The specs look impressive. It is always good to have competition.
They announced tapeout in October with planned dev boards next year. Vaporware is when things don’t appear, not when they are on their way (it takes some time for hardware).
It’s also strategically important for Europe to have its own supply. The current and last US administration have both threatened to limit supply of AI chips to European countries, and China would do the same (as they have shown with Nexperia).
And of course you need the software stack with it. They will have thought of that.
These kinds of things-- cheaper-than-NVIDIA cards that can produce a lot of tokens or run large models cheaply are absolutely necessary to scale text models economically.
Without things like these-- those Euclyd things, those Groq things, etc. no one will be able to offer up big models at prices where people will actually use them, so lack of things like this actually cripples training of big models too.
If the price/token graph is right, this would mean 2.5x more tokens, which presumably means actually using multiple prompts to refine something before producing the output, or to otherwise produce really long non-output sequences during the preparation the output. This also fits really well with the Chinese progress in LLM RL for maths. I suspect all that stuff is totally general and can be applied to non-maths things too.
The negativity doesn’t make much sense. The specs are strong, and the chip already taped out in October that’s a concrete milestone, not vaporware. Hardware of this class always takes months between tape-out and dev boards.
Official announcement: https://vsora.com/vsora-announces-tape-out-of-game-changing-...
Strategically, having a European AI inference chip matters. The US has already threatened export limits to Europe, and China has shown similar behavior (e.g., Nexperia). Building local supply is important.
Calling this vaporware makes no sense: tape-out + published roadmap = real, not slides.
> Calling this vaporware makes no sense: tape-out + published roadmap = real, not slides.
I agree that the comments here are surprisingly superficial in their complaints, but I guess it the typical bike-shedding, people don't have technical points to nitpick or the experience to judge the actual product, so from their US-based point of view, they find something else to hook on to, even when there are facts like concrete partnerships making it clear it isn't vaporware, they just have to say something.
If the company is working in isolation (this company is not), the company is continuously pushing back on time frames (they haven't, at least yet), the company never enter actual fabrication/production (which they just did) or if the company never release any details what so ever (they have), then it looks more like vaporware than just "It'll be coming soon".
If those things are true in ~6 months, then I'll join the crowd here who are overly pessimistic at this moment, but until then, as most of the time, I'll give them benefit of the doubt.
I don't believe labeling healthy skepticism and criticism as negativity to farm artificial sympathy in retaliation, does any good to anyone.
Humans have pattern recognition capabilities for a reason, and if a company is triggering that in them, then it's best expressed why(probably because they saw this MO before and got burned) instead of just cheerleading the unknown for fake positivity.
6th comment: "They haven't disclosed any release date, Lots of chip startups start as this kind of vaporware" (they did literally just enter fabrication it seems)
10th comment: "So far, they just talk about it."
Maybe it looks differently now, after 14 hours since the submission was made, but initially yesterday, most of the comments were unfounded (and poorly researched) criticism.
Impressive numbers on paper, but looking at their site, this feels dangerously close to vaporware.
The bottleneck for inference right now isn't just raw FLOPS or even memory bandwidth—it's the compiler stack. The graveyard of AI hardware startups is filled with chips that beat NVIDIA on specs but couldn't run a standard PyTorch graph without segfaulting or requiring six months of manual kernel tuning.
Until I see a dev board and a working graph compiler that accepts ONNX out of the box, this is just a very expensive CGI render.
That seems like not much compared to the hundreds of billions of dollars US companies currently invest into their AI stack? OpenAI pays thousands of engineers and researchers full time.
It is. The problem is latency. All these fields are moving very fast, and so it doesn't sound bad spending 6 months tuning something, but in reality what is happening is that during those 6 months the guy who built the thing you're tuning has iterated 5 more times and what you started on 6 months ago is now much much better than what you got handed 6 months ago whilst simultaneously being much worse than what that person has in their hands today. If the field you're working in is relatively static, or your performance gap is large enough it makes sense. But in most fields the performance gap is large in absolutely terms but small in temporal terms. You could make something run 10x faster, but you can't build something that will run faster than what will be state of the art in 2 months.
This 100x. I used to work for one of those startups. You need something crazy like a 10x performance advantage to get people to switch from Nvidia to some here-today-gone-tomorrow startup with a custom compiler framework that requires field engineer support to get anything to run.
The outcome is that most of custom chips end up not being sold on the open market; instead their manufacturers run them themselves and sell LLM-as-a-service. E.g. Cerebras, Samba Nova, and you could count Google's TPUs there too.
Always good to see more competition in the inference chip space, especially from Europe. The specs look solid, but the real test will be how mature the software stack is and whether teams can get models running without a lot of friction. If they can make that part smooth, it could become a practical option for workloads that want local control.
It needs a "buy a card" link and a lot more architectural details. Tenstorrent is selling chips that are pretty weak, but will beat these guys if they don't get serious about sharing.
Edit: It kind of looks like there's no silicon anywhere near production yet. Probably vaporware.
Nice wave they've been able to ride if it's vaporware, considering they're been at it for five years. Any guesses to why no one else seemingly see the obvious you see?
Look at the CGI graphics and indications in their published material that all they have is a simulation. A
It's all there without disclosing an anticipated release date. Even their product pages and their news page don't seem to have indications of this.
Also, the 3D graphic of their chip on a circuit board is missing some obvious support pieces, so it's clearly not from a CAD model.
Lots of chip startups start as this kind of vaporware, but very few of them obfuscate their chip timelines and anticipated release dates this much. 5 years is a bit long to tapeout, but not unreasonable.
I can ensure you it's not vaporware at all. silicon is running in the fab, application boards have finished the design phase, software stack validated...
Guessing you work at Vsora, been there or at least know someone there? Any interest in sending out demos/sample hardware to European-based hackers who could spread the word if it ends up actually being a really nice device and developer experience? I can sacrifice myself and serve up bug reports as well :)
It loaded fine for me, but that slash before the unit was a bit smelly. :| Just a tiny edit, but it's a rather core part of their message so they should probably notice and format it correctly before publishing.
I think it could be intended, there is a SI document that says something like "x /unit" is a common way to indicate the unit of a quantity, which a guy I know is using as basis for advocating for that ugly display standard.
The next generation will include another processor to offload the inference from the RISC V processors used to offload inference from the host machine.
An FP8 performance of 3200TFLOPS is impressive, could be used for training as well as inference. "Close to theory efficiency" is a bold statement. Most accelerators achieve 60-80% of theoretical peak; if they're genuinely hitting 90%+, that's impressive. Now let's see the price.
I'll believe it when I see it wishing them the best!
> To streamline development and shorten time-to-market, VSORA embraces industry standards: our toolchain is built on LLVM and supports common frameworks like ONNX and PyTorch, minimizing integration effort and customer cost.
I'm sorry I won't share much details, I don't think much is public on Vsora architecture and don't want to breach any NDA...
From their web page Euclyd is a "many small cores" accelerator. Doing good compilation toolchains for these to get efficient results is a hard problem, see many comments on compilers for AI in this thread.
Vsora approach is much more macroscopic, and differentiated. By this I mean I don't know anything quite like it. No sea of small cores, but several more beefy units. They're programmable, but don't look like a CPU: the HW/SW interface is at a higher level. A very hand-wavy analogy with storage would be block devices vs object storage, maybe. I'm sure more details will surface when real HW arrive.
Does anyone know why they brand it an "inference chip"? Is it something at the hardware level that makes is unsuitable for training, or is it simply that the toolchain for training is massively more complicated to program?
Very simplified, AI workloads need compute and communications and compute dominates inference, while communications dominate training.
Most start-ups innovate on the compute side, whereas the techno needed for state of the art communications is not common, and very low-level: plenty of analog concerns. The domain is dominated by NVidia and Broadcom today.
This is why digital start-ups tend to focus on inference. They innovate on the pure digital part, which is compute, and tend to use off-the-shelf IPs for communications, so not a differentiator and likely below the leaders.
But in most cases coupling a computation engine marketed for inference with state of the art communications would (in theory) open the way for training too. It's just that doing both together is a very high barrier. It's more practical to start with compute, and if successful there use this to improve the comms part in a second stage. All the more because everyone expects inference to be the biggest market too. So AI start-ups focus on inference first.
It doesn't have to compete on price 1:1. Ever since Trump took office, the Europeans woke up on their dependence on USA who they no longer regard as a reliable partner. This counts for defense industry, but also for critical infrastructure, including IT. The European alternatives are expected to cost something.
Probably because their software only supports inference. It's relatively easy to do via ONNX. Training requires an order of magnitude more software work.
Even if it's not vapourware, the website makes it look like one. Just look at those two graphs titled "Jotunn 8 Outperforms the Market" and "More Speed For the Bucks" (!) ; WTH?
The people who build landing pages for hardware startups are usually (almost always) not the same people who design and build the actual hardware, for better or worse. A lot of the times, it's a outsourced web design agency who receive a brief, often written by business people, and finally the website is reviewed by business people, with some feedback from technical people who complain about the graphs, accuracy and so on. Then the business people say "But we need to point out we're faster than everyone else" and the engineers reply with "Sure, ok, whatever, I have actual work to do, sounds good".
That's not an excuse. This company has more than a few handful of employees. Somebody saw those graphs without units, labels, tags and approved them. Have the CEO never opened their own webpage once and scrolled down?
Well, this has already taped-out whereas the entire reason people call the tachyum prodigy vapourware is that they keep missing their target dates for tapeout and keep delaying it.
I don’t get the negativity.
The specs look impressive. It is always good to have competition.
They announced tapeout in October with planned dev boards next year. Vaporware is when things don’t appear, not when they are on their way (it takes some time for hardware).
It’s also strategically important for Europe to have its own supply. The current and last US administration have both threatened to limit supply of AI chips to European countries, and China would do the same (as they have shown with Nexperia).
And of course you need the software stack with it. They will have thought of that.
https://vsora.com/vsora-announces-tape-out-of-game-changing-...
It's not just competition.
These kinds of things-- cheaper-than-NVIDIA cards that can produce a lot of tokens or run large models cheaply are absolutely necessary to scale text models economically.
Without things like these-- those Euclyd things, those Groq things, etc. no one will be able to offer up big models at prices where people will actually use them, so lack of things like this actually cripples training of big models too.
If the price/token graph is right, this would mean 2.5x more tokens, which presumably means actually using multiple prompts to refine something before producing the output, or to otherwise produce really long non-output sequences during the preparation the output. This also fits really well with the Chinese progress in LLM RL for maths. I suspect all that stuff is totally general and can be applied to non-maths things too.
The negativity doesn’t make much sense. The specs are strong, and the chip already taped out in October that’s a concrete milestone, not vaporware. Hardware of this class always takes months between tape-out and dev boards. Official announcement: https://vsora.com/vsora-announces-tape-out-of-game-changing-...
Multiple independent sources confirmed the tape-out: EE Times: https://www.eetimes.eu/vsora-tapes-out-ai-inference-chip-for...
L’Informaticien: https://www.linformaticien.com/magazine/infra/64028-vsora-me...
Solutions Numériques: https://www.solutions-numeriques.com/vsora-franchit-un-cap-a...
There’s also an industrial manufacturing partnership with GUC: https://www.design-reuse.com/news/202529700-vsora-and-guc-pa...
Strategically, having a European AI inference chip matters. The US has already threatened export limits to Europe, and China has shown similar behavior (e.g., Nexperia). Building local supply is important.
Calling this vaporware makes no sense: tape-out + published roadmap = real, not slides.
> Calling this vaporware makes no sense: tape-out + published roadmap = real, not slides.
I agree that the comments here are surprisingly superficial in their complaints, but I guess it the typical bike-shedding, people don't have technical points to nitpick or the experience to judge the actual product, so from their US-based point of view, they find something else to hook on to, even when there are facts like concrete partnerships making it clear it isn't vaporware, they just have to say something.
im guessing the negativity is caused by bad branding
> Vaporware is when things don’t appear, not when they are on their way
How do you tell the difference? Wait infinitely long and see if it appears?
If the company is working in isolation (this company is not), the company is continuously pushing back on time frames (they haven't, at least yet), the company never enter actual fabrication/production (which they just did) or if the company never release any details what so ever (they have), then it looks more like vaporware than just "It'll be coming soon".
If those things are true in ~6 months, then I'll join the crowd here who are overly pessimistic at this moment, but until then, as most of the time, I'll give them benefit of the doubt.
>I don’t get the negativity.
Where do you see the negativity?
I don't believe labeling healthy skepticism and criticism as negativity to farm artificial sympathy in retaliation, does any good to anyone.
Humans have pattern recognition capabilities for a reason, and if a company is triggering that in them, then it's best expressed why(probably because they saw this MO before and got burned) instead of just cheerleading the unknown for fake positivity.
> Where do you see the negativity?
First comment: "Looks expensive, I'm guessing"
Second comment: "Probably vaporware"
6th comment: "They haven't disclosed any release date, Lots of chip startups start as this kind of vaporware" (they did literally just enter fabrication it seems)
10th comment: "So far, they just talk about it."
Maybe it looks differently now, after 14 hours since the submission was made, but initially yesterday, most of the comments were unfounded (and poorly researched) criticism.
4 replies →
Impressive numbers on paper, but looking at their site, this feels dangerously close to vaporware.
The bottleneck for inference right now isn't just raw FLOPS or even memory bandwidth—it's the compiler stack. The graveyard of AI hardware startups is filled with chips that beat NVIDIA on specs but couldn't run a standard PyTorch graph without segfaulting or requiring six months of manual kernel tuning.
Until I see a dev board and a working graph compiler that accepts ONNX out of the box, this is just a very expensive CGI render.
Six months of one developer tuning the kernel?
That seems like not much compared to the hundreds of billions of dollars US companies currently invest into their AI stack? OpenAI pays thousands of engineers and researchers full time.
It is. The problem is latency. All these fields are moving very fast, and so it doesn't sound bad spending 6 months tuning something, but in reality what is happening is that during those 6 months the guy who built the thing you're tuning has iterated 5 more times and what you started on 6 months ago is now much much better than what you got handed 6 months ago whilst simultaneously being much worse than what that person has in their hands today. If the field you're working in is relatively static, or your performance gap is large enough it makes sense. But in most fields the performance gap is large in absolutely terms but small in temporal terms. You could make something run 10x faster, but you can't build something that will run faster than what will be state of the art in 2 months.
more like 100 developers for 2 years
1 reply →
Inference accelerators are not where Nvidia is maintaining their dominance afaik.
This 100x. I used to work for one of those startups. You need something crazy like a 10x performance advantage to get people to switch from Nvidia to some here-today-gone-tomorrow startup with a custom compiler framework that requires field engineer support to get anything to run.
The outcome is that most of custom chips end up not being sold on the open market; instead their manufacturers run them themselves and sell LLM-as-a-service. E.g. Cerebras, Samba Nova, and you could count Google's TPUs there too.
very good point leo_e
indeed no mention of PyTorch in their website...honestly it looks a bit scammy as well
Always good to see more competition in the inference chip space, especially from Europe. The specs look solid, but the real test will be how mature the software stack is and whether teams can get models running without a lot of friction. If they can make that part smooth, it could become a practical option for workloads that want local control.
It needs a "buy a card" link and a lot more architectural details. Tenstorrent is selling chips that are pretty weak, but will beat these guys if they don't get serious about sharing.
Edit: It kind of looks like there's no silicon anywhere near production yet. Probably vaporware.
Tapeout apparently completed last month, dev boards in early 2026: https://www.eetimes.eu/vsora-tapes-out-ai-inference-chip-for...
Nice wave they've been able to ride if it's vaporware, considering they're been at it for five years. Any guesses to why no one else seemingly see the obvious you see?
Look at the CGI graphics and indications in their published material that all they have is a simulation. A It's all there without disclosing an anticipated release date. Even their product pages and their news page don't seem to have indications of this.
Also, the 3D graphic of their chip on a circuit board is missing some obvious support pieces, so it's clearly not from a CAD model.
Lots of chip startups start as this kind of vaporware, but very few of them obfuscate their chip timelines and anticipated release dates this much. 5 years is a bit long to tapeout, but not unreasonable.
2 replies →
I can ensure you it's not vaporware at all. silicon is running in the fab, application boards have finished the design phase, software stack validated...
Guessing you work at Vsora, been there or at least know someone there? Any interest in sending out demos/sample hardware to European-based hackers who could spread the word if it ends up actually being a really nice device and developer experience? I can sacrifice myself and serve up bug reports as well :)
[flagged]
need to create an account to reply to people like you who pretend to know. I do not promise anything, I am just giving facts.
I love that the JS loads so slow on first load that it just says "The magic number: 0 /tflops"
It loaded fine for me, but that slash before the unit was a bit smelly. :| Just a tiny edit, but it's a rather core part of their message so they should probably notice and format it correctly before publishing.
I think it could be intended, there is a SI document that says something like "x /unit" is a common way to indicate the unit of a quantity, which a guy I know is using as basis for advocating for that ugly display standard.
Hopefully they do better than UK's Graphcore who seem to be circling the drain
288GB RAM on board, and RISC V processors to enable the option for offloading inference from the host machine entirely.
It sounds nice, but how much is it?
The next generation will include another processor to offload the inference from the RISC V processors used to offload inference from the host machine.
The next next generation will include memory to offload memory from the on chip memory to the memory on memory (also known as SRAM cache)
An FP8 performance of 3200TFLOPS is impressive, could be used for training as well as inference. "Close to theory efficiency" is a bold statement. Most accelerators achieve 60-80% of theoretical peak; if they're genuinely hitting 90%+, that's impressive. Now let's see the price.
hey, we (ZML) happen to know them very well. they are incredible.
The fact that I have to give them an email for details just feels immediately like a B2B-scam.
Hope they can figure out software, but what im seeing isn't super-promising
Esperanto tried to do the same but went out of business. https://www.esperanto.ai/products/
Looks like the design lives on. Wonder if it'll get any traction.
https://www.opensourceforu.com/2025/11/ainekko-turns-esperan...
I'll believe it when I see it wishing them the best!
> To streamline development and shorten time-to-market, VSORA embraces industry standards: our toolchain is built on LLVM and supports common frameworks like ONNX and PyTorch, minimizing integration effort and customer cost.
> This is not just faster inference. It’s a new foundation for AI at scale.
Did they generate their website with their own chips or on Nvidia hardware?
One has got to love the fact hat you only get more information if you submit your email address.
How does this compare to Euclyds product (another new EU AI chip company)?
https://euclyd.ai/
I'm sorry I won't share much details, I don't think much is public on Vsora architecture and don't want to breach any NDA...
From their web page Euclyd is a "many small cores" accelerator. Doing good compilation toolchains for these to get efficient results is a hard problem, see many comments on compilers for AI in this thread.
Vsora approach is much more macroscopic, and differentiated. By this I mean I don't know anything quite like it. No sea of small cores, but several more beefy units. They're programmable, but don't look like a CPU: the HW/SW interface is at a higher level. A very hand-wavy analogy with storage would be block devices vs object storage, maybe. I'm sure more details will surface when real HW arrive.
Does anyone know why they brand it an "inference chip"? Is it something at the hardware level that makes is unsuitable for training, or is it simply that the toolchain for training is massively more complicated to program?
Very simplified, AI workloads need compute and communications and compute dominates inference, while communications dominate training.
Most start-ups innovate on the compute side, whereas the techno needed for state of the art communications is not common, and very low-level: plenty of analog concerns. The domain is dominated by NVidia and Broadcom today.
This is why digital start-ups tend to focus on inference. They innovate on the pure digital part, which is compute, and tend to use off-the-shelf IPs for communications, so not a differentiator and likely below the leaders.
But in most cases coupling a computation engine marketed for inference with state of the art communications would (in theory) open the way for training too. It's just that doing both together is a very high barrier. It's more practical to start with compute, and if successful there use this to improve the comms part in a second stage. All the more because everyone expects inference to be the biggest market too. So AI start-ups focus on inference first.
They also have the 'tyr 4' [1].
It doesn't have to compete on price 1:1. Ever since Trump took office, the Europeans woke up on their dependence on USA who they no longer regard as a reliable partner. This counts for defense industry, but also for critical infrastructure, including IT. The European alternatives are expected to cost something.
[1] https://vsora.com/products/tyr/
Probably because their software only supports inference. It's relatively easy to do via ONNX. Training requires an order of magnitude more software work.
Even if it's not vapourware, the website makes it look like one. Just look at those two graphs titled "Jotunn 8 Outperforms the Market" and "More Speed For the Bucks" (!) ; WTH?
The people who build landing pages for hardware startups are usually (almost always) not the same people who design and build the actual hardware, for better or worse. A lot of the times, it's a outsourced web design agency who receive a brief, often written by business people, and finally the website is reviewed by business people, with some feedback from technical people who complain about the graphs, accuracy and so on. Then the business people say "But we need to point out we're faster than everyone else" and the engineers reply with "Sure, ok, whatever, I have actual work to do, sounds good".
That's not an excuse. This company has more than a few handful of employees. Somebody saw those graphs without units, labels, tags and approved them. Have the CEO never opened their own webpage once and scrolled down?
Not just hardware startups, alas.
The silly verbiage can be excused but not the graphs with completely unlabeled data points, IMO.
Yep that's what I mean - looks like AI slop to me.
reminds me of the famous tachyum prodigy vapourware https://www.tachyum.com/
Well, this has already taped-out whereas the entire reason people call the tachyum prodigy vapourware is that they keep missing their target dates for tapeout and keep delaying it.
[dead]