이미지: TechCrunch AI
Summary
- AI chip startup Groq has raised $350 million at a $3.5 billion valuation
- The lower valuation compared to $6.9 billion last September follows NVIDIA's $20 billion licensing deal that brought over founder Jonathan Ross and key personnel
- Instead of its own LPU chips, Groq has pivoted to a neocloud model renting out NVIDIA GPUs, now serving 6 million developers and companies through 13 data centers
- 신규 투자액
- 3억5000만달러
- 신규 기업가치
- 35억달러 (2025년 9월 69억달러 대비 하락)
- 주도 투자사
- Disruptive (엔비디아 참여 예정)
- 직전 사건
- 엔비디아가 창업자 조너선 로스 등 핵심 인력을 200억달러 라이선싱 계약으로 영입
- 데이터센터 규모
- 13곳 (북미·유럽·중동·아태), 2027년까지 54MW→200MW+ 확장 목표
- 이용자 규모
- 개발자·기업·AI 네이티브 기업 600만 이상
- 직전 라운드
- 2026년 6월 6억5000만달러 (피벗 착수)
A chipmaker now rents out someone else's chips
Groq originally set out to go head-to-head with NVIDIA. The company was founded on the idea of using its own custom LPU (Language Processing Unit) chips to shake up the AI inference market — the business of running already-trained models in real time. Now that same Groq has raised $350 million in fresh funding. The round was led by investment firm Disruptive, with NVIDIA reportedly set to participate as well. The valuation assigned was $3.5 billion, half of the $6.9 billion the company was valued at last September.
The company insists this isn't a down round. A Groq spokesperson told TechCrunch that the new valuation reflects "the version of Groq that exists after the NVIDIA licensing deal" — an acknowledgment that the company's identity has fundamentally changed over the past few months.
The spot NVIDIA took by hiring away the founder
A few months after that $6.9 billion valuation last September, NVIDIA struck a $20 billion licensing deal that brought over Groq founder and CEO Jonathan Ross along with key team members. In effect, NVIDIA absorbed the brain trust of a company that had been positioning itself as a direct rival. Having lost its star team, Groq abandoned its own chip development roadmap and pivoted toward operating cloud and data center infrastructure built on NVIDIA systems. What remains of Groq is no longer NVIDIA's competitor — it's now a customer.
The shift actually began back in June, when Groq raised $650 million as the first step of the pivot. This latest $350 million builds on that momentum, with the company saying the funds will support "demand for using medium-to-large scale NVIDIA-accelerated computing clusters for training and inference."
Toward 13 data centers and 200 megawatts
Groq currently operates 13 data centers across North America, Europe, the Middle East, and Asia-Pacific, serving more than 6 million developers, enterprises, and AI-native startups. The company aims to expand its power capacity from the current 54 megawatts to over 200 megawatts by 2027. Megawattage is a measure of how many GPU servers a data center can run simultaneously, meaning this target would represent nearly a fourfold expansion of Groq's infrastructure within three years.
Groq Chairman Alex Davies, who also heads Disruptive, said in a statement: "We are building Groq into the world's leading AI inference cloud." His view is that inference will become the largest and most critical layer of AI infrastructure.
Everyone's doing neocloud, and everyone's worried about it
The business of renting out NVIDIA GPUs to sell AI infrastructure is commonly referred to in the industry as "neocloud." Groq is far from the only one taking this path. NVIDIA supplies GPUs to companies like CoreWeave, Lambda, and Nebius while simultaneously investing billions of dollars in them — playing the dual role of supplier and investor, fueling the growth of the very companies that run on its chips.
The open question is whether this business can be profitable in the long run. CoreWeave recently posted strong revenue growth in its latest quarter and landed major contracts with Meta and Anthropic, but investors continue to flag concerns about the company's heavy capital expenditures, reliance on debt, and the risk of rapidly depreciating hardware. Whether growth can be converted into actual cash flow remains an unresolved question for the neocloud model as a whole. Groq's financials remain undisclosed, but this pivot has placed the company firmly inside NVIDIA's ecosystem.
Editor's take
What's more interesting about this round than the halved valuation is the story underneath it. A company that set out to beat NVIDIA with its own chips lost its founder to NVIDIA, and has now been reborn as a company that survives by selling NVIDIA's GPUs. It's a rare, literal enactment of Silicon Valley's familiar line: "If you can't beat them, join them." The company's insistence that this isn't a down round reads like a defensive move aware of exactly that narrative.
Looking at the trajectory of inference-chip startups over the past year or two, the outcomes tend to converge on a similar conclusion. Whether it's Cerebras or Groq, attempts to challenge NVIDIA's CUDA ecosystem with proprietary architectures have lost momentum against the wall of software compatibility and developer inertia. Meanwhile, simply renting out NVIDIA GPUs offers a lower barrier to entry — but also thinner margins. Groq's pivot shows what the most realistic alternative looks like when that wall proves too high to clear.
For teams in Korea preparing AI chip or infrastructure ventures, the lesson here is clear: if you intend to compete with your own chip, you need to budget far more time than you'd expect for building out the software stack and developer community alongside it. If that time can't be bought, pivoting to a business of renting out someone else's chips isn't a shameful fallback — it's a survival strategy. But even that business, as CoreWeave's case shows, only remains sustainable if cash flow can keep pace with capital expenditure.
In the coming weeks, Groq's next moves will likely include announcements of expanded data centers and new customers. At the same time, competition on price and capacity will intensify among the other NVIDIA-backed neoclouds, and whichever of them is first to report a path to profitability will serve as a test of credibility for the neocloud model as a whole.
