AMD’s rack business: how much growth reaches each share?

Moving from chips to rack-level offerings lets AMD sell more content into each deployment. It also increases engineering and funding obligations. The shareholder case rests on incremental gross profit, cash collection and dilution as much as on AI revenue growth.

Revenue grows faster than cash conversion

Whether AMD’s expanding rack business benefits common shareholders depends on gross-profit dollars left after R&D, system support, working capital and customer equity incentives. A modestly lower margin can accompany higher earnings. If upfront investment and the share count rise alongside sales, however, the per-share result needs to be recalculated.

The second quarter offers an encouraging but incomplete starting point. The table retains the financial statements' original units and distinguishes profit margins from profit dollars.

Source: AMD Q2 earnings release, August 4, 2026: income statements and cash-flow reconciliations. Gross margins are rounded as reported by the company. Free cash flow is operating cash flow from continuing operations less purchases of property and equipment.

Calculated from the reported figures, adjusted gross profit rose 14.1% sequentially, while free cash flow fell 39.3%. Both can be true: the point at which new business starts generating profit need not coincide with the payments required to fund the next stage of growth. One quarter of weaker cash flow does not invalidate product competitiveness, but rising revenue does not explain away every cash outflow either. Subsequent deliveries must turn prepayments, inventory and receivables back into cash.

The year-over-year improvement also reflects the comparison base. Q2 2025 included USD 0.8 billion in inventory and related charges associated with MI308 export restrictions. Adding back only that charge to the quarter's USD 3.326 billion of non-GAAP gross profit, then dividing by USD 7.685 billion of revenue, produces approximately 53.7%, well above the reported 43%. This is not a restatement of the accounts. It does show why the entire year-over-year gross-margin improvement cannot be attributed to pricing on new products. AMD earnings release: prior-year comparison footnote

The different economics of CPUs, GPUs and racks

AMD’s products do different jobs: EPYC supplies server CPUs, Instinct supplies accelerators, and Helios combines compute, networking and software in a rack. Its July configuration included 72 MI455X GPUs, 18 Venice CPUs, Pensando networking and ROCm. Customers gain a more complete deployment offering; AMD must turn the additional integration work into profit.AMD, July 23, 2026

Commercially, CPUs can earn revenue from both general-purpose servers and workloads supporting AI. GPUs address more of the expansion in AI compute, while rack-scale solutions may increase AMD's content per deployment and reduce the customer's own integration costs. These are not three separate revenue pools that can simply be added together. CPUs and GPUs inside Helios already belong to their respective product families. Nor can rack shipments be multiplied by an assumed selling price and treated as AMD revenue without checking the contracts.

A rack-scale strategy also does not mean AMD is becoming a server contract manufacturer again. It sold the ZT manufacturing business in October 2025 while retaining design-related teams and intellectual property, a distinction documented in its Q1 2026 Form 10-Q. AMD Q1 Form 10-Q, Note 5, May 6, 2026 A manufacturing partner's entire rack sales therefore cannot be assigned to AMD, and a conventional system manufacturer's gross margin cannot simply be applied to it. The scope of recognized revenue depends on the transaction's allocation of product control, procurement responsibilities and acceptance obligations.

Cost pressure remains real. As systems scale, memory, packaging, networking, validation and customer-support requirements may all increase. We distinguish at least two causes of lower gross margins: a shift toward lower-margin products, and deteriorating pricing or unit costs within the same product. The first can accompany higher total profit; the second generally deserves greater concern. Public segment disclosures do not allow the two to be separated precisely. An assumed Instinct margin must not be presented as a disclosed figure.

Customer economics and supplier margins are also different measures. A system that adds memory or networking components may deliver more useful compute per dollar and become more attractive to buyers, while requiring AMD to pay more for components upfront. That customer value becomes supplier profit only if the performance advantage supports pricing, increases volume or reduces delivery and servicing costs. Pricing and cost recovery still stand between performance leadership and gross-margin expansion.

CPU and GPU growth should not be assumed to diversify risk automatically. If both depend on the same large customers' budgets, packaging resources or data-center delivery schedules, a single delay can affect both product lines. A broader product portfolio reduces earnings volatility only when demand, supply and sources of cash collection are sufficiently diversified as well. Assessing the business mix requires asking how many independent purchasing decisions support growth, rather than simply counting product names.

On the August earnings call, CFO Jean Hu told Barclays analyst Tom O'Malley that the server business supports corporate gross margins, while data-center AI carries a margin slightly below the corporate average but adds gross profit dollars. Lisa Su scheduled initial Helios shipments for Q3, with a subsequent ramp. Q2 results therefore cannot yet establish the earnings profile of rack-scale deliveries at substantial volume. AMD official Q2 earnings-call transcript, pages 8–9

The operating leverage inside long-range EPS forecasts

A single analyst’s forecast can serve as a test of operating assumptions. Investing.com reported on September 10 that Piper Sandler’s David O’Connor expected 50% revenue and 65% EPS compound growth over FY2026–2030, reaching $53 of EPS in 2030. The full model was unavailable. The analysis below tests those assumptions rather than treating them as consensus.Report

The mathematics is worth examining. Across four annual intervals, revenue reaches 1.5⁴, or 5.06 times its starting level, while EPS reaches 1.65⁴, or 7.41 times. With an unchanged share count, the net profit margin would also need to improve by approximately 46.4% in relative terms to explain the difference. If the share count rises, the required margin improvement is greater. Supporting such a long-range earnings forecast takes more than evidence that racks can be sold. It also requires expenses to grow more slowly than revenue, unit economics to avoid material deterioration, and dilution to remain under control.

A different version of the bullish case is more willing to accept lower gross margins. In a September TipRanks article, Marty Shtrubel reported that Truist analyst William Stein cut his Q4 2026 gross-margin estimate from 55.4% to 55.0%, while his 2027 model assumed 54.5%. Stein remained positive on the data-center business. The article linked near-term pressure to an increasing mix of MI455 and rack-scale AI systems. An accompanying newsletter post was published on September 17; this was not a formal update to AMD's quarterly guidance. TipRanks article by Marty Shtrubel, TipRanks accompanying post

This is not a bull-versus-bear split between the two analysts. The distinction concerns the profit structure through which growth is expected to emerge. O'Connor's long-range growth rates require substantial operating leverage; Stein's model accommodates lower gross margins. Our counterargument to both is that system revenue may bring more purchased components, customer-specific engineering and after-sales obligations. Expenses may not scale as efficiently as they did under a chip-sales model. Revenue forecasts could be met while EPS falls short.

That counterargument has limits too. Engineering work that is completed once and reused can create genuine operating leverage. If customers adopt a common architecture, incremental revenue may not require a proportional increase in staffing. The greater concern is extensive new customization for each large customer, or persistently high support costs after delivery. R&D returns should be judged by how many subsequent orders a design can serve, rather than treating higher R&D spending as inherently negative.

Valuation must reflect these distinctions. Suppose future EPS is 20% below a model's estimate and the P/E multiple investors apply falls from 30 times to 25 times. Implied equity value becomes 0.8 × 25 ÷ 30 of the original model's value, a decline of roughly one-third. This is a valuation sensitivity, not AMD's current P/E or a price target. Without a same-date share price, a consistent earnings definition and a complete forecast model, an old reported multiple cannot establish whether the shares are cheap or expensive today.

How much growth offsets a lower margin?

AMD’s own third-quarter guidance offers a nearer reference: revenue of $13 billion plus or minus $300 million and a non-GAAP gross margin near 56%. The call added adjusted expenses of $3.65 billion, $55 million of other net income, a 13% tax rate and 1.66 billion diluted shares. These inputs allow a consistent test of margins, costs and EPS.Results; call, page 7

The table uses the guidance midpoint as its reference. The other three columns are our illustrative single-quarter scenarios, not formal forecasts for a particular quarter. Each holds the tax rate, net other income and share count at the assumptions above, excluding additional dilution. The simplified formula is: EPS = [revenue × gross margin − operating expenses + net other income] × (1 − tax rate) ÷ share count.

Source: The first column's inputs come from the company guidance cited above; the remaining scenarios are the author's assumptions and calculations. Calculated EPS is neither company-issued EPS guidance nor a GAAP forecast.

The result shows why a 55% gross margin is not, by itself, a bearish argument. With expenses unchanged, a decline from 56% to 55% requires revenue growth of just 56 ÷ 55 − 1 = 1.82% to preserve gross profit dollars. If expenses also rise by USD 0.25 billion, however, preserving the reference scenario's operating income requires revenue of (3.63 + 3.9) ÷ 55% ≈ USD 13.691 billion, approximately 5.3% above the reference level. The compensation required depends on expenses, not just gross margin.

The third scenario comes closest to the mechanism bulls want to see: revenue growth more than offsets lower gross margins and higher expenses. The fourth illustrates the downside. Revenue does not decline, yet EPS can still be approximately 17% below the reference case. These scenarios do not establish that either outcome is inevitable. They identify what to look for in subsequent results: incremental gross profit, incremental expenses and changes in share count, rather than orders or the number of product launches alone.

This model cannot directly establish free cash flow. R&D includes cash and non-cash expenses; cash taxes have their own payment schedule; and revenue may be recognized before cash is collected. Even an attractive EPS table can leave shareholders waiting longer because capital remains tied up in the business.

There is a further asymmetry that is easy to miss. If revenue falls below plan, expenses may not fall in tandem within the same quarter. Chip-design, software-maintenance and customer-support teams are already in place, and inventory may have been purchased against the original plan. A valuation downside case that reduces expenses proportionally with revenue can therefore understate risk. Equally, an upside case should not treat all incremental gross profit as net income: higher volumes often require additional testing, support and supply-chain investment.

Working capital and warrants shape per-share returns

Second-quarter free cash flow was about 56.4% of non-GAAP net income. In the first half, receivables, inventory, prepayments and other assets absorbed $2.505 billion, while payables and other liabilities provided $2.231 billion, leaving $274 million of net absorption. The cash-flow statement shows that growth relies on both collections and supplier terms; balance-sheet movements cannot directly replace these cash figures.Q2 10-Q

Supplier-payment timing is helping to support expansion. Reasonable payment terms are a normal operating tool, but cannot be extended indefinitely. If payables come due before inventory completes customer acceptance, the gap between earnings and cash can widen. Conversely, revenue and collections following an inventory build would demonstrate that procurement served actual demand. An increase in inventory is neither automatic evidence of unsold products nor a guarantee of future growth.

An independent illustration shows the significance of payment terms. Assume USD 13 billion in quarterly sales, evenly distributed across a 91-day quarter. A ten-day delay in average collections would tie up approximately USD 1.429 billion more in receivables, calculated as 13 ÷ 91 × 10. It would leave current-period earnings unchanged in this simplified example while directly reducing cash. AMD's actual sales need not be evenly distributed, and this is not a forecast of its collection period. It illustrates how a seemingly modest payment delay can affect near-term funding requirements more than a one-percentage-point change in gross margin.

Free cash flow is not necessarily all available for distribution. The Q2 Form 10-Q also disclosed subsequent conditional investment commitments of up to USD 5.0 billion, expected to extend through fiscal 2028, and long-term data-center leases with USD 9.5 billion in aggregate future payments, expected to commence in 2027–2028. AMD Q2 Form 10-Q, Liquidity and Capital Resources These are not current-quarter expenditures and should not be deducted in full from quarterly cash flow. They do, however, require valuation models to account for future uses of capital. Mechanically extrapolating current free cash flow would miss those subsequent funding requirements.

Adjusted EPS also needs to be tested against economic costs. Q2 GAAP and non-GAAP net income were USD 2.297 billion and USD 2.760 billion, respectively. The reconciliation adds back stock-based compensation and acquisition-related amortization, but also removes investment gains and incorporates tax effects. AMD net-income reconciliation Amortization does not represent a current-period cash payment, so excluding it can help assess operations. Stock-based compensation, however, can impose costs on shareholders through additional shares or the cash needed for repurchases. It cannot be permanently added back in the earnings analysis and then ignored in both share-count and cash-flow analysis.

Customer warrants introduce another risk to the EPS denominator. OpenAI and Meta each received warrants for up to 160 million shares at an exercise price of USD 0.01 per share, subject to purchasing, share-price and other conditions. None had vested or become exercisable as of June 27. AMD Q2 Form 10-Q, Note 12 That is not a current status statement as of October, and the warrants cannot simply be treated as shares already issued.

For a purely mechanical maximum-share illustration, assume both warrant packages add their full 320 million shares to a simplified base of 1.66 billion shares, with profit unchanged. EPS would become 1.66 ÷ 1.98 of its previous level, a decline of approximately 16.2%. An actual assessment must also include profit generated by the associated purchases and reflect the timing of vesting in the denominator. This is not a forecast of accounting diluted EPS. It highlights an economic issue: the cost of winning a customer can be paid in equity, so revenue contracts and equity incentives must be evaluated together.

Buybacks cannot eliminate that cost for free. Using cash to repurchase additional shares may stabilize the EPS denominator, but leaves less cash for R&D, debt repayment or distributions. An assessment of equity incentives needs a consistent method: reflect dilution in the share count, or reflect the cash needed to maintain that count in the cash-flow valuation. The cost should neither disappear nor be counted twice. Purchases tied to warrants and the cost of the equity awarded must likewise be assessed within the same scenario.

Delivery must produce earnings and collections

The next two reports would be most persuasive if Helios volumes, gross-profit dollars and collections rise together while expenses remain controlled. That combination can support per-share earnings even with a slightly lower margin. Revenue growth on its own would leave the case for EPS growing faster than sales unproven.

The countercase also has observable tests. First, revenue growth alongside stagnant gross profit would suggest weaker incremental economics than a simple mix-shift explanation implies. Second, if receivables and prepayments continue to absorb cash for two consecutive quarters, with weak cash conversion that cannot be explained by identifiable delivery timing, the cash-realization assessment should be lowered. Third, if customer equity awards increase without enough incremental profit to offset dilution, the contribution of growth to existing shareholders must be recalculated.

Product delivery must also be distinguished from operational customer compute capacity. A rack reaching the customer does not establish that power, cooling, networking and software workloads are all ready. If a contract ties acceptance or payment to those milestones, deployment delays can alter the timing of earnings and cash. This is a contractual risk to investigate, not a claim that AMD bears all of it in the absence of the relevant terms.

The valuation question is concrete: how much does an additional dollar of revenue leave in per-share earnings after components, R&D, capital tied up in operations and equity costs? Lower gross margins do not automatically invalidate the AMD case, while higher gross profit alone does not prove it. The valuation warranted by rack-scale growth should ultimately depend on the earnings and cash generated through sustained delivery.