AI chip funding splits from chip design: Armadin raises $255.5M for agentic security

Armadin announced the Series B on Thursday.

Andreessen Horowitz and Accel led the round, with Bain Capital Ventures, Redpoint, 8VC, Ballistic Ventures, Google Ventures, In-Q-Tel, Kleiner Perkins and Menlo Ventures joining, according to TechCrunch. The raise lands six months after a $190 million Series A in March, taking total funding past $445 million. Mandia's last company, Mandiant, sold to Google for $5.4 billion in 2022. His pitch now is different: instead of periodic penetration tests, Armadin runs always-on agentic swarms that chain vulnerabilities together to find holes in enterprise networks before attackers do. That is the company's framing, not an independently verified outcome.

Capital is still finding chip and infrastructure startups

The week around Armadin's round produced a cluster of chip adjacent deals. Broadcom agreed to lend Anthropic $42 billion to fund next generation AI chips, according to Ventureburn. Volantis raised $88 million for chip to memory interconnect, reported by The Economic Times. CScale emerged from stealth with $145 million for optical AI networking, with Nvidia and Intel backing, per Reuters. And Efficient Computer closed roughly $97 million at a $650 million valuation, according to SiliconAngle and Reuters.

Reuters put the Efficient figure at $97 million; SiliconAngle and Dealroom also reported $97 million, while The Economic Times headlined $100 million at the same $650 million valuation. The dossier does not reconcile the gap, so treat the round size as approximately $97 million to $100 million.

Two of those deals, Efficient and CScale, sit in the same part of the stack: moving data and saving power rather than designing a general purpose processor. That is a narrower bet than the headline numbers suggest. Efficient's own pitch, per the coverage, is drones and robots before data centres.

Huawei claims a China share shift

On the same day Armadin announced, Huawei's rotating chairman Eric Xu claimed its Ascend NPUs had overtaken Nvidia in Chinese market share.

The Register reported the Q&A release on 30 September. Xu also said Huawei lacks capacity to satisfy Chinese demand and has no plans to expand internationally in a fully fledged way. Nvidia's Q2 earnings call, cited in the same piece, showed H200 shipments to China at under 1 percent of data centre revenue. That claim is a company statement, not an audited figure. Ars Technica reported on 28 September that China's Ministry of Industry and Information Technology asked Alibaba and ByteDance to detail plans to buy Nvidia RTX Pro 5500 chips, citing The Information. Both things can be true: domestic substitution pressure alongside renewed foreign purchases.

The design layer is moving too. OpenAI and Synopsys signed a multi year partnership to build GPT-Synopsys, a model aimed at chip design and verification, The Decoder reported on 30 September. Synopsys is licensing its EDA tools into the project; early customer tests are underway, and the two companies will share revenue. Separately, Emergence AI said it is deploying neuroformal AI with fabless chipmakers, co-founder Satya Nitta telling EE Times that "you cannot make chips any faster, but what you can do is definitely get more chips per wafer yielding with AI."

Nuvacore, led by former Apple and Nuvia architects, said on 30 September it will delay instruction set selection for its WarpCore core IP as long as possible, per Tom's Hardware. That is a design methodology bet, not a funding event, but it points the same way: teams are trying to compress the front end of chip development.

Money is also going to the boring layers

MaxLinear announced its Puma 9 DOCSIS chip, claiming 30 to 50 percent cost reductions versus its predecessor and support for DOCSIS 3.1, 3.1+ and 4.0 on one SoC. Light Reading reported the launch on 28 September, with modems expected in 2027. The chip moves to ARM from Intel x86 and adds DDR5 support. Those are supply chain decisions as much as performance ones.

None of this proves an AI chip bubble, and the Armadin round is a security raise rather than a semiconductor one. But the pattern across the last 72 hours is consistent: capital is concentrating on tooling, interconnect, power and memory, the layers that determine whether more accelerators can actually be built and run.

The harder question is timing. Armadin's $2.5 billion valuation rests on agentic security demand that is still mostly described by the company itself. Huawei's share claim comes from its own chairman. Efficient's revenue is not public. Reporters can date the money. They cannot yet date the returns.