Theses›Datacentre & AI compute›AI accelerator silicon

AI accelerator silicon

General-purpose AI accelerator silicon (a new GPU, or a horizontal training/inference ASIC competing with NVIDIA / Google TPU / AMD / the well-funded ASIC challengers) is a real, enormous, fast-growing market whose durable value is captured by listed incumbents (NVIDIA…

  • Bear
  • Shut
  • $1163B by 2035

The callRestraints win: NVIDIA absorbs every generation's gains and a challenger needs $100M before revenue. The market goes to $1.16T and almost none of it is open.

D1D2R1R2R3C1foragainst D1D2R1R2R3C1foragainst
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 restraints.
  1. D1Inference splits by phase and workload, so one GPU design stops being right for all of it and a chip shaped for one job can win on cost per token.
  2. D2Routing layers stop forcing every model through CUDA, which erodes the software tax that keeps buyers on NVIDIA.
  3. R1NVIDIA absorbs each generation's gains into its own roadmap, so a challenger's advantage lasts one cycle.
  4. R2First silicon costs $100M or more before any revenue, and the hyperscalers already build their own designs.
  5. R3The durable private value in AI compute goes to the sidesteps: packaging, memory, interconnect and power, not the chip.
  6. C1EU Chips Act 2.0 political agreement with EU-added-value rules steering critical-sector chip buying toward European suppliers; it funds a challenger without making it competitive.Q2 2027
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
20272028202920302031today12watch 20272028202920302031today12watch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1Any European general-purpose AI accelerator startup founded 2024 or later raises €50M or more; that would reopen this.Dec 2027 · open
  2. 2Any such company reaches a venture-scale outcome on the merchant product; I expect not.Dec 2030 · open
  • ·A post-transformer architecture that devalues CUDA; the other reopen signal.
Expressionslisted, private, technologies
  • 688256.SSCambricon
    • full
     China's NVIDIA alternative.
  • AMDAMD
    • full
     MI300/MI350 is the only credible non-NVIDIA training silicon at scale.
  • AVGOBroadcom
    • cheap
     Custom ASICs for Google/Meta + networking silicon.
  • GOOGLGoogle
    • cheap
  • NVDANVIDIA
    • cheap
     Dominant share of AI training compute.
  • MRVLMarvell Technology
    • cheap
     Designs custom AI accelerator and networking ASICs for hyperscalers (AWS, Microsoft).
  • INTCIntel
    • full
     Sells Gaudi AI accelerators and datacentre GPUs to clouds and enterprises.
Private · 10 private companies, 7 of them in the US; none in continental Europe.

EVAS Intelligence · Tenstorrent · Esperanto Technologies · Tenspresso · Cerebras Systems · Groq · SambaNova Systems · d-Matrix · Etched · Graphcore

16 more private companies tracked, not shown

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

$8B raised by 7 companies, $6.3B 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 top-tier venture firm: Etched, Cerebras Systems, Graphcore, d-Matrix. Backed by a strategic investor: Groq, SambaNova Systems, Cerebras Systems, Tenstorrent, Graphcore, d-Matrix.

What changedthe record, newest first
  • 21 Jul 2026An investor in one of the big AI chip startups gives the commercial reality: of roughly fifteen companies that set out to challenge Nvidia, three remain, all were pushed off training onto inference, and as of early 2025 none had meaningful revenue. Deals with national labs and sovereign programmes should not be mistaken for commercial traction.
  • 17 Jul 2026The best-funded independent AI-chip challenger ended 2025 as a ~$20B asset sale to NVIDIA, not a standalone outcome: merchant-accelerator value keeps re-pooling at the incumbent.
  • 1 Jun 2026The 'agents need new silicon' thesis is no longer contrarian: the agentic-inference lane already has well-funded incumbents, so a new entrant's edge has to be architecture, not timing.
Reading
Show the reading