Theses›Datacentre & AI compute›Hyperscaler ASICs

Hyperscaler ASICs

As hyperscalers in-source AI silicon (Google TPU, AWS Trainium/Inferentia, Microsoft Maia, Meta MTIA), the AI-compute profit pool partially migrates off the merchant GPU.

  • Bull
  • Closing
  • $84B by 2035

The callCustom silicon becomes the normal way to buy AI compute; the profit pools at design houses and packaging chokepoints, and a software-native design company could take part of it.

D1D2D3R1R2R3C1C2foragainst D1D2D3R1R2R3C1C2foragainst
Fig. 1The balance. Drivers and catalysts against restraints, each block sized by its weight; catalysts hang dashed because they have not happened yet, sized by their odds. It tips toward the drivers.
  1. D1Every model lab, neocloud, carmaker and sovereign programme wants a chip shaped to its workload; Marvell puts its datacentre opportunity at $94B by 2028.
  2. D2Only a handful of co-design houses can take a chip through physical design and signoff, so everyone below the hyperscalers is priced out; software could do it in months.
  3. D3Hyperscaler programmes ramp 2026-2028 with CoWoS-class capacity binding now, which opens packaging, interposers, substrates and memory test to second sources.
  4. R1The margin may not stick: GUC's gross margin fell from 32% to 25% on turnkey mix, Alchip runs 26%, and Hock Tan says AI ASICs dilute Broadcom's margin.
  5. R2If NVIDIA holds the rack with NVLink and Spectrum-X, the merchant pool may not shrink much at all.
  6. R3The sizing is contested: a 15% owner share against the 5% ingredient cap moves the answer 3x, and a first product is 3nm silicon with CoWoS packaging before revenue.
  7. C1Google's 9th-gen TPU goes into production on Intel's EMIB-T packaging with MediaTek doing the co-design: the first flagship hyperscaler chip built outside TSMC CoWoS and outside Broadcom or Marvell.2027
  8. C2In-house ASICs at 35% or more of new AI compute and NVIDIA's merchant share under 65%: the checkpoint that says the share shift is on track.mid-2027
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
202720282029today123456predictionswatch 202720282029today123456predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Near-dated checkpoint: by end-2027 EITHER (a) the Google–Intel TPU advanced-packaging arrangement is confirmed in production (not just an order/MOU), OR (b) Broadcom + Marvell disclosed custom/AI-silicon run-rate exceeds the 2026 baseline by >50%.Dec 2027 · open
  2. 2Google's 9th-gen TPU in production on Intel EMIB-T with MediaTek co-design.Dec 2027 · open
  3. 3By end-2028 hyperscaler in-house AI ASICs account for ≥35% of new datacentre AI-accelerator compute capacity added that year, with NVIDIA's merchant share of new AI-accelerator deployments falling below ~65%.Dec 2028 · open
  4. 4Over FY2026-FY2028 the combined AI/custom-silicon revenue of the merchant design-service houses grows at a faster CAGR than NVIDIA data-center revenue.Dec 2028 · open
  5. 5By end-2028 at least one hyperscaler custom AI accelerator ships in volume with advanced packaging sourced OUTSIDE TSMC, breaking TSMC's effective sole-source on leading-edge AI packaging.Dec 2028 · open
  6. 6Broadcom plus Marvell custom revenue growing faster than NVIDIA's datacentre line; one hyperscaler chip shipping in volume with packaging from outside TSMC.Dec 2028 · open
Expressionslisted, private, technologies
Private · 7 private companies, 4 of them in the US; none in continental Europe.

Google Cloud · Intel Foundry · Alphawave Semi · SEMIFIVE · ASICLAND · Eliyan · Baya Systems

1 more private company tracked, not shown

Technologies
Who has already bet on itprivate rounds, from public sources

$454M raised by 5 companies, $241M of it in the last three years. Wave under way: Large rounds in the last three years; the category is funded and getting crowded.

Backed by a strategic investor: Eliyan, ASICLAND, Baya Systems.

What changedthe record, newest first
  • 22 Oct 2026Nvidia is funding the buyers who will never build their own chip, AI clouds and smaller labs, because its biggest customers are all moving onto their own silicon.
  • 22 Oct 2026One broker says Google''s next TPU is three weeks late on packaging yield at Intel, which would make packaging the gate at a second foundry, not just TSMC.
  • 22 Oct 2026If Intel''s packaging yield on a flagship external customer is 45-50% in late 2026, the second source of advanced packaging arrives later than the plans assume.
6 earlier
  • 22 Oct 2026Google''s warrant to Marvell pays out per $500M of purchases and commits to none of them.
  • 26 Jul 2026The demotion was an artefact of two bad inputs, not of the market.
  • 23 Jul 2026Google Cloud margin doubling to ~36% while revenue grew 82% is the owns-the-chip profit pool showing up in the P&L: make, do not rent.
  • 3 Jul 2026A lab designing its own chip is really renting a merchant design house and a foundry wafer — the custom-silicon wave pools rent at Broadcom/Marvell/Alchip and TSMC packaging, on schedule.
  • 23 Jun 2026The custom-AI-chip profit pool is real but contested: MediaTek just took Google's next big AI-chip order from Broadcom, a reminder that backing the chip-design houses is a sector bet, not a one-stock bet.
  • 21 Jun 2026The AI-compute trade is no longer just "own NVIDIA". As hyperscalers design their own chips, a parallel profit pool is forming in the houses that help them design those chips (Broadcom, Marvell) and in the packaging/memory chokepoints everyone still shares.