Groq stopped serving Kimi K2 (1T params) when they got aquihired by NVIDIA, so I guess NVIDIA took most of the hardware in addition to the employees. The largest model they serve now is the relatively minuscule gpt-oss-120b.
Before they removed it, I was using groq Kimi K2 model for a chat bot in small community site/chat. It was really good, seemed to have incredibly vast general world knowledge and the fast speed (400tok/s if I remember right) meant that chat users got a response instantly which was a much better experience compared to other SOTA models at the time.
Those were amazing times. You could vibe code an entire prototype in seconds (200 tps). With Qwen3.6-35B-A3B and MTP, you can program at that speed on a single GPU at home now, but Kimi K2 is of course much smarter at almost 30 times the size.
I'm also looking forward for the Cerebras Kimi K2.6 release, which should be even better at 1000 tps. It is hard to overstate how important speed is for programming. Instead of having to wait for a few minutes until a task is done, it is just done instantly, and you don't have to context switch from whatever else you were working on while waiting.
I hope they will make it available to regular customers.
I've was part of an acquire-to-remove-an-annoyance, including having the better product that didn't win for niche reasons - which vaguely matches the situation here. Yes, NVIDIA can wind down Groq (unambiguously a better product in this case) even though it doesn't make sense.
That being said, there's still a chance that NVIDIA engineering is in the process of stripping it for parts. Or the lawyers are - maybe they have too much momentum with GPUs and just want ASICs out of the market.
This kind of innovation stifling acquisition should have been blocked. NVIDIA is a serious monopoly threat.
Somehow I never got the hype around Groq. They served models with fast inference speed - that sounded great in theory - and as a user I was looking forward to use them. But, when I did try, I discovered that they are quantizing the models underneath. And they dont even disclose it. So I stopped using them.
The whole thing never made any sense to me - but I guess AI hype is a thing.
I believe despite quantisation they were still extremely fast, which is still incredibly useful if you don't need high precision/accuracy (which is good enough for many use cases)
I don't really get the value proposition of groq as a user, the performance is really poor for the token price. Data centres on the other hand are becoming a commodity, and I don't see any reason a priori to invest in groq specifically for something like that.
Groq is considerably faster and better at inference, they have a totally superior product to Nvidia for inference based tasks, which will be the dominant concern in the future.
Plausibly, take all the Nvidia hype and multiply that by a factor and that's what 'Groq' could be worth.
And there is no real commodification - there's Nvidia, Cerebras, Groq ... not many otheres.
I love Groq, but across every single line break in your post there is a glaring issue that is easy to refute with in 15 seconds, even without 300t/s of throughput.
You wasted all of your commentary on snark and sadly unfunny humour, and yet still managed to add nothing.
Groq is more performant for the growing categories of inference-based tasks, wherein Nvidia's advantage in inference depends bulk/batch processing which will make up a smaller category over time, in relative terms.
The future of AI Silicon is inference, and the cost structure of AI data centres is constrained around the current necessity to have 'high GPU utilization' otherwise, the cost / amortization of the chips doesn't work out.
That cost structure is a limitation of Nvidia architecture.
Groq serves a lot faster, and without the limiting batching requirement, which opens hosting arrangements common in most classical hosting scenarios aka without necessarily the high utilization requirements.
Groq has bespoke hardware, lack of CUDA, much lower memory desnsity obviously and they don't have the deep distribution networks and leverage over TSMC that Nvidia has - but pound for pound, were we to be able to 'fire up a server' for our inference needs, it would be Groq, not Nvidia that we'd turn to.
Were they not a later market entrant and didn't have those barriers to entry, they'd be gigantic.
Google has been releasing a new TPU generation every year since 2023 and the eight generation consists of a training and an inference optimized design.
Google's eight generation TPU inference chip has 384 MB of on-chip SRAM vs 500 MB for Groq's third generation LPU.
My company had a really terrible experience trying to use Groq, and I would NOT recommend anyone use their service if you need reliability. So many random errors, so many silly quirks.
I'd say they have a "Tier 2" inference stack, but that'd be Apple Silicon and AMD: their well below that.
In part because unlike those, no one can get their hands on the hardware, so they miss out on a massive amount of free development, testing, etc.
At this point there's years of complaints about open weight models performing worse on their platform, tool calling makes it especially easy to tell since it's so sensitive to these issues.
About a year back I tried to build a few latency-sensitive products on Groq, and found the response times would vary so wildly that it just about erased any benefit in their throughput. Is that still this case?
I don’t understand one part of the licensing here: if it was just a license, can’t they relicense the software and hardware of LPU3 to AMD? Or hire new software and hardware people?
The new designs were their main asset besides the amazing talent that went to NVIDIA, not the remaining DCs.
The most bizarre thing here is the reporters. They intentionally misrepresented what happened. All the news stories from last year claimed Nvidia "acquired" Groq - and in the same story, quoted what actually happened, and pretended it didn't. It's like the journalists had some kind of group psychosis, pretending it was an acquisition. It wasn't. It was a really white-glove product rental.
From the actual Groq PR release:
"Groq announced that it has entered into a non-exclusive licensing agreement with Nvidia" - "As part of this agreement, Jonathan Ross, Groq’s Founder, Sunny Madra, Groq’s President, and other members of the Groq team will join Nvidia to help advance and scale the licensed technology" - "Groq will continue to operate as an independent company with Simon Edwards stepping into the role of Chief Executive Officer" (https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusiv...)
Nothing about an acquisition there. It says Nvidia is licensing it, and that others can too. The execs work for Nvidia to integrate it into Nvidia's... something. And Groq the company remains the same as before.
There's also no official source for the amount Nvidia paid for the tech, or two unofficial ones. Journalistically speaking, this is some bullshit.
Why's the deal like this? No idea. Does it make sense? No idea. But it's not odd that Groq is continuing to raise more money, because they never stopped being a normally operating company.
If you want an explanation for why Nvidia would do this deal, my best offer is here (https://openrouter.ai/rankings): Of the top 10 fastest AI models, Groq is the provider of 4 of them. And of the price of those top 10 fastest AI models, Groq is #1, #2, #3, and #5. And you wonder why someone's giving them a measly half billion dollars? They're the fastest cheapest thing on the market. If you don't understand the value of that, you really don't understand AI.
> There's also no official source for the amount Nvidia paid for the tech,
Did you honestly believe that this kind of deal doesn't leave a trace in financial disclosures?
The February 2026 Nvidia 10-K has this:
Cash flows from investing activities: Groq. -$13B.
And this: "Total consideration consists of $13.0 billion paid at closing and $4 billion, inclusive of imputed interest, payable within one year included in Accrued and Other Current Liabilities on our Consolidated Balance
Sheets."
Deals like these are quite rare, would be nice to know more about how the funding worked, but for customers, what are they doing? Total silence from Groq for a very long time now.
what's the confusion. Groq offers a fast inference solution that is currently unique (why do you think Nvidia paid $8 billion to end-run around the SEC to acquire the technology). This is good news as it ensures that Groq customers can can be assured continuity to use their service.
You pay them money to run models. Their website doesn’t list them as offering any models that were released recently. For a “pay for inference” provider, questions like “do they have Deepseek or Qwen 2.6?” are germane
I’m confused by the confusion. Groq licensed their technology (sold part of their business) to Nvidia for a large amount of money and distributed the spoils to their investors. Seems quite normal? But then the Axios article says…
“Existing shareholders will receive the remaining cash distributions and then have the opportunity to invest into a new company”
New company? But Groq still exists and continued to exist.
“The bottom line: Don't be surprised if this becomes a new transaction template in the AI private markets.”
A transaction template? I don’t follow what was novel about this situation. The Meta not-acquisition-acquisition of Scale seems more novel.
I guess I feel like Zach’s confusion is because of the way Axios has presented what is happening to Groq. Looking at why actually happened with Groq, it seems like Axios are reporting it weird.
Unless Groq really is starting a new company in which case I am equally as confused.
The interesting thing here isn't "how, logistically, is the Groq corporate entity able to raise more money?". That's straightforward.
Rather, the interesting thing and the topic of most of the article is "how, after Nvidia hired most of Groq's team and licensed all their IP, did Groq manage to convince investors to invest in the remaining corporate entity?"
There's nothing normal at all about the Nvidia Groq deal, it's hard to read in terms of what it means. A straight licensing deal would have been easier to ingest.
I could be completely off the mark but I thought the non-exclusive license was necessary because Groq’s datacenter business uses the technology already? Nvidia acquired the assets but Groq needed to retain rights to use the technology for their own product.
Are stagnating models really a problem? Because if so, then how is "if we stop training we are profitable" an acceptable answer to the profitability question?
Groq don't train their own models, they serve open ones. They're a few versions behind the frontier open weights models. The top model they serve is GPT OSS 120B, released last summer.
I think you might be confusing Groq with another company. Groq is still independent — that's the whole premise of this thread (they're raising more money). They haven't been aquihired by NVIDIA.
As far as I know, Groq serves their own LPU hardware and still offers models like Llama 4, not just gpt-oss-120b. Do you have a source for the Kimi K2 removal being NVIDIA-related?
Groq stopped serving Kimi K2 (1T params) when they got aquihired by NVIDIA, so I guess NVIDIA took most of the hardware in addition to the employees. The largest model they serve now is the relatively minuscule gpt-oss-120b.
The community support forum is also getting retired and there haven't been any posts by support employees in forever anyway, so they are probably gone, too. Also, the number of issues have been piling up, suggesting that the developers are gone as well. https://community.groq.com/c/forum/4 (archive link for when it goes down https://web.archive.org/web/20260602064050/https://community...)
To me, it looks they are trying to raise 650M with a few remaining (ancient) LPUs and no employees.
Before they removed it, I was using groq Kimi K2 model for a chat bot in small community site/chat. It was really good, seemed to have incredibly vast general world knowledge and the fast speed (400tok/s if I remember right) meant that chat users got a response instantly which was a much better experience compared to other SOTA models at the time.
On the bright side it looks like Cerebras might be serving Kimi K2.6 at 1000tok/s soon https://www.cerebras.ai/blog/cerebras-kimi-k2-Enterprise
Those were amazing times. You could vibe code an entire prototype in seconds (200 tps). With Qwen3.6-35B-A3B and MTP, you can program at that speed on a single GPU at home now, but Kimi K2 is of course much smarter at almost 30 times the size.
I'm also looking forward for the Cerebras Kimi K2.6 release, which should be even better at 1000 tps. It is hard to overstate how important speed is for programming. Instead of having to wait for a few minutes until a task is done, it is just done instantly, and you don't have to context switch from whatever else you were working on while waiting.
I hope they will make it available to regular customers.
But too much of a speed doesn’t allow you to build up the context as the llm is working, it’s a two-edged sword.
Cerebras are only serving kimi for dedicated endpoint customers; for that you need a >$5m annual deal with them
Cerebras also seems to be killing off their regular APIs, they're deprecating models and GLM is still stuck on GLM 4.7, a whole 2 versions behind.
I was quite baffled they removed it and didn't double down on Kimi and serving the latest models instead.
Thanks for the tip, looks fire.
> The largest model they serve now is the relatively minuscule gpt-oss-120b
This model will run on any laptop with 128GB RAM, wow.
I've was part of an acquire-to-remove-an-annoyance, including having the better product that didn't win for niche reasons - which vaguely matches the situation here. Yes, NVIDIA can wind down Groq (unambiguously a better product in this case) even though it doesn't make sense.
That being said, there's still a chance that NVIDIA engineering is in the process of stripping it for parts. Or the lawyers are - maybe they have too much momentum with GPUs and just want ASICs out of the market.
This kind of innovation stifling acquisition should have been blocked. NVIDIA is a serious monopoly threat.
Somehow I never got the hype around Groq. They served models with fast inference speed - that sounded great in theory - and as a user I was looking forward to use them. But, when I did try, I discovered that they are quantizing the models underneath. And they dont even disclose it. So I stopped using them.
The whole thing never made any sense to me - but I guess AI hype is a thing.
I believe despite quantisation they were still extremely fast, which is still incredibly useful if you don't need high precision/accuracy (which is good enough for many use cases)
I don't really get the value proposition of groq as a user, the performance is really poor for the token price. Data centres on the other hand are becoming a commodity, and I don't see any reason a priori to invest in groq specifically for something like that.
> the performance is really poor for the token price
That doesn’t match my experience or the numbers:
https://openrouter.ai/openai/gpt-oss-120b?sort=throughput
Groq is considerably faster and better at inference, they have a totally superior product to Nvidia for inference based tasks, which will be the dominant concern in the future.
Plausibly, take all the Nvidia hype and multiply that by a factor and that's what 'Groq' could be worth.
And there is no real commodification - there's Nvidia, Cerebras, Groq ... not many otheres.
> they have a totally superior product to Nvidia for inference based tasks
They're not really competing with Nvidia because 1) Nvidia owns their chips now, and 2) Nvidia is not really an inference provider.
Groq is a slicon maker, the inference provider stuff is a path to market, it's not really the reflection of their market potential.
Nvidia doesn't own them or all their IP now, we don't quite know the terms of the deal.
AFAIK the terms were the chip-making + talent stuff went to Nvidia, and the api provider stuff gets to keep existing separately.
Define “totally superior”?
Was this comment created using quantized llama 3?
I love Groq, but across every single line break in your post there is a glaring issue that is easy to refute with in 15 seconds, even without 300t/s of throughput.
You wasted all of your commentary on snark and sadly unfunny humour, and yet still managed to add nothing.
Groq is more performant for the growing categories of inference-based tasks, wherein Nvidia's advantage in inference depends bulk/batch processing which will make up a smaller category over time, in relative terms.
The future of AI Silicon is inference, and the cost structure of AI data centres is constrained around the current necessity to have 'high GPU utilization' otherwise, the cost / amortization of the chips doesn't work out.
That cost structure is a limitation of Nvidia architecture.
Groq serves a lot faster, and without the limiting batching requirement, which opens hosting arrangements common in most classical hosting scenarios aka without necessarily the high utilization requirements.
Groq has bespoke hardware, lack of CUDA, much lower memory desnsity obviously and they don't have the deep distribution networks and leverage over TSMC that Nvidia has - but pound for pound, were we to be able to 'fire up a server' for our inference needs, it would be Groq, not Nvidia that we'd turn to.
Were they not a later market entrant and didn't have those barriers to entry, they'd be gigantic.
is groq still using 6 racks to serve Llama3-70B or is that old news?
The new chip isn't out yet so that's the only thing they could be doing.
Google has been releasing a new TPU generation every year since 2023 and the eight generation consists of a training and an inference optimized design.
Google's eight generation TPU inference chip has 384 MB of on-chip SRAM vs 500 MB for Groq's third generation LPU.
My company had a really terrible experience trying to use Groq, and I would NOT recommend anyone use their service if you need reliability. So many random errors, so many silly quirks.
Are you sure your post is about
> Groq, the AI chip company company that was acquired by Nvidia in December of last year, is raising $650M
? Could you provide details?
What's so surprising?
I'd say they have a "Tier 2" inference stack, but that'd be Apple Silicon and AMD: their well below that.
In part because unlike those, no one can get their hands on the hardware, so they miss out on a massive amount of free development, testing, etc.
At this point there's years of complaints about open weight models performing worse on their platform, tool calling makes it especially easy to tell since it's so sensitive to these issues.
as soon as i saw they switched to "call us for quotes" for the new models, i knew they are over.
About a year back I tried to build a few latency-sensitive products on Groq, and found the response times would vary so wildly that it just about erased any benefit in their throughput. Is that still this case?
I don’t understand one part of the licensing here: if it was just a license, can’t they relicense the software and hardware of LPU3 to AMD? Or hire new software and hardware people?
The new designs were their main asset besides the amazing talent that went to NVIDIA, not the remaining DCs.
Maybe Tom Ellis can give another AMA https://news.ycombinator.com/item?id=39429047 so we can find out.
The most bizarre thing here is the reporters. They intentionally misrepresented what happened. All the news stories from last year claimed Nvidia "acquired" Groq - and in the same story, quoted what actually happened, and pretended it didn't. It's like the journalists had some kind of group psychosis, pretending it was an acquisition. It wasn't. It was a really white-glove product rental.
From the actual Groq PR release:
"Groq announced that it has entered into a non-exclusive licensing agreement with Nvidia" - "As part of this agreement, Jonathan Ross, Groq’s Founder, Sunny Madra, Groq’s President, and other members of the Groq team will join Nvidia to help advance and scale the licensed technology" - "Groq will continue to operate as an independent company with Simon Edwards stepping into the role of Chief Executive Officer" (https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusiv...)
Nothing about an acquisition there. It says Nvidia is licensing it, and that others can too. The execs work for Nvidia to integrate it into Nvidia's... something. And Groq the company remains the same as before.
There's also no official source for the amount Nvidia paid for the tech, or two unofficial ones. Journalistically speaking, this is some bullshit.
Why's the deal like this? No idea. Does it make sense? No idea. But it's not odd that Groq is continuing to raise more money, because they never stopped being a normally operating company.
If you want an explanation for why Nvidia would do this deal, my best offer is here (https://openrouter.ai/rankings): Of the top 10 fastest AI models, Groq is the provider of 4 of them. And of the price of those top 10 fastest AI models, Groq is #1, #2, #3, and #5. And you wonder why someone's giving them a measly half billion dollars? They're the fastest cheapest thing on the market. If you don't understand the value of that, you really don't understand AI.
> There's also no official source for the amount Nvidia paid for the tech,
Did you honestly believe that this kind of deal doesn't leave a trace in financial disclosures?
The February 2026 Nvidia 10-K has this:
Cash flows from investing activities: Groq. -$13B.
And this: "Total consideration consists of $13.0 billion paid at closing and $4 billion, inclusive of imputed interest, payable within one year included in Accrued and Other Current Liabilities on our Consolidated Balance Sheets."
Deals like these are quite rare, would be nice to know more about how the funding worked, but for customers, what are they doing? Total silence from Groq for a very long time now.
what's the confusion. Groq offers a fast inference solution that is currently unique (why do you think Nvidia paid $8 billion to end-run around the SEC to acquire the technology). This is good news as it ensures that Groq customers can can be assured continuity to use their service.
Do they have any good models yet?
It’s not X AI, that would be Grok with a ‘k’.
You pay them money to run models. Their website doesn’t list them as offering any models that were released recently. For a “pay for inference” provider, questions like “do they have Deepseek or Qwen 2.6?” are germane
The list is:
-- allam-2-7b
-- canopylabs/orpheus-arabic-saudi
-- canopylabs/orpheus-v1-english
-- groq/compound
-- groq/compound-mini
-- llama-3.1-8b-instant
-- llama-3.3-70b-versatile
-- meta-llama/llama-4-scout-17b-16e-instruct
-- meta-llama/llama-prompt-guard-2-22m
-- meta-llama/llama-prompt-guard-2-86m
-- openai/gpt-oss-120b
-- openai/gpt-oss-20b
-- openai/gpt-oss-safeguard-20b
-- qwen/qwen3-32b
-- whisper-large-v3
-- whisper-large-v3-turbo
Thanks. These all came out a year ago? Except a few which came out two years ago.
I’m confused by the confusion. Groq licensed their technology (sold part of their business) to Nvidia for a large amount of money and distributed the spoils to their investors. Seems quite normal? But then the Axios article says…
“Existing shareholders will receive the remaining cash distributions and then have the opportunity to invest into a new company”
New company? But Groq still exists and continued to exist.
“The bottom line: Don't be surprised if this becomes a new transaction template in the AI private markets.”
A transaction template? I don’t follow what was novel about this situation. The Meta not-acquisition-acquisition of Scale seems more novel.
I guess I feel like Zach’s confusion is because of the way Axios has presented what is happening to Groq. Looking at why actually happened with Groq, it seems like Axios are reporting it weird.
Unless Groq really is starting a new company in which case I am equally as confused.
edit: when announced last year it was announced as an asset acquisition https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startu...
The interesting thing here isn't "how, logistically, is the Groq corporate entity able to raise more money?". That's straightforward.
Rather, the interesting thing and the topic of most of the article is "how, after Nvidia hired most of Groq's team and licensed all their IP, did Groq manage to convince investors to invest in the remaining corporate entity?"
I thought you wrote the convincing explanation:
“One could argue that Groq’s datacenters alone could make them worth billions of dollars.”
Groq is a successful datacenter business with a high-revenue cloud product. That’s a compelling investment in its own right, right?
> Groq Launches European Data Center Footprint in Helsinki, Finland
https://groq.com/newsroom/groq-launches-european-data-center...
That sounds like they are renting racks in a Equinix data centre. Do Groq have 4 data centers worth billions?
There's nothing normal at all about the Nvidia Groq deal, it's hard to read in terms of what it means. A straight licensing deal would have been easier to ingest.
I could be completely off the mark but I thought the non-exclusive license was necessary because Groq’s datacenter business uses the technology already? Nvidia acquired the assets but Groq needed to retain rights to use the technology for their own product.
The deal was probably structured the way it was due to concerns over regulatory approval.
These 'we get your executives' type of deals - aka Windsurf - are new, weird thing in M&A.
They could have sold the IP then licensed it back. The nonexclusive part was purely a fig leaf to dodge antitrust.
How well can a company continue to exist when the vast majority of its employees leave that company all at once via an acquihire?
i like groq but models seem to have stagnated - looks like the company isn't focused on b2c anymore?
Are stagnating models really a problem? Because if so, then how is "if we stop training we are profitable" an acceptable answer to the profitability question?
Groq don't train their own models, they serve open ones. They're a few versions behind the frontier open weights models. The top model they serve is GPT OSS 120B, released last summer.
I think you might be confusing Groq with another company. Groq is still independent — that's the whole premise of this thread (they're raising more money). They haven't been aquihired by NVIDIA. As far as I know, Groq serves their own LPU hardware and still offers models like Llama 4, not just gpt-oss-120b. Do you have a source for the Kimi K2 removal being NVIDIA-related?