China Gut-Punch Quarter: Data Center Growth Cratered to +14% as MI308 Charge Bites, But Q3 Guide Above Street Keeps the MI355 Bridge Intact — Maintaining Hold

Results vs. Consensus

Q2 2025 was the quarter where the optics caught up to the operating story. The headline revenue print was a record, but every line below revenue carried the weight of the $800M MI308 charge management telegraphed three months ago. We characterized that charge in our Q1 note as "the optics killer for the Q2 print" — that proved right. The relevant analytical question is no longer whether the charge would land (it did, exactly as previewed) but whether the underlying business excluding the charge is unimpaired. Our read of the print is: yes — the ex-MI308 non-GAAP gross margin was ~54%, in line with Q1's structural run-rate, and the Q3 guide returns reported GM to ~54% in a single quarter. The MI308 trough is one quarter, not a multi-quarter march.

Quality of Beat / Miss

• Revenue: Fully organic. The $7.685B revenue print — a record — came in $265M above consensus despite the MI308 China revenue removal. The composition is the story: Client and Gaming overshot the model badly on the upside (combined +69% YoY), more than offsetting the Data Center deceleration to +14%. This is a different mix of beat than Q1 (where Data Center carried the upside) but the magnitude is similar.
• Margins: Reported 43% non-GAAP gross margin landed exactly at the Q1 guide. The $800M MI308 charge ran through cost of revenue and produced the headline-ugly compression. Critically, ex-charge non-GAAP GM was ~54% — identical to Q1's structural run-rate — confirming our Q1 view that the underlying margin trajectory is unimpaired. The Q3 guide of ~54% is the test: if it prints, the trough is one quarter.
• EPS: The penny miss vs Street ($0.48 vs $0.49) is essentially noise around the $800M charge. The non-GAAP $0.48 is down 30% from Q2 2024's $0.69, but that compare is corrupted by the MI308 charge. Excluding the charge, run-rate EPS would have been roughly in line with Q1's $0.96. The GAAP EPS of $0.54 actually exceeds the non-GAAP figure — an inversion driven by a tax benefit and valuation allowance release that flattered the GAAP P&L. This is technically real income but is non-cash and non-recurring; the operating story is fully captured in the non-GAAP $0.48.

Segment Performance

Data Center: The Trough Quarter

This is where the $800M MI308 charge lives, and the visible damage is severe: revenue grew only +14% YoY (to $3.24B) versus +57% in Q1, and the segment swung from a $932M Q1 operating profit to a $(155)M operating loss. Excluding the charge the segment OI would be approximately $645M — a step down from Q1's $932M, reflecting the absent China revenue contribution. The growth deceleration is not a competitive issue: EPYC server CPU continued its share-gain march with what management called the "33rd consecutive quarter" of YoY share gains, and over 100 new AMD-powered cloud instances launched in the quarter. The deceleration is entirely the China hole plus a normalized comp.

"We expect Instinct revenue to grow year over year in the third quarter, driven by the ramp of MI350 at multiple customers." — Lisa Su, CEO — the single most important guide-related statement on the call

The MI350 series began volume production in June — ahead of schedule — and management reported that "7 of the top 10 model builders and AI companies use Instinct." On the call, Su characterized MI350 customer adoption as "a bit faster than we might have expected." The MI355 specifically was pitched as matching or exceeding Nvidia B200 in critical training and inference workloads, which if true positions AMD's high-volume FY25 SKU as competitive with Nvidia's current shipping platform.

Assessment: Data Center is the trough quarter. The MI308 charge is one-time and the China revenue loss is now baked into the comp. The bridge from here is MI350/MI355 production ramp through Q3 and Q4 — which the Q3 guide implicitly anticipates. CFO Hu's statement that "we continue to drive the market share up compared to Q1" suggests core EPYC + Instinct ex-China ramped sequentially, even as reported segment revenue was nearly flat. We don't change our Data Center 2H thesis on this print; it is uncomfortable but tracking with the Q1 framing.

Client & Gaming: The Upside Surprise

The combined segment printed $3.62B (+69% YoY, +20% QoQ) with $767M of segment OI — a stunning result given that almost no model had Gaming flipping from -30% to +73% YoY in a single quarter. Client at $2.50B is a record and represents the second consecutive quarter of mid-60s%-plus growth, on top of an already-elevated comp. Gaming's $1.12B is the cleanest inflection of the quarter: a 75-point YoY swing from Q1's -30% to Q2's +73%, driven by both Radeon discrete (which had been showing strength in Q1) and semi-custom (where the console SoC business apparently inflected sharper than even AMD expected).

"We don't think a whole lot of that is [tariff pull-forward]; end user consumption is actually quite strong." — Lisa Su, CEO — responding to Thomas O'Malley (Barclays)

The "tariff pull-forward" thesis that some sell-side desks pushed in Q1 has effectively been refuted by Q2's continuation of the strength. Two consecutive quarters of +60s%-plus Client growth, with mix-led ASP expansion (Ryzen 9000 series leading desktop), is a trend not a one-quarter pull-in. The Gaming inflection is more genuinely surprising and likely reflects a combination of next-gen console pre-builds and unusually strong Radeon 9070 series sell-through at retail.

Assessment: Client+Gaming is the offset that made the Q2 revenue beat possible despite the China hole. Both legs need to hold for the Q3 guide to print — and in Client we are now compressing into difficult comps. Watch sequential Client momentum in Q3 closely; the ASP-led story can only sustain mid-60s%-plus YoY growth for so long before the comp gets unforgiving. Gaming's recovery is more durable in our view because it is reflecting a real semi-custom cycle inflection, not just easy comps.

Embedded

Embedded printed $824M (-4% YoY) with $275M of segment OI — almost identical to Q1's $823M revenue / $328M OI. The expected sequential recovery did not materialize as cleanly as management framed three months ago, and the operating margin compressed modestly (~33% vs. 40% in Q1) on mix. Management continues to guide to inflection back to growth in 2H.

Assessment: Embedded is now the laggard relative to its Q1 guide. The recovery framing remains intact but the timing is slipping — not catastrophically, but visibly. Embedded carries 30%+ operating margins so any miss to the recovery cadence has outsized OI implications. We watch this for a Q3 inflection and would mark it down as a small-but-real demerit on this print.

Key Topics & Management Commentary

Overall Management Tone: Notably more confident than the Q1 call. Where Q1's tone was visibly defensive on China (the qualifier "despite" was doing heavy lifting), Q2's tone leans into the MI350 production ramp narrative and the breadth of the operating beat. Su led with "record server and PC processor sales" rather than with regulatory caveats. The tone shift is rational — the worst of the MI308 optics are now in the rearview — but it does set a high bar: management is now implicitly committing to MI355 execution.

The MI308 Charge: One Quarter, Then Behind Us

The $800M charge landed exactly as previewed in the Q1 call. Per CFO Hu, the bulk was work-in-process inventory rather than finished goods — "Majority of them are WIPs. We really don't have that on the shelf." This is a meaningful detail: WIP can in principle be re-spec'd or re-routed if licensing posture changes; finished MI308 destined specifically for China would have less optionality. Su signaled the company will pursue licenses — "We expect to resume MI308 shipments as licenses are approved" — but importantly, no MI308 China revenue is in the Q3 guide.

"Most of our inventory was not in finished goods, so it was work in process." — Jean Hu, CFO

Assessment: The MI308 P&L impact is one quarter. The asset disposition question is partially answered (mostly WIP, some optionality), and the licensing pathway is partially open (management is pursuing). We continue to model zero China MI308 contribution through year-end and treat any license-driven recovery as upside. The fact that the Q3 guide explicitly excludes China MI308 means a Q4 surprise could come from license approvals — that's a free option in the model.

MI350 Series: Production Ramp Started Ahead of Schedule

The single most important positive datum in the print: MI350 began volume production in June 2025, which is ahead of the mid-year schedule communicated in the Q1 call. Su characterized adoption as "a bit faster than we might have expected" and reported that 7 of the top 10 model builders and AI companies are now using Instinct GPUs. This is the variable that the FY25 outlook depends on, and it is tracking ahead of plan.

"MI355 matches or exceeds B200 in critical training and inference workloads." — Lisa Su, CEO

The B200 comparison is the most aggressive performance claim AMD has made on a public earnings call. If the workload-specific claims hold up in production-customer benchmarks over the next 1-2 quarters, the merchant-share narrative gets a meaningful boost — AMD becomes a credible second source for the B200 generation specifically, which is the volume Nvidia platform shipping today.

Assessment: MI350/MI355 is on track and arguably ahead of schedule. This is the central operating positive of the print and it is what justifies the constructive Q3 guide. The risk is the customary one with semis ramps: production execution at scale, yield, and customer qualification cycles can all slip silently between earnings calls. But as of August 6, the variable that mattered most coming out of Q1 is on plan.

Sovereign AI: A Quietly Material New Customer Vector

Vivek Arya (BofA) pressed Su on sovereign AI sizing, and her response was the most expansive AMD has been on this customer category to date: sovereign deployments are "additive" to the hyperscaler opportunity, with "over 40 active engagements globally." This was not in the Q1 framing — sovereign AI was mentioned only in passing — and the size of the engagement count (40+) is meaningfully larger than the bull case scenario most sell-side desks were carrying coming into the print.

"[Sovereign is] additive to the hyperscaler opportunity... over 40 active engagements globally." — Lisa Su, CEO — in response to Vivek Arya, BofA

Assessment: Sovereign AI is becoming a real customer vector for AMD in a way that wasn't fully appreciated three months ago. Government and quasi-government compute build-outs (think national AI initiatives across Europe, the Middle East, and parts of Asia ex-China) are typically dual-vendor by mandate, which structurally favors a credible #2 to Nvidia. The 40+ engagement number is a leading indicator for 2026 revenue rather than 2H 2025, but it does materially de-risk the FY26 picture.

EPYC and Client: The Underappreciated Compounders

Buried under the China optics: EPYC server CPU put up its 33rd consecutive quarter of YoY share gains, and Client posted its second consecutive ~67% YoY growth quarter (with record desktop CPU sales led by Ryzen 9000). Combined with Gaming's surprise +73% inflection, the non-GPU portfolio is the floor under the AMD investment case while the Instinct ramp plays out. The Q1 call had to defend Client growth against tariff-pull-forward skepticism; the Q2 print buries that skepticism with a second confirmation quarter.

Assessment: The EPYC + Client + Gaming complex generated $6.1B of revenue this quarter (~80% of the company) growing in aggregate at well above 30% YoY with strong margin contribution. Even if the Instinct ramp slipped, the floor under the company is meaningfully higher than it appeared three months ago. This is a genuine positive surprise from Q2 that we don't think the post-print drawdown is reflecting.

ZT Systems Integration and Sanmina Partnership

The ZT Systems integration is now manifesting as a concrete commercial structure: AMD announced Sanmina will be the lead AI rack manufacturing partner, taking the data-center systems integration ZT brought in-house and translating it into a production-scale rack-supply chain. This is the "rack-scale capability" we discussed in Q1 going from acquisition closed to commercial reality.

Assessment: This is a 2026/MI400-era story, not a near-term P&L lever. But the Sanmina announcement is a concrete operational marker that the rack-scale capability is real. We don't change near-term estimates on this; we mark it as a positive de-risking of the MI400 / Helios platform narrative for FY26.

MI400 / Helios: 10x Generational Performance Claim

Su called out the MI400 series and the Helios rack-scale platform as expected to deliver "up to a 10x generational performance increase." That is an aggressive headline number and customer enthusiasm continues to be characterized as "very strong" with active engagement on large-scale 2026 deployments.

Assessment: MI400/Helios is the FY26 thesis. Today's Q1/Q2 prints don't depend on it; FY26 revenue does. The 10x performance claim and the "expanded set of customers... for large-scale deployments in 2026" framing are constructive but unverifiable until the platform ships. We treat MI400 as a free option on the FY26 outlook.

Guidance & Outlook

The Q3 revenue guide of ~$8.7B ±$300M is the constructive surprise of the print. The midpoint implies +13% sequential growth and +28% YoY — well above the $8.3B Street consensus. That guide implicitly assumes (a) Data Center bouncing back as MI350 ramps through the quarter (no MI308 charge to suppress reported margin, modest ramp contribution from MI350 on top of the ex-China EPYC + Instinct base), (b) Client and Gaming holding the Q2 run-rate (~$3.6B+ combined), and (c) Embedded inflecting modestly. The math is tight but achievable if MI350 production tracks.

The ~54% non-GAAP gross margin guide is what we needed to see to confirm the Q1 framing of the MI308 charge as one-quarter, not multi-quarter. A clean snap-back to the structural run-rate — with the MI350 mix tilt arguably positive — means the operating leverage story for 2H is intact. CFO Hu credited "strong operational team" efficiency improvements as part of the bridge back to 54%.

Implied 2H ramp: With H1 2025 revenue of $15.13B (Q1 $7.44B + Q2 $7.69B) and a Q3 guide midpoint of $8.7B, full-year revenue tracks toward roughly $32.5–33.5B if Q4 sequential growth is in the high single digits. That places FY25 at approximately +25-28% YoY — consistent with the "strong double-digit" framing management committed to in Q1. The MI355 production ramp is the swing variable.

Street at: Pre-print FY25 consensus was around $32.5B revenue / $4.30 EPS. The Q2 in-line print and above-consensus Q3 guide should leave FY25 estimates roughly in place or slightly higher. The Q3 EPS implied by ~$8.7B revenue at ~54% GM and ~$2.55B opex is roughly $1.10–1.20 non-GAAP — a substantial sequential snap-back from Q2's $0.48 and likely above current Street.

Guidance style: Su's guidance posture has historically been measured. The fact that she chose to guide Q3 revenue ~$400M above Street consensus signals genuine confidence in the MI350 ramp and the Client/Gaming durability. This is a more aggressive guide than the Q1 print delivered, and it materially raises the bar for what Q3 needs to deliver.

Analyst Q&A Highlights

MI350 Production Ramp and Adoption

• Ross Seymore, Deutsche Bank: Asked about MI350 customer adoption pace. Su responded that "Adoption is a bit faster than we might have expected" with production starting in June.
  Assessment: This is the bullish data point of the call. Adoption "faster than expected" on a product that was the FY25 thesis-defining ramp is the best outcome we could have anticipated coming out of Q1.
• Timothy Arcuri, UBS: Pressed on MI350 revenue specifically and timing. Su pointed to production start in June and YoY growth in Q3 driven by MI350 ramp at multiple customers.
  Assessment: Multiple-customer framing rather than concentration risk on a single hyperscaler is constructive.
• C.J. Muse, Cantor Fitzgerald: Asked about the MI355-to-MI400 product handoff. Su confirmed MI355 ramps through 2H 2025 and MI400 is positioned for "significant revenue contribution" in 2026.
  Assessment: Roadmap cadence intact; MI400 is the FY26 swing variable now that MI355 looks tracking.

Sovereign AI and Customer Diversification

• Vivek Arya, Bank of America: Asked about sovereign AI sizing and how it sits relative to hyperscaler demand. Su called sovereign "additive" with "over 40 active engagements globally."
  Assessment: The 40+ number is materially larger than buyside expectations. Sovereign is becoming a real second-leg-of-the-stool for Instinct beyond the hyperscaler concentration that has dominated the Nvidia/AMD GPU narrative.

Client Sustainability

• Thomas O'Malley, Barclays: Asked whether the Client growth had any tariff pull-forward component. Su: "We don't think a whole lot of that is. End user consumption is actually quite strong."
  Assessment: Second consecutive quarter of management firmly rejecting the pull-forward narrative. The +67% Client growth is the second confirmation quarter; we now treat the share-gain dynamic as durable structural rather than cyclical pull-in.

Data Center Market Share

• Aaron Rakers, Wells Fargo: Pressed on Data Center market share progression given the segment's headline deceleration. CFO Hu responded that AMD continues to "drive the market share up compared to Q1."
  Assessment: Important read-through — the +14% Data Center growth includes the MI308 zero contribution; ex-China the underlying share trajectory is still progressing. The headline number is a distraction from the underlying competitive dynamic.

Developer Cloud and AI Software Ecosystem

• Joshua Buchalter, TD Cowen: Asked about the financial impact of the AMD Developer Cloud rollout. Su: Developer cloud "doesn't add meaningfully to revenue in the second half" but is part of the long-term ROCm software ecosystem investment.
  Assessment: A reasonable acknowledgment that ecosystem investment is opex without near-term revenue. This is the right kind of long-term spend for AMD given the CUDA/ROCm gap; we'd be more worried if management were puffing it as a near-term revenue lever.

What They're NOT Saying

1. FY25 explicit revenue guide: Management gave a Q3 number but did not reaffirm or update the "strong double-digit" FY25 framing from Q1, nor provide an implicit FY25 range. This is mildly unusual at this point in the year and suggests management wants flexibility on Q4 given MI350 ramp execution risk and the unresolved China MI308 licensing pathway.
2. MI308 inventory disposition for the WIP component: CFO Hu confirmed the bulk of the charge was WIP rather than finished goods, but did not articulate whether the WIP can be re-routed to non-China markets, re-spec'd into MI355 production, or written off. The optionality on this asset is a free call option on the AMD model that management is choosing not to value publicly.
3. Hyperscaler customer concentration on Instinct: Su disclosed "7 of top 10 model builders and AI companies use Instinct" but continues to avoid revealing the revenue concentration across those 7. The thesis quality differs sharply between an evenly-distributed franchise and one that depends on 1-2 large frontier-model labs. The opacity continues from Q1.
4. Embedded recovery slippage: Q1 framing was "sequential improvement in Q2 and Q3, return to growth in 2H." Q2 printed -4% YoY and roughly flat sequentially — the recovery is slipping. Management did not directly acknowledge the slippage on the call; they continued to point to 2H inflection without recalibrating the prior framing. Watch this.
5. Tax rate / valuation allowance commentary: The GAAP EPS of $0.54 exceeded non-GAAP EPS of $0.48 — an inversion driven by a tax benefit and valuation allowance release that flattered GAAP. Management did not provide a normalized tax rate forward outlook, leaving forward GAAP EPS modeling murky.

Market Reaction

• After-hours move (Aug 5 evening): Stock declined approximately 6.3% in after-hours trading following the press release and conference call. Initial reaction was driven by the EPS miss vs Street ($0.48 vs $0.49 consensus) and the optics of the $(155)M Data Center segment operating loss.
• Next-day intraday (Aug 6): Pre-market opened down ~6% and traded down 6%+ through early Wednesday trading, with the focus on Data Center growth deceleration to +14% (from +57% in Q1) being the dominant headline.
• Volume: Materially elevated vs 30-day average as the Street re-priced the trough quarter and the constructive Q3 guide.
• Pre-print context: AMD had rallied sharply into the print on broader AI-momentum buying — shares had appreciated ~44% off their April lows by the Tuesday Aug 5 regular session close. That higher entry meant the bar for the print to clear was elevated, and a "guide-as-expected with a Data Center optical miss" was always going to be met with selling. Compared to where AMD entered the Q1 print (down ~17% YTD), this print entered with much less buyer cushion.

The 6%+ drawdown is a rational reaction in our view, not an overshoot or an underreaction. The market is processing three things simultaneously: (1) the Data Center deceleration to +14% confirms the China hole bit harder than the bull case framing of "well contained" allowed for, (2) the EPS miss-by-a-penny is mostly noise but carries optical weight, and (3) the Q3 guide is constructive but not so far above Street that it changes the FY25 algorithm by a wide margin. Net: a modest derating that captures the bridge-quarter ugliness without writing off the MI355 ramp thesis. We think this is roughly the correct market read.

Street Perspective

Debate: Is the Data Center deceleration to +14% a temporary trough or a structural reset?

Bull view: The +14% print is entirely the MI308 China hole plus a normalized comp; ex-China the underlying segment is still growing at strong double-digits and CFO Hu confirmed share is still progressing. With MI350 production starting ahead of schedule and 7 of top 10 model builders on Instinct, the Q3 guide implies a clean snap-back as MI350 contributes and the China comp drag stops mattering.

Bear view: A 43-point growth deceleration in a single quarter is a stark visual that will weigh on multiple compression for at least 1-2 quarters regardless of the underlying explanation. The +14% print is also potentially flattering — it includes whatever ex-China Data Center GPU revenue AMD generated, and the MI308 zeroing was already in the Q1 guide. If MI350 doesn't ramp on time, Q3 Data Center could disappoint again and reset the multiple lower.

Our take: Bull view is closer to right on the underlying trajectory; bear view is right that the optical damage takes time to repair. The Q3 print will resolve the question definitively — if Data Center reaccelerates meaningfully (call it +25%+ YoY) the bear case loses its main exhibit; if Data Center prints flat-to-modest growth again, the structural-reset narrative gains traction.

Debate: Has the Client/Gaming surprise raised the floor under AMD?

Bull view: Client+Gaming printed $3.62B at +69% YoY with $767M of operating income — a record-large contribution from the non-AI portfolio. With Client now annualizing $10B+ and Gaming inflecting positively, AMD has a $14B+ non-GPU revenue base growing at strong double digits. That is a higher floor than the bull case carried into the print, and it materially de-risks the FY25 outlook even if Instinct ramp slips.

Bear view: The +69% Client+Gaming growth is unrepeatable. Client comps get materially harder in 2H, Gaming's +73% inflection has a meaningful semi-custom component that is inherently lumpy, and the structurally lower Q1/Q2 comp benefit goes away. Treat this segment's Q2 contribution as a peak rather than a baseline.

Our take: The bear's right that the YoY growth rate compresses, but they're wrong that the absolute dollar contribution is at risk. Client at $2.5B and Gaming at $1.1B may not grow 60-70% YoY in 2H, but at flat-to-up sequential they generate $7B+ of H2 revenue with healthy margins. That is a meaningful floor that wasn't on the bull case before.

Debate: Does the MI355-vs-B200 performance claim hold up?

Bull view: Su's claim that MI355 matches or exceeds B200 in critical training and inference workloads — combined with adoption "faster than we might have expected" — positions AMD as a credible second source for the Nvidia generation that is currently shipping at scale. That changes the merchant-share narrative meaningfully and supports a re-rating of Instinct revenue forecasts for 2026.

Bear view: Public performance claims on AI workloads are inevitably cherry-picked; the real test is sustained customer deployment at hyperscaler-scale and the software ecosystem (ROCm vs. CUDA) gap that has historically constrained AMD's ability to convert performance benchmarks into commercial wins. Wait for hyperscaler-disclosed deployment scale before taking the claim at face value.

Our take: Both have merit; we lean modestly toward giving Su the benefit of the doubt because (a) AMD has been increasingly accurate on roadmap delivery for 2 years running, and (b) the customer count (7 of top 10) and the production-start-ahead-of-schedule data points corroborate the performance claim with operational evidence. But the proof is in the Q3/Q4 prints — specifically whether MI355 contribution is large enough to swing reported Data Center growth materially higher.

Model Update Needed

Valuation impact: Net of the changes above, we'd characterize fair value as broadly unchanged from our Q1 framing, with a slight upward bias from the Client/Gaming and sovereign upside offsetting the Data Center disappointment. At a 22–26x forward P/E on FY26 estimates, fair value lands in the $130–160 range. We don't think the post-print drawdown takes the stock into clearly cheap territory; it removes some of the AI-momentum froth but the policy and execution overhangs are still active.

Thesis Scorecard Post-Earnings

Overall: The thesis is roughly unchanged in net economic value. The Data Center optical disappointment is offset by the Client/Gaming upside surprise and the sovereign-AI new vector. The MI308 charge landed exactly as flagged; the MI350/MI355 production ramp is tracking ahead of schedule; the Q3 guide is constructive. None of these are decisive thesis-shifters individually, and they roughly net out.

Action: Maintaining Hold. The Q2 print delivered exactly the ugly bridge we modeled in Q1 and the Q3 guide gives us the constructive setup we needed — but neither side got a decisive new data point to force a rating change. We continue to need either (a) tangible MI355 production ramp evidence in the Q3 print (Data Center re-accelerating to +25%+ YoY), or (b) constructive movement on US-China MI308 licensing posture, or (c) a 15-20% drawdown without thesis impairment, to upgrade to Outperform. None of those triggers fired this quarter; we stay patient.