Variant Perception

Where We Disagree With the Market

The sharpest disagreement, stated plainly. At ~22x EV/sales, consensus is pricing the Instinct franchise on a "scale-into-corporate-margin" template — the implicit model that merchant-GPU revenue ramps, Xilinx amortization runs off, and gross margin walks toward NVIDIA. The evidence on the page contradicts every leg of that template over the next 12-18 months. CFO Jean Hu has told the Street MI450 ramps below corporate-average gross margin for the next one-to-two years; the OpenAI penny-warrant is not just dilution but an ASC 606 contra-revenue mechanism that increases as Instinct ships volume; and AMD's MI450 unit ceiling is capped by a TSMC CoWoS allocation it does not control. Consensus FY27 EPS at $13.10 does not reconcile with what management has already disclosed. The variant is not that the AI story is wrong — it is that the Street has mis-modelled the shape of the cash that comes from being right. The watchlist number: Data Center segment GAAP gross margin in the Q3 FY26 print — the first quarter MI450 mix materially builds. Sub-52% is bear-confirming; 55%+ is bull-confirming.

Mapping Consensus — What the Market Actually Believes

Variant perception is only credible when "the market believes X" is anchored to a concrete consensus signal — a multiple, a target, a guide, a price reaction, a repeated narrative. The table below converts the sentiment into a testable underwriting assumption, because an assumption is something the next print either confirms or breaks. A vibe is not.

Two read-the-room points before the ledger. First, the bull and bear pages already capture the direction of disagreement — the variant edge here is sharpening the mechanism, especially the under-modelled ASC 606 contra-revenue path on the OpenAI warrant. Second, the most useful disagreement on this page is not a fresh argument for the bear; it is a reframing of which numbers the consensus model has wrong, and over what horizon they become legible.

The Disagreement Ledger — Where Evidence Disagrees With Consensus

Three disagreements survive all five tests (consensus exists, evidence contradicts, materiality is real, observable resolution path, falsifiable). Ranked by what would most update a PM's underwriting today.

Classifying the three variants against the eight high-quality buckets

None of the three lean on the banned weak forms ("high-quality but undervalued," "market too pessimistic," "execution risk remains"). Each is anchored to a mechanism the next two prints will either confirm or refute.

The Lead Disagreement — Walked Through

If the PM only reads one section, this is the one.

The chart makes the central observation visual. Revenue and gross margin variants are small in percentage terms — yet EPS is 16% below consensus because the tax-rate and share-count deltas magnify modest top-line and margin haircuts into a large bottom-line gap. This is why the variant is monetizable: even a directionally-correct revenue and margin view gets to a sharply different EPS once the second-order accounting items are layered in.

What consensus would say. "Mechanical Xilinx amortization run-off (~6pp over 5-7 years) closes most of the GAAP / non-GAAP gap; Instinct gross margin migrates to corporate average by FY28 as scale leverage kicks in; the warrant is one-time dilution priced into the diluted share count; FY25 tax rate is a discrete-item issue that doesn't recur." All four claims are directionally correct — and all four under-estimate the magnitude and timing in the way the bridge above quantifies.

What our evidence disagrees on. The CFO has explicitly committed to sub-corp-avg GM during the MI450 ramp; ASU 2025-04 was specifically clarified in 2025 to address share-based consideration to a customer as contra-revenue when vesting is probable (forensics-claude.md, research-claude.md). These are not bear interpretations of management language — they are management language plus FASB clarification. The bull's "mechanical amortization run-off" argument is real over 5-7 years but provides no offset in the FY26-27 window where consensus is being marked.

What the market must concede if we're right. That AMD's consolidated multiple — already inside NVDA's neighborhood at 22x EV/sales — is paying for margin expansion that is delayed at minimum 18-24 months past the consensus EPS path. That delay is sufficient to compress the forward P/E and trigger the same setup that produced the -17.3% Q4 FY25 reaction off a soft guide.

The cleanest disconfirming signal. Two consecutive quarters of Data Center segment GAAP gross margin holding above 55% while Instinct revenue scales meaningfully. The verdict-claude.md page named this as the bull-confirming signal; we agree. If Q3 FY26 prints DC GM at 55%+ with visibly higher MI450 mix, the lead variant breaks.

The Bullish Variant — EPYC + Embedded + Console Are Floor Franchises

This is the variant the bears miss, not the bulls.

The bear case names a downside target of ~$270 and the long-term-thesis bear scenario lands ~$165. Both pricing exercises under-value the three franchises above. EPYC alone — at the management TAM and conservative 15x segment OI — is worth ~$250-300B standalone; Embedded $25-35B; semi-custom contracted cash flow $30-50B PV. The floor is closer to ~$300-400B of equity value even where Instinct's multiple collapses to volume-only economics. That is roughly $180-240/share against today's $521. The asymmetry: the variant is not "AMD is cheap here" — it is "the floor is real, and the cleanest setup to underwrite is the next 30-40% drawdown when the bear gets paid on the lead variant while the floor franchises trade as cyclical drag."

What consensus would say. "Consolidated multiples are how the market prices conglomerates; isolated SOTPs don't pay you in a re-rating event." Fair point, but it is exactly the moment the consolidated multiple compresses that SOTP gets reactivated — see Intel post-FCF-collapse where the foundry was valued separately precisely because the consolidated multiple broke.

What would prove us wrong. A Q3 FY26 EPS disappointment that triggers a re-rating and coincident weakness in EPYC share data from Mercury Research — i.e., the floor franchises themselves start to crack at the same time the AI multiple compresses. If EPYC share rolls below 42% in any Mercury report before Intel 18A Panther Lake is in server volume (a thesis-breaker per long-term-thesis-claude.md F4), the floor argument loses the franchise that does most of the work.

The Quiet Variant — ASC 606 Contra-Revenue on the OpenAI Warrant

This is the accounting mechanic that consensus is uniquely positioned to miss because it sits at the intersection of revenue recognition and customer-equity treatment, with the relevant FASB clarification (ASU 2025-04) only codified in 2025.

The variant here is not "the warrant is bad" — bulls and bears both know the warrant exists. The variant is that the warrant's revenue/margin mechanic is asymmetric to the bull narrative: the more deeply the deal works on the demand side, the more the GAAP top-line and gross margin compress, and the more tranches vest on the back of the stock-price escalator. Consensus has it inversely: bulls treat the warrant as a one-time dilution cost paid in exchange for revenue; the accounting says the warrant compresses revenue and dilutes shares in proportion to the bull case landing.

What consensus would say. "Meta signed without a warrant — proof the warrant was a one-time cost of entry to anchor the first hyperscaler. The mechanic doesn't recur." That is the bull's strongest counter, and we partially agree: the Meta deal is a positive signal that AMD did not have to pay a recurring tax. But it does not unwind the existing OpenAI warrant — that accounting plays out through Oct 2030 regardless of what Meta did. The variant is about the modelled magnitude of the existing warrant, not the precedent for the next deal.

The disconfirming signal. A clean Q3 / Q4 FY26 10-Q footnote that quantifies the ASC 606 treatment in a smaller-than-feared range (sub-$0.5B / Q in first disclosure), OR an 8-K-disclosed third hyperscaler anchor at standard pricing that prices Instinct against a clean comp. Both would shrink this variant materially.

Evidence the PM Can Audit Fast

The strongest items in the report that move the probability of these three variants. The fragility column matters most — every piece of evidence has a way to become misleading, and an honest PM should know what it is.

Resolution Signals — What Settles the Debate

For each variant, the specific observable that proves it right or wrong, where to look, and on what timing. Every signal below is in a filing, an earnings call, a vendor disclosure, or an industry data publication — none rely on "execution" or "time will tell."

Red Team — What Would Make Us Wrong

The discipline of variant perception is to name the evidence that breaks the view before the market does. The following four reads of the same evidence base would make this page wrong.

Of the four, risk #1 is the most genuinely fair pushback — AMD's guidance posture has been mechanically conservative on the DC GPU dollar guide twice this cycle (the three-time-raised FY24 DC GPU $4B → $4.5B → $5B+ pattern is the receipt). A Q2 print at guided GM with strong Q3 outlook would meaningfully weaken the lead variant. Risk #2 is the second-most fair — probability assessment under ASC 606 has discretion, and AMD may stage recognition slowly. The variant survives risk #3 and #4 even if both materialise: tax planning at 6-8% buys back ~$0.50 of the bridge but the GM and warrant legs remain; hyperscaler capex resilience and ROCm progress support the long-term case but do not unwind the FY26-27 mechanical compression.

The Closing Pointer