> Today, Groq announced that it has entered into a non-exclusive licensing agreement with Nvidia for Groq’s inference technology. The agreement reflects a shared focus on expanding access to high-performance, low cost inference.
> 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.
> GroqCloud will continue to operate without interruption.
Indeed, as justincormack comments: ”It is not structured as an outright acquisition to avoid US Gov't anti trust scrutiny, but effectively it probably is”. “Non-exclusive” ? Ummmm, yeah, right, sure. You can probably bet there is an private understanding that Groq will no longer offer it's “top of the line” best technology to competitors of Nvidia. Some may see this as a clever, “slight of the hand” attempt for Nvidia to maintain it's perceived dominance & lead in GPU-TPU development.
“Non-exclusive” does not in any form or fashion spell out that all Nvidia's competitors can and will obtain the very top, cutting edge Groq technology as Nvidia will obtain . . .
> 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.
A really strange agreement where top executives of a company "join" another company for the benefit of the other company.
Are they buying them to try and slow down open source models and protect the massive amounts of money they make from OpenAI, Anthropic, Meta ect?
It quite obvious that open source models are catching up to closed source models very fast they about 3-4 months behind right now, and yeah they are trained on Nvidia chips, but as the open source models become more usable, and closer to closed source models they will eat into Nvidia profit as these companies aren't spending tens of billion dollars on chips to train and run inference. These are smaller models trained on fewer GPUs and they are performing as good as the pervious OpenAI and Anthropic models.
So obviously open source models are a direct threat to Nvidia, and they only thing open source models struggle at is scaling inference and this is where Groq and Cerberus come into the picture as they provide the fastest inference for open source models that make them even more usable than SOTA models.
Shy of an algo breakthrough, open source isn't going to catch up with SOTA, their main trick for model improvement is distilling the SOTA models. That's why they they have perpetually been "right behind".
They don't need to catch up. They just need to be good enough and fast as fuck. Vast majority of useful tasks of LLMs has nothing to do with how smart they are.
GPT-5 models have been the most useless models out of any model released this year despite being SOTA, and it because it slow as fuck.
Hard disagree. There are very few scenarios where I'd pick speed (quantity) over intelligence (quality) for anything remotely to do with building systems.
I agree with you for many use cases, but for the use case I'm focused on (Voice AI) speed is absolutely everything. Every millisecond counts for voice, and most voice use cases don't require anything close to "deep thinking. E.g., for inbound customer support use cases, we really just want the voice agent to be fast and follow the SOP.
For coding I don’t use any of the previous gen models anymore.
Ideally I would have both fast and SOTA; if I would have to pick one I’d go with SOTA.
There a report by OpenRouter on what folks tend to pay for it; it generally is SOTA in the coding domain. Folks are still paying a premium for them today.
There is a question if there is a bar where coding models are “good enough”; for myself I always want smarter / SOTA.
FWIW coding is one of the largest usages for LLM's where SOTA quality matters.
I think the bar for when coding models are "good enough" will be a tradeoff between performance and price. I could be using Cerebras Code and saving $50 a month, but Opus 4.5 is fast enough and I value the piece-of-mind I have knowing it's quality is higher than Cerebras' open source models to spend the extra money. It might take a while for this gap to close, and what is considered "good enough" will be different for every developer, but certainly this gap cannot exist forever.
> They don't need to catch up. They just need to be good enough
The current SOTA models are impressive but still far from what I’d consider good enough to not be a constant exercise in frustration. When the SOTA models still have a long way to go, the open weights models have an even further gap distance to catch up.
I am by no means an expert, but I think it is a process that allows training LLMs from other LLMs without needing as much compute or nearly as much data as training from scratch. I think this was the thing deepseek pioneered. Don’t quote me on any of that though.
Too bad, so sad for the Mister Krabs secret recipe-pilled labs. Shy of something fundamental changing, it will always be possible to make a distillation that is 98% as good as a frontier model for ~1% of the cost of training the SOTA model. Some technology just wants to be free :)
Then why are they spending $20 billion dollars to handicap an inference company that giving open source models a major advantage over closed source models?
Realistically groq is a great solution but has near impossible requirements for deployment. Just look at how many adapters you need to meet the memory requirements of a small llm. SRAM is fast but small.
I would guess their interconnect technology is what NVIDIA wants. You need something like 75 adapters for an 8b parameter model they had some really interesting tech to make the accelerator to accelerator communication work and scale. They were able to do that well before nvl 72 and they scale to hundreds of adapters since large models require more adapters still.
That's a non-charitable interpretation of what happened. The are not "spending $20 billion to handicap Groq". They are handing Groq $20 billion to do whatever they want with it. Groq can take this money and build more chips, do more R&D, hire more people. $20 billion is truly a lot of money. It's quite hard to "handicap" someone by giving them $20 billion.
> Groq added that it will continue as an “independent company,” led by finance chief Simon Edwards as CEO.
The $20B does not go to Groq's investors. It goes to Groq. You can say that Groq is owned by its investors, and this is the same thing, but it's not. In order for the money to go to the investors, Groq needs to disburse a dividend, or to buy back shares. There is no indication that this will happen. And what's more, the investors don't even need this to happen. I'm sure any investor that wants to sell their shares in Groq will now find plenty of buyers at a very advantageous price.
Yes, you are way off, because Groq doesn't make open source models. Groq makes innovative AI accelerator chips that are significantly faster than Nvidia's.
> Groq makes innovative AI accelerator chips that are significantly faster than Nvidia's.
Yeah I'm disappointed by this, this is clearly to move them out of the market. Still, that leaves a vacuum for someone else to fill. I was extremely impressed by Groq last I messed about with it, the inference speed was bonkers.
For inference, but yes. Many hundreds of tokens per second of output is the norm, in my experience. I don't recall the prompt processing figures but I think it was somewhere in the low hundreds of tokens per second (so slightly slower than inference).
Nvidia just released their Nemotron models, and in my testing, they are the best performing models on low-end consumer hardware in both terms of speed and accuracy.
I'd say that it's probably not a play against open source, but more trying to remove/change the bottlenecks in the current chip production cycle. Nvidia likely doesn't care who wins, they just want to sell their chips. They literally can't make enough to meet current demand. If they split off the inference business (and now own one of the only purchasable alternatives) they can spin up more production.
That said, it's completely anti-competitive. Nvidia could design a inference chip themselves, but instead the are locking down one of the only real independents. But... Nobody was saying Groq was making any real money. This might just be a rescue mission.
They need to vertically integrate the entire stack or they die. All of the big players are already making plans for their own chips/hardware. They see everyone else competing for the exact same vendor’s chips and need to diversify.
Idk- cheaper inference seems to be a huge industry secret and providing the best inference tech that only works with nvidia seems like a good plan. Makes nvidia the absolute king of compute against AWS/AMD/Intel seems like a no brainer.
The problem is, if everyone knows it going to curry favour and you're the odd man out - are you in Violation of your fiduciary duty to your shareholders?
as useful as it was before this administration when big tech was sucking up to whomever was running the country (e.g. “macho man” Zuck was getting ready to tattoo DEI on his forehead couple of years ago) or just now it’ll be magically useful?
It’ll go through. It’s not an acquisition, it’s an exclusive licensing deal. Same end result, but it lets them runaround the usual regulatory approvals for acquisitions.
I got curious about how many wheelbarrows of cash $20bn actually is.
Two ways to think about it: weight vs volume.
By weight (assuming all $100 bills):
$20,000,000,000 / $100 = 200,000,000 bills
Each bill is roughly 1g, so total mass is ~200,000 kg
A typical builder’s wheelbarrow can take about 100 kg before it becomes unmanageable
200,000 kg total
/ 100 kg per wheelbarrow
≈ 2,000 wheelbarrows (weight limit)
By volume:
A $100 bill is ~6.14" × 2.61" × 0.11 mm, which comes out to about 102 cm³ per bill
200,000,000 bills × 102 cm³ ≈ 20,400 m³ of cash
A standard wheelbarrow holds around 0.08 m³ (80 litres)
20,400 m³ total
/ 0.08 m³ per wheelbarrow
≈ 255,000 wheelbarrows (volume limit)
So,
About 2,000 wheelbarrows if you only care about weight
About 255,000 wheelbarrows if you actually have to fit the cash in
So the limiting factor isn’t how heavy the money is; it’s that the physical volume of the cash is absurd. At this scale, $20bn in $100s is effectively a warehouse, not a stack.
I think your volume per bill should be 6.14 * 0.0254 * 2.61 * 0.0254 * 0.00011 ≈ 1.137e-6 m³. That means about 227 m³ total volume, or about 2800 wheelbarrows.
The price is 40x their target revenue. That's twice the price to revenue multiplier applied to Anthropic in their most recent funding round, and really really hard to portray as a good deal.
I don't think it really helps Nvidia's competitive position. The serious competition to Nvidia is coming from Google's TPU, Amazon's Trainium, AMD's Instinct, and to a much lesser extent Intel's ARC.
Grow recent investors got back a 3x multiple and may now invest in one of Nvidia's other competitors instead.
The only thing I can think of here is that OpenAI’s DRAM land grab is going to stack on a non-NV target and NV need to hedge with an SRAM design that’s on the market NOW. Otherwise, I can’t see how NV couldn’t eat Groq’s lunch in one development cycle - it’s not like NV can’t attach a TPU to some SRAM and an interconnect. Either that or Groq closed a deep enough book to scare them, but 40x is a lot of scared.
That's an interesting take, it's plausible Nvidia wants to have an SRAM based product, but I am struggling to see why they would pay $20bn to have one /right now/. Even if DRAM prices make Groq's approach more economical, Nvidia can develop a competitive product before Groq could take any real market share.
Exactly. The only way this makes sense to me is if the board needed this product in <1 cycle. Which makes no sense for a market player like NV who already have the PDK, volume, and literally everything else in the universe. But here it is, so there is clearly a factor I have not considered :)
But if all of these large companies create a monopoly where they realize that if they start undercutting each other or innovating too extremely then long term if Nvidia's stock decreases in value by a huge margin, the whole market itself can get down
As an example: if google TPU (perfected?) itself to the point that it hurts nvidia sales (maybe mass production perhaps?) then all the companies stock price might decrease (in my opinion including google)
Honestly, I feel like there is going to happen something in the market which is gonna be very spooky soon regarding AI.
I feel like we are gonna drag this bubble really long and actually worsen all the pain which is gonna be caused by it long term.
There is also movement over of some key people to nvidia which is pretty significant: "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."
curious how the execs can honor the fiduciary duty to share holders assuming only they? get the 20 billion and the company is left headless and leaking all of its intellectual property to 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.
Legit feels like Nvidia just buying out competition to maintain their position and power in the industry. I sincerely hope they fall flat on their face.
> Legit feels like Nvidia just buying out competition to maintain their position and power
Well, I mean, isn't that exactly what they should be doing? (I'm not talking about whether or not it benefits society; this is more along the lines of how they're incentivized.)
Put yourself in their shoes. If you had all that cash, and you're hearing people talk of an "AI Bubble" on a daily basis, and you want to try and ensure that you ride the wave without ever crashing... the only rational thing to do is use the money to try and cover all your bases. This means buying competitors and it also means diversifying a little bit.
Dunno thought AGI would make everything obsolete and it's just around the corner? It looks rather like it dawns on everyone that transformers won't bring salvation. It's a show of weakness.
The bottleneck in training and inference isn’t matmul, and once a chip isn’t a kindergarten toy you don’t go from FPGA to tape out by clicking a button. For local memory he’s going to have to learn to either stack DRAM (not “3000 lines of verilog” and requires a supply chain which openai just destroyed) or diffuse block RAM / SRAM like Groq which is astronomically expensive bit for bit and torpedoes yields, compounding the issue. Then comes interconnect.
Look dude, this guy failed his Twitter internship and is not about to take on Jensen Huang. This isn't some young guy anymore and this isn't 200x where is he about to have another iPhone / Sony moment.
There's this curious experience of people bringing up geohot / tinygrad and you can tell they've been sold into a personality cult.
I don't mean that pejoratively, I apologize for the bluntness. It's just I've been dealing with his nonsense since iPhone OS 1.0 x jailbreaking, and I hate seeing people taken advantage of.
(nvidia x macs x thunderbolt has been a thing for years and years and years, well before geohot) (tweet is non-sequitor beyond bogstandard geohot tells: odd obsession with LoC, and we're 2 years away from Changing The Game, just like we were 2 years ago)
My deepest apologies, I can't parse this and I earnestly tried: 5 minutes of my own thinking, then 3 llms, then a 10 minute timer of my own thinking over the whole thing.
My guess is you're trying to communicate "tinygrad doesn't need gpu drivers" which maybe is transmutated into "tinygrad replaces CUDA" and you think "CUDA means other GPUs can't be used for LLMs, thus nvidia has a strangehold"
I know George has pushed this idea for years now, but, you have to look no further than AMD/Google making massive deals to understand how it works on the ground.
I hope he doesn't victimize you further with his rants. It's cruel of him to use people to assuage this own ego and make them look silly in public.
Not good. This shouldn't be allowed. What would be better is if groq and cerebras combined, and maybe other companies invested in them to help them scale. Why would the major cloud providers not lobby against this?
Usually antitrust is for consumers, but here I think companies like Microsoft and AWS would be the biggest beneficiaries of having more AI chip competition.
Market share wise, Groq is perhaps "tiny"? Nvidia may be paying a premium for Groq [0] since it eliminates competition (at least on the inference side).
That’s a loophole. Regulation hasn’t caught up to the innovation of non-exclusive licensing deal. Hopefully we’ll get some competence back in government soon-ish and can rectify the mistake
There isn't to be shared between the two techs, Groq's hardware is a like a railgun that installs all the weights into the optimal location before firing off an inference. Cerebras computer engineering more convention requiring the same data movement that GPUs struggle with optimizing.
Suspect Groq is complementary/superior to nvidia's GPUs, while it is unclear what Cerebras brings other then maybe some deals with TSMC.
There are others as well but NVidia is aggressive when it comes to punishing companies willing to buy non-NVidia products. As a result, they prefer to remain under the radar, at least until they have enough market leverage to be more widely known.
I imagine 2 big giants basically (Nvidia/google/amd basically influenced by a few select of people) vs (Chinese companies who have investments from the govt)
its sort of like proxy wars and this is sort of whats happening in software side of things with open source models but I think that the benefit of the proxy wars is going to be for the end consumers
But although on the other hand, having two large countries compete with each other while buying everything else and all feels like it astronomically increases the price if someone wants to compete with these two giants (any other country perhaps)
We definitely need a better system where it doesn't feel like we are seeing pacman eat everything up basically
Is it not? All this money is going into AI under the fear that China will win the race to AGI. China releases open-source models that keep OpenAI/Anthropic researching and training their models, which in turn creates demand for more Nvidia GPUs.
I just stopped my Groq API. Sad to see competition being eaten up by shitty Nvidia. I like their products but Jensen is an absolute mfer with deceitful marketing.
Honestly cerebras is good. I can recommend it. I talked to their team once on discord as a literal free user so that was something really nice personally
They were also the faster one compared to groq but they were always a little slow on adding new models compared to groq but not sure what changed right now.
Definitely recommend cerebras tho now that groq's been eaten up from inside basically
I do not understand this move by Nvidia, they are afraid of being out competed by this startup in their core competence of building chips for AI? They may be eliminating a competitor for now but this move will immediately many more AI chip startups to get founded
They're not eliminating a competitor, they're (effectively) acquiring a competitor. Nvidia's GPUs are great for training, and not bad for inference, but the custom chips are better for inference and Nvidia's worried about losing customers. Nvidia will no doubt sell custom Groq-like chips for inference now.
groq is a series E hardware startup founded in 2016. It took them this long to be a potential threat, I'm not sure they are even an actual threat.
Even if this purchase causes 100 new hardware startups to be funded tomorrow, nVidia is perfectly fine with that. Let's see how many survive 5 years down the line
the play is 10x faster inference leads to 100x demand give or take, which isn't a bad assumption at all if you ask me. the problem is actually fitting a good model onto hardware that fast.
This is smart as hell. I’ve long wondered how they’d combat ASIC’s without diluting their own benefits. This gives them a bit more time to figure out the moats, which is useful because Groq was going places. This juices Groq’s distribution, production, ability to access a wider range of skills where necessary.
I expect China to want to compete with this. Simpler than full-blown Nvidia chips. Cue much cheaper and faster inference for all.
They almost certainly plan to invest in the technology. One of the biggest threats to Nvidia is people developing AI-centric ASICs before they get there. Yes, Google has their TPUs and there are others around, but it's early on.
In some ways, it's not about eliminating a competitor. It's about eliminating all the competitors. Nvidia can use its resources to push AI ASICs farther faster than others, potentially cutting off a whole host of competitors that threaten their business. Nvidia has the hardware and software talent, the money, and the market position to give their AI ASICs an advantage. They know if they don't lean into ASICs that someone else will and their gravy train will end. So they almost certainly won't be abandoning the technology.
The acquisition price of a company usually comes at a premium to the last valuation. This applies even with publicly traded companies, which is why acquisition announcements cause stock prices to pop up to some number between the last trade price and the acquisition price, proportional to how much the market thinks the acquisition is likely to go through.
The premium can make sense to the acquirer because the acquired company is worth more when combined with all of the assets and power (brand name, distribution, patents, trade secrets) of the acquiring company.
This confuses a lot of people who think the valuation of a company is equivalent to the number that would be paid to acquire it at that instant, but it’s not.
Imagine a pharma with a weight loss drug that isn't approved yet; it's either worth $20B (if approved) or zero (if not approved).
Now imagine the LPUv2 ASIC. If it works it's worth $20B and if it doesn't it's zero. If investors think LPUv2 has a 1/3 chance of success they would buy in at $7B. Then the chip boots up and... look at that.
I got recommended this and I will watch it today, thank you. One of the comments points out
"They’ve literally told us that the plan is to get bailed out by the taxpayers"
This reminded me of how I think what's gonna happen/ is already happening is that they become too big to fail and get bailed out and the burden/loss becomes of taxpayers
So we are kind of living in a system which is reckless about finances/stability behind businesses where the system is such that all the profits are privatized but all the losses are shared/even funded by the average person
Combine in a mix of corruption in any political party to begin with and I am wondering why we don't have yet another french revolution.
They want the Trump Jr. and Chamath connections to be forgotten by the time the media and YouTubers get active again after New Year. It's a standard media tactic.
Honestly, it's basically an accepted fact that the best thing to do is just to never respond anything that media/people say. If you respond, the conversation will continue. If you just say nothing, they'll move on to the next topic, because everyone needs to keep their audience engaged. With no response, your audience isn't interested either, so you have to move onto something else.
Hopefully prices stay the same, I run my small apps on groq. I get good enough summarization and simple agents from gpt120b which is on 15 cents for a million input tks.
I think it’s pretty obvious at this point that Nvidia’s architecture has reached scaling limits - the power demands of their latest chips has Microsoft investing in nuclear fusion. Similar to Intel in both the pre-Core days and their more recent chips, they need an actual new architecture to move forward. As sits, there’s no path to profitability for the buyers of these chips given the cost and capabilities of the current LLM architectures, and this is obvious enough that even Nvidia has to realize it’s existential for them.
If Groq’s architecture can actually change the economics of inference and training sufficient to bring the costs in line with the actual, not speculative, benefits of LLMs, this may not be a buy-and-kill for Nvidia but something closer to Apple’s acquisition of P.A. Semi, which made the A- and M- class chips possible.
(Mind you, in Intel’s case they had to have their clocks cleaned by AMD a couple times to get them to see, but I think we’re further past the point of diminishing returns with Nvidia - I think they’re far enough past when the economics turned against them that Reality is their competition now.)
No path to profitability for the people using their products for their putative purpose, which seems like it might affect Nvidia’s bottom line at some point. Clarified.
I believe that this news kind of helps cerebras as groq and cerebras are the only two companies working extensively in this space
I feel as if Nvidia is eating up even companies which I thought had genuine potential or anything related to AI industry whether profitable or not
Nvidia's trying its best to take all major players and consolidate into one big entity from top to bottom.
The problem with this approach imo is that long term, nvidia's stock is extremely overvalued and its still a bubble which will burst and it will take nvidia first and foremost.
The issue is that when nvidia falls, it will take the whole literal industry from top to bottom, even those companies which I thought could survive an AI burst. Long term I feel like it will have really bad impacts if nvidia continues to gobble up every company.
I am pretty sure that Nvidia might be looking at cerebras too and if they offer them a shit ton of money and cerebras gets bought. I genuinely believe that Nvidia has sort of invested in literally all pockets of any hardware related investment for AI. And when OpenAI is unable to pay Nvidia, I feel like it can all come crashing down since this whole cycle is only being possible via external investor money.
Maybe the EU or individual states will sue under their own anti-trust laws will stop this - seems pretty clearly anti-competitive and probably a prelude of these over-valued companies using their stock to gobble up any possible competitor to consolidate even more.
In my opinion, Groq's technical decisions are unsound in a normal world. But being HBM-free may have some merit in a world where HBM supply is constrained.
Groq is very different from TPUs. The difference in memory should be a big clue. (Groq using SRAM built into each compute unit, vs TPUs using HBM more similarly to GPUs)
> Prior to founding Groq, Jonathan began what became Google’s Tensor Processing Unit (TPU) as a 20% project where he designed and implemented the core elements of the first generation TPU chip.
For nvidia increasing the numbers of their money-multiplying mutual investment ring is more important than the value of the deals. It's about involving more capital and people and making their grift too big to fail and keeping the stock numbers up. Nvidia has the ability to promise large amounts of money like this in announcements but I haven't read about any of them actually having money or good exchange hands yet.
No, it doesn't really matter if they pay in cash or stock. If you think NVDA has room to run you're welcome to use your buyout money to buy NVDA on the open market.
I'm not sure in this specific case. They could choose to pay the employees some portion of the funds.
If not, the owners are likely liable to be sued for "selling in effect" without paying equity holders.
Presuming the company becomes a defacto subsidiary of Nvidia (even if not legally so)
My guess, without researching it, is they will compensate existing equity holders to avoid that possibility. I mean the valuation multiple is enormous, it's worth it simply to derisk the legal aspect.
For vested RSUs it's likely that the Groq husk will pay out the $20B as a dividend or buyback or something. I don't know if unvested RSUs are accelerated or just canceled. Of course the employees will receive new RSUs when they join Nvidia.
groq was targeting a part of the stack where cuda was weakest: guaranteed inference time at a lower cost per token at scale. This was in response to more than just goog's tpus, they were also one of the few realistic alternative paths oai had with those wafers.
The last time I tested, I could swear they were doing some problematic quantization, because I was getting kind of random results with one or two models, which worked perfectly when I switched providers.
It was really disappointing too because Cerebras does not provide any service reliability on their cheap plans. So I came to the conclusion that unless I could convince the client to set up an enterprise contract or something, we could not use either provider for low-latency, which we need for voice calls. I think for organizations that can afford a hefty contract that guarantees service levels, Groq and Cerebras especially are basically cheat codes for meeting latency requirements for voice. But that might not be an option for really small businesses.. although maybe I am just not a good sales person.
> Groq is expected to alert its investors about the deal later on Wednesday. While the acquisition includes all of Groq’s assets, its nascent Groq cloud business is not part of the transaction, said Davis.
Wait, what? How is the cloud business supposed to run if Nvidia is acquiring the rights to the hardware?
It isn't, and the other companies that offer cloud AI that Nvidia has invested in can carry on happy they have one less competitor.
This is how business works in the 21st century - once one company has a dominant position and a massive warchest they can just buy any business that has any potential of disrupting their revenue. It's literally the thesis Peter Thiel sets out in Zero To One. It works really well for that one business.
That works fine with office buildings and stuff where a company is redistributing its risk profile, but not when the company it’s selling to has every incentive to kill the asset as a competitor.
> Groq raised $750 million at a valuation of about $6.9 billion three months ago. Investors in the round included Blackrock and Neuberger Berman, as well as Samsung, Cisco, Altimeter and 1789 Capital, where Donald Trump Jr. is a partner.
Makes it very hard not to think of this as a way to give money to the current administration.
I know, this sounds conspiracy theory grade, but 20b is too much for groq.
The value of Groq comes from its excellent price-to-performance ratio. Its inferencing speeds are faster than those of H200s, and it has the lowest costs in the industry. When running similar batch jobs across different providers compared to Groq, the processing speed can sometimes be more than 10 times faster. These figures are important for developing practical applications for production use. It's common for me to run workloads in Groq that cost less than $100, while the same workload can approach $1,000 on Bedrock or Gemini. They have tuned a set of OS models that can now deliver a full application. The speeds have allowed me to offload a lot of the functionality from heuristics to straight-up LLMs.
Can't wait for the abuse of the word Grok to die (bet none of these techbros even read the book). There was even an AI company that made a product called "Sophon". Talk about an overinflated sense of self-worth.
I like the Wright Brothers, they called the first plain, "Flyer".
this is genuinely sad, groq had really fast inference and was a legit alternative architecture to nvidia's dominance. feels like we're watching consolidation kill innovation in real time. really hoping regulators actually look at this one but not holding my breath
Will be interesting technically to see what develops from this. NVLink? Full CUDA feels maybe doubtful but who knows. Nvidia CUDA Tile feels like more of a maybe, their new much more explicit way of making workloads.
This does feel a bit sad for sure, worrying whether this might hold Groq and innovation back. Reciprocally, perhaps kind of cool to see Groq get a massive funding boost and help from a very experienced chip making peer. It feels like an envious position somewhat, even with the long term consequences being so hazy. From the outside yes it looks like Nvidia solidifying their iron grasp over a market with very limited competitive suppliers, but this could help Groq, and maybe it's not on the terms we think we want right now, but could be very cool to see.
I really hope some of the rest of the markets can see what's happening, broadly, with Nvidia forming partnerships all over the place. NVLink with Intel, NVLink with Amazon's Tritanium... there's much more to the ecosystem, but just connecting the chips smartly is a huge task, is core to inter-operation. And for all we've heard of CXL, UltraAccelerator Link (UALink) and UltraEthernet (UET) it feels like very few major players are taking it seriously enough to just integrate these new interconnects & make them awesome. They remain incredible expensive & not commonly used, lacking broad industry adoption, and reserved for very expensive systems: there's a huge existential risk here that (lack of) interconnect will destroy competitors ability to get their good chips well integrated and used. The rest of the market needs more clear alarm bells going off, and needs to be making sure good interconnect is available on way more chips, get it into everyone's hands ASAP not just big customers, so that adoption & Linux nerd type folks can start building stacks that open up the future. The market risks getting left behind, if NVLink is built in everywhere and the various other fabrics never become common-place.
…except they are a rather small hardware startup, there is a dozen other rather small hardware startups they did not buy, and there will be more such startups funded just on the news that NVIDIA bought one at a big premium.
Uh oh, not good that a major Nvidia competitor with genuine alternative technology will no longer be competing... Chances this tech gets killed post-acquisition?
I don't know specifically, but I think they're referring to the current USA administration's posture of approving anything, or pardoning anyone, in exchange for some cryptocurrency or similar big favour.
> Groq raised $750 million at a valuation of about $6.9 billion three months ago. Investors in the round included Blackrock and Neuberger Berman, as well as Samsung, Cisco, Altimeter and 1789 Capital, where Donald Trump Jr. is a partner.
That was impressive to see what you did there and the harsh reality that its true hits like a brick.
Don't forget that those forks of VScode are gonna be bought by Nvidia or chatgpt (OpenAI which gets invested by Nvidia) and everything else
Its all one large web connecting every investment and cross-investments and everything. The bubble image which got infamous recently is definitely popping up even more. Its crazy basically.
Groq press release: https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusiv...
> Today, Groq announced that it has entered into a non-exclusive licensing agreement with Nvidia for Groq’s inference technology. The agreement reflects a shared focus on expanding access to high-performance, low cost inference.
> 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.
> GroqCloud will continue to operate without interruption.
Another example of the growing trend of buying out key parts of a company to avoid any actual acquisition?
I wonder if equity holding employees get anything from the deal or indeed if all the investors will be seeing a return from this?
I wonder if such deals will create employee lawsuits. I'd certainly be looking at legal options if I was one of the founding employees.
The employees are getting paid twice.
wdym?
So it is not structured as an acquisition to avoid anti trust but effectively it probably is.
Indeed, as justincormack comments: ”It is not structured as an outright acquisition to avoid US Gov't anti trust scrutiny, but effectively it probably is”. “Non-exclusive” ? Ummmm, yeah, right, sure. You can probably bet there is an private understanding that Groq will no longer offer it's “top of the line” best technology to competitors of Nvidia. Some may see this as a clever, “slight of the hand” attempt for Nvidia to maintain it's perceived dominance & lead in GPU-TPU development. “Non-exclusive” does not in any form or fashion spell out that all Nvidia's competitors can and will obtain the very top, cutting edge Groq technology as Nvidia will obtain . . .
What generated this comment?
Probably a good old fashioned Mk 1 Human Brain given the use of "slight of hand"
GPT1 Nano
Grok... with a k
Yes I'm sure that "non exclusive" partnership is exactly that, wink wink!
> 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.
A really strange agreement where top executives of a company "join" another company for the benefit of the other company.
If it quacks like a duck...
Are they buying them to try and slow down open source models and protect the massive amounts of money they make from OpenAI, Anthropic, Meta ect?
It quite obvious that open source models are catching up to closed source models very fast they about 3-4 months behind right now, and yeah they are trained on Nvidia chips, but as the open source models become more usable, and closer to closed source models they will eat into Nvidia profit as these companies aren't spending tens of billion dollars on chips to train and run inference. These are smaller models trained on fewer GPUs and they are performing as good as the pervious OpenAI and Anthropic models.
So obviously open source models are a direct threat to Nvidia, and they only thing open source models struggle at is scaling inference and this is where Groq and Cerberus come into the picture as they provide the fastest inference for open source models that make them even more usable than SOTA models.
Maybe I'm way off on this.
Shy of an algo breakthrough, open source isn't going to catch up with SOTA, their main trick for model improvement is distilling the SOTA models. That's why they they have perpetually been "right behind".
They don't need to catch up. They just need to be good enough and fast as fuck. Vast majority of useful tasks of LLMs has nothing to do with how smart they are.
GPT-5 models have been the most useless models out of any model released this year despite being SOTA, and it because it slow as fuck.
> just need to be good enough and fast as fuck
Hard disagree. There are very few scenarios where I'd pick speed (quantity) over intelligence (quality) for anything remotely to do with building systems.
I agree with you for many use cases, but for the use case I'm focused on (Voice AI) speed is absolutely everything. Every millisecond counts for voice, and most voice use cases don't require anything close to "deep thinking. E.g., for inbound customer support use cases, we really just want the voice agent to be fast and follow the SOP.
For coding I don’t use any of the previous gen models anymore.
Ideally I would have both fast and SOTA; if I would have to pick one I’d go with SOTA.
There a report by OpenRouter on what folks tend to pay for it; it generally is SOTA in the coding domain. Folks are still paying a premium for them today.
There is a question if there is a bar where coding models are “good enough”; for myself I always want smarter / SOTA.
FWIW coding is one of the largest usages for LLM's where SOTA quality matters.
I think the bar for when coding models are "good enough" will be a tradeoff between performance and price. I could be using Cerebras Code and saving $50 a month, but Opus 4.5 is fast enough and I value the piece-of-mind I have knowing it's quality is higher than Cerebras' open source models to spend the extra money. It might take a while for this gap to close, and what is considered "good enough" will be different for every developer, but certainly this gap cannot exist forever.
> They don't need to catch up. They just need to be good enough
The current SOTA models are impressive but still far from what I’d consider good enough to not be a constant exercise in frustration. When the SOTA models still have a long way to go, the open weights models have an even further gap distance to catch up.
GPT 5 Codex is great - the best coding model around except maybe for Opus.
I'd like more speed but prefer more quality than more speed.
I get GPT 5.2 responses on copilot faster than for any other model, almost instantly. Are you sure they’re slow as fuck?
This. You can distill a foundation model into open source. The Chinese will be doing this for us for a long time.
We should be glad that the foundation model companies are stuck running on treadmills. Runaway success would be bad for everyone else in the market.
Let them sweat.
Confused. Is ‘fuck’ fast or slow? Or both at the same time? Is there a sort of quantum superposition of fuck?
It's an intensifier
well, it's not slow as fuck! it's quick as lightning and speedy as hell
Bullseye.
> their main trick for model improvement is distilling the SOTA models
Could you elaborate? How is this done and what does this mean?
I am by no means an expert, but I think it is a process that allows training LLMs from other LLMs without needing as much compute or nearly as much data as training from scratch. I think this was the thing deepseek pioneered. Don’t quote me on any of that though.
Too bad, so sad for the Mister Krabs secret recipe-pilled labs. Shy of something fundamental changing, it will always be possible to make a distillation that is 98% as good as a frontier model for ~1% of the cost of training the SOTA model. Some technology just wants to be free :)
NVIDIA release some of the best open source models around.
Almost all open source models are trained and mostly run on NVIDIA hardware.
Open source is great for NVIDIA. They want more open source, not less.
Commoditize your complement is business 101.
Then why are they spending $20 billion dollars to handicap an inference company that giving open source models a major advantage over closed source models?
Realistically groq is a great solution but has near impossible requirements for deployment. Just look at how many adapters you need to meet the memory requirements of a small llm. SRAM is fast but small.
I would guess their interconnect technology is what NVIDIA wants. You need something like 75 adapters for an 8b parameter model they had some really interesting tech to make the accelerator to accelerator communication work and scale. They were able to do that well before nvl 72 and they scale to hundreds of adapters since large models require more adapters still.
We will know in a few months.
> handicap
Your words.
Because it's very good tech for inference?
It doesn't even do training.
And most inference providers for Open Source models use NVIDIA eg Fireworks, Basten, TogetherAI etc.
Most NVIDIA sales go to training clusters. That is changing but it'd be an interesting strategy to differentiate the training and inference lines.
> to handicap an inference company
That's a non-charitable interpretation of what happened. The are not "spending $20 billion to handicap Groq". They are handing Groq $20 billion to do whatever they want with it. Groq can take this money and build more chips, do more R&D, hire more people. $20 billion is truly a lot of money. It's quite hard to "handicap" someone by giving them $20 billion.
Groq doesn't have any employees. They can't do R&D because there's no one to do it. The $20B goes to Groq's investors.
From the article:
The $20B does not go to Groq's investors. It goes to Groq. You can say that Groq is owned by its investors, and this is the same thing, but it's not. In order for the money to go to the investors, Groq needs to disburse a dividend, or to buy back shares. There is no indication that this will happen. And what's more, the investors don't even need this to happen. I'm sure any investor that wants to sell their shares in Groq will now find plenty of buyers at a very advantageous price.Let's bet on this shit. Where's the Polymarket.
they spending $20 billion dollars to handicap an inference company
Inference hardware company
Yes, you are way off, because Groq doesn't make open source models. Groq makes innovative AI accelerator chips that are significantly faster than Nvidia's.
> Groq makes innovative AI accelerator chips that are significantly faster than Nvidia's.
Yeah I'm disappointed by this, this is clearly to move them out of the market. Still, that leaves a vacuum for someone else to fill. I was extremely impressed by Groq last I messed about with it, the inference speed was bonkers.
If it's that good Nvidia can just keep selling it.
For inference, but yes. Many hundreds of tokens per second of output is the norm, in my experience. I don't recall the prompt processing figures but I think it was somewhere in the low hundreds of tokens per second (so slightly slower than inference).
Nvidia just released their Nemotron models, and in my testing, they are the best performing models on low-end consumer hardware in both terms of speed and accuracy.
I'd say that it's probably not a play against open source, but more trying to remove/change the bottlenecks in the current chip production cycle. Nvidia likely doesn't care who wins, they just want to sell their chips. They literally can't make enough to meet current demand. If they split off the inference business (and now own one of the only purchasable alternatives) they can spin up more production.
That said, it's completely anti-competitive. Nvidia could design a inference chip themselves, but instead the are locking down one of the only real independents. But... Nobody was saying Groq was making any real money. This might just be a rescue mission.
They need to vertically integrate the entire stack or they die. All of the big players are already making plans for their own chips/hardware. They see everyone else competing for the exact same vendor’s chips and need to diversify.
They acquired in order to have an ASICs competitor to Google TPU.
Idk- cheaper inference seems to be a huge industry secret and providing the best inference tech that only works with nvidia seems like a good plan. Makes nvidia the absolute king of compute against AWS/AMD/Intel seems like a no brainer.
How does this work considering the Nemotron models?
I don’t see how this isn’t anti trust but knowing this political climate, this deal will go through.
https://www.reuters.com/business/nvidias-huang-joins-tech-ti...
Good of them to make a list themselves, isn't it? It'll be useful in the future.
The problem is, if everyone knows it going to curry favour and you're the odd man out - are you in Violation of your fiduciary duty to your shareholders?
as useful as it was before this administration when big tech was sucking up to whomever was running the country (e.g. “macho man” Zuck was getting ready to tattoo DEI on his forehead couple of years ago) or just now it’ll be magically useful?
You miss my point. This is a list of people engaging in something flat-out corrupt. The ballroom is an inherently corrupt project.
It will prove to be simple corruption.
whats the punishment for corruption (especially when you have 100’s of billions of dollars) I wonder…
Why is that relevant if there is no one willing to prosecute and convict?
A forest can still exist despite people choosing to not see or look at it.
so corruption exists, that’s the pitch? learned something new today…
it is completely irrelevant but people still waste internet bandwidth with nonsense :)
If justice is served it'll be knocked down by the next admin, if it is ever built.
An American Oligarchy.
It’ll go through. It’s not an acquisition, it’s an exclusive licensing deal. Same end result, but it lets them runaround the usual regulatory approvals for acquisitions.
I got curious about how many wheelbarrows of cash $20bn actually is.
Two ways to think about it: weight vs volume.
By weight (assuming all $100 bills):
$20,000,000,000 / $100 = 200,000,000 bills
Each bill is roughly 1g, so total mass is ~200,000 kg
A typical builder’s wheelbarrow can take about 100 kg before it becomes unmanageable
200,000 kg total / 100 kg per wheelbarrow ≈ 2,000 wheelbarrows (weight limit)
By volume:
A $100 bill is ~6.14" × 2.61" × 0.11 mm, which comes out to about 102 cm³ per bill
200,000,000 bills × 102 cm³ ≈ 20,400 m³ of cash
A standard wheelbarrow holds around 0.08 m³ (80 litres)
20,400 m³ total / 0.08 m³ per wheelbarrow ≈ 255,000 wheelbarrows (volume limit)
So,
About 2,000 wheelbarrows if you only care about weight
About 255,000 wheelbarrows if you actually have to fit the cash in
So the limiting factor isn’t how heavy the money is; it’s that the physical volume of the cash is absurd. At this scale, $20bn in $100s is effectively a warehouse, not a stack.
I think your volume per bill should be 6.14 * 0.0254 * 2.61 * 0.0254 * 0.00011 ≈ 1.137e-6 m³. That means about 227 m³ total volume, or about 2800 wheelbarrows.
I think you’re off by about a factor of 100 on the volume of a single bill. So both cases it’s in the ballpark of 2000 wheelbarrows.
Something wrong about representing the weight of US dollars in metric units.
Then what do you say to 6.14" × 2.61" × 0.11 mm = 102 cm³
Your volume of a single bill is a bit off.
The price is 40x their target revenue. That's twice the price to revenue multiplier applied to Anthropic in their most recent funding round, and really really hard to portray as a good deal.
I don't think it really helps Nvidia's competitive position. The serious competition to Nvidia is coming from Google's TPU, Amazon's Trainium, AMD's Instinct, and to a much lesser extent Intel's ARC.
Grow recent investors got back a 3x multiple and may now invest in one of Nvidia's other competitors instead.
The only thing I can think of here is that OpenAI’s DRAM land grab is going to stack on a non-NV target and NV need to hedge with an SRAM design that’s on the market NOW. Otherwise, I can’t see how NV couldn’t eat Groq’s lunch in one development cycle - it’s not like NV can’t attach a TPU to some SRAM and an interconnect. Either that or Groq closed a deep enough book to scare them, but 40x is a lot of scared.
That's an interesting take, it's plausible Nvidia wants to have an SRAM based product, but I am struggling to see why they would pay $20bn to have one /right now/. Even if DRAM prices make Groq's approach more economical, Nvidia can develop a competitive product before Groq could take any real market share.
Exactly. The only way this makes sense to me is if the board needed this product in <1 cycle. Which makes no sense for a market player like NV who already have the PDK, volume, and literally everything else in the universe. But here it is, so there is clearly a factor I have not considered :)
Would a Groq chiplet be worth $20B? If you bundle it while no one else can, maybe..
But if all of these large companies create a monopoly where they realize that if they start undercutting each other or innovating too extremely then long term if Nvidia's stock decreases in value by a huge margin, the whole market itself can get down
As an example: if google TPU (perfected?) itself to the point that it hurts nvidia sales (maybe mass production perhaps?) then all the companies stock price might decrease (in my opinion including google)
Honestly, I feel like there is going to happen something in the market which is gonna be very spooky soon regarding AI.
I feel like we are gonna drag this bubble really long and actually worsen all the pain which is gonna be caused by it long term.
And Cerebras
Also Tsavorite scalable intelligence - their architecture seems to cover the broadest use cases and compatible with cuda
is this an ad? do you work there?
I'm following the chip industry on a daily basis and never heard of them...
Headline is incorrect.
NVIDIA isn't buying Groq.
It's a non exclusive deal for inference tech. Or am I reading it incorrectly?
There is also movement over of some key people to nvidia which is pretty significant: "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."
curious how the execs can honor the fiduciary duty to share holders assuming only they? get the 20 billion and the company is left headless and leaking all of its intellectual property to Nvidia?
Shareholders get the $20B so that's the fiduciary responsibility fully satisfied.
Non-exclusive licensing and hiring the team.
> 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.
Legit feels like Nvidia just buying out competition to maintain their position and power in the industry. I sincerely hope they fall flat on their face.
> Legit feels like Nvidia just buying out competition to maintain their position and power
Well, I mean, isn't that exactly what they should be doing? (I'm not talking about whether or not it benefits society; this is more along the lines of how they're incentivized.)
Put yourself in their shoes. If you had all that cash, and you're hearing people talk of an "AI Bubble" on a daily basis, and you want to try and ensure that you ride the wave without ever crashing... the only rational thing to do is use the money to try and cover all your bases. This means buying competitors and it also means diversifying a little bit.
No one is claiming that it's a bad move.
It's just an anti-competitive move that could be very bad for the consumer as it makes the inference market less competitive.
Dunno thought AGI would make everything obsolete and it's just around the corner? It looks rather like it dawns on everyone that transformers won't bring salvation. It's a show of weakness.
Stuff like tinygrad will change this. Geohot already made nvidia run on macs via thunderbolt.
Also: https://x.com/__tinygrad__/status/1983469817895198783
The bottleneck in training and inference isn’t matmul, and once a chip isn’t a kindergarten toy you don’t go from FPGA to tape out by clicking a button. For local memory he’s going to have to learn to either stack DRAM (not “3000 lines of verilog” and requires a supply chain which openai just destroyed) or diffuse block RAM / SRAM like Groq which is astronomically expensive bit for bit and torpedoes yields, compounding the issue. Then comes interconnect.
The main point is that it will not be an nvidia’s monopoly for too long.
This guy has the greatest dunning-kruger of all time. Lots of smoke and mirrors.
He’s no delusional: https://x.com/__tinygrad__/status/1983476594850283820
However, I would say you are wrong about it being only smoke
Look dude, this guy failed his Twitter internship and is not about to take on Jensen Huang. This isn't some young guy anymore and this isn't 200x where is he about to have another iPhone / Sony moment.
It is peak delulu.
Edit: His whole blog is 'hot take #n'. Not even anything serious. Basically podcast bro level stuff. https://geohot.github.io/blog/jekyll/update/2025/12/22/the-o...
And where do you think he’s wrong in that post?
There's this curious experience of people bringing up geohot / tinygrad and you can tell they've been sold into a personality cult.
I don't mean that pejoratively, I apologize for the bluntness. It's just I've been dealing with his nonsense since iPhone OS 1.0 x jailbreaking, and I hate seeing people taken advantage of.
(nvidia x macs x thunderbolt has been a thing for years and years and years, well before geohot) (tweet is non-sequitor beyond bogstandard geohot tells: odd obsession with LoC, and we're 2 years away from Changing The Game, just like we were 2 years ago)
Can you show any other thing that runs nvidia gpu under m-series macs?
Who cares? Nobody is building large scale inference services with macs.
Because this is exactly the demonstration of abstraction: the same stuff allows direct gpu communication so that even mac nvidia thing is possible.
It is not tied to nvidia as well.
This is the power of tinygrad
My deepest apologies, I can't parse this and I earnestly tried: 5 minutes of my own thinking, then 3 llms, then a 10 minute timer of my own thinking over the whole thing.
My guess is you're trying to communicate "tinygrad doesn't need gpu drivers" which maybe is transmutated into "tinygrad replaces CUDA" and you think "CUDA means other GPUs can't be used for LLMs, thus nvidia has a strangehold"
I know George has pushed this idea for years now, but, you have to look no further than AMD/Google making massive deals to understand how it works on the ground.
I hope he doesn't victimize you further with his rants. It's cruel of him to use people to assuage this own ego and make them look silly in public.
Re: has someone else does this? https://github.com/albertstarfield/apple-slick-rtx (May 2024, 19 months ago. didn't bother looking further than 4th google result for "apple silicon external gpu")
> Compute Workload Test: This will be add soon
Wonder what happened that it never came.
> Willy's got his i3-12100 Gen RTX3090 hosted on Ubuntu with Juice Server
E-gpu my ass
That's unfortunately what most acquisitions are.
which is exactly what a business should do.
it's not like Nvidia doesn't invest a ton into R&D, but hey, they have the cash, why not use it? like a good business.
In a normal world, this is where Nvidia gets trust busted. But that's long behind us now.
Not good. This shouldn't be allowed. What would be better is if groq and cerebras combined, and maybe other companies invested in them to help them scale. Why would the major cloud providers not lobby against this?
Usually antitrust is for consumers, but here I think companies like Microsoft and AWS would be the biggest beneficiaries of having more AI chip competition.
Groq is absolutely tiny. I don't think antitrust is an issue here.
20 billions is tiny?
That's the sales price of the company. Their marketshare, I imagine, is absolutely miniscule.
WhatsApp was a tiny team
Is nowadays
Market share wise, Groq is perhaps "tiny"? Nvidia may be paying a premium for Groq [0] since it eliminates competition (at least on the inference side).
[0] valued ~£6.5bn 2mo ago https://www.reuters.com/business/groq-more-than-doubles-valu...
It's a non-exclusive deal.
No reason for antitrust action whatsoever.
That’s a loophole. Regulation hasn’t caught up to the innovation of non-exclusive licensing deal. Hopefully we’ll get some competence back in government soon-ish and can rectify the mistake
It's a backdoor acquisition by looting the key talent.
>> if groq and cerebras combined
There isn't to be shared between the two techs, Groq's hardware is a like a railgun that installs all the weights into the optimal location before firing off an inference. Cerebras computer engineering more convention requiring the same data movement that GPUs struggle with optimizing.
Suspect Groq is complementary/superior to nvidia's GPUs, while it is unclear what Cerebras brings other then maybe some deals with TSMC.
They are both SRAM based solutions currently with the same benefits and pitfalls.
Damn. Was hoping Groq and Cerebras would give the giants a run for their money.
There are others as well but NVidia is aggressive when it comes to punishing companies willing to buy non-NVidia products. As a result, they prefer to remain under the radar, at least until they have enough market leverage to be more widely known.
Yes, Groq failed. But there will be others.
I guess I have some hope for Tenstorrent
There is still Modular
And China.
It would be interesting if it turned out that Chinese competition was the only thing that kept this market working!
I imagine 2 big giants basically (Nvidia/google/amd basically influenced by a few select of people) vs (Chinese companies who have investments from the govt)
its sort of like proxy wars and this is sort of whats happening in software side of things with open source models but I think that the benefit of the proxy wars is going to be for the end consumers
But although on the other hand, having two large countries compete with each other while buying everything else and all feels like it astronomically increases the price if someone wants to compete with these two giants (any other country perhaps)
We definitely need a better system where it doesn't feel like we are seeing pacman eat everything up basically
> Nvidia/google/amd
One of these things is not like the others
Is it not? All this money is going into AI under the fear that China will win the race to AGI. China releases open-source models that keep OpenAI/Anthropic researching and training their models, which in turn creates demand for more Nvidia GPUs.
Which China players are doing inference hardware? As indeed that is a good space for them.
Huawei
I just stopped my Groq API. Sad to see competition being eaten up by shitty Nvidia. I like their products but Jensen is an absolute mfer with deceitful marketing.
I literally said “oh no” out loud when I read the headline.
was the API good
fast but furiously expensive
Honestly cerebras is good. I can recommend it. I talked to their team once on discord as a literal free user so that was something really nice personally
They were also the faster one compared to groq but they were always a little slow on adding new models compared to groq but not sure what changed right now.
Definitely recommend cerebras tho now that groq's been eaten up from inside basically
I do not understand this move by Nvidia, they are afraid of being out competed by this startup in their core competence of building chips for AI? They may be eliminating a competitor for now but this move will immediately many more AI chip startups to get founded
They're not eliminating a competitor, they're (effectively) acquiring a competitor. Nvidia's GPUs are great for training, and not bad for inference, but the custom chips are better for inference and Nvidia's worried about losing customers. Nvidia will no doubt sell custom Groq-like chips for inference now.
groq is a series E hardware startup founded in 2016. It took them this long to be a potential threat, I'm not sure they are even an actual threat.
Even if this purchase causes 100 new hardware startups to be funded tomorrow, nVidia is perfectly fine with that. Let's see how many survive 5 years down the line
the play is 10x faster inference leads to 100x demand give or take, which isn't a bad assumption at all if you ask me. the problem is actually fitting a good model onto hardware that fast.
There was an AMA here last year https://news.ycombinator.com/item?id=39429047
This is smart as hell. I’ve long wondered how they’d combat ASIC’s without diluting their own benefits. This gives them a bit more time to figure out the moats, which is useful because Groq was going places. This juices Groq’s distribution, production, ability to access a wider range of skills where necessary.
I expect China to want to compete with this. Simpler than full-blown Nvidia chips. Cue much cheaper and faster inference for all.
Hopefully they plan to invest in the technology and not just eliminate a competitor.
They almost certainly plan to invest in the technology. One of the biggest threats to Nvidia is people developing AI-centric ASICs before they get there. Yes, Google has their TPUs and there are others around, but it's early on.
In some ways, it's not about eliminating a competitor. It's about eliminating all the competitors. Nvidia can use its resources to push AI ASICs farther faster than others, potentially cutting off a whole host of competitors that threaten their business. Nvidia has the hardware and software talent, the money, and the market position to give their AI ASICs an advantage. They know if they don't lean into ASICs that someone else will and their gravy train will end. So they almost certainly won't be abandoning the technology.
But that doesn't mean that it'll be good for us.
I remember when Google acquired YouTube in 2006 for $1.65 billion in stock.
Media said it was crazy back then, well I think this sounds a lot crazier but hindsight is 20/20.
I thought the Skype deal back then was worse: it sucked balls back then already.
And now we’re told that bet was such a sure thing that regulators should have blocked it!
This doesn't make much sense- In September, Groq was valued at $7B. How is it that in 4 months it is being bought for $20B?
Can someone with better understanding dumb this down for me please?
Acquisition premium: https://www.investopedia.com/terms/a/acquisitionpremium.asp
The acquisition price of a company usually comes at a premium to the last valuation. This applies even with publicly traded companies, which is why acquisition announcements cause stock prices to pop up to some number between the last trade price and the acquisition price, proportional to how much the market thinks the acquisition is likely to go through.
The premium can make sense to the acquirer because the acquired company is worth more when combined with all of the assets and power (brand name, distribution, patents, trade secrets) of the acquiring company.
This confuses a lot of people who think the valuation of a company is equivalent to the number that would be paid to acquire it at that instant, but it’s not.
It only has to be overvalued by a lower multiple then NVidia; not undervalued.
Imagine a pharma with a weight loss drug that isn't approved yet; it's either worth $20B (if approved) or zero (if not approved).
Now imagine the LPUv2 ASIC. If it works it's worth $20B and if it doesn't it's zero. If investors think LPUv2 has a 1/3 chance of success they would buy in at $7B. Then the chip boots up and... look at that.
Or it's just a massive bubble.
eli5 lpu please
Groq calls their inference chips “Language Processing Unit”: https://groq.com/blog/the-groq-lpu-explained
It's an AI blackhole https://www.youtube.com/watch?v=21e5GZF3yx0
I got recommended this and I will watch it today, thank you. One of the comments points out
"They’ve literally told us that the plan is to get bailed out by the taxpayers"
This reminded me of how I think what's gonna happen/ is already happening is that they become too big to fail and get bailed out and the burden/loss becomes of taxpayers
So we are kind of living in a system which is reckless about finances/stability behind businesses where the system is such that all the profits are privatized but all the losses are shared/even funded by the average person
Combine in a mix of corruption in any political party to begin with and I am wondering why we don't have yet another french revolution.
Trump Jr. entered at $7 billion. In the meantime Nvidia got permission to sell GPUs to China.
All-In pundit Palihapitiya is invested in Groq as well. It is going well for friends of David Sacks.
Biggest surprise for me is that they chose to announce this on Christmas Eve.
I’d love to have been in the room when that was decided. The big, exciting news doesn’t typically get announced during a major holiday week.
They want the Trump Jr. and Chamath connections to be forgotten by the time the media and YouTubers get active again after New Year. It's a standard media tactic.
Honestly, it's basically an accepted fact that the best thing to do is just to never respond anything that media/people say. If you respond, the conversation will continue. If you just say nothing, they'll move on to the next topic, because everyone needs to keep their audience engaged. With no response, your audience isn't interested either, so you have to move onto something else.
What's your angle on the Chamath link? Is Nvidia greasing Sack's favour through Chamath?
Hopefully prices stay the same, I run my small apps on groq. I get good enough summarization and simple agents from gpt120b which is on 15 cents for a million input tks.
Can you list any internet facing apps you have that use Groq? Looking into them myself because of the speed
The absolute best case I can make for this:
I think it’s pretty obvious at this point that Nvidia’s architecture has reached scaling limits - the power demands of their latest chips has Microsoft investing in nuclear fusion. Similar to Intel in both the pre-Core days and their more recent chips, they need an actual new architecture to move forward. As sits, there’s no path to profitability for the buyers of these chips given the cost and capabilities of the current LLM architectures, and this is obvious enough that even Nvidia has to realize it’s existential for them.
If Groq’s architecture can actually change the economics of inference and training sufficient to bring the costs in line with the actual, not speculative, benefits of LLMs, this may not be a buy-and-kill for Nvidia but something closer to Apple’s acquisition of P.A. Semi, which made the A- and M- class chips possible.
(Mind you, in Intel’s case they had to have their clocks cleaned by AMD a couple times to get them to see, but I think we’re further past the point of diminishing returns with Nvidia - I think they’re far enough past when the economics turned against them that Reality is their competition now.)
NVIDIA and "no path to profitability" don't belong in the same zip code.
I read it as path to profitability for the AI companies buying Nvidia's chips.
No path to profitability for the people using their products for their putative purpose, which seems like it might affect Nvidia’s bottom line at some point. Clarified.
there’s no path to profitability for the buyers of these chips given the cost and capabilities of the current LLM architectures
Didn't Anthropic say inference is already profitable?
Presuming that’s why they raised 3 times this year.
Inference being profitable doesn’t mean that they’re selling enough inference to offset their other costs.
I believe that this news kind of helps cerebras as groq and cerebras are the only two companies working extensively in this space
I feel as if Nvidia is eating up even companies which I thought had genuine potential or anything related to AI industry whether profitable or not
Nvidia's trying its best to take all major players and consolidate into one big entity from top to bottom.
The problem with this approach imo is that long term, nvidia's stock is extremely overvalued and its still a bubble which will burst and it will take nvidia first and foremost.
The issue is that when nvidia falls, it will take the whole literal industry from top to bottom, even those companies which I thought could survive an AI burst. Long term I feel like it will have really bad impacts if nvidia continues to gobble up every company.
I am pretty sure that Nvidia might be looking at cerebras too and if they offer them a shit ton of money and cerebras gets bought. I genuinely believe that Nvidia has sort of invested in literally all pockets of any hardware related investment for AI. And when OpenAI is unable to pay Nvidia, I feel like it can all come crashing down since this whole cycle is only being possible via external investor money.
Maybe the EU or individual states will sue under their own anti-trust laws will stop this - seems pretty clearly anti-competitive and probably a prelude of these over-valued companies using their stock to gobble up any possible competitor to consolidate even more.
> Groq will continue to operate as an independent company with Simon Edwards stepping into the role of Chief Executive Officer.
Kindof feel bad for Simon Edwards, lol. I wonder what the plan is for the future of Groq
I would assume this is a very well paid position, and with basically nothing to do.
I would assume he is a fiduciary... I am curious what happens with investors here?
He has to keep the company in business for 2 more years
Are there consequences if he fails to do so?
In my opinion, Groq's technical decisions are unsound in a normal world. But being HBM-free may have some merit in a world where HBM supply is constrained.
Given that Groq is basically the TPU spin out Google should have done years ago it shows what a valuable asset TPUs are in Google.
Still this they should spin that out though!
Groq is very different from TPUs. The difference in memory should be a big clue. (Groq using SRAM built into each compute unit, vs TPUs using HBM more similarly to GPUs)
The founder created TPUs at Google.
> Prior to founding Groq, Jonathan began what became Google’s Tensor Processing Unit (TPU) as a 20% project where he designed and implemented the core elements of the first generation TPU chip.
I hope non executives and founders get something for their equity.
Chamath is a bonafide scammer, but he makes good investments(gets good returns for himself)
He is good at scamming others
nVidia is being scammed here? Seems unlikely…
For nvidia increasing the numbers of their money-multiplying mutual investment ring is more important than the value of the deals. It's about involving more capital and people and making their grift too big to fail and keeping the stock numbers up. Nvidia has the ability to promise large amounts of money like this in announcements but I haven't read about any of them actually having money or good exchange hands yet.
Big win for Chamath!
Congratulations to Chamath
For failing?
I could not figure out if „cash“ means literally cash or figuratively cash in the sense of “no trade in shares”.
Will there be a truck full of paper money or not?
Implies they will pay cash value to equity holders as opposed to issuing NVDA shares.
(Electronically)
Is this to somehow screw the employees with RSUs or what?
No, it doesn't really matter if they pay in cash or stock. If you think NVDA has room to run you're welcome to use your buyout money to buy NVDA on the open market.
Well, this isn't framed as a buyout/takeover, so I was curious how existing RSUs would be cashed out?
This deal is framed as IP transfer and talent transfer without owning the full company. Probably to skirt anti trust, among other things.
I'm not sure in this specific case. They could choose to pay the employees some portion of the funds.
If not, the owners are likely liable to be sued for "selling in effect" without paying equity holders.
Presuming the company becomes a defacto subsidiary of Nvidia (even if not legally so)
My guess, without researching it, is they will compensate existing equity holders to avoid that possibility. I mean the valuation multiple is enormous, it's worth it simply to derisk the legal aspect.
For vested RSUs it's likely that the Groq husk will pay out the $20B as a dividend or buyback or something. I don't know if unvested RSUs are accelerated or just canceled. Of course the employees will receive new RSUs when they join Nvidia.
Unless one is worried about the government, no one ever means paper money.
It's the latter. They'll send a wire.
groq was targeting a part of the stack where cuda was weakest: guaranteed inference time at a lower cost per token at scale. This was in response to more than just goog's tpus, they were also one of the few realistic alternative paths oai had with those wafers.
The last time I tested, I could swear they were doing some problematic quantization, because I was getting kind of random results with one or two models, which worked perfectly when I switched providers.
It was really disappointing too because Cerebras does not provide any service reliability on their cheap plans. So I came to the conclusion that unless I could convince the client to set up an enterprise contract or something, we could not use either provider for low-latency, which we need for voice calls. I think for organizations that can afford a hefty contract that guarantees service levels, Groq and Cerebras especially are basically cheat codes for meeting latency requirements for voice. But that might not be an option for really small businesses.. although maybe I am just not a good sales person.
is this in response to the threat from Google tpu
> Groq is expected to alert its investors about the deal later on Wednesday. While the acquisition includes all of Groq’s assets, its nascent Groq cloud business is not part of the transaction, said Davis.
Wait, what? How is the cloud business supposed to run if Nvidia is acquiring the rights to the hardware?
It isn't, and the other companies that offer cloud AI that Nvidia has invested in can carry on happy they have one less competitor.
This is how business works in the 21st century - once one company has a dominant position and a massive warchest they can just buy any business that has any potential of disrupting their revenue. It's literally the thesis Peter Thiel sets out in Zero To One. It works really well for that one business.
That's the neat trick - it isn't...
Sell the asset and then lease it from the buyer.
That works fine with office buildings and stuff where a company is redistributing its risk profile, but not when the company it’s selling to has every incentive to kill the asset as a competitor.
From the press release, Nvidia now has a non exclusive license to the hardware.
Groq will continue to operate as an independent company with Simon Edwards stepping into the role of Chief Executive Officer.
GroqCloud will continue to operate without interruption.
I do not consider this good news; I had hopes Anthropic or so would buy them, not the main AI hardware people.
What are the other top AI silicon vendors?
Graphcore
Tenstorrent
SambaNova
Rivos
From the comment below:
> Groq raised $750 million at a valuation of about $6.9 billion three months ago. Investors in the round included Blackrock and Neuberger Berman, as well as Samsung, Cisco, Altimeter and 1789 Capital, where Donald Trump Jr. is a partner.
Makes it very hard not to think of this as a way to give money to the current administration. I know, this sounds conspiracy theory grade, but 20b is too much for groq.
The value of Groq comes from its excellent price-to-performance ratio. Its inferencing speeds are faster than those of H200s, and it has the lowest costs in the industry. When running similar batch jobs across different providers compared to Groq, the processing speed can sometimes be more than 10 times faster. These figures are important for developing practical applications for production use. It's common for me to run workloads in Groq that cost less than $100, while the same workload can approach $1,000 on Bedrock or Gemini. They have tuned a set of OS models that can now deliver a full application. The speeds have allowed me to offload a lot of the functionality from heuristics to straight-up LLMs.
It's not a conspiracy theory. It's basically a bribe to Chamath, Don Jr. and Sacks.
Can't wait for the abuse of the word Grok to die (bet none of these techbros even read the book). There was even an AI company that made a product called "Sophon". Talk about an overinflated sense of self-worth.
I like the Wright Brothers, they called the first plain, "Flyer".
this is genuinely sad, groq had really fast inference and was a legit alternative architecture to nvidia's dominance. feels like we're watching consolidation kill innovation in real time. really hoping regulators actually look at this one but not holding my breath
Will be interesting technically to see what develops from this. NVLink? Full CUDA feels maybe doubtful but who knows. Nvidia CUDA Tile feels like more of a maybe, their new much more explicit way of making workloads.
This does feel a bit sad for sure, worrying whether this might hold Groq and innovation back. Reciprocally, perhaps kind of cool to see Groq get a massive funding boost and help from a very experienced chip making peer. It feels like an envious position somewhat, even with the long term consequences being so hazy. From the outside yes it looks like Nvidia solidifying their iron grasp over a market with very limited competitive suppliers, but this could help Groq, and maybe it's not on the terms we think we want right now, but could be very cool to see.
I really hope some of the rest of the markets can see what's happening, broadly, with Nvidia forming partnerships all over the place. NVLink with Intel, NVLink with Amazon's Tritanium... there's much more to the ecosystem, but just connecting the chips smartly is a huge task, is core to inter-operation. And for all we've heard of CXL, UltraAccelerator Link (UALink) and UltraEthernet (UET) it feels like very few major players are taking it seriously enough to just integrate these new interconnects & make them awesome. They remain incredible expensive & not commonly used, lacking broad industry adoption, and reserved for very expensive systems: there's a huge existential risk here that (lack of) interconnect will destroy competitors ability to get their good chips well integrated and used. The rest of the market needs more clear alarm bells going off, and needs to be making sure good interconnect is available on way more chips, get it into everyone's hands ASAP not just big customers, so that adoption & Linux nerd type folks can start building stacks that open up the future. The market risks getting left behind, if NVLink is built in everywhere and the various other fabrics never become common-place.
This is the most blatant buy the competition move if i've ever seen one....
"2.7M Developers and Teams"
So, about ~$1,000/each? Seems pricey, even assuming all of them still use it every week/month.
Is any part of this because Google has the TPU and Groq has the LPU?
There's definitely a narrative that ASICs/TPUs/LPUs are more efficient than GPUs and thus Nvidia "needs" an ASIC. Whether this is true is debated.
Well that sucks.
Following the age old playbook of monopolies. https://www.arte.tv/en/videos/103517-001-A/capitalism-in-ame... (Use a vpn if outside EUR)
A free market is a regulated market. Otherwise you will end up with monopolies and a dead market.
…except they are a rather small hardware startup, there is a dozen other rather small hardware startups they did not buy, and there will be more such startups funded just on the news that NVIDIA bought one at a big premium.
Uh oh, not good that a major Nvidia competitor with genuine alternative technology will no longer be competing... Chances this tech gets killed post-acquisition?
It may be more likely that Nvidia sells the LPUv2 at a price that doesn't threaten Rubin.
Zero. Non-zero only if someone says something deemed “woke”.
Great choice and what a great deal.
Quite obvious that Groq would get acquired. [0]
[0] https://news.ycombinator.com/item?id=39438820
They should have bought nbis
How can this pass antitrust régulation ?
There is no "antitrust regulation" in the US in 2025. (Until 2029)
States are "not allowed" to regulate AI companies.
There also weren't any antitrust regulations before, let's not kid ourselves.
There was an attempt under Lina Khan.
The regulations are still in place. They are not enforced, however.
Care to give more details?
I don't know specifically, but I think they're referring to the current USA administration's posture of approving anything, or pardoning anyone, in exchange for some cryptocurrency or similar big favour.
https://www.bbc.co.uk/news/articles/crmddnge9yro
> Groq raised $750 million at a valuation of about $6.9 billion three months ago. Investors in the round included Blackrock and Neuberger Berman, as well as Samsung, Cisco, Altimeter and 1789 Capital, where Donald Trump Jr. is a partner.
They made Jimmy Carter sell his peanut farm…
That's the thing though -- no one made Jimmy Carter sell his farm[0].
But Jimmy Carter was an honorable human, and, well...there are fewer people fitting that description sitting behind the Resolute desk, today.
[0] He didn't sell it, he put it into a blind trust. He should have sold it. When he left office, the farm was $1MM in debt.
I doubt Nvidia will be regulated in their home jurisdiction. America tends to protect it's cash cows, for better or worse.
Nvidia. Please stop. Just stop it already.
Put Groq and Nvidia execs in prison, blatant anti-trust.
Next they would acquire and kill Cerebras. I hate every part of Nvidia
That's great, but LLMs are still not generating revenue.
I pay $20/mo for Gemini, so they're generating at least that much in revenue!
All depends on how much it costs them to service your $20/month sub in OPEX and how much it cost them in capex to buy and maintain that hardware.
They’re generating tons of revenue, just not necessarily profits
They are generating revenue, profit is the dubious thing.
Related on the business side, and from the last two years:
AI Chip Startup Groq Raises $750M at $6.9B Valuation - https://news.ycombinator.com/item?id=45276985 - Sept 2025 (5 comments)
Groq Raises $640M to Meet Soaring Demand for Fast AI Inference - https://news.ycombinator.com/item?id=41162875 - Aug 2024 (34 comments)
AI chip startup Groq lands $640M to challenge Nvidia - https://news.ycombinator.com/item?id=41162463 - Aug 2024 (12 comments)
Groq CEO: 'We No Longer Sell Hardware' - https://news.ycombinator.com/item?id=39964590 - April 2024 (149 comments)
From $6.9b to 20 in a few months, not bad…
Almost as good as forking VSCode, impressive.
That was impressive to see what you did there and the harsh reality that its true hits like a brick.
Don't forget that those forks of VScode are gonna be bought by Nvidia or chatgpt (OpenAI which gets invested by Nvidia) and everything else
Its all one large web connecting every investment and cross-investments and everything. The bubble image which got infamous recently is definitely popping up even more. Its crazy basically.
Acquisitions generally come in at a significantly higher price.
Even in public markets, acquiring all the shares of a company will require an offer that is a significant step above the current trading price.