Breaking

Wafer Raises $40 Million Series A Above $200 Million After Declining Buyouts

September 1, 2026 at 12:29 EDT

Funding & M&A
AI Agents
Infra & Chips

Wafer, a year-old inference company that uses autonomous agents to speed up large language models across Nvidia, AMD, and other accelerators, has raised $40 million in Series A funding at a valuation of more than $200 million and turned down acquisition offers along the way.

September 1, 2026 · Wafer

A $40M bet on inference software

Wafer closed a $40 million Series A at a valuation above $200 million, declining acquisition offers to keep building agents that continuously retune AI inference.

$40M

Series A, co-led by Marathon and Chemistry

$200M+

Reported post-round valuation

$4M

Seed round in April 2026

Nordlys clustering kernel speedup

Drawn to scale: 8x faster routing kernel

Latency on a production clustering kernel fell from about 180 microseconds to 21.7 microseconds after profile-guided optimization.

Before · ~180 µs

After · 21.7 µs

One stacked block per multiple — height is the speedup, not a slogan.

The loop Wafer wants to automate

1 · TRAFFIC

Learn workload patterns and constraints

→

2 · TUNE

Retune kernels, engine, and model path

→

3 · PLACE

Pick hardware, including AMD and Nvidia

→

4 · REPEAT

Keep optimizing after deployment

WHY BACKERS LEANED IN

Investors describe unusually fast gains on open-model serving, including AMD work at more than 2x lower cost than Blackwell on a published GLM5.2 test.

THE OPEN QUESTION

Agent-written kernels can game benchmarks — Wafer caught a fake 104.89x result — and Nvidia’s software path is still easier than day-zero AMD stacks.

Same-week backdrop

Nvidia put $3.5 billion into MediaTek convertibles so custom chips plug into NVLink Fusion — a hardware lock-in next to Wafer’s software-optimization bet.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.