2Q26: Servers Carry the Quarter, Instinct the Burden

Thesis status: Staging affirmed. Conversion untested. Framework mis-aimed.

I. Thesis Refresher

This is the first update to the AMD second-platform primer published 4 June 2026. Second-quarter results, reported 4 August 2026, give the first quarter against which the primer's conditions can be tested. The question under evaluation is whether AMD ($AMD) converts the Instinct accelerator cycle into a durable second-platform position in data centre infrastructure at a pace consistent with what the share price requires. The primer answered it through a sequential chain in which supply execution gates the MI450 ramp, and the ramp in turn gates both ROCm adoption and the OpenAI commercial conversion.

The horizon remains FY2026 through FY2028, with FY2029 as the sustained-scale validation point. This memo assesses the quarter against the six watch conditions and the categorical China condition. It also records a framework problem the quarter exposed. The primer set EPYC aside as settled, on the stated basis that its trajectory was not in contention and that the entire gap between the value of management's guidance and the share price belonged to Instinct. The beat and the second-half raise came from EPYC.

II. 2Q26 in Brief

The year-on-year profitability comparisons are distorted, and the two 2Q25 items pull against each other rather than compounding. Revenue, free cash flow and the share count are not mechanically touched by the inventory charge, though the revenue base does carry a China-constrained Instinct quarter. The register below sizes each one-off against the headline it moves.

Management's own comparable figure for non-GAAP EPS growth is +82%, against the +246% the unadjusted arithmetic produces. The GAAP line is distorted in both directions at once. The $800M charge depressed the year-ago operating result, which flatters the growth rate, while the $853M tax release lifted the year-ago net result to $872M on a $134M operating loss, which suppresses it. Strip the release out and the 2Q25 base falls to roughly a cent, so the reported +156% understates the underlying movement rather than overstating it. The sequential comparison carries neither effect: 1Q26 held no equivalent one-off, and against it gross margin added 100 bps and operating margin 290 bps.

One line resists the operating read. Free cash flow fell sequentially on a quarter in which revenue rose 13%, which is a working-capital and capital-intensity story rather than an earnings one, taken up in Section V.

III. The Engine the Conditions Do Not Track

AMD does not disaggregate Instinct from EPYC within the Data Center segment, so the segment's $3,478M year-on-year increase cannot be split from the disclosure. What can be established is the direction of the surprise. Hu attributed segment growth first to EPYC, above 70% year-on-year with record enterprise sell-through. Su put cloud and enterprise server sales each above 70% and stated the result exceeded the outlook given at 1Q26. Instinct more than doubled on MI350-series silicon, against a 2Q25 base constrained by the MI308 export restrictions rather than by the charge itself. MI450 shipped nothing.

The margin record is what matters for the framework. Data Center operating margin clears the FY2024 peak, which is the level C5's commercial-impact test asks FY2027 to reach, and it clears it with Helios not yet shipped and Instinct still on the prior generation. Hu characterised data centre AI gross margin as running slightly below the corporate average, and attributed segment strength to unit and ASP growth on the Zen 5 mix. That points to the CPU line as the principal source of the improvement rather than the accelerator line. The condition built to detect rack-scale integration economics has had its threshold cleared by the business the primer declared settled.

Supply reaches the same conclusion. Su stated the server CPU chain was tight through 1H2026 because much of the demand was unforecasted, and named added wafer, back-end, substrate and component capacity as a reason the second-half outlook could be raised. C2 tracks Instinct supply, and neither of its named mechanisms was a constraint. The constraint that cost revenue sits in a product line C2 does not mention.

The markers split by engine. Su attributed the FY2027 segment doubling to data centre AI ramping and put the server contribution above 70%, so the forward promise remains Instinct-led even where the current beat is not. Agentic AI as a server demand vector did not exist in the primer, and management now expects it to be the largest and fastest-growing slice of the server TAM by 2030. The EPS row reaches into the primer's quantitative section, and Section VI takes it up.

Five of seven conditions read Developing in the assessment below, and in four of them the binding reason is that a named mechanism cannot be read rather than that it is deteriorating. A framework that cannot see the variable setting the results will keep producing that answer.

Adding a condition mid-cycle belongs to the primer restatement rather than to a memo. The overlay below runs from this memo forward, carrying no status and no thresholds, until the FY2026 restatement decides whether any line in it becomes a condition.

Server CPU growth is management commentary rather than a disclosed line, and is carried as stated. Inventory days on total cost of sales and a 91-day quarter. The payables adjustment is derived in Section V.

IV. Watch Condition Assessment

Trend is measured against the primer, since this is the first update. Green denotes thesis support rather than signal direction, so a warning condition that is not firing reads green.

C1 — MI450 production ramp enters volume in 2H2026

🟡 Developing ↑ The schedule the condition names is being met. The revenue leg it also names has no read, and the deployment scale has moved into FY2027.

MI450 Series and Helios launched at the Advancing AI event in July and are in production, with initial shipments guided to begin in 3Q26 and ramp through 4Q26 and into FY2027. The anchor-customer set is the material change. The primer carried one gigawatt-scale customer, and management now describes multi-generation gigawatt-scale deployments with OpenAI and Meta ($META) alongside a new Anthropic agreement for up to 2GW, with Anthropic, Cirrascale, HUMAIN, Meta, Microsoft ($MSFT), OpenAI, Oracle ($ORCL), Tensorwave and Vultr named as Helios deployers.

The cap is the timing. The primer framed the first OpenAI gigawatt as an observable second-half revenue step rather than an FY2027 event. Anthropic's first gigawatt is dated to 1H2027, OpenAI's is dated nowhere in this disclosure, and Su characterised 3Q26 as the very beginning of the ramp, 4Q26 as a step up and 1Q27 as a further one. The condition's schedule holds while the volume behind it arrives later than the primer assumed. One trigger inside the condition also cannot be read in any quarter, because the primer asks for unit or revenue commentary on MI450 specifically and AMD reports Instinct and EPYC as a single segment line. That trigger is permanently qualitative and belongs in the lifecycle log rather than in the grade.

C2 — Advanced node yield and HBM allocation support the FY2026 commitment draw-down

🟡 Developing ↑ HBM visibility reads clean, and MI450 remains on schedule without a disclosed yield problem. Neither the yield outcome nor the commitment draw-down is directly readable at quarterly cadence.

Gross margin cleared the 52.8% threshold the primer set and is guided higher again. HBM was addressed directly rather than by absence, which is the stronger form of evidence for a trigger written on management not citing a constraint. Answering a question on Helios carrying roughly 50% more HBM than competing racks, Su stated there is very good visibility into HBM allocation for 2027, and disclosed a lever the primer did not carry: AMD can modify the memory footprint where total cost of ownership does not justify it. The yield leg is weaker evidence than it first appears. The 2nm and chiplet commentary concerns Venice, which is a server CPU, and MI450's node is not disclosed, so what supports the Instinct side is indirect: the product is in production, shipments hold their schedule, no advanced-node problem has been disclosed, and no charge was recorded against a $100M annual bar. Su also expects Helios yields to improve over the first few ramp quarters, which describes a ramp in progress rather than one already cleared.

The conversion outcome is the problem. Both the $8.5B FY2026 purchase-commitment draw-down and inventory by process node are annual disclosures, so the quantity the condition is named after cannot be observed this period. The one visible proxy moved against, with inventory building ahead of a ramp, which the primer's own noise clause anticipates and which management attributes to data centre demand. The test window for that trigger opens in 2H2026, so 2Q26 neither passes nor fails it. The condition also missed the supply failure that did occur, for reasons set out in Section III. Two of three strengthening triggers cleared and no weakening trigger fired, which is why the trend improves while the status does not.

C3 — ROCm parity reduces enterprise migration friction through FY2027

🟡 Developing ↑ The first specific milestones in this condition's life. All are first-party, and not one comes from an enterprise.

The primer asked for a ROCm capability milestone mapping to an independently verifiable benchmark, and this is the first quarter to produce any. More than 3 million models now run out of the box, leading open models launch with day-zero Instinct support, open-source contributions have risen more than tenfold over the year, and ROCm.ai delivers better than twice the training and three times the inferencing performance of ROCm 7 while letting developers port through Claude, Codex and Cursor. Su characterised the platform as having reached an inflection point.

Every one of those figures is AMD's own. Several are reproducible in principle, which is what the primer's trigger asks for, and none has been shown to have been tested by an independent party. The distinction matters for a condition built to substitute observable proxies for a migration cost nobody discloses, so the trigger is half met. The milestones exist and are specific. Independent validation does not.

The named mechanism still has no read, and the evidence that exists comes from the customer segment least able to stand in for it. Every data point is frontier-lab co-optimisation, including the Anthropic collaboration to accelerate ROCm development using Claude and joint work with OpenAI and Meta. Frontier labs hold the engineering capacity that enterprises lack, which is the entire reason migration cost binds in the enterprise and not at the labs. A tenfold rise in contributions from organisations that write their own kernels says little about a customer who cannot.

Neither of the condition's other two triggers produced anything. No analyst raised enterprise migration friction, so the Q&A test was silent in both directions, and the trigger comparing Data Center growth against hardware-only procurement tailwinds has no denominator and never will.

C4 — OpenAI warrant tranches begin vesting in the FY2026 or FY2027 annual filing

⏸ Dormant → No read on the vesting test, which is structurally unavailable in three quarters of every four. The commitment it proxies widened materially.

Diluted shares of 1,659M carry no warrant shares, and the dilutive impact is attributed entirely to employee equity grants under the treasury-stock method. The 160 million shares at $0.01 remain a 10-K note, exercisable through 5 October 2030 against a six-gigawatt window, with none vested at FY2025 year-end.

The commercial substance moved even though the test could not. Arya put roughly 3GW of 2027 line of sight and at least $15B per gigawatt to management, sourcing the latter to warrant disclosures. Su declined the gigawatt figure and held the revenue marker at double-digit billions per gigawatt, which is the same marker the primer carried. Anthropic extends the template one step further, and the reported AMD investment of up to $5B in Anthropic would invert it by placing AMD's equity in the customer rather than the customer's in AMD. That arrangement appears in no AMD disclosure and is not relied on here.

C5 — Helios rack-scale platform reaches hyperscale deployment readiness in FY2026

🟢 Affirmed ↑ The FY2026 readiness test cleared five months before its deadline. The FY2027 margin test it is paired with can no longer discriminate what it was built to measure.

Helios launched in July combining EPYC Venice, MI450 Series, Pensando networking and ROCm, is in production, and ships from 3Q26, with claimed advantages of up to 15% more throughput at equal rack power and up to 30% more tokens per dollar. Microsoft will deploy it at scale on Azure for frontier model inferencing across its own services, its AI customers and Azure AI, which satisfies the hyperscaler-reference trigger outright. A Cerebras collaboration pairs Helios with the Wafer-Scale Engine for ultra-low-latency inference from 4Q26. Su described customer pull as tracking ahead of initial forecasts and clarified under questioning that the statement concerns volumes rather than yields, with yields expected to improve over the first few ramp quarters on a product she characterised as highly complex. The primer set the readiness deadline at FY2026 year-end and expected confirmation in 2H2026 earnings commentary, so the 4 August call delivered it a quarter ahead of the earliest window the condition contemplated.

The grade covers the leg that was due. The commercial-impact leg is an FY2027 test and its metric has already been cleared by the CPU line, so when FY2027 arrives the segment margin will not separate Helios integration economics from EPYC mix. That is a measurement failure rather than a thesis failure, and it is logged as a refinement rather than allowed to change the status of a milestone that passed.

C6 — Client unit share reaches and holds at or above 40% through FY2027

🟡 Developing ↓ The threshold has no read at all, and the one readable trigger failed.

No unit-share figure appears in the release, the slides or the transcript. Management referenced continued share gains without a percentage, and the condition is written on independent survey data at or above 40% for two consecutive quarters, which requires a third-party source this disclosure does not contain. The margin trigger did produce a read and it failed, at 15.2% against a 19.6% bar, which Hu attributed to go-to-market investment and an expanding product roadmap. That threshold compares a quarter against a full-year ratio, and the equivalent-period comparison against 21.2% a year ago points the same way, so the basis mismatch does not rescue it.

The second-half backdrop also turned, on a cause the condition does not contemplate. Management is planning for a softer PC market as higher memory and component costs weigh on demand, with AMD expected to outperform that market, and Su noted the market has held up better than most expected. Component cost is neither a share problem nor a product problem, and the primer's trigger asking whether AI PC demand is incremental rather than pull-forward went unaddressed.

The status stops short of Weakening because the primary share threshold is unreadable, while Client revenue still grew 23% on record mobile and Ryzen PRO sales up more than 50%, and consolidated liquidity remains ample. The margin signal nevertheless moved against the condition, and the drag inside the segment is Gaming, on semi-custom at this point in the console cycle.

C7 — China regulatory resolution without further inventory charges

🟡 Developing → No further charge, which is the leg that reads clean. The regulatory resolution the condition also names has no read, which caps the status.

No new inventory or related charge was recorded, no reference was made to the 15% revenue-share expectation hardening into the cost structure, and China does not appear in the transcript. On the other half of the condition, no guidance restatement incorporated China AI GPU revenue, so the licence environment has neither tightened nor resolved. AMD continues to exclude that revenue from formal guidance, which the primer identified as both the appropriate discipline and the monitoring mechanism. The condition bundles two independent tests and only one of them can be read, so the multi-component rule caps it at Developing rather than letting the absence of deterioration report as a milestone met. A split is proposed below.

One basis distinction matters for anyone comparing this memo to the primer. The charge taken in 2Q25 was $800M gross, and it netted to $440M across FY2025 after licence-dependent recovery. Both figures are correct on their own basis and they are not interchangeable.

V. Cash Conversion and the Capital-Light Claim

Free cash flow rests on the payables line. Receivables, inventory, prepaid assets and accrued liabilities absorbed $3,004M between them against a $2,274M cash contribution from the increase in accounts payable, and payables have risen from $2,929M at FY2025 year-end. Had payables been unchanged in 2Q26, operating cash flow would have been $92M and free cash flow -$716M. Payables expanding alongside a component build is ordinary and is what a rack-scale ramp should produce, and the movement is consistent with purchase volume and invoice timing rather than evidencing any change in supplier terms. What the bridge does establish is that reported free cash flow depends heavily on payables financing the build. If that source pauses while receivables and inventory keep rising, quarterly free cash flow could turn negative.

Every figure in the bridge is reported. AMD publishes three-month cash-flow columns alongside six-month ones, so none of the quarterly components above is derived by subtracting 1Q26 from a half-year statement.

The balance-sheet positions those flows land in moved further than capex alone records.

Capital intensity may be changing character. The primer described a fabless business concentrating capital at the architectural layer, with R&D as the primary capital commitment and no capital expenditure base to carry. Capex has since roughly doubled as a share of revenue, and 1H2026 ran 142% above 1H2025. Whether that is a step change or a ramp-timing effect cannot be settled from two quarters. What is already established is that the capital-light framing was load-bearing in the primer's account of how AMD funds an annual product cadence against a competitor with greater scale, and that no watch condition carries the question.

Two items sit outside the bridge. Other non-current assets rose $2,452M over six months against $1,197M of capital expenditure and $844M of long-term investment purchases, a movement whose composition is a 10-Q disclosure and is not assessed here. If a material part of it is capacity prepayment, the cash committed to the ramp exceeds reported capex by a wide margin. And a $243M working-capital settlement on the ZT Systems divestiture sits in discontinued investing, outside the free cash flow definition, so the $1,558M headline understates the period's total cash consumption.

Free cash flow to GAAP EBIT stands at 1.19x on a trailing basis, above parity because GAAP EBIT is struck after roughly $2.2B a year of Xilinx intangible amortisation and after stock compensation, neither of which consumes cash. The amortisation rolls off around 2032. Capex is rising. Cash tax scales with profit from a 13% base. Each of the three supports decays across the horizon the primer set, which is why the model behind Section VI steps conversion down rather than holding it at the trailing level.

VI. What the Price Now Embeds

This section characterises the financial conditions the current enterprise value implies. It is not a price target, a fair-value estimate, or a view on whether the shares are mispriced. The reverse solve maps the territory. The watch conditions track whether it is being crossed.

The enterprise value has barely moved while the base beneath it grew 40%, compressing the multiple by roughly a third. That statement needs no discount rate, no terminal assumption and no account of investor behaviour, which makes it the most robust output of the quarter.

The path between those two columns was not quiet, and AMD discloses enough to establish it. The company puts its own average share price for 2Q26 at $398 and its average for 29 June to 24 July at $534, in the share-count appendix it uses to guide dilution. The price has returned to within 0.8% of the primer's strike by way of a wide excursion in both directions.

Beta is the whole of the discount-rate change that matters. Debt is 0.4% of market capitalisation, so the cost of equity is effectively the entire weighted cost of capital, and the price move itself does not touch it. The Blume adjustment has also become load-bearing in a way it was not at the primer: carried unadjusted at 2.49, the weighted cost of capital reaches 15.2% and the reverse solve returns a required terminal margin above 140%.

On the base-case driver path the reverse solve returns a required terminal EBIT margin of 95%. A figure past NVIDIA ($NVDA)'s own operating margin forecasts nothing. It reports that on a base-case growth path margin cannot carry this price, and that growth is the lever the valuation rides on.

The base-case path above is what the reverse solve inverts. The curve below broadens the sensitivity by letting revenue growth and terminal margin trade against one another, rather than resting on the single 95% implied-margin result. It remains conditional on the same linear margin ramp, conversion assumption, ten-year horizon, terminal growth rate and discount rate.

Every row is a consistency result rather than a decomposition. Varying one input at a time identifies a combination the price is compatible with, and establishes nothing about what any part of the price is paying for. Read that way the curve says what the single figure cannot, which is that no point on it is comfortable. The row nearest AMD's own demonstrated ceiling still asks revenue to compound above 30% for a decade. Handing AMD the entire margin structure of the company it is taking share from still requires ten years above 20%.

The primer's cross-check has moved out from under it. Management characterised greater than $20 in annual EPS as a medium-term objective, and the primer discounted that outcome to $180 to $195 per share at its central rate. Management now expects to significantly exceed the target without stating by how much, which removes the anchor rather than raising it.

The discount rate does not resolve the tension. At 10%, the rate the primer labelled its low-risk-execution case, the required terminal margin is 46%, still well beyond anything AMD has recorded. The required margin passes 100% at a 13.41% weighted cost of capital, and the re-struck rate now sits 27 bps below that boundary. Beyond it the solve is degenerate, which is a fact about the model rather than about the company.

The results were released after the close on 4 August 2026, when the shares stood at $525.00, and the shares opened on 5 August 2026 at $494.16, which is the mark this section strikes. The 5.9% fall is recorded and not interpreted, because two sessions establish nothing and neither index nor sector moves are controlled for here.

VII. Thesis Standing

The sequential chain the primer built is intact and, on its product legs, running ahead of schedule. Supply execution has not failed on either named dimension. MI450 is in production on the guided quarter. Helios cleared a readiness test the primer did not expect before FY2026 year-end, with a hyperscaler committed at scale. ROCm has produced its first specific, externally reproducible milestones. Nothing in the chain has broken, and the chain has been staged faster than the primer assumed.

None of it has been tested at volume. MI450 has recognised no revenue, no warrant tranche has vested, and no enterprise has been observed migrating. The primer's schedule holds, with the material deployment volume sitting in FY2027 rather than 2H2026. The quarter established that the staging is real. It could not establish that the conversion is.

The valuation burden has fallen without the thesis being validated, which is the more awkward result. The multiple compressed by roughly a third because the earnings base grew 40% while the price stood still, and that growth came substantially from the business the primer set aside. An investor who bought at the primer's price now pays a lower multiple on the strength of a franchise the primer declared not in contention, while the Instinct outcome the multiple was struck against remains ahead.

The reverse solve cannot say which part of the enterprise value is paying for which business, and does not try to. What it does establish is narrower and sufficient. Demonstrated economics, which are now substantially EPYC economics, do not clear the price at any point on the trade-off curve, which leaves the unvalidated Instinct ramp as a necessary component of the case rather than an optional upside to it.

Proposed refinements

Seven refinements and three candidate additions go forward to the restatement: a server CPU condition, capital intensity and cash conversion, and agentic AI as a server demand vector. No condition is resolved or retired this period.

VIII. What to Watch

Three disclosures sit in the 10-Q rather than the results release, and none is assessed in this memo: the composition of the $2,452M increase in other non-current assets, any movement in the FY2026 purchase commitments, and any change to the warrant note.

Source: AMD second quarter 2026 earnings release, conference call transcript and earnings slides, 4 August 2026, for the period ended 27 June 2026. FY2025 Form 10-K for purchase commitments and warrant terms. Damodaran implied equity risk premium, 1 August 2026. US ten-year Treasury yield, 5 August 2026. Daily news digest items, 1 to 28 July 2026.