Groq Stayed Independent at Half Its 2025 Valuation

By August 2026, Groq remained independent at a $3.5 billion valuation, roughly half its $6.9 billion valuation in September 2025. In between, Nvidia licensed Groq’s inference technology, hired its former chief executive and announced a rack built around 256 Groq LPUs. The legal entity survived, but the competitive system no longer fit inside it.

Groq entered public view in 2017, when engineers from Google’s Tensor Processing Unit project raised $10.3 million to build it. By 2021, the company had raised a $300 million Series C at a valuation above $1 billion. Groq kept its architecture, engineers, hardware roadmap, capital and commercialization capacity inside one corporate boundary. Investors could use the company’s name as shorthand for the whole system, but only while that boundary held.

Nvidia separated Groq’s technology, talent and company

Nvidia and Groq signed a non-exclusive licensing agreement for Groq’s inference technology. Groq said it would continue operating independently and keep GroqCloud running. Under the same agreement, CEO Jonathan Ross and several other senior executives joined Nvidia.

A later report put the transfer at about 90% of Groq’s employees and said vested shares would be paid in cash. Groq did not confirm the workforce figure, so it belongs in the reported column rather than the contractual one. The confirmed executive departures still matter because a chip architecture does not commercialize itself. Ross and the departing staff carried design judgment, implementation knowledge and operating relationships that a patent schedule cannot enumerate.

Reports tied shareholder payouts to a $20 billion valuation, while Nvidia denied that it was acquiring Groq. Each statement answered a different question. The license determined who could use the technology. The employment agreements determined where many of the people carrying its know-how would work. Groq’s statement established that the corporation and cloud service would continue to exist.

By licensing the design and hiring key people who built it, Nvidia reproduced much of an acquisition’s strategic effect without extinguishing Groq. Investors, customers and competition officials must therefore trace which productive assets—leadership, engineering capacity, IP rights, customers, hardware roadmap and capital—remain beyond the buyer’s control.

Nvidia made the challenger part of its catalog

In March 2025, Groq and Cerebras stood among the companies challenging Nvidia in inference. Groq’s architecture offered a different route through a workload that GPUs dominated but did not optimize for every cost and latency profile.

By March 2026, Nvidia had announced the Nvidia Groq 3 LPX, a server rack containing 256 Groq 3 LPUs, 128GB of SRAM and 40 petabytes per second of SRAM bandwidth. Nvidia scheduled the rack for the second half of 2026, placing Groq’s design inside its own distribution system.

Nvidia later said it would sell AWS one million GPUs and a broader mix of chips, including its new Groq chips, by the end of 2027. It can now place Groq-derived hardware beside its GPUs, networking products and data-center systems inside an existing hyperscaler relationship. A design that once differentiated an outside challenger has become another option in the incumbent’s portfolio.

The non-exclusive license lets Groq continue commercializing the technology itself. The agreement also does not establish that Nvidia owns all of Groq’s IP or business. Even so, a cloud buyer can now procure Groq-based hardware alongside Nvidia’s GPUs, networking and support through Nvidia’s channels.

Nvidia paired the licensed architecture with engineers who knew how to implement it, creating a hardware talent moat. A schematic records what engineers built; the engineers remember why one implementation survived timing, memory and manufacturing constraints while another failed. Moving both transfers more capability than either action would move alone.

Groq is rebuilding at a $3.5 billion valuation

Groq’s latest financing makes the remaining company visible in money rather than terminology. Groq raised $350 million at a $3.5 billion valuation, down from $6.9 billion in September 2025, before Nvidia licensed the technology and hired much of the team.

Groq now must retain customers, replace leadership, recruit engineers, maintain GroqCloud and finance a hardware-and-infrastructure roadmap after Nvidia gained technology rights and absorbed senior personnel.

Groq’s annual revenue was reportedly near $100 million when the Nvidia deal occurred. Whether Groq keeps those customers will reveal more than the transaction headline. GroqCloud’s continued operation establishes availability; customer renewals would establish retention.

Groq has also raised $650 million and set a target of 200 megawatts of data-center capacity by the end of 2027. Reaching that target requires sites, power, equipment and service capacity on a scale that makes Groq part of the emerging market for contracted AI capacity. Groq must then convert that capacity into demand, showing how quickly its new team can turn financing into deployed service.

A hot inference market rewards designs early

Barclays projected inference capital expenditure would reach $208.2 billion in 2026 and surpass training within two years. Documents reviewed by the Wall Street Journal showed OpenAI and Anthropic reporting inference costs above half of revenue. Those costs give model providers and chip incumbents a reason to secure promising inference designs before the startups behind them establish durable standalone distribution.

ByteDance and InnoStar were also reportedly developing a low-cost inference chip modeled after Groq’s LPUs. Groq’s architecture therefore had strategic value beyond GroqCloud’s reported revenue and Groq’s ability to finance its own expansion. Nvidia could obtain access to the design and its creators without assuming every obligation attached to a conventional acquisition.

Other inference challengers remain independent. SambaNova raised a $1 billion Series F at an $11 billion valuation and signed JPMorgan as a customer, while Cerebras continued to secure platform and infrastructure partnerships with OpenAI and AMD. Specialized accelerators can still attract capital and customers as standalone businesses.

AMD still used a conventional acquisition for Taalas, paying an undisclosed sum for the inference-chip startup. Strategic buyers can now choose between acquiring the corporate boundary and securing the design and engineers that make it valuable.

A six-line asset ledger shows what the label misses

A post-deal review should begin with the systems that produce competitive capability rather than the transaction label.

• Leadership: Which executives still set the company’s technical and commercial priorities, and which now work for the strategic buyer?
• Engineering capacity: How many people remain capable of maintaining the current product and producing the next one? Groq’s confirmed executive departures and reported workforce transfer require separate treatment.
• IP rights: Is the license exclusive, non-exclusive or transferable, and which implementation rights remain with the startup? Groq’s non-exclusive agreement leaves commercial room that an outright assignment would remove.
• Customers and revenue: Which contracts, workloads and recurring payments stayed with the company? GroqCloud’s continued operation establishes availability; only renewals establish retention.
• Roadmap and distribution: Which party can turn the architecture into supported hardware at scale? Nvidia’s Groq 3 LPX and AWS agreement provide concrete answers on Nvidia’s side.
• Capital and capacity: Can the company finance fabrication, deployment and data-center power? Groq’s fundraising and 200-megawatt target make that question measurable.

Reviewers can use the same ledger outside semiconductors. Google’s Windsurf license-and-acquihire structure left rank-and-file employees facing uncertainty over leadership and compensation. Such deals can reward selected executives, move a technical core and leave the remaining workers to discover which operating system is still theirs.

GroqCloud kept the Groq name, Ross moved to Nvidia, and 256 Groq LPUs reappeared inside an Nvidia rack. The incorporation papers stayed put; the valuation fell from $6.9 billion to $3.5 billion as the load-bearing members changed buildings.