AI and data center semiconductors. AMD / NVDA. Run 23 September 2026. Built from AMD 10-Ks FY2021, FY2023, FY2025, 10-Qs Q1 and Q2 2026 and 8-Ks through 17 Aug 2026; NVIDIA 10-Ks FY2022, FY2024, FY2026, 10-Qs Q1 and Q2 FY2027, 8-K 17 Aug 2026 and the Q2 FY2027 earnings call (26 Aug 2026); plus cited web sources. Events swept through 23 Sep 2026; most recent events checked: NVIDIA Q2 FY2027 call and outlook (26 Aug 2026), NVIDIA SB Energy / OpenAI guarantees and AMD $4.75bn notes (both 17 Aug 2026). Figures in USD; AMD fiscal year = calendar year, NVIDIA fiscal year ends late January (about one month offset, treated as comparable). Not a valuation and not a recommendation.
Scale note: NVIDIA's last fiscal year revenue ($215.9bn) is 6.2x AMD's ($34.6bn), so the page compares ratios and mix, not absolutes, wherever size would decide the answer by itself.
Three cells carry NVIDIA's next five years. First, hyperscale AI systems, where each generation lifts the dollars per gigawatt (management puts Grace Blackwell near $25bn and Vera Rubin near $40bn per gigawatt, a company claim) and Rubin already makes up about 20% of Q3 FY27 data center revenue. Second, the AI clouds, sovereign and enterprise cell (+138% in Q2 FY27), made of buyers who do not design their own chips and whom AMD barely reaches. Third, networking attached to every rack. The FY2028 guide of about +70% is limited by supply, and $279bn of supply and capacity commitments stand behind it (Q2 FY27 10-Q and call). AMD's case is real but conditional. Server CPUs are guided to grow more than 80% in H2 2026 and more than 70% in 2027. Helios racks with MI450 start shipping late in Q3 2026 into commitments of 6GW from OpenAI, 6GW from Meta and up to 2GW from Anthropic, of which only the first gigawatt from OpenAI and Meta is binding (8-Ks; Q2 2026 call summary). If MI450 lands on time, AMD's growth rate in 2027 could beat NVIDIA's. That is why this answer is narrow.
NVIDIA's growth engine is also its highest-margin line: the Compute & Networking segment earned a 71.0% operating margin in Q2 FY27, on price-setting for full systems and networking sold with every rack. AMD's growth engine is its lowest-margin line. Management says AI GPU gross margins will sit below the corporate average during the MI450 ramp, so the server CPU business has to hold the mix up (Q2 2026 call summary). AMD is improving: GAAP gross margin rose from 46.1% to 49.5% over three years and data center segment margin reached 31.3% in Q2 2026. But the gap in points is not closing. The test is NVIDIA's memory reset: it guides gross margin down to 71-72% in Q4 FY27, then back to 72-73% in FY28 once announced price increases take effect. A supplier that can do that sets the price.
NVIDIA's exposure is the capex cycle and the financing loop around it. Data center is 92.5% of revenue, three direct customers took 16%, 15% and 13% of H1 FY27 revenue, and its $279bn of supply commitments are fixed while demand is not. It also holds $99bn of equity in its own ecosystem and guarantees leases for an OpenAI site of about 4.25GW, and management says AI labs it backs make up roughly a quarter of next year's business (Q2 FY27 10-Q, 8-K of 17 Aug 2026, call). AMD's exposure is its competitor. When NVIDIA's supply loosens, buyers cut the second source first. AMD's GPU commitments are staged by milestone and paired with warrants, which gives buyers option-like terms. With a 17% EBIT margin, AMD has little room to absorb a price war. NVIDIA's 60% margin does. In a digestion, AMD's AI ramp stalls first. NVIDIA's reported numbers fall harder but stay highly profitable.
These eight cells cover essentially all of both companies' revenue. Neither company reports product by region, so geography enters where it changes the call (hyperscale vs. other buyers, China). Every score is argued in the tabs below.
| Cell (product x region) | AMD | NVIDIA | Why (one clause, sourced) |
|---|---|---|---|
| AI accelerators, hyperscale (US-HQ CSPs, global deployment) | 3 | 5 | NVIDIA: hyperscale DC revenue $48.7bn in Q2 FY27, +102% (Q2 FY27 10-Q). AMD: MI350 in volume, MI450 anchor deals with OpenAI, Meta, Microsoft, Oracle, but ~6-8% share (approx., unverified) and warrants attached (8-Ks). |
| AI accelerators, AI clouds, sovereign, enterprise (NVIDIA "ACIE") | 2 | 5 | NVIDIA ACIE $40.3bn, +138% (Q2 FY27 10-Q). AMD sub-scale here; revenue ND (inferred: no rack-scale product shipped until Helios, H2 2026). |
| Data center networking (scale-up and scale-out) | 2 | 5 | NVIDIA networking $31.4bn FY26, +142% (FY26 10-K). AMD Pensando revenue ND, first rack networking ships with Helios (Q2 2026 10-Q). |
| Server CPUs (global) | 4 | 2 | AMD 46.2% revenue share, 33.2% units, up 6 pts in a year (Mercury Q1 2026 via Tom's Hardware). NVIDIA Grace TTM above $5bn, Vera standalone just launched (Q2 FY27 call). |
| PC client CPUs (global) | 4 | 0 | AMD client units +34%, ASP -6% in Q2 2026 (Q2 2026 10-Q); unit share 29.6% vs 24.1% (Mercury). NVIDIA no PC CPU revenue disclosed. |
| Gaming GPUs and consoles (global) | 2 | 4 | NVIDIA gaming $16.0bn FY26, +41% (FY26 10-K). AMD gaming $3.9bn FY25 then -31% in Q2 2026 on semi-custom (Q2 2026 10-Q). |
| Embedded, industrial, auto and edge (global) | 3 | 3 | AMD embedded $3.5bn, 36% segment margin, back to +19% in Q2 2026. NVIDIA pro viz + auto + OEM $6.2bn, Edge +27% Q2 FY27 (10-K, 10-Q). |
| China data center (both) | 1 | 1 | NVIDIA: no China DC compute in outlook, Hopper below 1% of DC revenue (Q2 FY27 call). AMD: MI308 licensed shipments only; MI325 licenses carry a 25% tariff path (Q2 2026 10-Q). |
Scores are per cell and are not summed. The three lenses get their verdicts in the Three Answers above.
Normalization. AMD reports Data Center (EPYC, Instinct and Pensando, not split), Client, Gaming and Embedded. NVIDIA reports Compute and Networking inside Data Center, then Gaming, Pro Visualization, Automotive and OEM. From Q1 FY2027 it regroups these into Hyperscale, AI Clouds, Industrial & Enterprise (ACIE) and Edge Computing. Geography cannot be matched: AMD reports by customer billing location, NVIDIA by customer headquarters, a basis it adopted in Q3 FY2026. On those bases AMD's revenue is 32.8% United States, 22.4% China including Hong Kong, 15.0% Taiwan and 12.4% Singapore. NVIDIA's is 69.3% United States, 19.6% Taiwan and 9.1% China, and NVIDIA estimates 76% of the data center revenue from Taiwan-headquartered customers ends up in the US and Europe (FY2025 and FY2026 10-Ks). Much of AMD's China revenue is PC and gaming billed to Chinese manufacturers (inferred: Client and Gaming sold through OEMs, ODMs and distributors per the FY2025 10-K).
| Cell | AMD rev FY25 (% of total) | AMD 2y CAGR | NVIDIA rev FY26 (% of total) | NVIDIA 2y CAGR | Margin signal / leader |
|---|---|---|---|---|---|
| DC compute | 16,635 (48.0%)* | +60% | 162,361 (75.2%) | +104% | NVIDIA. Compute & Networking operating margin 67.3% FY26 vs AMD DC 21.7% (24.3% before the $440m MI308 net charge) |
| DC networking | ND | ND | 31,376 (14.5%) | +91% | NVIDIA. Attached to every rack; AMD's first rack ships H2 2026 |
| PC client CPUs | 10,640 (30.7%) | +51% | 0 | n/a | AMD. Client and Gaming segment margin 19.6% FY25 |
| Gaming | 3,910 (11.3%) | -21% | 16,042 (7.4%) | +24% | NVIDIA. Graphics segment margin 40.8% FY26 |
| Embedded and edge | 3,454 (10.0%) | -19% | 6,159 (2.9%) | +44% | Split. AMD embedded segment margin 36.0%, best in AMD |
USD m. *AMD DC includes server CPUs, Instinct GPUs and networking. 2y CAGR = FY2023 to FY2025 (AMD) and FY2024 to FY2026 (NVIDIA). Sources: AMD FY2025 10-K segment note; NVIDIA FY2026 10-K segment note and end-market table.
AI accelerators. NVIDIA ships on a yearly cadence: Blackwell Ultra now, and Vera Rubin, which began production shipments in Q3 FY27. It adds CPUs (Vera), five kinds of networking and the Groq-licensed LPX inference accelerator to each gigawatt (Q2 FY27 10-Q and call). AMD runs on anchor contracts. OpenAI and Meta have each committed to 1GW binding out of up to 6GW, and each holds a warrant for 160m AMD shares at $0.01 that vests with purchases and with AMD share price hurdles rising to $600 (8-Ks of 6 Oct 2025 and 24 Feb 2026). Anthropic plans up to 2GW from 1H 2027 (Q2 2026 call summary). Both are organic. From AMD's ZT Systems purchase ($3.2bn cash plus shares), it kept the design team and sold the manufacturing arm to Sanmina for $2.4bn (FY2025 10-K).
Server CPUs. Agentic AI is lifting CPU demand. AMD's server revenue share is 46.2% against 33.2% of units, a higher selling price than Intel's (Mercury Q1 2026). Management expects server CPU revenue up more than 80% in H2 2026 and more than 70% in 2027, with supply tight this year (call summary). NVIDIA's Grace passed $5bn over the trailing twelve months and Vera is now sold standalone (Q2 FY27 call). PC client. AMD's gains here are cyclical: Intel's supply problems added share (client units 29.6% vs 24.1%, Mercury), but ASP fell 6% in Q2 2026 and memory prices are hurting PC demand.
NVIDIA sells complete systems through ODMs and OEMs to hyperscalers and AI clouds. It has announced price increases for Q1 FY28 and gives investment-grade buyers extended payment terms (DSO 60 days). AMD sells chips. It licenses the Helios rack design but "does not manufacture or sell the completed Helios rack systems" (Q2 2026 10-Q), and in the PC channel its distributor agreements include price protection (FY2025 10-K). No AMD customer reached 10% of revenue in FY2024 or FY2025.
| Input | AMD | NVIDIA |
|---|---|---|
| Wafers / packaging | TSMC, GF for 12/14nm; limited IC package suppliers | TSMC, Samsung; CoWoS; Foxconn and others |
| Memory | Limited suppliers, industry shortage flagged | SK hynix, Micron, Samsung; "extreme pricing conditions" |
| Capacity locked | $1.0bn supply prepayments in H1 2026 | $279bn supply and capacity commitments (from $119bn a quarter earlier) |
| Inventory | $8.5bn | $31.6bn |
Sources: AMD FY2025 10-K and Q2 2026 10-Q; NVIDIA FY2026 10-K, Q2 FY2027 10-Q and call.
In AI accelerators, NVIDIA holds about 81% of 2025 revenue, AMD 6-8%, and hyperscaler custom silicon (Google TPU, AWS Trainium, Microsoft Maia and Meta MTIA, largely built with Broadcom and Marvell) about 15% (siliconanalysts.com, approx., unverified). Custom silicon competes with both, but mostly inside the hyperscale cell. In server CPUs, Intel still ships about two-thirds of units (Mercury). The China data center market is effectively closed to both, and Chinese policy pushes buyers to domestic suppliers (NVIDIA Q1 FY27 10-Q).
AI accelerators: NVIDIA is more exposed to the financing loop: equity stakes, cloud-capacity backstops and lease guarantees tied to buyers whose growth, in management's words, outruns their balance sheets. AMD is more exposed to execution. Helios is its first rack, and supply of third-party rack components can delay it (Q2 2026 10-Q risk factors). Server CPU: AMD alone is exposed, to Arm designs (Graviton, Grace/Vera) capping x86 share. Client: AMD alone, a memory-driven PC slowdown with ASP already down.
x86 server franchise (High durability): 46.2% server revenue share on 33.2% of units shows it prices above Intel (Mercury Q1 2026). Breadth (Medium): the only merchant vendor with leading CPUs, GPUs, FPGAs from Xilinx and DPUs from Pensando. ROCm open software (Low-Medium): improving, but no outside evidence of parity with CUDA (inferred).
CUDA ecosystem (High): 7.5m+ developers and $76.7bn of cumulative R&D (FY26 10-K). Replicating it takes years, not money. Rack-scale system plus networking (High): third generation of NVL72 and $31.4bn of networking revenue. Supply priority (Medium-High): $279bn of commitments that only its cash flow can fund.
AMD had no customer above 10% of revenue in FY2024 or FY2025. To win its two largest AI GPU customers, it issued each a warrant for up to 160m shares at $0.01: 320m shares in total, 19.6% of the 1,632m outstanding. The warrants vest only as OpenAI and Meta buy toward 6GW each and as AMD's share price clears hurdles up to $600 (8-Ks; Q2 2026 10-Q). NVIDIA is far more concentrated: three direct customers took 16%, 15% and 13% of H1 FY27 revenue. It pays for demand from cash flow instead of shares: $99bn of equity investments plus $25bn committed, nearly $50bn of it in frontier AI labs, cloud-capacity backstops with revenue sharing, and residual value guarantees on about 4.25GW of leases for an OpenAI site (Q2 FY27 10-Q; 8-K 17 Aug 2026; call). NVIDIA still raises prices. AMD's AI GPU margins sit below its corporate average.
Both depend on TSMC, CoWoS packaging and three HBM makers, so single-source risk is the same. Their ability to absorb shocks is not. When the same export rule hit both in April 2025, NVIDIA took a $4.5bn H20 charge and AMD about $800m (later $440m net) for MI308, and NVIDIA still ended the year at a 60.4% EBIT margin (FY2026 and FY2025 10-Ks). On memory, NVIDIA guides gross margin to a 71-72% trough and then passes the cost through with a two-quarter lag. AMD holds its non-GAAP gross margin guide at about 56% for Q3 2026 and has not disclosed how it will pass memory costs through (ND).
| Cell | Route control | Pocket price | Continuity | Outcome (share / margin) | Confirming KPI |
|---|---|---|---|---|---|
| Hyperscale AI accelerators | NVIDIA racks via ODMs; AMD chips plus licensed Helios design | AMD below NVIDIA (inferred from the ~25 pt GM gap and management's comment) | NVIDIA shipping; AMD contingent on MI450 in late Q3 2026 | NVIDIA share flat to slowly down, margin flat; AMD share up, margin below corporate average | AMD DC operating margin at or above 30% through the ramp |
| AI clouds, sovereign, enterprise | NVIDIA full stack plus financing; AMD sub-scale | NVIDIA sets the price | NVIDIA strong; AMD ND | NVIDIA share up, margin up | ACIE at or above 45% of NVIDIA DC revenue |
| Server CPUs | AMD direct to OEMs and CSPs; NVIDIA Grace/Vera mostly attached to its systems | AMD above Intel | AMD supply tight in 2026 | AMD share up, margin up | AMD server revenue share at or above 50% |
Tension: AMD grows fastest exactly where it has the least pricing power (AI GPUs), and its highest-power cell (server CPUs) is the one NVIDIA is now entering.
1. Software ecosystem, Major. CUDA vs ROCm. Closing it takes years of developer adoption, and money alone does not buy that. 2. System and networking attach, Major. Helios is AMD's first rack. NVIDIA is on its third NVL72 generation. Realistically two to three years to reach parity, and only if Helios succeeds. 3. Capacity to finance customers, Moderate. NVIDIA funds investments and guarantees from a 60% EBIT margin. AMD substitutes equity warrants. That gap closes only if AMD's margins rise.
Both companies report costs by function under US GAAP, so the five ratios are clean. AMD's cost of goods includes amortization of acquired intangibles booked in cost of sales. Its opex amortization ($1.9bn, $1.4bn, $1.2bn) and 2024 restructuring ($186m) sit outside R&D and SG&A but inside EBIT. Windows: AMD FY2023-25, NVIDIA FY2024-26. One-offs are left in: NVIDIA's $4.5bn H20 charge costs about 2.1 points of FY2026 gross margin, AMD's $440m net MI308 charge about 1.3 points of FY2025.
| % of sales, 3y avg | AMD | NVIDIA | Gap | What drives it |
|---|---|---|---|---|
| COGS | 51.7 | 27.1 | 24.6 | Price per unit of silicon, not unit cost: same foundry, same memory |
| R&D | 24.8 | 10.9 | 13.9 | Scale: NVIDIA's $18.5bn is 2.3x AMD's $8.1bn on a 6.2x revenue base |
| SG&A | 10.9 | 3.1 | 7.8 | AMD sells into PC and distribution channels; NVIDIA sells big systems to a few buyers |
| Gross margin | 48.3 | 72.9 | -24.6 | Pricing power in AI systems and networking |
| EBIT margin | 6.6 | 59.0 | -52.4 | Gross margin plus operating leverage; AMD also carries Xilinx amortization |
| Year | AMD COGS | R&D | SG&A | GM | EBIT | Year | NVIDIA COGS | R&D | SG&A | GM | EBIT |
|---|---|---|---|---|---|---|---|---|---|---|---|
| FY23 | 53.9 | 25.9 | 10.2 | 46.1 | 1.8 | FY24 | 27.3 | 14.2 | 4.4 | 72.7 | 54.1 |
| FY24 | 50.6 | 25.0 | 10.6 | 49.4 | 7.4 | FY25 | 25.0 | 9.9 | 2.7 | 75.0 | 62.4 |
| FY25 | 50.5 | 23.4 | 12.0 | 49.5 | 10.7 | FY26 | 28.9 | 8.6 | 2.1 | 71.1 | 60.4 |
% of sales. Raw values (USD m): AMD revenue 22,680 / 25,785 / 34,639; EBIT 401 / 1,900 / 3,694. NVIDIA revenue 60,922 / 130,497 / 215,938; EBIT 32,972 / 81,453 / 130,387.
The most persistent gap is gross margin: about 25 points in every year, 24.6 on average. The mechanism is the one in the Power Map. NVIDIA sells a full system with networking and CUDA, so it prices by the work done. AMD sells into markets where it is the challenger or the second source, so it prices below the leader. The cost engine therefore confirms the power map. The pricing power NVIDIA shows in the customer section appears directly in its gross margin. The second gap looks like AMD overspending on R&D (24.8% vs 10.9%), but it runs the other way: AMD spends twice the share of revenue and still less than half NVIDIA's dollars, and the dollar gap widened to 2.8x in the latest quarter. AMD has little self-inflicted cost to cut, because amortization is non-cash and the rest is structural. It is improving (gross margin +3.4 points and EBIT margin +8.9 points over three years), but not closing the gap in points.
| Metric | Threshold | By when | If it hits, it favors | Where published |
|---|---|---|---|---|
| AMD Data Center segment operating margin during MI450 ramp | At or above 30% (31.3% in Q2 2026) | Q4 2026 and Q1 2027 prints | AMD | AMD 10-Q / 10-K segment note |
| OpenAI or Meta first warrant tranche vests (1GW shipped) | Vesting disclosed | By Q3 2027 | AMD | AMD 10-Q equity note |
| AMD server CPU revenue share | At or above 50% (46.2% in Q1 2026) | Q4 2027 | AMD | Mercury Research, via press |
| NVIDIA gross margin after memory reset | 72-73% in FY2028 (71-72% trough in Q4 FY27) | Q1-Q2 FY2028 (May-Aug 2027) | NVIDIA | NVIDIA 10-Q |
| NVIDIA FY2028 revenue growth vs guide | About +70%, with ACIE at or above 45% of DC | FY2028 prints through Feb 2028 | NVIDIA | NVIDIA 10-Q / 10-K |
| NVIDIA ecosystem exposure vs revenue | Equity investments plus guarantees growing faster than revenue | FY2027 10-K (Feb 2027) | AMD (weakens NVIDIA) | NVIDIA 10-K notes |