---
title: "MetaX Is Two Steps Into Six. The Chip Has Not Been Taped Out."
date: 2026-11-17
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-smic-c700-fp4-1117
language: en
---

# MetaX Is Two Steps Into Six. The Chip Has Not Been Taped Out.

> At the third-quarter earnings briefing on November 16, Peng Li listed all six in one breath: design, verification, tape-out, software adaptation, customer testing, volume production ramp. Peng is MetaX's chief technology officer and chief h

At the third-quarter earnings briefing on November 16, Peng Li listed all six in one breath: design, verification, tape-out, software adaptation, customer testing, volume production ramp.

Peng is MetaX's chief technology officer and chief hardware architect. Those six words turned the C700's remaining road into a grid. Then she said the first two were "largely complete."

"Largely" is where the questioning should stop.

## Six steps in one breath

Her exact wording: the C700 began in April 2025 and is now in design and development; core chip design and functional verification are largely complete, and deeper performance tuning is under way.

Watch the verbs. Core design is "complete." Verification is "largely complete." Tuning is "under way." Three degrees of completion, one fact: the first two steps are not closed. Behind verification sit process-corner validation, joint debugging with the software stack, and the move to silicon.

Any one of them can send the design back for a revision.

By her own ordering, the next box is tape-out. A CTO who recites six steps is handing the audience a progress bar. Anyone can count to the second box.

Nineteen months in, that is not slow. Nvidia took two years from Hopper to Blackwell, and domestic GPUs need longer because they route around certain toolchains and IP.

Not slow is not the same as nearly there.

## What C600 cost in time

The control group is in the same filing.

The Xiyun C600 is MetaX's first high-performance general-purpose GPU on a fully domestic process. It entered volume delivery in the first half of 2026, and in May 2026 its first batch passed the security and reliability evaluation run by the China Information Security Evaluation Center and the State Secrecy Science and Technology Evaluation Center.

MetaX's site calls it a closed domestic loop, from design through manufacturing and packaging to the software stack. That is not marketing. It is the cost line: every domestic link gets validated from scratch.

C600 spanned a full calendar year from announcement to volume sales. The industry calls "runs, samples, ships in small volume" risk production, not volume production. Reaching monthly batch delivery, stable yield and controlled cost takes another six months.

## After tape-out there are no shortcuts

A tape-out takes months from submission to wafers back, then packaging, test, board assembly and first light. If version one has a bug, the masks are revised and the run starts again, at tens of millions of dollars a set.

There is no "just tweak it" on this path. There is "it works" and there is "run it again."

Which makes tape-out the most expensive bet a design house places all year.

The C700 uses a domestic advanced process, which means supply chain autonomy and a queue for capacity. Every domestic link competes for domestic capacity; a design house controls its own schedule, not the foundry's.

First-half R&D expense was 525 million yuan across 747 R&D staff, over 70% of all employees. People are not the constraint. Capacity windows are.

## The quarter the cards turned

The C700 timetable also depends on how long C600 can keep earning.

First three quarters: revenue 2.824 billion yuan, up 128.45%; profit attributable to shareholders 639 million yuan, or 259 million excluding non-recurring items.

Split by quarter and the story changes. First-half attributable profit was 612 million yuan, the adjusted figure negative 49 million. Subtract, and the third quarter shows 27 million yuan attributable and 308 million adjusted.

Selling cards earned 300 million yuan. Reported profit was 27 million.

The 280 million difference went to fair value changes. A first-half gain carried attributable profit to 612 million while the adjusted figure was negative; in the third quarter the order reversed.

The core business only turned positive this quarter. C700 has to build on it.

## Who FP4 is for

One term first: FP4.

Peng sketched the direction. C700 will extend support for low-precision compute including FP4, outperform C600 overall, benchmark against mainstream international GPUs, and go deeper into internet applications. Not one performance figure appears. Parameters wait for silicon.

FP4 compresses each number to 4 bits, trading precision for throughput and memory footprint. It suits inference best, especially long context and large batch.

Internet customers do not buy peak compute. They buy cost per token.

That reframes C700's position. It is not a lab training part but a tool for cloud providers to compress unit cost. Such customers check two things: whether software migrates cleanly, and whether a cluster runs three months without falling over. MetaX's MXMACA stack supports over 40 AI frameworks, 1,000 models and 6,000 open source projects. A ticket, not a victory.

## The last step is not technical

Volume production ramp is the easiest to overlook.

It is commercial and supply chain: whether the foundry schedules you, whether the packager gives you capacity, how many quarters yield takes to climb, who gets the first batch.

Work backwards from C600. If C700 tapes out in the first half of 2027, with six months each for software and customer testing and another six for ramp, volume lands in 2028.

An estimate, not company guidance. MetaX has never published a production timetable for the C700.

Investors asked about domestic compute share, not timing. Chairman Chen Weiliang said rising share is a long-term trend that could accelerate soon. True, and it does not answer when C700 can be sold.

"Verification complete" is a milestone in engineering.

In investing it is a comma.

Four gates remain — tape-out, software, customers, capacity — and each can cost an extra six months. Nineteen months bought two steps; the other four, on the smoothest schedule, still land in 2028.

Anyone who understands "largely complete" knows it is good news. They also know it is not the finish line.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-smic-c700-fp4-1117
Free to quote with attribution and a link to the original.
