Theses›Robotics & physical AI›Edge AI inference accelerators

Edge AI inference accelerators

Robots need big models on a battery and today's chips waste energy moving weights, and the one device approach that computes inside stacked memory has thin evidence so far; early.

  • Bear
  • Early
  • $9.6B by 2035

The callThe restraints win today: the robot socket is new but NVIDIA's Jetson prices it, the digital compute-in-memory companies are funded at 40x the memcapacitor bet, and the device evidence is a 2021 lab result. If this is wrong the market goes from $2.5B to about $9.6B by 2035, with about $1.3B of it open to newcomers.

Disclosure: SEMRON, a company developing this technology, pays me to advise it and to write sponsored editorial about it.

D1D2D3R1R2R3C1foragainst D1D2D3R1R2R3C1foragainst
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. D1Humanoids and mobile robots must see, plan and act on board on a battery, and a phone NPU is too small while a Jetson is priced like a small GPU.
  2. D2Both spend most of their energy moving weights between memory and compute, so a chip that computes where the weights are stored wins on energy per task.
  3. D3The humanoid market alone is forecast at $38B by 2035, a new buyer of inference silicon.
  4. R1The memcapacitor evidence is thin: 156 devices at 50x50 micron measured in 2021, one energy figure, and analogue-to-digital conversion at every block caps performance.
  5. R2The market has voted for digital compute-in-memory with 3D stacking: d-Matrix has about $450M raised against roughly €10M for the memcapacitor bet.
  6. R3Car sockets are locked for years, phone NPUs are in-house, and Jetson prices the robot socket, so the open part is robots and cameras only.
  7. C1Mobile and humanoid robots reach perception-module volume, opening a merchant inference socket outside NVIDIA worth about $160M at the start.2027-28
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. 1A memcapacitor maker tapes out a product-grade CapRAM chip with a named partner or raises a Series A.Jun 2027 · open
  2. 2EDGE PART: CapRAM wins a named edge always-on evaluation or design engagement judged on the edge KPI by end-2028 — an OEM, MCU vendor, or module maker publicly evaluating/designing-in CapRAM at mJ-at-task metrics.Dec 2028 · open
  3. 3A high-volume design win or 1M units a year, and a named third party evaluating CapRAM for an always-on edge socket.Dec 2028 · open
  4. 4DATACENTRE SLICE: by end-2029 CapRAM demonstrates a full-SYSTEM tokens-per-watt or TCO-per-million-tokens result, under a stated interactivity SLA, competitive with the then-shipping SRAM decode benchmark.Dec 2029 · open
  5. 5€10M of revenue and a full-system tokens-per-watt figure against shipping SRAM silicon; d-Matrix taking the socket first would change the view.Dec 2029 · open
  6. 6Memcapacitor remains a credible standalone device class through 2030, not subsumed by digital-CIM + 3D-stacking (d-Matrix Raptor/3DIMC) before it reaches volume.Dec 2030 · open
Expressionslisted, private, technologies
  • WBT.AXWeebit Nano Limited
    • speculative
  • MBLYMobileye
    • cheap
     EyeQ ADAS and automated-driving perception SoCs sold to tier-1 suppliers and carmakers.
  • AMBAAmbarella
    • full
     Edge-AI vision SoCs (CV series) sold to security-camera, automotive and robotics device makers.
  • BZAIBlaize
    • not screened
     Graph-streaming AI inference processors and edge servers sold to system integrators and public-sector buyers.
  • NVDANVIDIA (Jetson / DRIVE)
    • cheap
     Jetson robot/edge modules and DRIVE automotive computers sold to robot, edge-box and car makers.
  • QCOMQualcomm
    • not screened
     Integrated NPU SoCs (Snapdragon Ride, IoT, Cloud AI 100 edge) sold to car, device and camera makers.
Private · 11 private companies, 6 of them in the US.

Rain AI · Mythic · Syntiant · EnCharge AI · Horizon Robotics · Black Sesame Technologies · Hailo · Axelera AI · SiMa.ai · DEEPX · Kneron

2 more private companies tracked, not shown

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

$4B raised by 9 companies, $1.6B 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: EnCharge AI. Backed by a strategic investor: Horizon Robotics, Hailo, DEEPX, EnCharge AI, Syntiant, Kneron.

1 more funded company in the market.

What changedthe record, newest first
  • 14 Aug 2026When a whole category of chip startups starts failing at once, the founder still standing will tell you the market has vindicated his approach. Sometimes that is right. But the post-mortem that actually holds — the models outgrew the on-chip memory, so the data had to move again and the advantage vanished — points at whoever solves capacity cheaply, not at whichever storage device is most elegant. That is still an unanswered manufacturing-cost question, and a rival's funeral does not answer it.
  • 10 Jul 2026Datacentre chips are bought on cost per token, edge chips on energy inside a thermal envelope. A device pitched at both is really two bets with two different falsifiers; the physics may transfer but the proof points do not.
  • 10 Jul 2026A device's headline efficiency number can carry a 149x gap between the simulated best case and the same paper's no-assumptions case, before any chip-level tax. The measured number, on a device 1,000x larger than production geometry, is the only fact; everything else is a stated assumption you are being asked to underwrite.
2 earlier
  • 17 Jun 2026A novel low-power compute device can have the right physics and still lose the edge socket, because rivals on duller, already-manufacturable memory reach volume in the same market first. The differentiated bet then narrows to one question: is the manufacturing-cost edge real, and can it ship before the boring options take the design-ins.
  • 16 Jun 2026When a market moves to the "boring" binary version of a memory technology, that is a verdict on the whole analog-multilevel approach, not an opening for whichever analog device is least flawed. A rival storage element that is also analog inherits the same headwind, plus its own scaling wall, and wins only if a separate manufacturing edge (cheap vertical stacking) turns out to be real.
Reading
Show the reading