Theses›Defence & security›FHE acceleration silicon

FHE acceleration silicon

Fully homomorphic encryption is moving from ~10,000x-slower-than-plaintext toward practical via dedicated silicon (FHE accelerators), the one confidential-computing primitive that needs new chips rather than incumbent-owned TEEs.

  • Bear
  • Shut
  • $6.8B by 2035

The callEncrypted AI needs a chip that moves ciphertext far faster, but Intel, NVIDIA and Apple are folding FHE into silicon they already sell before a standalone market forms.

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. D1AI is moving onto patient records, bank transactions and government files whose owners cannot hand them to someone else's model in the clear; enclaves still require trusting the chip vendor.
  2. D2FHE runs about 10,000x slower than plaintext because every operation drags ciphertexts through memory, so a processor built for it has a real job to do.
  3. R1Intel, NVIDIA and Apple fold FHE into silicon they already sell: Heracles graduated straight out of DARPA DPRIVE, NVIDIA does it in software, Apple ships it in iOS 18.
  4. R2Standard GPUs are closing the gap, with Zama at 1,000 confidential transfers a second on eight H100s.
  5. R3Every standalone market figure is under $1B and estimate-grade, no FHE acquisition or listing above $50M is on record, and no regulation mandates FHE.
  6. C1Niobium's production FHE silicon on Samsung 8nm ships to a hyperscaler or sovereign cloud at 1,000+ units, the first standalone accelerator volume.end-2028
Rests onthe beliefs underneath
Supported by · confidence
What would prove itdated tests, open and resolved
20272028202920302031today123456predictionswatch 20272028202920302031today123456predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1The EU AI Act high-risk date, and whether any regulated-data customer commits to FHE-accelerated production at scale.Dec 2027 · open
  2. 2>=1 of Nvidia / Intel / a top-3 hyperscaler ships a GA FHE-acceleration product materially undercutting standalone-accelerator economics (~65%).Dec 2027 · open
  3. 3NVIDIA, Intel or a top-3 hyperscaler shipping a general-availability FHE acceleration product that undercuts standalone accelerators; I put that at about 65%.Dec 2027 · open
  4. 4No independent FHE-accelerator startup reaches a $1B valuation or a $500M+ exit (~70%).Dec 2028 · open
  5. 5A photonic FHE result at production precision, or Niobium shipping 1,000+ units; either would change the view.Dec 2028 · open
  6. 6The standalone HE hardware market stays under $1B annual revenue through 2030 (~75%).Dec 2030 · open
Expressionslisted, private, technologies
  • AAPLApple
    • expensive
  • INTCIntel
    • full
     Foundry comeback bet (18A, 14A).
  • NVDANVIDIA
    • cheap
     Dominant share of AI training compute.
  • MSFTMicrosoft
    • full
     The open-source SEAL homomorphic encryption library and confidential-computing services on Azure.
  • IBMIBM
    • full
     HElib and the HE toolkit plus FHE consulting services for banks and healthcare.
Private · 7 private companies, 5 of them in the US.

Zama · Fabric Cryptography · Cornami · Chain Reaction · Niobium Microsystems · Niobium · Duality Technologies

1 more private company tracked, not shown

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

$331M raised by 5 companies, $119M of it in the last three years. Wave building: Money is arriving: recent rounds, but no runaway leader yet.

Backed by a strategic investor: Duality Technologies.