“The custom AI ASIC state of play”

Tom's article

TrendForce puts ASIC-based AI servers at 27.8% of shipments in 2026, the highest share since 2023, with custom ASIC shipments growing 44.6% year-over-year against 16.1% for merchant GPUs. GPUs still carry the bulk of the market at 69.7%. Luke James's Tom's Hardware survey gets those figures right and frames the stakes correctly: the question isn't whether custom silicon takes share from Nvidia, but how fast it erodes Nvidia's pricing power.

The spine of the piece is Broadcom and Marvell. Broadcom's side is airtight. The $8.4 billion in Q1 FY2026 AI revenue (up 106%), the $73 billion backlog, and Hock Tan's “$100 billion in 2027” line all match the 8-K and the earnings call verbatim. Worth adding the context the article skips: Mizuho models roughly $40.4 billion in FY2026 AI revenue, so the 2027 target is about a 2.5x jump off the full-year base, not the 3x you'd infer from annualizing the Q2 guide. Either way it's a stretch goal Tan put on the record.

The article says Marvell “projects up to $11 billion in AI ASIC revenue for 2026.” That isn't supported. Marvell reported $8.195 billion in total FY2026 revenue and guided Q1 FY2027 to $2.4 billion plus or minus 5%; the ~$11 billion figure in analyst coverage is a company-level FY2027 trajectory across optical, switching, and custom silicon, not ASIC revenue. Marvell doesn't break out custom ASIC cleanly, but Oppenheimer pegs it at roughly $2 billion in 2026, led by AWS Trainium, with a second large program (Microsoft Maia, a reported $2 billion purchase order) ramping late in the year. Conflating that with whole-company guidance overstates Marvell's ASIC scale several times over, and it matters because the next sentence leans on it.

That next sentence is the “Broadcom and Marvell control roughly 95% of the co-design market” claim, and it deserves a flag too. Public estimates vary widely and don't converge on 95% as a settled number. Counterpoint Research, as cited by The Next Web, is read as roughly 60% Broadcom and 25% Marvell by 2027; Counterpoint's own published figure has Broadcom at about 60% design-partner share with Marvell falling to around 8% by 2027. Those don't reconcile to 95%, and the spread between them is itself the point: this is a Broadcom-dominated duopoly, but the precise division is unsettled, and the article presents a high-end round number as fact. The rest of the survey is solid reporting. The architectural details (Ironwood's 4,614 FP8 TFLOPS and 192GB HBM3E, Trainium3 as AWS's first 3nm chip, Maia 200's delayed and bumpy path through OpenAI-requested design changes) all track with the primary coverage. The Dojo postmortem is handled honestly, including Musk's “all paths converged to AI6” line.

Which brings up the thread the article gestures at but doesn't land, and it's the one that should be the thesis: TSMC packaging allocation, not design-partner share, is the practical governor on how fast any of this happens. James notes CoWoS capacity is now a more binding constraint than wafer fab, with Nvidia holding ~60% of allocation and Broadcom ~15%. Every chip in the piece, Ironwood, Trainium, MTIA, Maia, needs CoWoS or CoWoS-L for HBM integration, and they're all drawing from the same constrained pool Nvidia already dominates. The TCO math (up to 65% advantage for inference at scale) explains why hyperscalers want custom silicon. The packaging allocation explains why the 44.6% growth figure is a forecast and not a guarantee. Demand makes the case for custom silicon; packaging decides the pace.

Drafted with AI assistance against parallel reporting.

Sources:

• Tom's Hardware Premium, “The custom AI ASIC state of play (May 2026),” Luke James, 21 May 2026 (primary).
• Broadcom Q1 FY2026 8-K and earnings release (SEC filing): corroborates the $8.4B AI revenue, $73B backlog, and $100B 2027 figures; Mizuho FY2026 estimate via secondary coverage.
• TrendForce, January 2026 AI server shipment forecast and October 2025 “NVIDIA Scale-Up” insight: the 27.8% ASIC share, 69.7% GPU share, and 44.6% vs 16.1% growth figures.
• Oppenheimer note via Investing.com, plus Marvell FY2026 results and Q1 FY2027 guidance: the ~$2B custom-ASIC vs ~$11B company-level trajectory distinction.
• Counterpoint Research (via The Next Web, and Counterpoint's own published design-share figure): the competing reads on the Broadcom/Marvell split that undercut the 95% claim.