Corsair by d-Matrix: status and timeline

Corsair

Inference accelerator using compute-in-memory for the decode phase. With NVIDIA GPUs for prefill: 10x faster inference, 80% lower power.

DeployStage 4 of 5

Full production June 2026; shipping in volume to hyperscalers and frontier labs.

Updated 9 Jun 2026Checked 10 Oct0 updates this week

Milestones

Next · Volume shipments to hyperscalers and neoclouds

Company founded2019Complete.
Seed fundingMar 2021Complete.
Series B funding ($110M)6 Sep 2023Complete.
Series C funding ($275M)12 Nov 2025Complete.
Corsair enters full production9 Jun 2026Complete.
Volume shipments to hyperscalers and neocloudsTarget Q3 2026Current milestone.

Most important updates

Upcoming

Q3 2026Volume shipments to hyperscalers and neoclouds (next)

Current obstacles

Software ecosystem integrationInference stacks built for GPUs. Inserting Corsair requires compiler and runtime changes.
Deployment cost and complexitySecond accelerator class increases operational burden: firmware, drivers, power, thermal.

Physics limits

Not for compute-bound prefillCorsair specializes in memory-bound decode. Prefill (processing the prompt) is compute-bound, where GPUs or specialized training hardware (like Trainium) excel.
Capacity memory latencyWhen model size exceeds SRAM, Corsair must use LPDDR5 capacity memory, which is ~10x slower. Large models or high batch sizes can hit this penalty.

How it works

3 parts

Decode bottleneck

Why decode needs different hardware

Prefill reads prompt in parallel (GPU-friendly). Decode generates one token, reading cache repeatedly; memory bandwidth limits.

Compute-in-memory

Math inside the SRAM array

Multipliers and accumulators in SRAM rows. Operations happen where data lives; no movement to CPU.

Heterogeneous inference

GPU prefill + Corsair decode

Request: GPU prefill → transfer KV cache to Corsair → Corsair decode → return tokens.

Spec sheet

Spec	Corsair
SRAM bandwidth per card	~150 TB/s^R (reported)
LPDDR5 capacity per card	Up to 256 GB^R (reported)
Decode speedup with GPU prefill	~10x vs. GPU-only^R (reported)
Power savings vs. GPU-only	~80%^R (reported)
Manufacturing process	TSMC N6^R (reported)

^R reported by the company

Update log

4 updates

About d-Matrix

The team behind Corsair

d-Matrix

AI inference accelerator, in-memory compute, Corsair

d-Matrix builds Corsair, an inference accelerator using compute-in-memory (CIM) architecture. Paired with NVIDIA GPUs for prefill, Corsair delivers 10x faster inference on decode workloads while using one-fifth the energy.

Corsair: SRAM-based in-memory compute. 10x faster inference vs. GPUs; 80% lower power.
June 2026: Full production; shipping to hyperscalers, neoclouds, and frontier labs.
Pairs with NVIDIA Blackwell GPUs: GPUs do prefill, Corsair excels at decode phase.

d-matrix.ai (opens d-matrix.ai)Company profile →

Founded: 20197 yrs
Headquarters: United States
Status: Private
Valuation: private
Staff: ~326Aug 2026
Raised: $392M3 rounds
Last round: Series C · $275M12 Nov 2025
Works in: ComputingAI accelerators & CPUs
Coverage: 1 program · 4 updateslatest 9 Jun 2026checked 10 Oct
Lead investors: Bullhound CapitalTriatomic CapitalTemasek
Partners: NVIDIA (opens nvidia.com)
Suppliers: TSMC (opens tsmc.com)Alchip Technologies (opens alchip.com)
People: Sid Sheth (opens linkedin.com)CEO & Co-founder

Spec

Corsair

SRAM bandwidth per card

~150 TB/s^R (reported)

LPDDR5 capacity per card

Up to 256 GB^R (reported)

Decode speedup with GPU prefill

~10x vs. GPU-only^R (reported)

Power savings vs. GPU-only

~80%^R (reported)

Manufacturing process

TSMC N6^R (reported)

Corsair

Milestones

Most important updates

Upcoming

Current obstacles

Physics limits

How it works

Why decode needs different hardware

Math inside the SRAM array

GPU prefill + Corsair decode

Spec sheet

Update log

Tue 9 Jun

Jun 2026

Wed 12 Nov 2025

Wed 6 Sep 2023

About d-Matrix

Current obstacles

Physics limits

How it works

Why decode needs different hardware

Math inside the SRAM array

GPU prefill + Corsair decode

Spec sheet

Update log

Tue 9 Jun

Jun 2026

Wed 12 Nov 2025

Wed 6 Sep 2023

About d-Matrix