Will Marvell have at least 3 custom XPU programs in volume production by Q2 FY2027?

The Q4 FY2026 earnings provide materially positive evidence. Custom revenue hit $1.5B (doubled as guided), socket count grew to 20+, and 'several XPU attach programs' are ramping in FY2027. The critical question is the definition of 'custom XPU programs in volume production.' The resolution criteria says 'at least 3 distinct custom XPU programs generating production-level revenue.' If we interpret this BROADLY to include XPU attach programs (NIC, CXL), which are custom silicon programs tied to custom XPU customers, then 3+ is very likely by Q2 FY2027 — there are already 2 core XPU programs in production plus multiple attach programs ramping to ~$400M in FY2027. If we interpret NARROWLY as only core XPU silicon designs, then we need the 3rd core program to reach volume production by August 2026. The lead customer is transitioning to next-gen XPU (which means the 1st gen continues in production while next-gen ramps — this is one customer with potentially 2 programs). The new Tier 1 XPU program ramps in FY2028, not FY2027. I lean toward a broad interpretation being reasonable since the resolution says 'custom XPU programs' and attach programs (NIC/CXL for XPUs) are custom programs for XPU customers. 68% reflects increased confidence from earnings data but significant definitional uncertainty.

I take a narrower interpretation of the resolution criteria. 'Custom XPU programs in volume production' most naturally refers to the core XPU silicon designs — the 100B+ transistor chips at TSMC 3nm/5nm that are the flagship custom products. NIC and CXL attach programs are accessories to the XPU, not XPU programs themselves. Under this narrow reading: 2 core XPU programs are in production. The lead customer's next-gen XPU is 'transitioning' — this could mean production shipments have begun (which would count) or it could mean the transition is in progress with volume production still months away. The new Tier 1 XPU program explicitly ramps in FY2028, not FY2027. So the question reduces to: (a) does the next-gen XPU transition count as a distinct 3rd program (it's the same customer, next generation), and (b) is it in 'volume production' by Q2 FY2027? CEO said 'stronger second half due to program transition' which suggests the transition happens mid-FY2027, and Q2 FY2027 ends in August 2026 — that IS the second half. This is genuinely uncertain. 55% reflects the narrow interpretation with the transition timing creating a near coin-flip.

Let me think about this from the management disclosure perspective. The resolution source is 'Marvell Technology earnings calls and/or 10-Q filings.' The question resolves YES if management 'confirms at least 3 distinct custom XPU programs are generating production-level revenue.' Management has strong incentive to frame their disclosures positively. If there are 2 core XPU programs in production, a next-gen XPU transitioning, and several attach programs ramping, management will likely present this as 3+ custom programs in production. They already upgraded socket count from 18+ to 20+ and doubled custom revenue. The CEO's comments about XPU attach programs ramping in FY2027 and the 'firm volume requirements' for the new Tier 1 program suggest management views these as distinct programs in their portfolio. Custom revenue growing >20% in FY2027 with 'biased higher' commentary suggests management expects to demonstrate broadening production across programs. I weight this at 72% — the earnings evidence is materially positive and management's disclosure patterns favor counting programs broadly.

The prior prediction of 58% was based on '2 in production, 1 ramping, 15+ in development.' The Q4 FY2026 earnings update shifts the landscape meaningfully. Key new evidence: (1) Custom silicon hit $1.5B — the doubling target was met, confirming execution. (2) Socket count grew to 20+ from 18+. (3) Lead customer transitioning to next-gen XPU — this is the 1st or 2nd program deepening, not necessarily a new 3rd program. (4) Several XPU attach programs ramping FY2027 — these are production programs generating revenue, but whether they count as 'XPU programs' under resolution criteria is debatable. (5) New Tier 1 XPU ramps FY2028 — too late for this market. The net effect is positive but bounded by definitional uncertainty. The attach programs generating ~$400M in FY2027 are substantial production revenue, which makes it harder to argue they don't count as 'production-level revenue' for distinct programs. I move from the prior 58% to 65%, reflecting the positive earnings data while acknowledging the interpretation risk.

I want to carefully parse the resolution criteria: 'at least 3 distinct custom XPU programs generating production-level revenue (not NRE or development revenue).' The word 'distinct' is important — it means 3 separately identifiable programs, not 3 generations of the same program. Current state: (1) Lead customer XPU Gen 1 — in production. (2) 2nd customer/program XPU — in production. (3) Lead customer next-gen XPU — transitioning. (4) XPU attach programs (NIC, CXL) — ramping. (5) New Tier 1 XPU — FY2028. If the next-gen XPU for the lead customer counts as 'distinct' from Gen 1, that's 3 right there (Gen 1, Gen 2 for lead customer, plus 2nd customer). But are successive generations of the same customer's XPU 'distinct programs'? In semiconductor industry practice, each generation IS a distinct tape-out and program, even for the same customer. So I lean toward counting them as distinct. The transition timing is the risk — 'stronger second half' comments suggest Q2 FY2027 (which is the first half of the calendar year but second half of fiscal year) may see initial volume production of next-gen. 62% probability.

I take the broadest reasonable interpretation. The resolution says 'custom XPU programs generating production-level revenue.' XPU attach programs (NIC, CXL) are: (a) custom-designed for specific hyperscaler XPU architectures, (b) generating production-level revenue (~$200M FY2026, doubling to ~$400M FY2027), and (c) distinct programs with separate development cycles. The XConn acquisition adding PCIe Gen 6 and CXL 3.1 with 20+ customers further validates these as real production programs. Under this interpretation: 2 core XPUs in production + at least 1 XPU attach program in volume production = 3+ by Q2 FY2027 with high confidence. The ~$400M FY2027 attach revenue across 'several' programs means at least some are generating meaningful production revenue NOW, with volumes increasing through FY2027. The custom silicon business hitting $1.5B and growing >20% provides the revenue validation. I assign 73% — the broad interpretation is defensible given the resolution language, and the attach programs are generating unambiguous production revenue.

Strong positive update from Q4 FY2026 earnings. Custom revenue $1.5B (doubled), 20+ sockets, multiple attach programs ramping. The prior assessment had '2 in production, 1 ramping' — the earnings confirm the pipeline is converting. Several XPU attach programs generating production revenue in FY2027 makes 3+ programs likely under a reasonable interpretation. The new Tier 1 XPU is FY2028 so it doesn't help this market, but it validates pipeline expansion. Key risk: narrow interpretation excluding attach programs would leave us dependent on the next-gen XPU transition timing, which is uncertain for Q2 FY2027. 70% reflects strong positive momentum with definitional risk.

Maintaining the prior probability level despite positive earnings. Here's why: The original question was designed to test whether 3 CORE custom XPU programs (the 100B+ transistor TSMC chips) reach volume production. The prior context explicitly distinguished between the core XPU programs and accessory products. Attach programs (NIC, CXL) are important revenue streams but they're not 'XPU programs' — they're companion chips. Under this strict reading: 2 core XPUs are in production. The 3rd is either the next-gen for the lead customer (which is transitioning, not yet in volume production by Q2 FY2027 based on 'stronger second half' timing) or the new Tier 1 (which is FY2028). So 3 core XPU programs in volume production by August 2026 is genuinely uncertain. The positive earnings data confirms execution on EXISTING programs but doesn't clearly accelerate the THIRD core program. 58% — essentially unchanged from prior, because the new data doesn't resolve the core question.

Splitting the difference between broad and narrow interpretations. The resolution criteria asks about 'custom XPU programs' — note it says 'custom XPU programs,' not just 'XPU programs.' This suggests any custom program related to XPU architectures could qualify, which would include attach programs designed custom for specific XPU implementations. But the criteria also says 'distinct' programs, and some attach programs (e.g., NIC variants for the same customer's XPU) might not be 'distinct' enough. Looking at management's likely disclosure: they have been counting sockets (20+), not distinguishing between XPU and attach. If management on the Q1 or Q2 FY2027 earnings call says 'we now have 3 or more custom programs in production,' that would resolve YES regardless of the internal distinction. Given management's tendency to present production breadth positively, and the revenue trajectory supporting such a claim, I assign 66%.