Theses›Datacentre & AI compute›Temporal compute

Temporal compute

Time-domain compute stays a lab technique: no temporal-logic design ships a qualified production part or exits above $100M by 2031 (watch)

  • Bear
  • Shut
  • $20B by 2035

The callPulse-timing arithmetic wins on min, max and sorting, which inference is not made of; inference waits on memory, so free arithmetic barely moves tokens per watt.

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 moves to the edge, billions of always-on devices running inference on a battery, and digital NPUs spend energy swinging voltages for every bit of every operation.
  2. D2Time-encoded arithmetic runs on a mature 28nm-plus node, so a first product is cheap: a design on a shuttle.
  3. R1Inference waits on memory, so free arithmetic moves tokens per watt by about 1.03x, and every hybrid that closes the multiply-accumulate gap gives the gain back.
  4. R2Time-domain logic was first shown in 2014 and there is still no shipped part; as of 2 Oct 2026 the leading company cannot give tokens per second or per dollar.
  5. R3Any use ships as a block inside Arm, Cadence or Synopsys IP, and Graphcore, a much bigger novel-compute bet, sold for a reported ~$600M.
  6. C1Time-domain logic does dense multiply-accumulate at competitive accuracy without falling back to digital, which would make it a general edge block.2031
Rests onthe beliefs underneath
Supported by · confidence
Threatened by · confidence
What would prove itdated tests, open and resolved
202720282029203020312032today1234567predictionswatch 202720282029203020312032today1234567predictionswatch
Fig. 2Each ring is a dated test: dashed while open, struck through when overdue, ticked or crossed once resolved.
  1. 1The UK AI Hardware Plan's £150M advance purchase awards: whether any goes to a time-domain design rather than photonic, in-memory or digital.Jun 2027 · open
  2. 2No time-domain design demonstrates a material gain on a BUYER metric against a digital baseline at equal accuracy, before end-2029.Dec 2029 · open
  3. 3Any time-domain design shows more than 1.5x on tokens per second per user or per batch against a digital part at equal accuracy.Dec 2029 · open
  4. 4No time-domain / race-logic design ships a qualified production part before end-2031.Dec 2031 · open
  5. 5No company whose core architecture is time-domain compute is acquired, or marked, above $100M before end-2031.Dec 2031 · open
  6. 6Any commercially deployed time-domain arithmetic appears as a block inside a conventional digital chip (an ingredient) rather than as a standalone architecture, through 2031.Dec 2031 · open
  7. 7A qualified production part, an acquisition or round above $100M, or a standalone processor rather than a block.Dec 2031 · open
Expressionslisted, private, technologies
  • GFSGlobalFoundries
    • cheap
  • NVDANVIDIA
    • cheap
     Dominant share of AI training compute.
  • SNPSSynopsys
    • full
     EDA duopoly with Cadence.
  • UMCUnited Microelectronics
    • cheap
  • QCOMQualcomm
    • not screened
     Hexagon NPUs inside Snapdragon phone, PC, automotive and IoT SoCs, sold to device makers.
  • INTCIntel
    • full
     NPUs in Core Ultra PC processors and edge inference silicon, sold to PC and industrial OEMs.
  • GOOGLGoogle
    • cheap
     Edge TPU (Coral) and Tensor phone SoC NPUs, for its own devices and embedded developers.
  • 2454.TWMediaTek
    • cheap
     APU neural accelerators inside Dimensity phone and IoT SoCs, sold to device makers.
Private · 6 private companies, 2 of them in the UK.

Temporal Computing · Mythic · Rain AI · Graphcore · Hailo · Axelera AI

4 more private companies tracked, not shown

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

$766M raised by 2 companies, $375M 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: Hailo.

What changedthe record, newest first
  • 27 Jul 2026A chip that makes arithmetic 16x cheaper and leaves the memory path alone moves the customer's number by roughly nothing. In decode, arithmetic is about 3% of the energy. Ask which of the two buyer metrics a claim moves before you read the efficiency ratio.