Financial Shenanigans

Financial Shenanigans

Verdict — Watch (32/100). The reported numbers look like a faithful representation of economic reality: no restatement, no SEC investigation, no auditor flag, clean unqualified E&Y opinion. But FY2025 GAAP earnings carry unusually high one-time content — an $853M release of uncertain tax positions flipped the effective tax rate negative, a $360M reversal of a prior MI308 inventory write-down padded Q4 gross margin, and a $577M "other income" line is dominated by mark-to-market gains on long-term investments. The 13-test scan returns ten green and three yellow categories — no red — but underwriting should haircut FY2025 earnings quality before sizing.

Headline forensic KPIs

The cash-conversion picture is the strongest single piece of clean evidence: CFO has run 1.8× net income on a 3-year stack and receivables grew 2% against 34% revenue growth, so headline strength is not being built on credit terms. The principal yellow flags are not in the operations — they sit in tax, "other income," and the inventory-reserve reversal, all of which boost FY2025 GAAP earnings in ways unlikely to repeat in FY2026.

The 13-shenanigan scorecard

Zero red, three yellow, ten green. The yellows cluster around a single theme: FY2025 headline GAAP earnings benefit from low-quality items (tax-position release, reserve reversal, mark-to-market gains) and from a one-time CFO boost from the seven-month ZT Manufacturing round-trip. None constitute manipulation; all reduce the run-rate signal in the reported numbers.

The three things flattering FY2025 earnings

Three discrete items added an estimated $1.5–2.0B of after-tax benefit to GAAP earnings that are unlikely to repeat. Stripped out, FY2025 net income converges toward the underlying run-rate rather than the reported $4.27B.

The $853M tax benefit is the single largest item: in April 2025, the IRS granted reasonable-cause relief on dual consolidated losses, allowing AMD to release uncertain-tax-position reserves. The management discussion is explicit that this is a one-time effect — the FY2026 outlook in the Q4 call assumes a 13% non-GAAP effective tax rate, vs the FY2025 GAAP rate of negative 2.5%. The Q2 MI308 inventory was written down $800M; Q4 brought back $360M as licenses came through. AMD's own critical-accounting language acknowledges this asymmetry: "if in any period we are able to sell inventories that had been written down… related revenue would be recorded with a lower or no offsetting charge to cost of sales resulting in a net benefit to our gross margin." The $577M "Other income" line was up $396M YoY on unrealized gains in the private-investment portfolio — a real GAAP item but one with no economic linkage to AI/CPU operating performance and zero predictability.

Cash flow quality — strong but the headline overstates it

CFO of $7.71B in FY2025 looks exceptional, and most of it is real, but two adjustments matter. First, $1.22B of that CFO comes from discontinued operations — ZT Manufacturing was held inside AMD for roughly seven months before being sold to Sanmina, and its operating cash inflow during that window flows through consolidated CFO. Second, the cash invested to buy ZT ($2.0B net of cash acquired) hits investing activities, while the proceeds from the sale ($1.4B net) hit investing activities for discontinued operations. The cleanest economic measure for FY2025 is continuing-operations CFO ($6.49B) minus continuing capex ($1.01B) = $5.48B underlying FCF, and FCF after all acquisitions of $4.94B.

The walk matters because the headline gap between FY2024 FCF ($2.41B) and FY2025 FCF ($6.70B) implies 178% growth, while the underlying acquisition-adjusted figure ($1.86B → $4.94B) is 166% growth — still excellent, but should be the benchmark for FY2026 modeling, not the headline.

Receivables, inventory, and the working-capital signal

Receivables growth running well below revenue growth is the single most positive forensic data point in the FY2025 file. DSO compressed from 75 days at FY2024 year-end to 66 days at FY2025 year-end, indicating customers paid faster, not slower — the opposite of channel-stuffing pattern. Inventory days, however, tell a different story: DIO is 142 days, the highest in our 10-year series, driven by an intentional build to support MI350 ramp and advanced-node CPUs. That is operationally rational but creates write-down risk if AI demand softens or if export controls bite again.

The FY2024 spike in AR growth (43% vs revenue +14%) was the single ugliest period — and it was concentrated in Q4 FY2024, where the management discussion says "$1.9 billion increase in accounts receivable driven primarily by higher revenue in the last month of 2024." The pattern reversed in FY2025, which is what a non-manipulator looks like: when collections catch up, they catch up. If management were stretching, FY2025 AR would have grown with revenue or worse.

Goodwill, intangibles, and the Xilinx legacy

The post-Xilinx balance sheet (Feb 2022 close, $48.8B all-stock deal) carries $25.1B of goodwill and $16.7B of net intangibles. That is 54% of total assets in soft form — material exposure to a future impairment trigger if the Embedded segment (Xilinx's home) continues to underperform. The Embedded segment posted a 3% revenue decline and a $178M operating-income drop in FY2025.

AMD's impairment test is "qualitative" rather than quantitative — the company concluded the carrying value of each reporting unit does not exceed fair value, but did not publish headroom calculations. This is standard practice and PCAOB-compliant. It is also where the next material write-down, if one ever comes, will surface. Embedded operating income of $1.24B against an estimated allocated soft-asset base of $10B+ implies very modest cushion. A 20% decline in Embedded long-term cash-flow expectations could trigger a quantitative test.

Non-GAAP hygiene — material gap, recurring exclusions

AMD reports both GAAP and non-GAAP results. The non-GAAP gap is unusually wide and most of it is not genuinely non-recurring.

The $2.3B amortization of acquired intangibles is real economic cost — AMD paid $48.8B for Xilinx and the intangibles amortization is part of the recovery of that price. Excluding it makes adjusted earnings look stronger than the deal economics support. The $1.64B SBC is a recurring real cost — the equity granted in FY2025 will dilute shareholders unless offset by ongoing buybacks, and AMD's net dilution (SBC $1.64B vs buybacks $1.92B) is roughly balanced in FY2025 but was negative in earlier years. The published non-GAAP EPS of $4.17 should not be the anchor for valuation work; building a "core" EPS that adds back only the truly one-time items (DCL release, MI308 reversal, mark-to-market) is more useful and lands roughly between GAAP $2.67 and non-GAAP $4.17 — call it $3.20–3.50 of underlying run-rate FY2025 EPS.

The OpenAI warrant — pending dilution and revenue-recognition complexity

In October 2025, AMD issued OpenAI a warrant for up to 160 million shares (~10% of share count) at a $0.01 strike, vesting against deployment milestones for the MI450 series and AMD stock-price targets. None vested in FY2025; AMD says the warrant had no impact on FY2025 financials. Two forensic considerations follow.

First, ASC 606 generally requires share-based consideration paid to a customer to be recorded as a reduction of revenue once vesting becomes probable. As deployment milestones approach, AMD will likely recognize a non-trivial revenue reduction (or contra-revenue charge) tied to the vested portion. The accounting will be GAAP-compliant but optically painful — non-GAAP could exclude it, widening the GAAP vs non-GAAP gap further. Second, the 160M-share maximum dilution against a ~1.63B float is roughly 9.8% — a meaningful future GAAP EPS drag if the program scales fully. Nothing here is shenanigan; it is, however, a known future earnings-quality headwind worth pricing today.

Stock-based compensation cadence

SBC at $1.64B in FY2025 is 4.7% of revenue and 21% of CFO. Buybacks ($1.92B) just offset SBC dilution in FY2025, but the trend on net dilution has been mixed across the post-Xilinx period.

Net share repurchase is roughly neutral relative to SBC at recent run-rate, but if the OpenAI warrant begins to vest while buybacks stay around $1.5–2.0B/year, net dilution will turn negative. Underwriting should model this explicitly.

Breeding ground — governance and incentive read

The breeding-ground signals are mixed-to-favorable and dampen rather than amplify the yellow flags above.

The structural conditions for accounting manipulation are mostly absent: independent audit committee with strong financial credentials, no related-party transactions, no auditor issues, no reverse-merger or shell-company history. The two yellow tints — combined Chair/CEO and equity-heavy comp tied to non-GAAP — are common features of every other US semiconductor mega-cap, not anomalies. The breeding ground here does not raise the probability that the three yellow scorecard items mask something more serious.

Peer-relative read

AMD's cash-conversion ratios sit in the healthy middle of the peer set — better than NVDA (which holds heavy operating profit close to net income), worse than MRVL/INTC (whose CFO/NI ratios are flattered by depressed net income). Soft-asset weight at AMD (54%) is lower than AVGO (78%) and MRVL (69%) but materially higher than NVDA (~4%) — reflecting the Xilinx acquisition. This is a sector-normal profile, not an outlier.

What to underwrite next

The work to track over the next four quarters, in order of materiality to the forensic verdict:

What would downgrade the grade (toward "Elevated"): any FY2026 reversal of a previously released tax position, a quantitative goodwill impairment on the Embedded reporting unit, a new restricted-customer write-down without disclosure of the offsetting reserve mechanics, or insider sales by Su or Hu accelerating into a positive guidance event.

What would upgrade the grade (toward "Clean"): FY2026 effective tax rate landing at the 13% guide without further one-time items, inventory days normalizing toward 110–120, no contra-revenue surprise on the OpenAI warrant, and continuing-operations CFO/NI staying above 1.5×.

Bottom line for the PM. The accounting risk is not a thesis breaker. It is also not a footnote. The right place for it is in earnings normalization — model FY2025 underlying earnings at roughly $3.20–3.50 of "core" EPS rather than the reported GAAP $2.67 or the reported non-GAAP $4.17, and benchmark FY2026 against that base. The risk does not warrant a position-size cap, but it does warrant skepticism of the headline FY2025 growth rate when projecting FY2026 — much of the upside in the GAAP line was tax, not operations.