Thesis. Data center revenue doubled to $6.7B and is now 58% of the company, with Anthropic committing 2 GW of MI450 and management guiding the segment to double again in 2027 — the credible second accelerator source, priced for exactly that outcome to land.
Reviewed 19 Sep 2026, after Q2 FY2026 results on 4 Aug. Next review after Q3 FY2026 earnings on 3 Nov 2026.
What it does
AMD is a fabless designer — TSMC builds everything it sells. Three engines: Data Center (EPYC server CPUs plus Instinct AI GPUs), Client (Ryzen, Radeon) and semi-custom silicon for game consoles. It is a two-front company: clear #2 in AI accelerators and the ascendant #2 in server CPUs against a struggling Intel. CEO Lisa Su.
Helios is AMD's rack-scale system — a full rack of MI450 GPUs sold as one unit, the same shift NVIDIA made with NVL72. It is what turns AMD from a chip vendor into a systems vendor, and it is the thing that has to ship on time.
CUDA is the reason this is hard. NVIDIA's software stack is what customers actually build on; AMD's ROCm has closed most of the inference gap and much less of the training gap.
A second source is only worth something if it arrives on schedule. Almost every commitment AMD has announced lands in H1 2027 — that is when this thesis is settled, not before.
Bull case
- Q2 FY2026 revenue $11.5B, +50% YoY — a record. Non-GAAP EPS $1.66, +82%
- Data Center $6.7B, +107% YoY, now 58% of revenue (was ~42%). Segment operating income $2.1B at a 31% margin, against −$0.2B a year earlier
- Non-GAAP gross margin 56%, +200bps YoY. Operating income $3.1B at 27%. FCF $1.6B, cash $13.1B against $3.2B of debt
- MI400 series and Helios launched — CDNA 5, first GPUs on TSMC 2nm, 432GB HBM4
- Anthropic: up to 2 GW of MI450 in Helios racks, first gigawatt H1 2027 — a third anchor alongside OpenAI (6 GW, potentially ~$100B) and Meta (6 GW)
- Microsoft expanded Helios at scale on Azure for frontier-model inference, plus two new EPYC VM series and broader Pensando DPU use. CBRS (Cerebras) pairing Helios with the Wafer-Scale Engine for low-latency inference in H2 2026
- Guidance is aggressive: server CPU revenue >80% growth in H2 2026 and >70% in FY2027; data center segment to more than double in 2027; data center AI "well over 100%." Q3 guided ~$13B ±$300M at ~56% gross margin
- EPYC share gains are durable, high-margin and independent of the GPU race
Bear case
- Data center AI gross margins run BELOW the corporate average — management disclosed this at Q2. The fastest-growing segment is the least profitable, so the server-CPU ramp has to carry the blend, and server CPU is itself supply-constrained
- The Anthropic deal is vendor financing. AMD committed up to $5B of investment in Anthropic to secure the 2 GW — equity rather than guarantees, but the same circular pattern seen at AVGO
- Everything lands in H1 2027. Helios is shipping initial units only; Anthropic, Cerebras and the Microsoft expansion all sit beyond this fiscal year
- Valuation is the bear case. ~64x FY26E consensus EPS of $7.51 and ~35x FY27E of $13.87 at $482. The stock fell 13% from $552 despite a record quarter, on a GAAP gross margin miss (54% vs ~56% expected) caused by Helios ramp costs
- CUDA's moat still holds for training. ROCm 7 closed most of the inference gap; training remains NVIDIA's
- Memory cost inflation is now pressuring Client and Gaming — Micron's pricing power arriving as AMD's cost. Gaming was −31%
- China ~20% of revenue under tightening export controls. Warrant dilution of ~10%+ from the OpenAI structure
- Perpetual #2 chasing an annual product cadence: any MI400 slip or N2 yield issue de-rates the stock hard
Major customers
- OpenAI — 6 GW deployment agreement, potentially ~$100B of revenue, plus a warrant for up to 160M shares (final tranche vesting at $600)
- Meta — 6 GW
- Anthropic — up to 2 GW of MI450 in Helios, first gigawatt H1 2027, with AMD investing up to $5B in Anthropic
- Microsoft — Helios on Azure at scale, EPYC VM series, Pensando DPUs. Also a semi-custom console customer
- Oracle — named Instinct customer
- Dell, HPE, Lenovo, Supermicro — server OEM channel for EPYC and Instinct
- Sony — semi-custom console silicon, on a cycle unrelated to AI
- Client and PC OEMs for Ryzen
⚠️ Customer names come from public announcements; revenue percentages are not disclosed.
Read-through. AMD's datacentre line tracks the same hyperscaler capex as NVDA's but from a much smaller base, so it is a share-gain story rather than a demand story — the useful test is whether hyperscalers name AMD alongside NVIDIA in capex commentary, not whether capex is rising. The Helios ramp drives the same suppliers NVIDIA's does: VRT for thermal, the photonics names for interconnect, and MU for HBM4 — and that memory relationship runs both ways, since Micron's pricing power is showing up on AMD's own cost line. MI400 being first to TSMC 2nm makes AMD an early read on TSM's N2 yields. The console segment moves on a separate cycle and mutes the AI signal in the headline numbers.
What would change the view
- Helios ramp through Q3 into the Q4 step-up, plus yield commentary — the single execution variable
- Blended gross margin against the ~56% guide, given that the growth segment dilutes it
- Customer diversity beyond the three anchors, in Q4 and Q1
- Server-CPU supply easing, since that is what has to carry the margin blend
- Any softening of the 2027 framework language — the guidance is specific enough that hedging it would be a signal
- ROCm gaining traction in training, not just inference
- China licence status and export-control changes
Update log
19 Sep 2026 — Converted to the standard template. Removed the superseded 6 July section, which was still carrying Q1 figures, a $552 price, a "~39x forward" multiple and an obsolete thesis-tracker table alongside the current Q2 view. Noted a discrepancy: the 7 September note quoted a forward P/E of ~83x, which does not reconcile with the consensus table on the same page ($482 ÷ FY26E $7.51 = ~64x). Using 64x FY26E / 35x FY27E and flagging it rather than repeating the 83x. Dominant Risk property updated accordingly.
7 Sep 2026 — Verdict unchanged: Constructive. Read-through refreshed: Anthropic, Cerebras and Microsoft all land in H1 2027, which is when the thesis is tested.
5 Aug 2026 — Page rewritten after Q2. New disclosure that data center AI gross margins run below the corporate average, and that memory cost inflation is pressuring Client and Gaming.
4 Aug 2026 — Q2 FY2026 reported. Revenue $11.5B (+50%), Data Center $6.7B (+107%, 58% of revenue), non-GAAP GM 56%, non-GAAP EPS +82%, operating income $3.1B. GAAP GM 54% versus ~56% consensus — the miss that drove a 7% selloff. Q3 guided ~$13B. Anthropic 2 GW announced with up to $5B of AMD investment. 2027 framework made public. TAMs raised: accelerators to ~$1.4T and server CPU to ~$220B by 2030.
Consensus EPS path
Estimates, not results. Analyst count falls from 30 to 10 by 2029 and the 2030 range is a 1.9x spread — treat 2026–27 as estimates and 2028+ as a scenario. None of these models embed a downturn.
Year | EPS avg | YoY | Range | P/E @$482 | #analysts |
2025 | $3.96 | — | $3.92–4.13 | 122x | 32 |
2026E | $7.51 | +90% | $6.76–8.53 | 64x | 30 |
2027E | $13.87 | +85% | $8.58–19.14 | 35x | 29 |
2028E | $19.46 | +40% | $12.97–28.41 | 25x | 16 |
2029E | $27.11 | +39% | $19.83–38.42 | 18x | 10 |
2030E | $38.56 | +42% | $28.21–54.65 | 12.5x | 10 |
Five-year CAGR ~58%. Consensus crosses management's ">$20 EPS" target in 2028. Always check which base year a forward multiple uses — $552 ÷ FY27E gave "~39x" in July while $482 ÷ FY26E gives "~64x" now, and they describe the same company.
Probabilities: strengthen ~55% / weaken ~40% — the widest weaken probability in the Bellwethers, on the rich multiple, single-point MI400 execution, the CUDA moat and China.
Research and education only — not investment advice.
Made with Bullet