |
| Foundry | $17.8B, ~98% internal | ND | none | n/a | Intel Foundry margin -57.9% |
| All other | $3.6B (7%) | n/m | $6.2B (3%) | n/m | Mobileye and Altera (to Sep 2025) vs pro visualization, automotive, OEM |
Sources: Intel FY2025 10-K Note 3 (FY2023-FY2025 recast) and Q2 2026 10-Q Note 2; NVIDIA FY2026 10-K revenue by market platform and segment note; Q2 FY2027 call for CPU revenue. Percent of total uses each company's consolidated revenue; Intel's segments sum above 100% before intersegment eliminations.
AI compute (NVIDIA): volume of gigawatts deployed multiplied by revenue per gigawatt, which rises each generation because NVIDIA adds CPU, networking and now Groq inference hardware to the rack (Q2 FY2027 call). Management's preliminary FY2028 view is revenue growth of about 70%, with supply the bottleneck "at least through the end of fiscal year 2028" (Q2 FY2027 call). Growth is organic: the December 2025 Groq agreement bought a licence and staff, not revenue, and the $11.9B Hugging Face deal has not closed (FY2026 10-K; 8-K 3 Sep 2026).
Server CPU (both): this is the only cell where the two meet head on. Intel's DCAI revenue rose 59% in Q2 2026, mostly on ASP (Q2 2026 10-Q), and Intel is co-developing custom x86 data-center CPUs with NVLink for NVIDIA, which invested $5B in Intel at $23.28 a share in December 2025 (NVIDIA release, 18 Sep 2025; closing reported by Quartz). NVIDIA meanwhile sells its own Grace and Vera CPUs and says it sees about $20B of server-CPU demand (Q2 FY2027 call). Intel's fastest-growing cell increasingly lives inside a platform NVIDIA defines.
PC client (both): Intel ships about seven in ten x86 PC processors but AMD's client unit share rose to 30.3% and Arm-based chips to 15.3% (Mercury Research via HotHardware and The Register, Aug 2026). NVIDIA enters PC processors this fall with RTX Spark, an Arm chip co-designed with MediaTek, through Dell, HP, Lenovo and others (NVIDIA release, 31 May 2026).
Both sell on cancellable purchase orders through OEMs, ODMs and cloud providers, and both net rebates and channel incentives from revenue (FY2025 10-K; FY2026 10-K). The difference is who sets the price. NVIDIA is passing a memory cost spike through, with price increases executed for Q1 FY2028, and still guides gross margin to 72-73% (Q2 FY2027 call). Intel raised ASPs in 2026 only while "market demand exceeded our available product supply", and says pricing contributed "to a lesser extent" than mix (Q2 2026 10-Q); in FY2025, with supply adequate, server ASPs fell 4% "driven by a competitive environment" (FY2025 10-K).
| Input | Intel | NVIDIA |
|---|---|---|
| Wafers | Own fabs (Intel 7, 3, 18A); some leading products made "exclusively" at TSMC with no long-term contract | TSMC and Samsung; secured via $279B of supply and capacity commitments |
| Memory and substrates | Industry shortages expected to persist into 2027 | HBM from SK hynix, Micron, Samsung; "extreme pricing conditions" cut the margin guide |
| Packaging and assembly | In-house, plus EMIB/Foveros offered externally | CoWoS; Hon Hai, Wistron, Fabrinet |
| Capex (latest FY) | $14.6B plus $3.0B vendor-financed; 2026 "more than $20B" | $6.0B |
Sources: Intel FY2025 10-K (cash-flow statement, risk factors), Q2 2026 10-Q, Q2 2026 prepared remarks; NVIDIA FY2026 10-K (Business; cash-flow statement), Q2 FY2027 10-Q Note 10 and call.
AI accelerators: GPUs are expected to be 69.7% of 2026 AI-server shipments and custom ASICs 27.8%, the highest ASIC share since 2023 (TrendForce, Jan 2026). Broadcom's AI semiconductor revenue was $16.7B in its latest quarter, up 221%, and AMD's data-center segment $6.7B, against NVIDIA's $89.0B of Data Center revenue (company releases, Aug-Sep 2026). NVIDIA is the incumbent, and custom silicon is the share-taker. x86 CPUs: AMD holds 34.5% of server and 30.3% of client units and 46.2% of server revenue in Q1 2026; Intel is the share donor (Mercury Research). Foundry: TSMC took 72.5% of top-10 foundry revenue in 2Q 2026; Intel Foundry is not ranked (TrendForce, Sep 2026).
AI compute (NVIDIA exposed): hyperscale was 55% of Q2 FY2027 Data Center revenue, and those buyers are named competitors building their own chips (Q2 FY2027 10-Q; FY2026 10-K). Inference is the natural first workload to move, and every ASIC generation that clears "good enough" moves hyperscale volume off NVIDIA at the margin (inferred). PC and server CPUs (Intel exposed): each design win AMD or an Arm licensee takes is usually built at TSMC, so it removes Intel product revenue and fab loading at once, raising unit cost on what remains (inferred). Foundry (Intel exposed): 14A proceeds without a committed external customer; Intel said in January 2026 it expected customer decisions from the second half of 2026 (Tom's Hardware, Jan 2026).
x86 licence duopoly with AMD and the Windows and enterprise installed base: High durability in PCs and on-premise servers, Medium and eroding in cloud, where Arm reached 13.6% of server units (Mercury, Q2 2026). The only US company running leading-edge logic R&D and volume manufacturing, with the US government holding about 10%: Medium, a policy asset that does not yet earn a return (FY2025 10-K; 8-K 25 Aug 2025).
CUDA plus "hundreds of" libraries and SDKs, built on more than $76.7B of cumulative R&D: High, since a rival needs years of developer adoption, not money (FY2026 10-K). Rack-scale co-design of GPU, CPU, NVLink and networking on an annual cadence, now in a third NVL72 generation: High (Q2 FY2027 call). Scale in networking, number one in data-center Ethernet switching: Medium-High (IDC 1Q26).
Both depend on a few buyers. Intel's three largest customers took 43% of FY2025 revenue (19%, 12%, 12%) (FY2025 10-K); NVIDIA's three largest direct customers took 44% of first-half FY2027 revenue (16%, 15%, 13%), down from one customer alone at 22% in FY2026 (Q2 FY2027 10-Q; FY2026 10-K). The difference is in exercised pricing power: NVIDIA's buyers accept price increases while supply is short through FY2028; Intel's buyers are OEMs who can switch to AMD on the same socket economics, and did so as supply allowed. NVIDIA's weakness is that its biggest buyers are also named competitors.
Intel's integration gives it continuity on its own nodes but not on its best products, which rely on TSMC without a long-term contract, and its Intel 7 output in Israel is flagged as uninsured war exposure (FY2025 10-K; Q2 2026 10-Q). NVIDIA depends on three memory makers and on TSMC's CoWoS, and memory pricing has just cost it about three points of guided gross margin; its answer is scale, with $279B of commitments that make it the priority buyer (Q2 FY2027 10-Q and call). Pass-through is partial for both: Intel raised ASPs "in part to offset higher input costs", NVIDIA executed price increases. NVIDIA's floor after pass-through is a 71-72% gross margin; Intel's ceiling in a shortage is about 41%.
| Cell | Route control | Pocket price | Continuity | Outcome (share / margin) | Confirming KPI |
|---|---|---|---|---|---|
| AI compute | NVIDIA direct and via ODMs/CSPs; Intel absent | NVIDIA far above ASIC alternatives, holding | NVIDIA via commitments | NVIDIA share flat to slightly down vs ASICs; margin down 2-3 pts in FY2028, then stable | NVIDIA Hyperscale growth vs Broadcom AI semiconductor growth |
| PC client | Intel via OEMs; NVIDIA via add-in boards and now Arm PCs | Intel at or above AMD while supply is short | Intel own fabs; NVIDIA TSMC | Intel share down, margin up near term then down; NVIDIA share up | Mercury client share; Intel CCPG margin |
| Server CPU | Intel via OEMs/CSPs; NVIDIA inside its own racks | Intel shortage premium (ASP +48%); NVIDIA bundled | Intel constrained on Intel 7 and 3 | Intel share down; NVIDIA share up from a small base | NVIDIA CPU revenue vs Intel DCAI revenue |
The tension is in the third row: Intel's fastest-growing cell, DCAI at +59% in Q2 2026, is exactly where it has the least pricing power once supply normalizes and where NVIDIA controls the attach point through NVLink (Q2 2026 10-Q; Q2 FY2027 call).
Three causes explain most of the difference. First, platform ownership: CUDA against x86 in a maturing market (Major); Intel cannot close it, as the failure of its AI accelerators so far shows (FY2025 10-K). Second, asset-light versus fixed-cost manufacturing (Major); Intel can close it only by filling its fabs with outside customers, a 14A story whose volume ramp Intel itself dates to 2028 (Q2 2026 prepared remarks). Third, customer price-setting (Moderate); this flips only if hyperscaler ASICs reach parity. None of the three can be closed with money in 18 months.
Both report by function under US GAAP, so the mapping is clean. Intel's "marketing, general and administrative" is treated as SG&A. Intel itemizes restructuring ($7.0B in FY2024, $2.2B in FY2025); it is excluded from the functional lines and included in reported EBIT, with the ex-restructuring figure shown. Intel's cost of sales carries fab depreciation plus $3.3B (FY2024) and $950M (FY2025) of impairments and accelerated depreciation; NVIDIA's carries TSMC's margin and a $4.5B H20 export charge in FY2026 (FY2025 10-K; FY2026 10-K). Fiscal years are offset by about four weeks.
| % of sales, 3y avg | Intel FY23-25 | NVIDIA FY24-26 | Gap | What drives it |
|---|---|---|---|---|
| COGS | 64.2 | 27.1 | 37.1 | System pricing and variable wafer cost vs fixed fab cost on under-loaded nodes |
| R&D | 29.0 | 10.9 | 18.1 | Scale: NVIDIA spends more dollars on a far larger base |
| SG&A | 9.8 | 3.1 | 6.7 | Few, very large system orders vs a broad OEM channel |
| Gross margin | 35.8 | 72.9 | 37.1 | Pricing power and asset-light model |
| EBIT margin | -8.7 (ex-restr. -3.0) | 59.0 | 67.7 | All of the above plus Intel Foundry losses |
| Year | COGS | R&D | SG&A | Gross margin | EBIT margin |
|---|---|---|---|---|---|
| Intel FY2023 | 60.0 | 29.6 | 10.4 | 40.0 | 0.2 |
| Intel FY2024 | 67.3 | 31.2 | 10.4 | 32.7 | -22.0 |
| Intel FY2025 | 65.2 | 26.1 | 8.7 | 34.8 | -4.2 |
| NVIDIA FY2024 | 27.3 | 14.2 | 4.4 | 72.7 | 54.1 |
| NVIDIA FY2025 | 25.0 | 9.9 | 2.7 | 75.0 | 62.4 |
| NVIDIA FY2026 | 28.9 | 8.6 | 2.1 | 71.1 | 60.4 |
Raw values (US$M): Intel revenue 54,228 / 53,101 / 52,853; COGS 32,517 / 35,756 / 34,478; R&D 16,046 / 16,546 / 13,774; MG&A 5,634 / 5,507 / 4,624; EBIT 93 / (11,678) / (2,214). NVIDIA revenue 60,922 / 130,497 / 215,938; COGS 16,621 / 32,639 / 62,475; R&D 8,675 / 12,914 / 18,497; SG&A 2,654 / 3,491 / 4,579; EBIT 32,972 / 81,453 / 130,387.
Gross margin carries the most persistent gap: 32.7, 42.3 and 36.3 points across the three paired years, never below 30. The mechanism comes from both earlier lenses: NVIDIA prices a system against the value of AI compute and buys wafers as it needs them, while Intel prices CPUs against AMD in maturing markets and carries factory cost whether or not the nodes are full. That confirms the power map; the company with pricing power shows it in gross margin. The second gap is quieter and matters more for the next decade: NVIDIA's R&D is now larger in dollars ($18.5B vs $13.8B in the last full year, $13.4B vs $6.7B in the last half-year), so the roadmap budget runs against Intel, whose 17% R&D cut in FY2025 lifted margins by shrinking it (10-Ks; 10-Qs). Intel's opex cuts are real self-help, visible in a Q2 2026 operating margin of 11.1%, but they cannot touch a gap that sits in cost of sales.
| Metric | Threshold | By when | If it hits, it favors | Where published |
|---|---|---|---|---|
| NVIDIA gross margin | At or above 72% in each FY2028 quarter | Q1-Q2 FY2028 (Apr-Jul 2027) | NVIDIA; below 70% narrows the gap | NVIDIA 10-Q |
| NVIDIA Hyperscale growth vs Broadcom AI semiconductor growth | Hyperscale year-on-year growth at or above Broadcom's | Through Q2 FY2028 (Jul 2027) | NVIDIA | NVIDIA 10-Q; Broadcom release |
| Intel gross margin after shortages ease | At or above 40% GAAP | Q2 2027 (Jul 2027) | Intel (pricing proves structural); below 37% confirms borrowed pricing | Intel 10-Q |
| Intel Foundry external revenue and 14A customer | Named 14A customer; external revenue above $1B a quarter | Mid-2027; Q4 2027 | Intel | Intel 8-K, 10-Q |
| Intel x86 server unit share | At or above 60% | 2Q 2027 | Intel; below 60% confirms erosion | Mercury Research via trade press |