Unfortunately, there's some real physical constraints: IIRC, manufacturing a wafer takes on the order of a month, start to finish, for the physical processing.
Maybe once LLM improvements asymptote further?
The bigger issue seems to be that these chips can’t hold that many weights at the moment.
(I’m curious if chips with large weights in them would be more tolerant or less to yield issues. If you flip a few bits in the weights, does it really matter at scale?)
Basically a https://en.wikipedia.org/wiki/Gate_array. (The non-field-programmable kind.)
It's not that it would be the best forever, it's that it would be useful for plenty long enough to be worthwhile, even if there was better stuff available. In exactly the same way that this computer I'm typing this message on is not the latest and hottest cutting edge stuff. A 7 year old CPU, 7 year old Intel integrated graphics, an older NVMe disk, a mere 32GB of RAM... ok, that's one spec that's still pretty modern although it is slower RAM... but it's still plenty fast enough to comment on HN, even these seven years after it was cutting edge.
the youd have to buy a new one to get a better model is a FEATURE not a bug.
like if im apple... and i can put a sol level llm in an iphone, market it as privacy first you own your data personal assistant, integrate it all over the os... and then when there is a better model/siri make all the users buy a new phone... thats how they "win" ai.
the old standbys of better screens thinner cameras and batteries arent enough anymore. its basically tapped out. all modern phones are as thin as they need as big as they need as fast as they need and last all day on a battery...
apple needs a new number to up thing that people can actually feel/see. model generations could be it... every year faster, smarter, more capbilities and integrations.
While it’s still too early to tell, I don’t think that’s how intelligence scales. Better models get you better solutions even to trivial problems. The ceiling for getting it done better is very high even if you’re not doing anything complicated. And difficulty isn’t uniformly distributed anyway - it seems to me that “mostly simple” tasks often have annoying 1% tails that low-intelligence models struggle with. I think we’ll see people chasing the top models for quite a while, or indefinitely - depending on the cost curve.
I have yet to saturate the 1M context of Gemini, for example.
The context size is still not nearly as big as it needs to be to store all the code an enterprise needs. And apparently as you increase the context size, there are more defects. So none of this is a solved problem. We are still in early days and there is a lot to be done.
I'm sure there are marketing people who will say "coding is solved" and other such snake oil but none of this is done, far from it!
That being said there is still enormous value in older models especially with tool calling which will let them access the latest data. I feel like we need to be a little more careful and the tools should cache in a smart way to avoid rework but clearly if we could have opus 4.8 level of work for like a one time payment of a system for local LLM it will have value for years into the future.
So I agree in a weird way that sol is good enough for certain tasks but really there is a long road ahead.
https://www.eetimes.com/taalas-specializes-to-extremes-for-e...
https://www.turingpost.com/p/taalas
https://cambrian-ai.com/taalas-launches-hardcore-chip-with-i...
Part of the key is that by moving even from 6nm to 3-4nm one could embed a 20-30B model as part of a MoE (or only a subset of activated layers) on a single reticle die (note B300s are already multi-reticle), with a separate predictive/dispatch model controlling them each on a separate chip. This is without even stacking CiM ROM die. Moving the layer activations (and KV cache etc) between die requires relatively high speeds (and low latency), but distributed with multiple die in parallel might well be doable even with standard multilane PCIe. Of course KV cache prefill could also be handled by external GPUs. I'm sure AMD will make some reasonable choices.
If none of that is baked into the chip as now then all the chips are running the latest weights every time.
Even if you could ignore the stuff built into the chip when the time came, at that point you just wasted money on silicon that’s useless in 2-3 months.
Would you use it?
I don't know why anyone would use Haiku.
If a ROM rack running a near frontier agent model at >10ktoken/sec costs <$1M (rather than $4-8M for NVL72s) and draws only 10-20kW (rather than 100-200kW), and doesn't require a completely new cooling and power system every time you update? There'll be lots of demand for GLM5.3 in a year.
What these don't do is TRAINING, they only do INFERENCE, but they could do it pretty well.
This would be a stupidly bad failure rate, basically the worst business decision you could make, especially if you're somehow on the hook for eating those losses (which seems to be the implication?). Is there a linkable source on this?
The only thing I could find is SemiAnalysis claims that 15% of Blackwells end up RMAed[1]. That appears to be a total failure rate, though, and if you're RMAing them, you're getting replacements. So that appears to be a pretty different state of things.
[1] https://www.dwarkesh.com/p/dylan-patel#:~:text=GPUs%20are%20...
I've personally heard this from several sources in the data centers (installers, training, network). It's not uncommon for 10% of racks to fail on delivery. I hear that's improved somewhat from GB200 to GB300, but the number of FW updates from the time they ship, until they're commissioned is >>10. If an HBM or GPU or backplane supply/cooling fails, it is basically not swappable or repairable. You have a "dead" rack, and deliveries are on allocation so you don't get a replacement for months (eg some "RMAs" for early delivered parts in late 2025 are still dead racks 9 months later). "Tray" swaps are technically possible, but still quite rare, perhaps because debugging takes as much time as commissioning a new rack.
I don't want to out anyone, but these are similar comments:
https://www.linkedin.com/posts/neelmaster1_aiinfrastructure-...
https://www.hostzealot.com/blog/news/nvidia-gb200-nvl72-is-n...
https://introl.com/blog/gb200-nvl72-deployment-72-gpu-liquid...
Given the current rate of change, it would be hard to guess either way. By some measures the cost at fixed quality score goes down vastly faster than that:
A similar trend is evident in the cost of models scoring above 50% on GPQA, a substantially more challenging benchmark than MMLU. There, inference costs declined from $15 per million tokens in May 2024 to $0.12 per million tokens by December 2024 (Phi 4).
- https://hai.stanford.edu/assets/files/hai_ai-index-report-20...15/0.12 -> factor of 125 cost reduction in 7 months.
But that may well be an extreme case. To show how broad the range is, another quote from the same publication:
Depending on the task, LLM inference prices have fallen anywhere from 9 to 900 times per year.A 6 months old model that can run at 1/10th hardware and much faster too, can be much more capable than a sota model when you don't have unlimited budget.
If SOTA models haven’t peaked, then the SOTA model companies would still be churning out better and better intelligence.
If companies working on ML-specific chips was evidence that large transformer models have fully saturated their potential, the field would have been done circa GPT-2.
Chips are another axis for improvements in training and inference. Orgs large enough to explore the space have been doing it for at least a decade now. This is just a silly line of reasoning based on the faulty assumption that somehow, looking for increases in efficiency in training/inference means teams have reached some theoretical limit in model capability.
This way newly post-trained model can be loaded and served the same day.
Maybe! But it also doesn't require the rate of improvement to slow down. As long as some current model is eventually "good enough" for general use, it could still be a market-killer at a very low marginal price thanks to ASIC. Even if slower, much more expensive models are 10x better, that doesn't actually diminish the utility of the ASIC model, as long as it's "good enough".
Yeah, with any luck it would put pressure on Nvidia to charge less, and not just to OpenAI. With a little more luck, we would see all the other players do the same thing, driving down the price of actual GPUs from GPU manufacturers.
It appears that to have working ASIC with the LLM baked into it we need to place and route macroblocks, and not a great variety of them. These macroblocks can be pre-placed-and-routed, available as masks already and shared between different LLMs.
Thus it appears that the tapeout delay can be substantially lower than a year.
That's quite a few parameters shy of today's trillion-weight behemoths, but it is fast.
Let's say a magic set of chips comes along to host this. Maybe it's 2-3x more efficient in size and power. You're still talking a form factor that's a good chunk of a rack, draws tens of kilowatts, and could actually be sold at a similar if not higher price point because the OPEX is so much lower.
It may be useful but it's certainly uneconomic to spend >$1m to self host the model, plus ongoing power and maintenance costs, plus the cost to adapt whatever building you're in to be able to power it.
Something like the next iteration of Cerebras hardware paired with HBM for KV cache + HBF for weights could be incredibly strong here and much more likely to see away to make into a product with some lifetime compared to "let's bake a old model into a very, very large number of custom chips, design all the interconnects from scratch, and pray". Maybe in 10-15 years once this all matures.
Now if that works out that means in '29/'30 we could easily see a run on NAND that's even worse than the current DRAM price issues, on top of the current increases. Fun times if that happens.
Cooling might be an issue though...
Taalas needed a giant chip (6nm) for an 8B model.
At best you could use a more advanced node to try to put a MoE model across several chips working together, but you can’t have GPT Sol size models on a single chip like that.
The other thing is, a lot of the time, model performance is improved with more 'thinking' time.
The thinking time is just more tokens... but instead of say 1000 tokens or 10,000 tokens worth of thinking its 1,000,000... how does that improve model performance? Could a 128B model hit levels of GPT Sol?
The more problem like these they solve the more they will look like GPU.
On some models a large context can be a notable proportion of the size of the weights themselves.
For example, qwen 3.8 27b uses ~64kb/token for the kv cache - so for a 256k token context that's ~16gb of the kv cache for a ~54gb model (assuming 2 bytes-per-param/f16 for both).
So if the current non-baked-in chip is already memory bandwidth bound, as is often the case for current hardware and models, and the "only KV cache in HBM" chip has the same total memory bandwidth, it can only ever be (54/16)=~3.4x faster for the baked in-silicon model.
EDIT: I guess actually (54+16)/16=~4.3x faster, as the current implementation would need to read that KV cache too :)
You're phrasing it like it was kind of an inherent technical limitation with this kind of burning weights into silicon. Which is also not new, it goes back to the 1980s with fixed function digital signal processors and little linear regressions or hardware classifiers for industrial control systems, all are the same basic principle.
It's just usually not worth it to go super small process node, because most models people thought to turn into silicon were pretty small parameter sizes. We're talking 10-100 weight regression or at most 2-4k weight neural net, used in some instrument or factory equipment. You can do a decent MNIST OCR with a 4k weight neural net. For this, 180/130nm is fine.
Or you might think it's required with their special 4-bit as transistor thing (plausible). It's more that when you're experimenting and iterating, TSMC 6nm is their advertised path for rapid prototyping at cost for proof of concepts. And that's already in hot demand, while good luck if you're a startup trying to break in with 3/4nm as your first run.
It is.
As I said, they could have shrunk it with a smaller process node, but that's not at all close to what would be required for a GPT Sol size model.
If you believe that, then you should expect to get Sol-level performance out of a Luna-cost model within six months or a year. If you have a system with the weights baked in, that means you're going to end up serving that Sol-class model several times more expensively than it will take someone who comes along a few months later. (such as what recently happened with DeepSeek's update.)
And under that assumption of continuing advancement, baking things in doesn't make sense in general - it's a play you'd make if you think things are slowing down a lot. Which may be right but it's not OpenAI or anthropic's play.
Their valuation does make sense if you believe: 1) they can retain a massive user base and 2) a massive user base can be monetized. Future value is almost always pulled forward these days for high growth tech companies.
An LLM the size of Google search in users is even more valuable than Google search. The ad market for LLMs will be even larger than search was (no matter what HN prefers).
The monetization part is the easier part. Silicon Valley understands extraordinarily well how to build ad networks. If OpenAI maintain their gigantic user base, a $100 billion ad network is a given bolt-on. They'd have to screw that up in an epic way to not get there.
Facebook - Insta - WhatsApp is an absolute dogshit tandem with a gigantic user base. $228 billion in ad sales and still expanding 10% per year.
Google knows this is what's happening, that's why they don't care about chasing Anthropic very much. They're busy completely remaking how their core search business works.
Even if the balance was net positive, you would also not be able to train them against new tools/harnesses or knowledge. How many years do you expect to keep using them?
Beyond the model, when would you freeze processor performance, such that it was good enough? Because that's exactly what freezing on Talaas is premised around.
The semiconductor technology will also continue to improve. You lose twice. Talaas is one of the dumbest ideas I've seen in semiconductors in decades.
I’d like to think that most parents would be weary of handing their children what basically amounts to a tape recorder that siphons all the data off to a large corporation.
OTOH, a completely local one (LLM + VAD + Speech Rec) would be a fun little thing to build.
but you trade updatability, which I don't think is worth it yet.
I suspect the answer to both of these questions is yes right now, but I agree it’s borderline.
Like he is optimizing to keep providing a vanilla token factory when weighted chips are coming and local models will supplement.
My head canon is savvy chip execs will be etching architecture his OpenAI pioneered into their flagship products while trying to minimize how much foothold he can get in hardware. Murica done offshored it. Not ours to control.
I think people underestimate how much of a revolution having an always-on, privacy-preserving personal notetaker / secretary would be.
Every time a company has spent resources doing this, a competitor innovated on the software and made the custom hardware irrelevant.
This often happened, though not always.
An important counterexample are 3D graphics cards, which basically put the OpenGL/DirectX fixed-function pipeline into silicon. It took a long time and many iterations to make the pipeline more programmable until the 3D graphics cards turned into modern GPUs.
Even today, GPUs live on as separate hardware in a computer instead of having become integrated into, say, the CPU. Intel's attempt to do something like this with the Larrabee project [1] was discontinued.
---
[1] https://en.wikipedia.org/w/index.php?title=Larrabee_(microar...
Only when they became more general with shaders, and then added support for GPGPU, did it truly take off.
I think this generality is the lesson here, not the fact that GPUs are not CPUs.
For example, GPT Sol baked into a custom chip run for $100M that runs 10x as fast and 10x as cheap should pay for itself as long as the chip is useful for long enough.
While 2 years ago nothing was useful more than 1 year long, there are many older models in use now (e.g. Haiku 4.5, GPT-OSS 120b), and I expect this trend to continue.
I know this is what Taalas was doing (acquired by AMD), here was their demo, https://chatjimmy.ai/ which is based on Llama 3.1 8B. It feels like this should start to happen soon.