The unit is one chip, designed in-house, fabricated by TSMC, and sold to hyperscalers, OEMs and distributors after rebates and price protection are netted off. For AI accelerators the contracting unit is now the gigawatt, worth “double-digit billions” of revenue each.
| What the business is | AMD designs processors and has them made by TSMC and other contract manufacturers: EPYC server CPUs, Instinct AI accelerators, Ryzen PC processors, Radeon graphics, custom console chips for Sony and Microsoft, and Xilinx FPGAs. It sells them to hyperscalers, server and PC makers, console makers and distributors. FY2025 revenue was $34.6B with about 31,000 employees, and Data Center was 48% of revenue in FY2025 and 58% in Q2 2026 (FY2025 10-K; Q2 2026 10-Q). |
| Industry | Merchant semiconductors for computing: x86 CPUs, AI accelerators and programmable logic. Profit concentrates at the bottlenecks: in the latest quarter NVIDIA earned a ~66% GAAP operating margin and TSMC 60%, against ~17% at AMD (company releases, web). |
| How it makes money | AMD books revenue when chips ship (custom console chips over time, on a cost-plus basis), after deducting price protection, rebates, return allowances and market-development funds, and keeps the spread between that net price and the cost of TSMC wafers, memory, packaging and test, which in FY2025 left 50 cents of gross profit per dollar to fund $12.2B of R&D and marketing, general and administrative spending. |
| Unit of economics | One chip sold. AMD discloses no unit counts or prices (unknown), only Client direction: FY2025 processor units +31% and ASP +15% (FY2025 10-K); each revenue dollar earned 54 cents of GAAP gross profit in Q2 2026 (Q2 2026 10-Q). For AI accelerators the contracting unit is now the gigawatt, at "double-digit billions" of revenue per GW (Q2 2026 call, web), with up to 14 GW under framework agreements. |
| What protects it | The royalty-free Intel cross-licence that makes x86 a two-company market and terminates on a change of control (AMD 8-K 2009, web); a decade of chiplet design and lead-customer standing at TSMC (first HPC product on N2); and multi-generation co-design with OpenAI, Meta and Anthropic at gigawatt scale. Nothing comparable protects it against NVIDIA's CUDA ecosystem in AI accelerators. |
| What drives earnings | Data Center mix, meaning Instinct GPU and EPYC CPU volumes; gross margin after TSMC and high-bandwidth-memory costs; operating leverage on an opex base growing 30–40% a year; share gains from Intel in server and PC CPUs. |
| What to watch | Delivery of the first MI450/Helios gigawatt to OpenAI and Meta, and the warrant tranches that vest with it (AMD 8-Ks and 10-Q warrant note); Mercury Research x86 server share; unconditional purchase commitments ($30.3B at Q2 2026) against Data Center revenue. |
| Cycle exposure | Data Center AI: high, early in its ramp and never tested by a down year. Server CPU: medium, at record share. Client: high, entering a PC downturn (IDC 2026 forecast −11.3%) after a record 2025. Gaming: high, near the trough of a seven-year console cycle. Embedded: medium, recovering from the 2024–25 correction. |
US$. AMD's fiscal year ends on the last Saturday of December; latest full year FY2025 (ended 27 Dec 2025), latest quarter Q2 2026 (ended 27 Jun 2026). (inferred) marks a conclusion reasoned from sourced facts, not a disclosed figure; (web) marks evidence from outside the filing archive.
Every data centre, PC and games console needs processors, and buyers want the most computation per dollar and per watt. AMD sells that computation as chips it designs but does not manufacture. Data Center holds EPYC server CPUs, Instinct AI GPUs, Pensando networking chips and rack-scale designs; Client and Gaming holds Ryzen PC processors, Radeon graphics and the custom chips inside PlayStation 5 and Xbox; Embedded holds Xilinx FPGAs, adaptive SoCs and embedded x86 parts (FY2025 10-K).
The unit is one chip. AMD designs it, with R&D of $8.1B in FY2025 or 23% of revenue, and TSMC fabricates every CPU and GPU wafer at 7nm and below, while GlobalFoundries makes older-node parts. Assembly and test happen mostly in China, Malaysia and Taiwan at two Tongfu joint ventures in which AMD holds 15%, and at SPIL and KYEC; AMD bought $2.0B from the joint ventures in 2025 (FY2025 10-K). Standard chips are recognised as revenue on shipment and paid on net 30–60 day terms; semi-custom console chips are recognised over time on a cost-plus basis, and revenue recognised over time moved with console sales: 25% of the total in 2023, 8% in 2024, 9% in 2025 and 4% in Q2 2026 (FY2025 10-K; Q2 2026 10-Q).
A good part of the list price never reaches revenue. Distributors get price protection on unsold stock and stock rotation, OEMs get rebates and return rights, and marketing-development funds are netted off; accrued marketing programmes stood at $1.45B at FY2025, up from $1.06B, and marketing and advertising spending doubled to about $2.4B in 2025 (FY2025 10-K). Each revenue dollar left 50 cents of GAAP gross profit in FY2025, or about 52.5 cents before the amortisation of Xilinx-era intangibles charged to cost of sales (inferred from FY2025 10-K).
AMD now contracts AI accelerators by the gigawatt of data-centre power rather than by the chip. OpenAI (October 2025) and Meta (February 2026) each intend to deploy up to 6 GW of Instinct GPUs, but only the first gigawatt is a binding commitment in either case (8-Ks 2025-10-06, 2026-02-24). Each received a warrant for up to 160M AMD shares at $0.01, vesting in tranches as purchases scale from 1 GW to 6 GW and as AMD's share price crosses thresholds rising to $600. Anthropic has signed for up to 2 GW from the first half of 2027, with AMD committing up to $5B of equity investment into Anthropic (AMD release 22 Jul 2026, web). Management puts revenue per gigawatt at "double-digit billions" (Q2 2026 call, web).
AMD does not sell the Helios rack itself. It licenses the rack design to customers and their manufacturers and supplies the chips inside it, yet it "may remain responsible to the customer for the full system" under warranty (Q2 2026 10-Q). This is why AMD bought ZT Systems for $4.4B in March 2025 and sold ZT's manufacturing arm to Sanmina for $3.0B including an earn-out of up to $450M in October 2025: it kept the rack-design engineers and shed the low-margin server assembly (FY2025 10-K; 8-Ks 2025-05-19, 2025-10-27).
| Segment ($M) | FY2025 revenue | FY2025 op. income | Op. margin | Q2 2026 revenue | YoY |
|---|---|---|---|---|---|
| Data Center | 16,635 | 3,603 | 21.7% | 6,718 | +107% |
| Client and Gaming | 14,550 | 2,855 | 19.6% | 3,841 | +6% |
| of which Client | 10,640 | n/d | n/d | 3,062 | +23% |
| of which Gaming | 3,910 | n/d | n/d | 779 | −31% |
| Embedded | 3,454 | 1,243 | 36.0% | 977 | +19% |
| All Other (amortisation, stock comp.) | — | (4,007) | — | — | — |
| Total | 34,639 | 3,694 | 10.7% | 11,536 | +50% |
Source: FY2025 10-K Note 4; Q2 2026 10-Q Note 4. n/d = not disclosed since Client and Gaming were combined in Q1 2025. FY2025 Data Center operating income absorbed ~$440M of net MI308 export-control charges. Margins inferred.
The mix is moving toward the segments that earn more. In Q2 2026 Data Center earned a 31% segment margin and Embedded 40%, against 15% for Client and Gaming (inferred from Q2 2026 10-Q). Gaming, the lowest-return line, has fallen from $6.2B in 2023 to $3.9B in 2025 and dropped a further 31% in Q2 2026 on lower semi-custom sales (FY2025 10-K; Q2 2026 10-Q).
The industry turns transistors into computation. Value passes from instruction-set and design-tool owners (Arm, Synopsys, Cadence) to chip designers (NVIDIA, AMD, Broadcom, Marvell, Qualcomm), to the foundry (TSMC), to memory makers supplying high-bandwidth memory (SK hynix, Micron, Samsung), to packaging and server assemblers, and finally to hyperscalers, which increasingly design their own chips as well. The profit sits where a customer cannot route around the supplier.
| Layer / company | Latest period | GAAP op. margin | Why |
|---|---|---|---|
| NVIDIA (merchant accelerators) | Qtr to 26 Jul 2026 | ~66% | CUDA software standard, NVLink, rack systems |
| SK hynix (HBM/DRAM) | Q2 2026 | 76% | HBM sold out; memory shortage |
| TSMC (foundry) | Q2 2026 | 60.3% | Sole leading-edge source |
| Broadcom (custom ASIC, networking) | Qtr to 2 Aug 2026 | ~54% | Designs hyperscalers' own accelerators |
| AMD (merchant CPU/GPU/FPGA) | Q2 2026 | ~17% (27% non-GAAP) | Challenger in AI; share-taker in x86 |
| Intel (design + fab) | Q2 2026 | ~11% | Foundry arm lost $2.1B in the quarter |
Sources (web): NVIDIA Q2 FY2027 release; SK hynix Q2 2026 results; TSMC 2Q26 release; Broadcom Q3 FY2026 release; AMD Q2 2026 release; Intel Q2 2026 release. Margins for NVIDIA, Broadcom, AMD and Intel inferred from reported operating income and revenue.
AMD sits between the two most powerful suppliers in the chain, TSMC and the memory makers, and the most profitable competitor. That placement explains why a company growing 50% still earns a GAAP operating margin a quarter of NVIDIA's. AMD competes in three markets with very different structures, and its position differs sharply across them.
x86 is in practice a two-company market: AMD's rights rest on a royalty-free 2009 cross-licence with Intel that ends if either party changes control (AMD 8-K 2009-11-11, web), and no third supplier sells x86 server or PC chips at scale (inferred). Inside that duopoly AMD has been taking share: Mercury Research puts it at 34.5% of x86 server units in Q2 2026, up about 7 points in a year, and 46.2% of server revenue in Q1 2026; in PCs it held 30.3% of client units (Mercury via The Register, HotHardware, Tom's Hardware, web). The mechanism is manufacturing: AMD's chiplet parts ride TSMC's process lead while Intel reports "one of the most severe supply constraints in its history" (Intel Q2 2026 remarks, web). The threat comes from outside x86, as Arm-based chips reached 13.6% of servers, led by AWS Graviton and NVIDIA Grace (Mercury, web).
NVIDIA's data-centre revenue was $89.0B in its quarter to 26 July 2026; AMD's whole Data Center segment, CPUs included, was $6.7B, about 7.5% of that (NVIDIA release, web; Q2 2026 10-Q, inferred). AMD does not disclose Instinct revenue after 2024's ">$5 billion" (NOT FOUND in filings). Custom chips are the larger alternative: Broadcom's AI semiconductor revenue was $16.7B in its latest quarter, up 221%, and Amazon says its own chip business runs at over $25B a year (Broadcom release; Amazon Q2 2026 call, web). In discrete graphics NVIDIA holds 90% of add-in-board shipments to AMD's 8% (Jon Peddie Research Q2 2026, web).
The barrier that binds in accelerators is software. NVIDIA reports over 7.5M CUDA developers (NVIDIA FY2026 10-K, web), and in the June 2026 MLPerf Training round AMD made its first multi-node submission at 512 GPUs while NVIDIA partners scaled to 8,192 (MLCommons, web). That barrier binds the long tail of developers but much less the frontier labs that write their own low-level code, which is exactly where AMD's gigawatt deals sit. Scale also binds: NVIDIA spent $18.5B on R&D in its FY2026, against AMD's $8.1B spread across CPUs, GPUs, FPGAs and networking (NVIDIA release, web; FY2025 10-K).
AMD calls itself "the market share leader in semi-custom game console products" (FY2025 10-K), and outside evidence supports it: it supplies both PlayStation 5 and Xbox Series and has a multi-year partnership on the next Xbox (Xbox Wire, web). In FPGAs the main rival, Altera, was spun out of Intel with Silver Lake buying 51% at an $8.75B valuation (Intel release, web); independent share figures were NOT FOUND.
NVIDIA invested $5B in Intel in September 2025, and the two will build x86 CPUs with NVIDIA's NVLink interconnect and PC chips pairing x86 with RTX graphics (NVIDIA release, web). That gives NVIDIA's AI systems an x86 host that is not EPYC, and sets an Intel-plus-RTX laptop chip against Ryzen's integrated-graphics advantage; AMD's own 10-K warns it "may result in increased competition and pricing pressure" (FY2025 10-K). NVIDIA is also entering Windows PCs with an Arm chip, RTX Spark, from fall 2026 (CNBC, web).
| AMD claim | Outside evidence | Verdict |
|---|---|---|
| Server CPU performance and TCO leadership | Server unit share 27%→34.5% in a year; 46% of revenue (Mercury) | Supports vs Intel; partial vs Arm |
| Open ROCm ecosystem gaining momentum | Frontier labs co-develop ROCm; first multi-node MLPerf at 512 vs 8,192 GPUs | Partially supports |
| The alternative to NVIDIA at scale | Only merchant GPU at GW scale, but custom ASICs are larger by revenue | Contradicted as stated |
| Full-stack rack-scale supplier (Helios) | Commitments from Oracle, OpenAI, Meta, Anthropic, Microsoft; shipments start late Q3 2026 | Unproven until 2027 |
Winners in this industry own a bottleneck customers cannot route around: a software and system standard, leading-edge manufacturing, or scarce memory. AMD is that kind of company in x86, where a licence limits entry and it out-executes the incumbent. In AI accelerators it is a well-run challenger that pays for position with equity warrants and customer investment, which is rational for a number two but is not the behaviour of a price-setter (inferred).
Revenue rose from $4.3B in 2016 to $34.6B in 2025, about 26% a year (inferred from FY2018 and FY2025 10-Ks). The engine changed twice. From 2016 to 2021 it was CPU share gains with Ryzen and EPYC plus the console cycle; in 2022 it was the Xilinx acquisition; from 2024 it has been Data Center, first EPYC and then Instinct GPUs.
Revenue ($B): 2016 4.3, 2017 5.3, 2018 6.5, 2019 6.7, 2020 9.8, 2021 16.4, 2022 23.6, 2023 22.7, 2024 25.8, 2025 34.6. Data Center under the current segment structure: 2020 1.7, 2021 3.7, 2022 6.0, 2023 6.5, 2024 12.6, 2025 16.6; not reported before 2020 (FY2018–FY2025 10-Ks; 2016–2017 restated for ASC 606).
| Year | Reported growth | Acquired contribution | Organic (inferred) |
|---|---|---|---|
| 2022 | +43.6% | Xilinx $4,612M revenue; Pensando immaterial | ~+15.5% |
| 2023 | −3.9% | Full-year Xilinx lifted Embedded +17% | Negative; Client −25% |
| 2024 | +13.7% | Silo AI, immaterial | ≈ reported |
| 2025 | +34.3% | ZT design team "not material"; ZT manufacturing in discontinued operations | ≈ reported |
| H1 2026 | +44.1% | $129M of tuck-ins, immaterial | ≈ reported |
Sources: FY2022 10-K (Xilinx contribution), FY2025 10-K, Q2 2026 10-Q. 2022 organic = (23,601 − 4,612) / 16,434 − 1; it also includes a 53rd week.
FY2025 revenue rose $8.9B. Data Center added $4.1B from EPYC and MI350 demand despite the MI308 charge; Client added $3.6B, with processor units up 31% and prices up 15%; Gaming added $1.3B; Embedded fell $0.1B (FY2025 10-K, inferred splits). In Q2 2026 the $3.9B increase was 90% Data Center, and Client grew on volume alone, with units up 34% and average prices down 6% on mix (Q2 2026 10-Q). Management says server CPU growth has "double-digit growth in both units and ASPs, but … more unit" (Q2 2026 call, web); Instinct prices and volumes are NOT FOUND.
The largest lever. Demand comes from AI data-centre spending, but AMD's share of it is partly bought with the warrants and customer investment described in Section 2. Management guides Q3 2026 revenue to about $13.0B and expects Data Center to "more than double" in 2027 (Q2 2026 release and call, web).
Share gains inside a licensed duopoly, about 7 points of server units in the latest year (Mercury, web), now helped by AI agents that raise the number of CPUs per GPU deployment. Management expects server CPU revenue to rise more than 80% year on year in the second half of 2026 and more than 70% in 2027 (Q2 2026 call, web). Part of the gain is borrowed from Intel's supply problems.
The tide under both levers above. 2026 guides: Alphabet $195–205B, Amazon ~$220B, Meta $130–145B; Microsoft spent ~$165B in its fiscal 2026 and Oracle guides $90–95B for fiscal 2027, roughly $0.8T combined (company calls, web; total inferred). Amazon says higher memory prices are part of its increase, so some of this is price rather than units.
AMD's share rises inside a market that is turning down: IDC forecasts 2026 PC shipments −11.3% as memory costs push prices up 17% (IDC, web). Share gains held Client at +23% in Q2 2026, but the market is now a headwind.
Embedded fell 33% in 2024 as customers worked down inventory, and grew 19% in Q2 2026 (FY2024 10-K via FY2025 10-K; Q2 2026 10-Q). Management expects more than $18B of new design wins in 2026 (Q2 2026 call, web).
The current console generation is in its seventh year; management expects semi-custom revenue to "come down significant double-digit" in 2026 (Q4 2025 call, web). A next-generation Xbox partnership resets the cycle later, on unknown timing.
Gross margin rose from 23% in 2016 to 48% in 2021 as the mix moved from low-end PC chips and consoles to Ryzen and EPYC (FY2016–FY2021 10-Ks). From 2022 the GAAP figure carries $0.9–1.4B a year of Xilinx-era amortisation in cost of sales; before it, gross margin was about 51% in 2022, 53% in 2024 and 52.5% in 2025 (inferred from FY2022–FY2025 10-Ks). Q2 2026 reached 54% GAAP and 56% non-GAAP, helped by the absence of 2025's MI308 charge (Q2 2026 10-Q; release, web). The CFO says Data Center AI gross margins are "slightly below company average" (Q2 2026 call, web), so a rising Instinct mix grows gross-profit dollars faster than the margin percentage.
GAAP gross margin, 2016–2025: 23%, 34%, 38%, 43%, 45%, 48%, 45%, 46%, 49%, 50%. Before acquisition amortisation in cost of sales, 2022–2025: 51%, 50%, 53%, 52.5% (inferred from FY2018–FY2025 10-Ks).
Wafers, memory and packaging are variable; R&D and marketing are the fixed base, and they are growing ahead of revenue. R&D rose 25% in FY2025 and 33% in Q2 2026, and marketing, general and administrative 52% and 41% (FY2025 10-K; Q2 2026 10-Q). The gap between GAAP and cash earnings is large: FY2025 GAAP operating margin was 10.7%, against about 22% non-GAAP, because $2.25B of amortisation and $1.64B of stock compensation sit in between (FY2025 10-K; non-GAAP $7.77B from release, web). In Q2 2026, of each revenue dollar, about 46 cents went to cost of sales, 22 cents to R&D, 12 cents to marketing, general and administrative, and 2 cents to amortisation in operating expenses, leaving 17 cents of GAAP operating profit (inferred from Q2 2026 10-Q).
As a fabless company AMD spends little on fixed assets: capex was $974M in FY2025, 2.8% of revenue, so free cash flow of about $5.5B from continuing operations was 1.3 times GAAP net income (inferred from FY2025 10-K). Working capital is where cash goes: inventory rose from $5.7B at FY2024 to $8.5B at Q2 2026, and first-half 2026 operating cash flow of $5.3B was helped by a $2.2B rise in payables that the 10-Q attributes to payment timing (Q2 2026 10-Q).
“We have entered and may continue to enter into long-term purchase commitments and prepayment arrangements with some of our suppliers … these risks have increased as our purchase obligations and prepayments have grown.”AMD Q2 2026 10-Q, risk factors
The "asset-light" label is getting less accurate. Non-cancellable purchase commitments rose from $12.2B at FY2025 to $30.3B at Q2 2026, with about $1.0B of supplier prepayments in the half, while customers can generally cancel standard orders more than 30 days before shipment (FY2025 10-K; Q2 2026 10-Q). Add $4.5B of data-centre and other real-estate leases not yet started and $9.5B of new leases signed after the quarter (Q2 2026 10-Q). AMD is committing capital upstream to secure supply and taking on long-dated obligations of its own (inferred).
| FY2016 | FY2020 | FY2023 | FY2025 | |
|---|---|---|---|---|
| Revenue ($B) | 4.32 | 9.76 | 22.68 | 34.64 |
| Gross margin (GAAP) | 23.2% | 44.5% | 46.1% | 49.5% |
| Operating income ($B) / margin | (0.37) / −8.6% | 1.37 / 14.0% | 0.40 / 1.8% | 3.69 / 10.7% |
| Free cash flow ($B) | 0.00 | 0.78 | 1.12 | 5.52 |
| Net cash / (debt) ($B) | (0.50) | 1.95 | 3.27 | 7.30 |
| Diluted shares (M) | 835 | 1,207 | 1,625 | 1,636 |
Sources: FY2018 10-K (2016, restated for ASC 606), FY2022 10-K (2020), FY2025 10-K (2023, 2025); balance sheets from each year's 10-K. Not fully comparable: 2016 includes a $340M wafer-agreement charge; from 2022 figures include Xilinx and $2.3–3.5B a year of acquisition amortisation; 2023 was a PC-market trough; FY2025 free cash flow is continuing operations. Free cash flow = operating cash flow − capex; net cash = cash and short-term investments − debt principal (inferred).
Over FY2016–FY2025, cash went, in order: $8.6B to buybacks, all since 2021; $4.3B gross to acquisitions but only about $0.2B net of cash acquired with Xilinx, the ZT manufacturing sale and the 2016 sale of 85% of the assembly-and-test operations; $4.2B to debt repayment, more than matched by $4.5B of new borrowing; and $3.8B to capex; no dividends (inferred from each year's cash flow statement). The 2016–2021 deleveraging came from a 2016 equity raise, a 2019 warrant exercise and convertible-note conversions rather than from operating cash (FY2016–FY2021 10-Ks).
The big decisions were paid in shares. Xilinx cost $48.8B, almost all in 429M AMD shares (FY2022 10-K), and the OpenAI and Meta warrants could add 320M more. Diluted shares rose 96% from 2016 to 2025, so revenue per diluted share grew about 4.1 times while revenue grew 8 times (inferred). Buybacks of $8.6B only modestly exceeded $6.7B of stock compensation over the decade, so they have offset employee dilution rather than shrunk the count. Management treats its equity as a currency for acquisitions and for winning customers.
At Q2 2026 AMD held $13.1B of cash and short-term investments against $3.25B of debt, about $9.9B net (Q2 2026 10-Q, inferred). In May 2026 it replaced its revolver with a $5.0B facility carrying no financial covenants and raised its commercial paper limit to $5.5B. After the reported figures, in August 2026, it issued $4.75B of notes due 2029–2036 at 4.60–5.50%, taking debt principal to about $8.0B before repaying the $875M 2026 notes (8-Ks 2026-05-15, 2026-08-17; inferred). It also committed up to $5.0B of investments through FY2028; the 10-Q does not name the counterparty, but it matches the Anthropic investment AMD announced on 22 July 2026 (Q2 2026 10-Q; AMD release, web; inferred).
| Segment | Share of H1 2026 revenue | What drives the cycle | Where it sits now |
|---|---|---|---|
| Data Center | 57% | Cloud and AI capital spending; product ramps | Record; early in AI ramp |
| Client | 27% | PC replacement and memory costs | Record share; PC market turning down |
| Gaming | 7% | Console generations (~7 years) | Near trough |
| Embedded | 8% | Industrial and communications inventory | Recovering |
Source: Q2 2026 10-Q (segment revenue, shares inferred); market position from Q2 2026 10-Q, IDC and AMD's Q2 2026 call (web).
AMD has been hit hard in every downturn but has never been hit everywhere at once. In 2016 it lost money at the operating line; in 2023 revenue fell 4% with Client down 25% and operating margin 1.8%; in 2024 Gaming fell 58% and Embedded 33% while Data Center grew 94% (FY2018, FY2025 10-Ks). Each time a new product cycle, first Ryzen and EPYC and then Instinct, carried the total. Data Center AI has not yet faced a down year; the nearest analogues are NVIDIA's FY2023, when its gross margin fell 8 points, and Intel's data-centre revenue falling 20% in 2023 (NVIDIA and Intel releases, web).
Against its own history AMD is at a peak on revenue and near one on gross margin. Q2 2026 revenue of $11.5B is a record, and its 54% GAAP gross margin matches Q4 2025 and sits well above the full-year peaks of 48% in 2021 and 49.5% in 2025 (Q2 2026 10-Q; FY2025 10-K; Q4 2025 release, web). Operating margin is not at a peak: 17% in Q2 2026 against 22% in 2021, because opex is growing ahead of revenue and acquisition amortisation now runs through the income statement (inferred).
Four constraints bind. TSMC is the sole source for every leading-edge CPU and GPU wafer, and a Taiwan disruption could materially interrupt deliveries (FY2025 10-K). High-bandwidth memory comes from three suppliers that are sold out, and the memory shortage is raising the cost of PCs, consoles and AI systems alike (Q2 2026 10-Q; SK hynix, web). US export controls cost ~$800M in Q2 2025 when MI308 needed a licence, $440M net for the year; the government's expected 15% share of licensed China revenue has not been written into regulation, and MI325 licences granted in 2026 carry a 25% tariff because units must be inspected in the US first (FY2025 10-K; Q2 2026 10-Q). Management forecasts no China data-centre revenue after Q1 2026 (Q4 2025 call, web), so China is now optional upside rather than a base-case driver.
The downside mechanism is the mismatch of commitments. If hyperscalers paused spending, AMD would hold $30.3B of non-cancellable purchase commitments and $8.5B of inventory against customer orders that can be cancelled with 30 days' notice, while its opex base is growing 30–40% a year (Q2 2026 10-Q; inferred). Unvested warrant tranches would then be the visible sign that promised gigawatts did not arrive.
| Indicator | Why it matters | Where published |
|---|---|---|
| Data Center revenue; Helios step-up in Q4 2026 | Tests the AI ramp | AMD 10-Q, earnings release |
| OpenAI and Meta first-GW delivery; warrant tranche vesting | Only the first GW is binding | AMD 8-Ks; 10-Q warrant note |
| x86 server unit and revenue share; Arm share | Durability of the CPU franchise | Mercury Research, via trade press |
| Purchase commitments and inventory vs Data Center revenue | Size of the commitment mismatch | AMD 10-Q commitments note |
| Broadcom AI revenue; NVIDIA data-centre revenue | Custom-chip substitution; relative share | Company releases |
| 2027 hyperscaler capex guidance | The tide | Alphabet, Amazon, Microsoft, Meta, Oracle calls |
| PC units and prices; DRAM and HBM contract prices | Client cycle and cost | IDC, Gartner; TrendForce |
Before forming a thesis, an investor would need to resolve two things: what Instinct earns per gigawatt after memory, rack and warrant costs once Helios ships in volume, and whether server CPU share gains hold once Intel's 18A products ramp.