The news
AI infrastructure startups working beyond the graphics processing unit (GPU) are drawing big checks. In September 2026, Positron raised up to $875 million for inference chips, Celero Communications $275 million for AI cluster networking and Cornelis Networks $205 million for AI network fabrics.
Cornelis, based in Wayne, Pennsylvania, announced its round on September 14 with a Qualcomm (QCOM) collaboration and a new architecture it calls Active Compute Fabric. The design combines lossless data transport with in-network acceleration, so the network processes data rather than simply moving it between chips. IAG Capital Partners invested; partner Joel Whitley said open-standard scale-up and scale-out networking, which links AI chips inside rack-scale systems and across them, is an opportunity worth more than $55 billion by 2030.
Cornelis chief executive Lisa Spelman said “faster endpoints alone are not enough” and that the network must become an active part of AI systems. Tony Pialis, who runs Qualcomm's data center business, said better utilization and AI economics require an integrated approach across compute, memory and networking.
The pattern shows up across the September deal lists. Crunchbase's roundup for the week ended September 11 included Celero, at a $3 billion valuation, and Fab2, which is designing its own fabrication tools, components and software for semiconductor manufacturing and raised a $500 million Series A at $3.7 billion. The week before, Gimlet Labs raised $300 million at $3 billion for an inference cloud that spreads workloads across different chip types, and Lyte AI raised $165 million for custom silicon for robotic perception.
Positron's round, announced September 10, combined a $375 million Series C with up to $500 million more, at a $5 billion post-money valuation. Its pitch is a memory-first design that uses commodity LPDDR5X memory instead of high-bandwidth memory (HBM), which it says avoids supply constraints. It is deploying more than 50 racks of its first-generation Atlas systems at Oracle (ORCL) Cloud Infrastructure, the company said.
The numbers
- Positron round
- Up to $875 million at $5 billion post-money
- Celero Communications round
- $275 million at $3 billion (Crunchbase)
- Cornelis Networks round
- $205 million
- Gimlet Labs round
- $300 million at $3 billion (Crunchbase)
- Open-standard AI networking opportunity
- More than $55 billion by 2030 (IAG Capital Partners)
- Chip startup funding, 2026 through early June
- $10.7 billion (Crunchbase)
Why CEOs should care
For CIOs and heads of infrastructure, the message is that AI costs are increasingly set by networking, memory and power, not only by GPU prices. When planning 2027 capacity, ask cloud providers and hardware vendors for measured utilization and cost per token on your own workloads, not benchmark results. Vendors built on open standards may reduce lock-in, but verify that they work with the systems you already run.
For CFOs, a wave of well-funded challengers is a pricing lever. Even companies that never buy startup hardware directly benefit when credible alternatives in inference chips and networking give them room to negotiate with incumbent suppliers and cloud providers. Many of these products are still early, though: Positron's next-generation Asimov chip is not due in production until the second half of 2027, so commitments should be tied to delivery milestones.
For boards, the investor lists show where strategic players see bottlenecks. Qualcomm is collaborating with Cornelis, and Cisco Investments and Naver Ventures invested in Positron. Directors should ask whether the company's AI plan assumes current GPU pricing and supply, and how the plan changes if inference costs fall faster than expected.
The bigger picture
Chip startups raised $10.7 billion in 2026 through early June, Crunchbase reported, with large rounds for MatX, Ayar Labs and Etched, and Cerebras Systems raised more than $5 billion in its May IPO. Networking startups Nexthop AI, with a $500 million Series B, and Eridu, with more than $200 million at Series A, raised big rounds in March, according to Crunchbase.
Large buyers are moving too. Crunchbase counted Qualcomm's $4 billion acquisition of Modular among the standout startup deals of the second quarter, and the AI cloud providers these chips would serve keep raising: Crusoe raised $3 billion at a $30 billion valuation and Fluidstack $1.5 billion at $18 billion in the week ended September 4.
What’s next
Watch whether the Cornelis and Qualcomm collaboration turns into shipping products, whether Positron's Asimov chip tapes out on schedule at the end of 2026, and whether large clouds and independent AI cloud providers such as Crusoe and Fluidstack commit to non-GPU inference hardware at scale. Those purchase orders, more than funding rounds, will decide which of these startups become lasting suppliers.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
