Groq raises $650 million and pivots to the neocloud: the survival of the former AI chip darling after Nvidia poached its soul for $20 billion
🔎 $650 million for a company whose death was announced six months ago
On Monday, June 23, 2026, Groq confirmed a $650 million funding round to reinvent itself as an AI inference neocloud. It's a spectacular rebound for a company everyone thought was finished.
Six months earlier, Nvidia paid $20 billion for a deal the press described as a "not-acqui-hire." The LPU architecture — Groq's core — landed at the giant in green. CEO Jonathan Ross, president Sunny Madra, and most of the technical team followed suit.
Groq was nothing more than an empty shell with a brand. Except it wasn't. The company just proved the opposite with this round led by Disruptive and Infinitum, and a new management team parachuted in from xAI and Meta.
Why now? Because the AI inference market is exploding and even without its original chips, Groq still has something others don't: an ultra-fast API ecosystem, a loyal developer base, and a brand associated with inference speed.
The key points
- Groq raises $650M (June 2026, check on groq.com) to pivot from hardware to an inference neocloud, led by Disruptive and Infinitum.
- Nvidia poached the LPU architecture, CEO Jonathan Ross, and key engineers for $20B six months earlier — a "not-acqui-hire" deal that stripped Groq of its technical soul.
- New CEO Adam Winter arrives from xAI/Meta, accompanied by a CTO and a CPO who are founders of cloud companies, to build a neocloud without proprietary chips.
- Groq must now compete with Cerebras, SambaNova, and hyperscalers in the cloud inference arena — a market where the margin for error is practically zero.
Recommended Tools
| Groq Cloud | Very low latency AI inference | Free / on quote (June 2026, check on groq.com) | Developers looking for raw speed |
|---|---|---|---|
| Best Free LLMs | Comparison of LLMs accessible without a subscription | Free | Beginners and quick testing |
| Free AI APIs | Free inference APIs (Groq, Google, OpenRouter) | Free | Personal projects and prototyping |
The $20 billion deal: how Nvidia "not-acqui-hired" Groq
Nvidia didn't buy Groq. That's the crucial point. A classic acqui-hire involves an acquisition followed by the closure of the target. Here, Nvidia paid $20 billion to gain access to the LPU architecture and hire key talent — without buying the legal entity.
Founding CEO Jonathan Ross is gone. President Sunny Madra too. The majority of the engineers who designed the LPU chips have joined Nvidia's offices.
According to TechCrunch, this deal left Groq in an "operational shell" state — a company with a name, servers, and very few people to keep them running.
This is arguably Jensen Huang's most aggressive move since the acquisition of Mellanox in 2020. Except here, the goal wasn't a portfolio of networking patents, but the outright elimination of a bothersome competitor in the inference space.
This maneuver fits into Nvidia's massive investment strategy. The company is pouring $40 billion into AI in 2026, and investing $150 billion a year in Taiwan to secure its supply chain. The Groq deal adds to this logic of domination by any means necessary.
What Groq was before the drama: LPUs and the promise of speed
Groq was not a minor player. Founded in 2016 by Jonathan Ross — a former member of Google's TPU team — the company had developed an inference chip radically different from Nvidia's GPUs.
The LPU (Language Processing Unit) was designed for one thing only: running language models as fast as possible. No graphics, no generic scientific computing. Just the sequential inference of transformers.
The result? Inference speeds up to 24x higher than those of conventional GPUs for certain workloads. Groq had installed clusters of LPU chips in its datacenters and offered free APIs via its cloud.
Developers love speed. Groq became one of the pillars of the low-latency inference ecosystem, to the point of featuring prominently in comparisons of free AI APIs and best free LLMs.
The promise was simple: replace the bruteforce of GPUs with the elegance of specialized silicon. It worked on paper. On the ground, Groq hit the same walls as all hardware challengers: manufacturing, costs, and the monopoly of Nvidia's CUDA ecosystem.
The new Groq: a management parachuted in from xAI and Meta
After the departure of Ross and Madra, they had to rebuild from scratch. Groq did so by drawing from the two most aggressive ecosystems right now: xAI and Meta.
Adam Winter takes the helm as the new CEO. He comes from xAI — Elon Musk's company — where he oversaw large-scale compute infrastructure. Before that, he spent years at Meta on cloud platforms.
The new COO is a veteran from the same xAI/Meta combination. The CTO and the CPO are cloud company founders who have already built and scaled infrastructure of this kind.
This is no coincidence. Groq is not hiring hardware people. It is hiring cloud people. The signal is clear: the company is not going to design new chips. It is going to buy compute power wherever she can and resell it as an inference service.
According to Startup Fortune, this team was assembled in record time — less than four months — which suggests that conversations had begun even before the Nvidia deal was finalized.
The neocloud pivot: inference without proprietary chips
That is the crux of the problem. Groq was a chip company. Its chips are now at Nvidia. What is left?
The cloud remains. Groq has datacenter infrastructure, contracts with compute providers, and above all, an API that thousands of developers are already using. The pivot consists of leveraging this software asset without depending on in-house silicon.
Concretely, Groq is going to buy chips — probably Nvidia GPUs, ironically — and optimize them for inference. The game is no longer on the hardware but on the software layer: orchestration, intelligent routing between models, caching, batching optimization.
This is the model of Cerebras, SambaNova, and a half-dozen neoclouds emerging in 2026. The difference is that these competitors still have their chips. Groq must prove that one can win on software alone.
FourWeekMBA analyzes this pivot as a "bet on the commoditization of AI hardware" — the idea that chips will become interchangeable and that the real value will lie in the abstraction layer above.
It is a compelling thesis. But it has yet to be proven by anyone at scale.
The inference market in 2026: an overcrowded battlefield
Groq is not entering an empty market. AI inference has become the most fiercely contested segment of the entire tech industry.
Cerebras sells inference on its WSE chips with a promise of speed comparable to what Groq used to offer. SambaNova pivoted to a neocloud model after years of disappointing hardware sales. The hyperscalers — AWS, Google Cloud, Azure — all offer optimized inference services.
And then there are the models themselves. In June 2026, the LLM rankings are dominated by Gemini 3.1 Pro (Google, score 92), GPT-5.5 (OpenAI, 91) and Claude Opus 4.7 Adaptive (Anthropic, 90). xAI's Grok 4.1 comes in at 90. These models already run on their creators' infrastructures.
In agentic, GPT-5.5 dominates with 98.2, followed by Gemini 3 Pro Deep Think at 95.4 and Claude Opus 4.7 at 94.3. Moonshot AI's Kimi K2.6 stands out in self-host at 88.1 — a remarkable score for an open-weight model that could very well run on neocloud infrastructure like the one Groq is building.
The question for Groq is not "can we run these models?" but "why would a developer choose Groq over OpenAI or Google's native API?".
The 650 million: what is it actually for?
A $650 million raise is considerable. But in the world of AI inference in 2026, it doesn't last long if misinvested.
According to the Groq press release, the funds are allocated to three priorities:
First, the expansion of cloud infrastructure. Groq must deploy compute clusters in multiple regions to offer acceptable latency on a global level. Each cluster costs tens of millions in hardware alone.
Second, massive hiring. The company has lost most of its technical team. It must hire system engineers, model orchestration specialists, and inference optimization experts. With the AI job market as it is in 2026, salaries are astronomical.
Third, the development of the software layer. This is the real potential differentiator. Groq is building a routing system that automatically directs each request to the most suitable model and infrastructure — a bit like what OpenRouter does but with much deeper optimization at the system level.
650 million is about 18 months of runway if Groq spends at a "neocloud" pace. Not enough to build an empire, but enough to prove the model or die trying.
Cerebras vs SambaNova vs Groq : a comparison of inference neoclouds
The specialized inference neocloud market is structuring itself around three major players. Each has a different strategy.
| Player | Main advantage | Major drawback | Business model |
|---|---|---|---|
| Cerebras | Proprietary WSE chips, raw speed | Dependence on in-house hardware, R&D costs | Neocloud + hardware sales |
| SambaNova | Reconfigurable architecture, flexibility | Recent pivot, uncertainty around scaling | Neocloud primarily |
| Groq | Strong brand, developer base, existing API | No more proprietary chips, dependence on third-party GPUs | Pure neocloud, software layer |
Cerebras has the silicon advantage. Its WSE-3 chips are the largest ever built, and the company controls the entire stack. SambaNova pivoted earlier and already has enterprise customers in production.
Groq is the only one of the three that no longer has a hardware secret sauce. Its bet is that orchestration software can create more value than custom silicon. It is bold. It is also unreasonably risky.
It should be noted that these three players are tiny compared to hyperscalers. The Chinese Lineshine supercomputer, which dethroned El Capitan in the June 2026 Top500, alone represents more computing power than all the inference neoclouds combined. Scale favors the giants.
The role of the ecosystem: why developers might stay
Despite everything, Groq has an asset that money can't buy: developer loyalty.
For two years, the Groq API was one of the few to offer free inference on quality models. Thousands of open-source projects, demos, and prototypes were built on this API. Switching inference means modifying code, retesting, redeploying. Many developers prefer to stay if the service remains good.
Groq understood this. The strategy is to maintain a generous free tier — exactly as described in the free AI APIs guides — to retain this user base during the transition.
The company is also betting on model diversity. Rather than being tied to a single model provider, Groq already offers OpenAI's GPT-5.4, Anthropic's Claude Sonnet 4.6, Google's Gemini 3.1 Pro, and open-weight models like Moonshot AI's Kimi K2.6 and DeepSeek V4 Pro.
This position as a "multi-model inference broker" is exactly where the market is heading. Developers no longer want to lock themselves into a single provider. They want a single entry point that gives them access to the best model for each task.
If Groq executes this vision well, the loss of LPU chips could end up being an advantage. Being free to choose the best hardware for each model is a flexibility that neither Cerebras nor SambaNova has.
The shadow of xAI: the new CEO and the Musk strategy
Recruiting Adam Winter from xAI is not insignificant. xAI is one of the most aggressive players in the inference market, with Grok 4.1 posting a score of 90 in general and 79 in agentic — solid scores but ones that also show the limits of an all-Musk approach.
Winter knows both worlds: hyperscalers (Meta) and aggressive AI startups (xAI). He knows how to scale infrastructure quickly and how to navigate the internal politics of a company in survival mode.
But he also brings cultural baggage. xAI is known for its raw execution speed, even at the expense of reliability. Meta for its open-source and Scaling culture. Mixing the two could give Groq a very different tempo from Jonathan Ross's meticulous approach.
The risk? Groq becomes a sort of mini-xAI without the proprietary model. A soulless infrastructure, chasing inference margins in a market where everyone is chasing the same margins.
The opportunity? Winter likely has deep connections at xAI and Meta that could translate into major inference contracts. In the neocloud space, two or three large enterprise contracts can change the life of a startup.
The geopolitical context: why China is accelerating and what it means for Groq
Groq's story is not happening in a vacuum. While the company rebuilds itself, China is massively accelerating in the AI race. Moonshot AI raised $2 billion and its Kimi K2.6 model dominates the open-weight segment with a score of 84 overall and 88.1 in agentic in self-host.
For a neocloud like Groq, Chinese open-weight models represent a concrete business opportunity. Kimi K2.6, DeepSeek V4 Pro, Z.AI's GLM-5.1 — all these models need third-party inference to reach Western developers.
Groq could become the entry point for these models in Europe and North America. It's a niche that no one is really claiming, and it perfectly matches the multi-model broker position the company is building.
But it is also a geopolitical risk. Hosting Chinese models on American infrastructure, even via API, could attract regulators' attention. Groq will have to navigate this terrain carefully.
Furthermore, the Taiwanese market remains the epicenter of chip manufacturing. Nvidia invests $150 billion a year in Taiwan, and the entire industry — including Groq's suppliers — depends on this supply chain. A geopolitical shock in Taiwan would affect Groq just as much as all the players in the sector.
Inference as a commodity: the thesis that makes or breaks Groq
Groq's pivot rests on a fundamental hypothesis: AI inference will become a commodity, and value will migrate to the orchestration layer.
This is the same thesis that drove cloud computing in the early 2010s. Servers became interchangeable, and Amazon, Google, and Microsoft won by offering the best abstraction layer.
But the analogy has its limits. AI inference is not classic cloud computing. Performance depends intimately on the coupling between the model and the hardware. A model optimized for Nvidia GPUs will not run as well on AMD chips or Google TPUs, even with the best orchestration layer in the world.
Groq is betting that this dependency will weaken with the standardization of model formats and the improvement of compilers. This is possible. Frameworks like MLX, TVM, and ONNX Runtime are making constant progress in hardware abstraction.
But in June 2026, the reality is that the best inference scores are achieved by models running on their creator's hardware. GPT-5.5 on Nvidia GPUs, Gemini 3.1 Pro on Google TPUs, Claude Opus 4.7 on Anthropic's infrastructure. The commodity is not there yet.
Groq may be right in the long term. But the long term is expensive, and 650 million dollars do not buy much patience from investors.
Possible scenarios for the next 18 months
Three trajectories are taking shape for Groq starting in June 2026.
Scenario one: the rebound. Groq executes its pivot perfectly. The API remains fast, the free tier attracts new developers, and the enterprise contracts signed by the xAI/Meta management generate recurring revenue. Eventually, Groq becomes an acqui-hire target for a hyperscaler looking for a boutique inference layer. Exit price: $3-5 billion.
Scenario two: slow asphyxiation. Compute costs explode, hyperscaler competition crushes margins, and developers migrate to native APIs. Groq burns through its $650 million without finding product-market fit. Shutdown or fire sale within 24 months.
Scenario three: the unexpected pivot. Groq discovers a specific niche — edge inference, specialized models for finance or healthcare — that justifies a pricing premium. The company shrinks but becomes profitable. No spectacular exit, but survival.
The most likely scenario? Something between one and three. The new management isn't the type to die slowly. But scenario two remains the default plan in this market.
❌ Common mistakes
Mistake 1: Confusing the pre- and post-Nvidia deal Groq
The company that designed LPU chips no longer exists. What remains is a cloud startup with a great name and fresh cash. Analyzing Groq as if it still had its hardware advantage is missing the point.
Mistake 2: Thinking $650M is enough to compete with hyperscalers
AWS spends $80 billion a year on infrastructure. Google Cloud and Azure are in the same ballpark. $650 million is seed money in this context — not enough to build a global inference network.
Mistake 3: Believing inference is a zero-sum market
The AI inference market is experiencing explosive growth. Even if Groq takes 1% of this market, it could represent hundreds of millions in revenue. The game isn't to kill Nvidia, but to find a profitable niche.
❓ Frequently asked questions
Does Groq still manufacture LPU chips?
No. The LPU architecture was transferred to Nvidia as part of the $20 billion deal. Groq now buys compute from third-party providers, likely Nvidia GPUs initially.
Who is the new CEO of Groq?
Adam Winter, formerly of xAI and Meta. He was hired for his experience in scaling large-scale cloud infrastructures, not for his hardware expertise.
Can you still use the Groq API for free?
Yes. Groq maintains a free tier to attract developers, comparable to what other AI API providers offer. It's a pillar of its retention strategy during the pivot.
Is Groq a good option for hosting open-weight models like Kimi K2.6?
This is precisely the niche Groq is targeting. The company positions itself as a multi-model broker capable of running Chinese and Western open-weight models through a single API. The quality of the service remains to be evaluated in production.
✅ Conclusion
Groq is the best-funded zombie in the AI industry — a dead company that refuses to stay down. With $650 million and management from xAI and Meta, it is attempting the market's most counterintuitive gamble: becoming an inference neocloud without proprietary chips, after handing its hardware soul to Nvidia for $20 billion. It could work. It could also become the most expensive cautionary tale of the AI war. The answer will come in the next 18 months — and it will be interesting to watch via their API.