Remember when GPUs were the dominant tech to mine crypto, then ASICs and other spepcialized purpose-built hardware came out and made GPUs irrelevant for mining, wont that also happpen in regards to ai training/inference. like aws is attemppting to do with their trainium/inferentia chips. Sure they suck now, but won't they get better and thus lose the dependency on GPUs?
It's already happened hasn't it. Google's TPUs are apparently much more efficient at inference workloads. However the hardware isn't available to the public like Nvidia chips are.
I'm unsure about the energy efficiency of Tenstorrent cards, but they are at the moment much more cost effective than Nvidia. But software support is far behind, I suspect if it wasn't the cost would be more equal.
Modern NVIDIA cards are essentially TPUs - especially those spec'd out for AI data centres. TPUs aren't anything particularly special - they are just large vector processors, with hardware specialised for the data + instruction fetch patterns that are seen when running AI models. The graphics cards that NVIDIA sells for data centres are just large vector processors with hardware specialised for the data + instruction fetch patterns that are seen when running AI models.
Google has two different types of TPUs now. One for inference and one for training.
Do we actually know they're more efficient per watt or for the power used and heat produced ? Or are they just easier to scale (and manufacturer) en masse ?
There's also the CUDA moat which is hard to get around
Cuda is a moat for traditional data science, where you have a torch model and you want to easily run it on a GPU. CUDA is the happy path for doing this.
Nvidia's moat for LLMs isn't Cuda, it's installed capacity, arrangements with frontier labs, TSMC bandwith, and just the fact that it currently has no real competition.
Are they modifiable after creation? If not that seems dangerously rigid until we reach a much more stable place with LLMs. It could be great for well worn architectures and models but building one now for LLMs seems like it'd be foolish when it'll be outdated in a few months.
Why would it be dangerous? It'd effectively be a stamped model at some point in time. The bleeding edge moves on every few months, but even now it seems to me that there's only a tenuous value proposition that's progressing for me. I'd rather have an instantaneous GPT 6, Opus 5.3, or Gemini Hydrogen that I own, rather than always leasing models I have no practical application for.
The better question is, would someone be able to turn a profit offering a deepseek chip for a few hundred dollars? I wouldn't even care about the energy draw (within reason) having that running entirely locally at better than a few hundred TPS without going bankrupt to obtain VRAM would be a game changer.
Basically capacity issues. NVidia + Apple have bought up a significant portion of TSMC's manufacturing capacity. The primary supplier for Google's TPUs is Broadcom, with TSMC as a second source, and they're trying to ink contracts with other suppliers like MediaTek, but there're only so many semiconductor fabs in the world and TPUs need a fairly modern process.
Within Google itself the TPUs are definitely preferred for AI workloads, you just can't always get them.
Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not really the same markets. NVidia is selling chips that AI runs on, Google is selling services that use AI. Google's play is to make better services than the competition can because they control the full infrastructure stack, thus preventing their markets from being commoditized. NVidia's play is to commoditize the AI providers so that the number of services expands rapidly, with NVidia chips powering all of them. Arguably it's Google's strategy that is the market failure, as it's designed to keep competition out, but...
Google's used this strategy a few times before, and it frequently doesn't work out. They kept MapReduce and GFS proprietary so they could build better big-data services than everyone else, but Doug Cutting from Yahoo clean-roomed it, and then eventually you had interviewees going "MapReduce? That's like Hadoop, right?" They had a long string of internal ML frameworks well before the rest of the industry (I remember in the '00s and early '10s Google was by far the leader in large-scale machine-learning) but had to open-source TensorFlow in the face of external competition, and then lost in the marketplace to PyTorch because TensorFlow was filled with internal WTFs and footguns needed to support specific Google projects. They had a big lead in datacenter technology over everyone else in the industry in the early 2010s, but kept it internal and proprietary until they were forced to sell it as a service by the huge growth of AWS and Azure, then had to play catch-up as a distant 3rd because they lacked the customer relationships and developer mindshare.
Both are fabless customers of TSMC, which is the choke point and center of the high-efficiency chip market.
Unlike gp, I do not believe its a market failures as much as the cutting-edge process node being a natural monopoly. Once TSMC processes got really good, TSMC started capturing most of revenue dollars and reinvesting them, which made TSMC products even better - a positive feedback loop that makes it impossible for Samsung or Texas Instruments to catch up. Intel rode on a similar wave for decades
Packaging is the choke point in AI. It's still TSMC but chip production is doing fine. It's the Chip-on-Wafer-on-Substrate build chain that is severly constrained.
Why is no one going to Intel then for this kind of stuff ? Or even a legacy producer like IBM or Ti ? I would imagine they could buy some slightly older ASML litho machines and make something on a slightly older process (or package it using the older process for the other stuff and just keep the main processor latest gen) ?
From what I’ve read it’s that TSMC is just so much better at it… much higher yields and much more reliable than any of the other fabs. I wanna say something like 2-3 years ago, before Blackwell, NVidia tried to use Samsung for example but had nothing but issues. Intel hasn’t proven its fabs capable of anything really useful in nearly a decade.
I think even if you have the capital to acquire something from ASML… by the time it’s setup and tuned correctly TSMC will likely have enough capacity to eat your lunch. If it takes 2-3 years to get a fab online TSMC will likely have a new node and their old ones will be available.
TSMC has really proven over the past 5-10 years that the fab is its own beast and requires its own immense level of expertise and investment. When you have the worlds largest chip companies shoveling all their manufacturing dollars through one single org, that seemingly does the right thing, it’s going to be pretty much impossible to beat them.
BigTable and MapReduce are differennt use-cases. The former is a database; the latter is a framework for running batch jobs. The (internal) successor to MapReduce is Flume (2010), which is quite a bit more capable, but didn't come out until Hadoop was already well-established. The external equivalent is Apache Spark (2009), but the gap between internal and external technologies had already shrunk to be nearly indistinguishable by then.
> Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not buying up "all the supply to build an inferior product". They're buying it up build a different, high-demand product.
TPUs are a systolic array. It's not comparable to the hardware architecture that backstops CUDA, nor do they serve the same markets.
The writing was on the wall for TSMC ever since the 5nm node; Apple proved that anyone with enough liquid cash could block competitors from fab access.
Modern NVIDIA GPUs contain tensor cores that are optimised for MMA operations as well though right? The standard CUDA execution path isn’t comparable, but the tensor core path very much is?
I'm no expert but I think we're dealing with two different ends of the spectrum here. Nvidia GPUs are likely cutting edge within the individual board / processor itself. Their main limiting factor right now seems in effectively networking them together (which is why they bought Mellanox or whatever awhile back to further develop Infiniband).
TPUs are significantly less flexible and I would guess at an individual level not as efficient or fast as Nvidia. But they're designed specifically to be bundled together in large arrays with similarly large data pipes to storage and interconnects to cache, ram and other compute. I would guess Google didn't bother trying to sell them until recently because you needed to be deploying them at large scale for it to be even worth also developing the software for.
What we really need is innovation in super fast (bandwidth and latency) and reliable interconnects. I suspect the days of slotted in GPU and RAM will be over soon and we'll see a move everywhere to pin based sockets (like we've always had for CPUs). Stuff like SXM and LPCAMM.
We've started to vertically stack ram out of necessity but have yet to really move on from the 2D plane for the motherboard itself. Slots allow more devices to connect to one flat board. But if the board itself had more surface area (by being maybe an L shape or box) you could slot in a lot more chips physically
It's logical that company like TSMC sells capacity to the highest bidder.
The main thing is that chip manufacturing is the most highly capitalized industry in the world. Each manufacturing step has gargantuan costs and so no company wants to manage (and pay for) all the steps - so TSMC buys chip manufacturing equipment from ASML and sets up a system where corporations buy capacity. It's essentially a very, very, very expensive print shop. The print shop doesn't make its own printers nor it print a newspaper.
And naturally, with very, very objects like leading-edge chips, "inferior" or "superior" is a ridiculous simplification. Nvidia chips are incredibly powerful and useful. Nvidia has been more committed than any other company to general purpose GPU programming.
But Nvidia also is doing it's best to sell at high, monopoly prices. That's why they buy a lot of capacity. The thing about the situation is these companies did spend a lot of money reaching this monopoly situation so from their viewpoint, this monopoly pricing might be fair. But I personally wouldn't care and hope some regulation stops it. But that kind of regulation would still have to get these companies significant profit on these investments, so they'd keep making them.
Interesting. What manufacturing process is Broadcom producing them at ? I'm guessing they're not newly as small (lithography wise) as the best of the best yet but probably good enough
Broadcom isn't a chip foundry so calling it Google's supplier doesn't make sense. TSMC has been the primary manufacturer of TPUs but Google's hoping to use Samsun and Intel also going forward (and these are the only leading edge foundries in the world). Broadcom is co-developer of the chip.
I don't know details at this level but I know that TPUs like GPUs need to be made on the most advanced available nodes and TSMC, Samsung and Intel are exactly the only enterprises pursuing those advanced nodes. The reason is that this is
I mean, Google tell me it's TSMC's 3 and 2 mm process - exactly what you'd expect. But you can Google as well as I can.
Presenting broadcom as an alternative to tsmc is like wet roads cause rain. The average knowledge of computer hardware development cycle in hn comment sections is frustratingly low.
Broadcom is just a chip designer or “mercenary” as I read the other day. TSMC is still the fab for most of the TPU, with Mediatek, and Intel used for auxiliary/supporting components.
google doesn't use them, they buy them to make them available for google cloud customers that require CUDA or other nvidia-only workloads. All training and inference at Google is done on TPUs.
You have basically managed to sum up why NVDA keeps dominating, no on can "just buy a few TPUs" and that has been the case since the first TPUs were announced years ago (yes yes GPU shortages and so on, but no one questions general availability of GPUs), and the current Tenstorrent model support page looks like an invitation to beta test HW that might (or might not) be able to run current and future models at some point.
But the next gen of Tenstorrent HW looks great on paper - problem is, I could have written that before the current gen was released.
All of the allegedly competing HW looks like that and has looked like that for years, either a lack of availability, capability, or both.
Relevent article from september 26. To me it’s obvious gpu are a « first response » solution. But I think nvidia can sell a lot before AI R&D slow down (because of the demand or the tech, infra, energy… or whatever reason it will be)
To be fair GPUs have already changed a lot in response to the AI demands. A lot of silicon these days are tensor cores which have very few applications outside of deep learning.
Inference, yes, see https://taalas.com/ and https://chatjimmy.ai/ As expected, efficiency boost is extreme. But you've got to pick the model you want to enshrine in silicon months or maybe years before you actually get that silicon (and maybe there are additional difficulties if the model is large. Maybe in a few years everybody will just want to use the same four or five models, with no tectonic shifts on the horizon. Then it'll be a no-brainer to make ASICs for those models.
Training? I'm not so sure. You need something flexible so that it can be trained. Burning the weights into the board is the opposite of that.
I read into Taalas when they first were putting out info and they also are wiring up massive CPU cache to the thing. It's not just the burned in models.
As much as I love the idea of Taalas, their chatjimmy demo is a hard-wired (and probably very quantized) 8b LLM producing garbage extremely fast (17k tokens/s). Extrapolating from the picture of their PCIe board in their demo, I would expect a frontier model to require an IC not smaller than my desk.
At these dimensions, you are better off producing general-purpose inference ICs that can run any model, instead of baking in a single fixed LLM.
So, Taalas (now AMD) will probably target small models with these (maybe something like qwen3.8-27b, which seems too large already), if at all.
Something that could benefit from throughput/Wh but wouldn't benefit from flexibility and being cheaper than using bespoke chips feels like such a narrow set of uses though
It still won't make normal GPUs irrelevant like they did with crypto, there will always be demand for training and SOTA models while there is no demand for GPU mining
> there will always be demand for training and SOTA model
Some, but how much? Once you have a mass produced ASIC with a model that's "good enough" much (but obviously not all) of the demand for the cutting edge could evaporate quite rapidly I think.
NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto, but now it seems the will drop the entire customer graphics market long term they just haven’t committed themselves to it. They can’t go back to GPUs anyway there are not enough gamers to reach the revenue numbers they need now and soon gaming/graphics revenue will be irrelevant for them. Most US companies won’t maintain a boutique of products that don’t have synergy or are actively competing for internal resources long term they put all their resources where the margins and growth are best and for NVIDIA that’s not gaming anymore.
> NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto
But Bitcoin mining moved GPUs to ASICs many many many years ago and Ethereum switched from proof-on-work using GPUs to proof-of-stake: GPUs are pointless now for Ethereum. And Ethereum was the big one for GPU mining. All the others cryptocurrencies are too tiny to register.
Ethereum switching from proof-of-work to proof-of-stake has been the death blow to GPU mining at scale. I'm not saying some have not been profitable mining other less important coins: I'm saying the market for GPU mining shrank to next to nothing.
Nvidia has multiple B2B customers that require the consumer graphics featureset, like Nintendo. There's sufficient motivation in keeping the GPGPU/raster capabilities on the same card even if PC gaming dies as a market.
There are also the opposing architectural directions to extend toward:
one in the "Cerebras" direction of ever-larger scale general purpose HPC silicon running inference workloads. Cerebras may have skipped ahead to the last act of that scaling story, notwithstanding any further wafer-diameter enlargements past 300mm. Whether the ending is happy or not is still unclear. Lot of iterative room to grow into it further and they have a formidable patent moat waiting for conventional die-sized AI semi vendors when they inevitably run up against limitations of single-die designs.
the other is in the "Taalas" direction of freezing a specific model into something like high speed ROM / FPGA and getting orders of magnitude inference speed improvements - for that specific model. Like Cerebras, Taalas may also have jumped directly to the architectural telos of that philosophy, with only iterative improvements in cost per work unit left to grind, though like CBRS quite a lot of grinding left to do especially in getting cost per custom tapeout down to near-disposable (hopefully, recyclable) levels.
I don't know. But my assumption is that they are. The crypto ASIC market was way smaller and less ambitious than this. It would have been just a relatively small cash grab for them to get into that market, whereas this market could represent an actual strategic direction.
Yeah, that's my thesis since the beginning of this madness. I see it as a temporary blip (albeit it's already much much longer than I would expect !) and once the bubble pops, Nvidia will be back to their normal cruising speed. Which is a good company, making great products, but their valuation will be 1/10 of what it is now.
And if the boom continues, there is no way either that their clients won't develop a product of their own. It just doesn't make sense to buy chips by the 10s of billions $ on which Nvidia makes >50% margin.
So will people buy gaming GPUs fron Nvidia again after how if abused them this time ? Or will the AI obsession be their downfall & example for others what not to do and how not to alienate the customers who get you where you are now?
This is why I think their pivot/venture to general robotics was smart. However, I doesn't seem like they will become profitable on that front because they were outcompeted pretty hard in most measures I can conjure.
Yes. Half dozen companies are designing special chips to bypass/replace the GPU.
Bitcoin was lucky for Nvida, then AI was lucky. Third times the charm?
On other hand, luck favors the prepared. Beyond luck, they did make their chips accessible, easier to program than other GPU's. So bitcoin miners and then AI could utilize them.
There are still a lot of people that want to run regular liner algebra.
Nvidia cards were the worst hashrate/purchase price of any cards. AMD were the lucky ones there if you consider anyone lucky in this situation. Consumers definitely weren't lucky because the GPU market hasn't recovered since the Etherium boom.
I think things changed, in the early days of BTC AMD (ATI at the time) was king. Later GPU made a comeback with Nvidia rigs doing well during Ethereum era.
Mining ASICs are bound to a single cryptographic hashing algorithm, meaning they do one thing and they do it very fast. Something similar may make sense for running models locally, but I believe the more multifaceted Nvidia chips make more sense for training.
Taalas takes a similar approach but at a much higher level of complexity - instead of an array of nnn tiny specialized cryptographic cores running the same algo (or a mix of several with interlocking dependencies), it freezes a single entire LLM on to one die, running inference both far faster and cooler than any general purpose compute but with the same one trick pony obsolescence drawback. Question remains whether high volume fab tech can be wrangled into a form that allows this approach to drop unit costs near "disposable/recyclable" levels of cheapness per chip. This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
>This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
Yeah I was thinking this will probably be the case. We may eventually settle into a slower release schedule as well, where a new model only comes every year or two.
If you look at the sessions from 2026's "Hot Chips" there's like a dozen competitors and possibilities and new architectures and the like, yes. Some very compelling. And OpenAI and Google and others are also making their own chips.
The thing is, it's only NVIDIA really that has a proven product in high volumes pumping through TSMC with the existing software stack that the community has already proven. That's a huge difference. Everything else is riskier if you're building out in volume. For now.
Check out Taalas, recently acquired by AMD. They printed a model directly to silicon and it generates 16,000 tokens per second. You can even try a demo here:
Whats amazing is that this is around 5x faster than the top end Cerebras wafer-scale silicon, at much less power/cooling cost per token. Of course Cerebras is a general purpose design, but perhaps Taalas could license CBRS wafer-scale patents to create a single monster chip that held either a single massive trillion parameter-scale model, or could an MoE design with a half dozen smaller specialized models on the same wafer with a shared SRAM pool.
I wonder if cooling such a beast, with very non-uniform hot spot distribution, would be too complicated to bother with though.
Cerebras don't need to customise their chips. To my knowledge they are crazy fast but struggle to fit larger models and it's not clear they're economically favourable.
Ethereum's work function was designed to be memory bandwidth limited, which is something gous are already quite well optimised for. An asic would be faster but not by enough to be worth the investment, at least while gpu margins were less crazy.
The main thing is the actual number crunching and memory interfaces for LLMs are not crazy hard to implement, and the potential scale is huge, so NVIDIA's moat is not likely to be big enough to make their current margins sustainable in the longer run.
Not to mention, many ASICs were outclassed by a few well-tuned GPUs in terms of power-efficiency. There were only a few coins that bucked that trend, usually just altcoin/shitcoin experiments that weren't worth mining in the first place.
I'm kinda bewildered, it's bizarre seeing people cheer on quickly-obsolete accelerators instead of demanding their OEM to adopt GPGPU. CUDA will be relevant for much longer than any ASIC can ever be, even (or especially) if the transformer gets replaced overnight with a better architecture.
Yes, ASICs are application-specific e.g. "fixed". But, I think it's possible to target algorithm implementation that is inherently flexible, only limited by HW constraints (bit widths, RAM depth, etc) and thus cost. Yet, still fully pipelined and much higher performance than alternatives.
My background is ASIC design and more recently 3rd-party IP development for integration into ASICs. In our last design I made heavy use of resource sharing (RAM and holding regs for intermediary values, counters, LUTs, complex multipliers, adders, etc, even state machines).
That got the wheels turning. I believe such an approach can be generalized to support a wide range of computation and thus algorithm development. Even state machines can be configurable if designed that way. All controllable by programmable registers.
Just don't look for AI to take on such a task. Or go ahead and try, it already fails miserably when it comes to RTL design (verilog, vhdl). LOL
Post hoc rationalization. Nothing in here was unknown on 10/29 when the scandal broke, yet the stock market value was pretty durable before this and dropped a ton after.
Common saying. I always wonder how the stock market would have priced pick sellers? Would they have baked in 10 years of continued revenue growth from the sale of picks?
Look at Cisco. Only this year did its stock return to the price it reached at its peak in 2000. Since then the amount of people on the internet has grown ten-fold.
Excellent point. I would guess gold mining profit for Caterpillar is similar now than selling shovels to gold rush in 1849 though. Point being, Nvidia may grow revenue 5x during that margin compression.
Given that Nvidia currently has ~80% market share, that implies the whole market also has to grow 5x. And if Nvidia starts losing market share to their equivalent of Komatsu, Volvo, John Deere etc, then that means market growth of 20x or more. I don't see it. I don't think the world can absorb data center growth at that rate.
I don't think anyone here is claiming that LLMs will go away. However I do think there are probably a lot of opportunities for commoditization around LLMs, and if/when that happens it will signal an end to the "gold rush" days.
I think a lot of people do think somehow it will go away aka the individual I Was responding to.
I don't think demand for Frontier AI will go away either - even as in the near future 90% of AI will be 'commodity AI' - the Frontier foundries will still have competitive differentiation to offer.
Just like SaaS - technically almost all SaaS should be 'commodity' ... and yet it's not.
I think we can all agree that we are 'sick of the hype' but that doesn't mean the engine will diminish.
> It's peak delusion to think that AI is going away.
Even if it's not going away, it'll be a breath of fresh air when it's not shoved into every possible place as a marketing tactic, even when it's not necessary or beneficial. AI does have it's place, I just hope we realize soon how every place is not that place.
I think the ultimate role of AI is probably going to be "every place", just not... foregrounded.
It will follow the same pattern that technologies like phones, computers themselves, the internet, and (I guess) cloud storage now touch on nearly everything.
Example: even if you're an old-fashioned boutique leather artisan who makes handmade bespoke horse saddles in 2026 like it was 1783... your business also probably uses all that other stuff in some capacity. You probably have some computers for design and payroll and answering emails, a website that lives somewhere in the cloud, etc.
On a closely related note, if you're tired of hearing "AI, AI, AI" everywhere, there's reason to feel encouraged. Ten years ago, nobody would shut the fuck up about "the cloud." It didn't go away, it just matured and now it's how most people do many things.
I don't think LLMs are as world-changing as the internet, that is a very, very high bar.
They're useful certainly and won't be going away after any financial bubble bursts, whether they are so useful they need to be rammed into every facet of our lives remains to be seen.
One big difference between the internet and LLMS - people enthusiastically adopted the internet and found all sorts of uses for it. LLMs companies are mandating usage by employees and trying to force end users to use it, which seems really strange, why would they have to do that if it's so useful? They're giving it away for free at present.
The best usage I see so far (which people like, though there are problems) is to replace search, so it certainly could be profitable, though it would require google-like infrastructure as google isn't going to offer those search results for free to bots forever.
AI will have about 'on par' affect as the Internet.
Companies are 'mandating' use of AI because they see the potential in it.
In 1997, people were not sophisticated enough to grasp how broad the web would be - if they did - managers would have been pushing everyone to adopt the browser and email.
That said, in the 1990's they were actually pushing everyone to adopt 'Windows' etc..
We have a better grasp of how tech affects the economy, and so now there's a lot of hype. Probably too much.
AI is just in 'Phase 1'. The current Chat Bots are basically like the 'Commodore 64' - a 64K computer.
The hype is annoying - I use AI all day and find it useful and drives me crazy but at the same time 'it's real at the baseline'.
It will not 'explode' I think any time soon, it's going to take years, but in 5 years we will look back and see a huge amount of change.
In 5 years, up to 1/2 of AWS services will be AI based for example. All sorts of new features.
An example: RAG is being embedded into most DBs now, that's a big deal and taht's just 'embedding' not even really AI. In a few years, regular 'javascript jockeys' will be used to that kind of thing.
I'm by no means AI-pilled, and I'm skeptical of all these shenanigans ... but 'it's real'. This is happening on about the scale of the Web Browser.
I really don't think we can say what impact LLMS will have yet - AI is such a broad term and is much broader than the LLM craze, which is very specific and led to this dramatic demand for GPUs.
AI has already had a massive impact (machine learning to recognise pictures, spot patterns in data etc) and will continue to do so.
We're talking about LLMs here and the huge amount of money that has been pumped into them by companies like Nvidia in circular deals with their customers. Personally I think the jury is still out on whether LLMs will develop much further than they have. Comparing Opus 5.5 to Fable 5.2 this week I've seen little difference. We'll see in time though.
It does make me wonder if this is a concerted press attack by short sellers.
All the stuff about Jensen's affairs, NVidia sales to China, circular financing, increasing competition in the market, etc. was known months ago. Indeed, much of it came out in research notes by major short-selling funds. And NVDA is an extremely attractive target for shorts, having been bid up high by the AI bubble and then having a lot of retail interest that could easily be pushed the other way. Entirely possible that certain funds have taken a large short interest and are now trying to give it that push to profit.
1) short sellers do not control the press. The best they can do is material leaks, like those 'Heisenberg' or whtever research companies, but you have to drop a real bomb for that
2) Trump is generally pro Jensen and most of the Cabinet and tech people are, and Dems don't have a specific fight with him
3) Jenses is not a likely face of 'tech evil' like Musk could be.
Most likely what you are witnessing is how ugly and nasty the press can be.
I've seen so, so many articles fro 'resepectable news outlets' with tawdry and slacious headlines - and almost no real information.
The 'timeline' of Mega Speed ? Had to go to AI to find it, and it's far more nuanced than what the headlines are saying.
The press loves waves and hype - it brings them revenue.
I think there is pent up anti-AI populism and the Jensen story was just the thing to pop first.
So he got it.
Also - the news just kept leaking aka 'yet another woman he slept with' so that kept the story alive.
There also might be things we don't know aka stuff that the Board has been made aware of that we don't know.
I think it's sad because I think Jensen is a reasonable and decent person that owes an apology to his wife, and not really anyone else.
I think it's an underreported detail in the english language press & internet that openai and anthropic are visibly lobbying to get the US government to ban open weight models or de facto make them difficult or impossible to use.
And Huang was by far the most powerful person outside of china taking the other side of that. I'm not going full conspiracy here, I genuinely don't think it was necessarily a coordinated takedown or anything like that. But if you look at whose public positions incentivize them to rile people up around this it is definitely worrying.
If you read amodei's remarks over the last few years he very clearly considers llm development to be a US vs china cold war 2 type deal that justifies dramatic action that would otherwise be morally abhorrent to him.
Now here's the conspiracy brain part. Say you're ideologically committed to impeding chinese labs' ability to develop open models, and willing to do anything short of literal bloodshed for it. We know tech companies consider manipulating public sentiment for their own purposes to be within the bounds of fair play. And in this hypothetical you have unlimited unrestricted access to claude mythos. What could you do with it?
> I'm not going full conspiracy here, I genuinely don't think it was necessarily a coordinated takedown or anything like that.
Unless it's some 4d chess tactic, the timeline doesn't have much sense, the ANT IPO is paying for this takedown.
And generally speaking, if you want to control you adversary you try to "leak" tech and information in order to control how much and what exactly he knows. Going full ban doesn't make any sense in the long term.
Epstein files, the American soldier who was arrested in UK for spying for Russia (https://www.bbc.com/news/articles/c6d7q77xgw0vo ), the blue wave, and I am sure I am missing a few others.
Nvidia's chip business seems certain to survive. They have also made themselves an "AI financing company" and the question is whether that enterprise could sink the Nvidia the corporation. Notably, Nvidia has "$530 billion in gross off-balance-sheet guarantees..." [1]. Nvidia total capitalization is 3.77 trillion and its yearly revenue is 200-300 billion. As long as things chugging along, the chip business can support the finance business but if you have a sudden shortfall, things may be tighter.
Which is very typical. It's one thing to recognize a bubble; it's another to predict when it's going to pop.
There were lots of traders who lost a lot of money betting against the housing bubble pop in 2005. The bubble was front page news on the Economist in 2005. Michael Burry made a lot of money because it took him a couple of years to assemble the money to be able to bet against the market and so he got lucky on the timing.
Or as the aphorism goes: the market can stay irrational longer than you can stay solvent.
Isn't there a bit of popular wisdom that goes like this? "If you want to time the market you have to get two bets almost perfectly right. When to get out and when to get back in"
The share price slide happened before the resignation. Uncertainty, sure, but if the only change were that Jensen had not yet resigned as of today, that uncertainty would still not be resolved.
The story of the affair(s) and the Chinese spy were, yes. My point is that those stories and uncertainty seemed to have done most of the share price damage so far. The actual act of pushing out Jensen resolved some of that uncertainty. If everything had played out the same but the board had not pushed him to resign, I'm not sure things would be resolved any better currently or in the near future.
Of course, you’re describing risk. Which is how players in the markets make investment decisions. Nvidia are now a more risky investment, until proven otherwise.
The markets saw Nvidia as an increasing risk before the actual decision to fire him, so that doesn't seem to be the single incident causing uncertainty. That risk and uncertainty already existed whether or not he resigned or was fired.
I never claimed it was a single incident. Risk profiles are an ongoing metric that take in myriad factors. Some weighted more than others. Losing the co-founder and CEO of 30+ years would have a heavy weighting, causing uncertainty until a new CEO has proven the ship is steady. Of course, the potential AI bubble is a factor too. But, losing your experienced CEO just before a potential popping of said bubble is also an extra factor to add to the risk metric.
People have claimed that competitors are catching up for a long time. But these journalists never understood why Nvidia was so dominant. And it seems they still don't. The Jensen story only made investors unhappy because of the high volatility and uncertainty. Noone who's well informed dumped Nvidia because of technological fears of competitors.
OTOH developers keep confusing R&D vs production. Nvidia is going to keep its dominance in R&D. But the vast majority of Nvidia chips are used for training and inference. Using Nvidia for that saves you 3 days of work porting from R&D to production.
It’s over if they don’t put a technocrat in charge a bean counter isn’t gonna keep them on top. See Apple when Jobs was fired or the clown lineup of bean counters at Intel Intel…
Maybe they wouldn't be the laughing-stock of AI if someone better than Tim Cook was leading them. Siri is pretty bad, "Apple Intelligence" is a misnomer.
Apple is currently the most valuable company in the world and unlike many of the companies that aren't "laughing-stock of AI" it actually has a defensible moat for its products for the foreseeable future.
I'm generally not an Apple customer at all (I own an old Mac Mini to do occasional iOS development, but rarely use it), but think they are a better bet long term than any of the AI labs.
Nvidia biggest problem going forward they aren’t as versatile now essentially a one trick pony and now that they don’t have their CEO, they will be rudderless. That latest team up with Microsoft was a dud and a bean counter CEO isn’t gonna be able to fix that.
To be fair, ram really is cyclical, and a commodity.
The Nvidia ecosystem… not so much. Most all of these pieces fail to recognise their already dominant position in the edge inference space - and most all of these pieces fall into the trap of AI=LLMs=ChatBots, and really, who wants to chat with a computer any more.
I don’t mind. This has been a great week to stock up on the equity, and there was enough blood in the water before the salacious spy sex story that I bought a bunch of very, very cheap $150 puts, as hysterical overreaction is the plat du jour, which I sold on Thursday.
TLDR: LPUs are very efficient, quite a lot more efficient and faster tok/s/u than anything else at this years Hot Chips.
It's my personal opinion that with the way models change you don't want silicon that is too specialised but it needs to be specialised enough. LPUs are designed to strike this balance such that they remain flexible if model architectures change drastically.
In my personal opinion as well "specialized enough" chips are exactly the right way to go - though disclaimer, also At Nvidia, working on spec dec/external drafts :)
What's your opinion of high bandwidth flash? It seems ideal for storing and retrieving model weights. Do you see that dramatically changing inference chip design over the next couple of years?
Media are basically a reporting machine trying to get views. I have known that for quite some time, but this is another experience I will remember.
On the article, most of the tech discussions make zero sense as many of the comments already pointed out. Some of those challenges is also one of the reason why Nvidia has accelerated the whole roadmap and doing iteration of GPU every 18 months of so to stay competitive. Capitalising on leading edge node early and securing larger wafer quantities than even Apple. Large commitments so TSMC is also building Fabs faster than ever. The whole thing filtered out to all the downstream supply chain. Financing it as well making sure the capital is well positioned rather than forcing their partner to raise capital on additional capacity. Same playbook as old Apple. And in case anyone is wondering, it is Nvidia that is now pushing and funding the development of leading edge node along with all the other leading edge tech. Without them, TSMC A14 wouldn't happen as early. PCIe 6.0 would be a year later, Optical interconnect is 2 years early if not more.
It is that Datacenter are put on halt that is the biggest concern. And that didn't happen before "Jensen's mess".
This "trouble" is just silly, and mostly misdirected
(a) making chips is not making datacenters (esp. the memory fabric). Nvidia is working up the value chain (racks, datacenters), backwards (by supporting customers), and sideways (by lowering the cost of related technologies). No one else is close.
(b) Nvidia profit margins are so huge they can destroy competition
(c) The main cited competition is from their main customers, who are just trying to get a better deal from them.
However, the China trouble is real: Huawei is making chips for model-building and inference that the Chinese labs are happily using (to copy models?) that are good enough to dump models and chips to make a monopoly. Stirring up anti-tech, anti-datacenter concerns at the start of this race makes that a real concern, hardly mentioned in the article. It is not possible to face down a cheaper substitute with infinite resources.
Not good that people are stirring false worries when there are real concerns.
One of Tim Apple's [sic] big innovations was to take Apple's giant pile of cash and use it for vendor financing. Building a new fab or process or factory is expensive. Borrowing money is expensive and risk. But Apple gave favorable terms in exchange for a period of exclusivity to whatever the vendor produced. This was huge for the iPhone for years.
NVidia is kind of in that position now. One can say that there is circular financing between NVidia and the big AI companies but is it really that different to Apple's approach? I'm not entirely convinced that there's an imminent domino effect collapse coming as a result of all of this. And, even if it does, I don't see it taking down NVidia.
Boeing (and Airbus) is the other company NVidia reminds me of. The business of selling planes is complicated. Boeing gets to throw its weight around and demand concessions. Launch partners for new planes get favorable delivery dates but that's also a chain. Qatar Airways famously went to war with Boeing and lost its A321 orders. That's the kind of power NVidia has now over the AI providers. It gets to decide who gets to buy and when.
As for competitors, there are always wannabees. Remember the decade of Google killers? I do. What happened to them all?
It surprises me on HN of all places how many people think you can vibecode a GPU competitor to NVidia, basically. I'm sorry but no. Chip design is harder than that. Managing relationships with suppliers, vendors, the fabs (particularly TSMC), etc is hard. NVidia has been doing it a long time, has a lot of built up expertise and has a lot of volume and weight it can throw around.
Jensen Huang was correct about China (IMHO). Restricting the chipmaking from ASML is one thing. Restricting the chips was a mistake in foreign policy terms. It has created captive domestic demand for Chinese chips. And it's too late to unring the bell. Xi Jinping realizes that chipmaking and AI is of national security interest to China. Yes there'll be some smuggling in the meantime but I fully expect China's EUV to be producing commercial chips by 2030. China simply doesn't want to be beholden to Western companies in the exact same way that most US tech companies were banned in China. I believe history will vindicate Xi's long-term strategic thinking in this area (yes, blocking the likes of Facebook and Uber predated Xi, I know).
I also think AI isn't going to be ASICed anytime soon like crypto mining was. Crypto had a fixed algorithm. AI does not, at least not yet. It's constantly changing. Even in recent years we've tried different quantization methods (eg fp4, int4). Model architecture is still evolving at a rapid rate. It's way too early to fix that into hardware.
Polymarket says 7% and hasn't budged at all. Jensen looks like he will be returning as well, turns out it was just a coordinated attack from the media.
There's a simpler explanation for the stock price drop here. The price had almost gone up 5x in five years, and bad news caused people to stop and take profits. You don't have to assume their business model is bullshit or that the AI bubble is popping or that there is fraud to justify a 10% pull back after a run up like that.
There's so much projection and wishful thinking about completely normal price movements in hyped stock.
> You don't have to assume … that the AI bubble is popping
I’m not saying it is, but what is a bubble popping, if not a bunch of people saying “oh this might be the highest we go for now, I’m gonna take my profits”, which causes downward pressure for those who come slightly later who also want to take profits, etc?
The question is just if enough people stop to take profits to trigger that chain reaction. After all, it’s basically the same as saying that the seller thinks the current value in cash is more valuable than the future value of the company, otherwise people don’t really sell
Because people think the stock is worth holding 150 (or 140, or wherever it bounces from) but not at 180. It doesn't need to go _all the way down_. A normal overshoot and correction is not a "bubble".
A bubble would be if there's nothing backing it, it's not the same as the price temporarily overshooting, which happens _all the time_.
People who don't watch stocks desperately want to tell a story about what this stock price is telling them. If you watch tech stocks, this kind of thing is completely normal when a stock is on a bull run. Just look at the huge swoon nvidia had last year when it dropped from 140 down to 100.
Oh no help, I have one of the most in demand products in the world that everyone is paying a huge premium on. I am in trouble, please save me. Total horse shit. If you do a sensitivity analysis to a total collapse of data center demand it needs to be like a 75% reduction before its much of a problem to fundamentals.
Shorting is really really hard. And the end goal of shorting, sure for someone it's self enrichment, but why it's tolerated as a whole is to help identify and pop bubbles before they get out of hand.
Journalists plausibly have the same goal: speak truth to power to protect society. To be an investigative reporter you take on immense amounts of personal risk, and on very rare occasions get a big payoff in terms of recognition and career prestige.
So Wired is "short" Nvidia only in the sense that they benefit from pointing out its problems, and mostly only if they are right and get the timing right, otherwise they just lose status.
It's a legal fiction to insulate the real people from having their personal lives destroyed in court; they were just following orders type thing.
I never said shorting was easy; don't put words in others mouths.
Doesn't stop people from trying. This forum is naive af about a the world; there are billions of people who don't know you exist making choices that, like a spider-web, shake the various social webs of people and impact you.
They won't be fitting into your absurd reductionist take on human behavior.
Well, I don't have a subscription to Wired, so it would seem I can't read this, but the lesson I draw from this is that no one person should be so important to your company's value that them making human mistakes drags the stock value down that much. No single human is that smart or that valuable. It's the job of boards to prepare for eventualities like this. Nvidia was in a better position than most to be able to do something about Huang's conduct, but now they're having to make up a succession plan as they go along.
If you own shares in a company like that, you need to evaluate how that can be changed - either your ownership of the shares, or the corporate governance of the company.
Well before any of this happened one report that ranked boards of large publicly traded companies had NVIDIA’s board 609th out of 650 or so, by far the weakest out of peer companies. Their board tenure is more than twice as long as comparable companies (which is not a good thing), and the members do seem to cater towards Huang - such as one of his mentors being on it, who also owns a company being shoveled billions by NVIDIA.
I don’t think the report and rankings should be taken as gospel but it’s an interesting read to consider given the clear failures of the board.
> It's the job of boards to prepare for eventualities like this.
The board is the one who ousted him! I agree that they should have been better prepared, but from the outside this looks like a long-running power struggle. Jensen didn't accidentally wind up with 60 reports. The board finally found an opportunity to correct the issue.
To me, this smacks of having to cater to Huang's desire to be the "special boy". The board had actual power over him (a thing that many founders have unfortunately learned since to avoid) but at the same time, they never actually made him sit down and set up a succession plan, and never set up a board president/chairman until last week.
If they had forced him to do either of those things, it might have suggested that he was... replaceable.
It shouldn't, but if you look at the filings of both Meta and SpaceX, they both say that a CEO search would take an indeterminate amount of time and be unlikely to satisfy the value proposition of the business. Patrick Boyle just had a video about this on the YouTube.
What we have here are a bunch of boards that can't do anything, won't do anything, or both. SV founder culture is a bunch of people who believe themselves to be the special boy and being completely unwilling to deal with anyone who might tell them otherwise.
And here I thought the lesson was control 51% of the vote like Zuckerberg, Karp, and Musk do, before letting your freaky flag fly.
But as this situation evolves, the spying allegations were BS which is why nothing moved on it for a year, and it looks like they need him back to manage Asia in exchange for tweaking the org chart and coming up with a succession plan. That org chart that led to his shenanigans is the same one that let him adapt faster to the moving target of AI than any of his peers. Good luck figuring out the best path forward here.
> And here I thought the lesson was control 51% of the vote like Zuckerberg, Karp, and Musk do, before letting your freaky flag fly.
That's the lesson lots of aspiring founders with untreated personality disorders will take from it, yes. The ones who don't are the ones who realize they're human, prone to mistakes, and will want to protect the legacy they've built with a system that constrains them.
Also, can't wait for the downvotes on this one. Every time I suggest that boards should have actual control over the c-suite, particularly founder CEOs, I get downvoted.
Why would I downvote you when I broadly agree with you in principle? It's just that reality gets in the way and you can either LARP the game you want to play or play the game by its rules.
Sorry, I should have clarified. I didn't mean to blame you for the downvotes.
Whenever I post something on here suggesting that perhaps we shouldn't give single individuals complete control over corporations worth trillions of dollars, regardless of whether or not they founded the company, I see downvotes. It makes sense; HN is the discussion forum for YC and my thoughts directly conflict with a sizable portion of the user base here, many of whom want to be the next Steve Jobs, just without the years in the wilderness.
https://archive.is/yclaB
Remember when GPUs were the dominant tech to mine crypto, then ASICs and other spepcialized purpose-built hardware came out and made GPUs irrelevant for mining, wont that also happpen in regards to ai training/inference. like aws is attemppting to do with their trainium/inferentia chips. Sure they suck now, but won't they get better and thus lose the dependency on GPUs?
It's already happened hasn't it. Google's TPUs are apparently much more efficient at inference workloads. However the hardware isn't available to the public like Nvidia chips are.
I'm unsure about the energy efficiency of Tenstorrent cards, but they are at the moment much more cost effective than Nvidia. But software support is far behind, I suspect if it wasn't the cost would be more equal.
Modern NVIDIA cards are essentially TPUs - especially those spec'd out for AI data centres. TPUs aren't anything particularly special - they are just large vector processors, with hardware specialised for the data + instruction fetch patterns that are seen when running AI models. The graphics cards that NVIDIA sells for data centres are just large vector processors with hardware specialised for the data + instruction fetch patterns that are seen when running AI models.
Google has two different types of TPUs now. One for inference and one for training.
Do we actually know they're more efficient per watt or for the power used and heat produced ? Or are they just easier to scale (and manufacturer) en masse ?
There's also the CUDA moat which is hard to get around
Cuda is a moat for traditional data science, where you have a torch model and you want to easily run it on a GPU. CUDA is the happy path for doing this.
Nvidia's moat for LLMs isn't Cuda, it's installed capacity, arrangements with frontier labs, TSMC bandwith, and just the fact that it currently has no real competition.
"There's also the CUDA moat which is hard to get around"
Is it really, though? Porting to a new architecture is an ideal use case for LLMs.
clearly they did not read the article
Inference has two steps, prefill and decode, with very different requirements on the compute to memory-bandwidth ratio.
Already these are run on different hardware, at least as I understand, and the computational needs could further drive differentiation.
there's also stuff like Talaas, where the ASIC is the actual model weights
Which is now owned by AMD
https://www.forbes.com/sites/jonmarkman/2026/08/09/amd-buys-...
Are they modifiable after creation? If not that seems dangerously rigid until we reach a much more stable place with LLMs. It could be great for well worn architectures and models but building one now for LLMs seems like it'd be foolish when it'll be outdated in a few months.
Why would it be dangerous? It'd effectively be a stamped model at some point in time. The bleeding edge moves on every few months, but even now it seems to me that there's only a tenuous value proposition that's progressing for me. I'd rather have an instantaneous GPT 6, Opus 5.3, or Gemini Hydrogen that I own, rather than always leasing models I have no practical application for.
no, and you're right that it's rigid. they call them "hardened models".
I'd gladly buy a $50 Opus chip pulling 5 watts.
I doubt they'd ever offer that until Opus was massively out of date if it's even possible to squeeze Opus into a chip at that power and budget.
The better question is, would someone be able to turn a profit offering a deepseek chip for a few hundred dollars? I wouldn't even care about the energy draw (within reason) having that running entirely locally at better than a few hundred TPS without going bankrupt to obtain VRAM would be a game changer.
DeepSeek Flash would require 1.3 trillion transistors. at 3bit weights so it would need roughly an entire wafer.
I think it would be on the order of ~1000 of them to fit opus.
Is 5 watt achievable?
Doesnt google themselves use nvidia chips. why do they need nvdia chips then?
Basically capacity issues. NVidia + Apple have bought up a significant portion of TSMC's manufacturing capacity. The primary supplier for Google's TPUs is Broadcom, with TSMC as a second source, and they're trying to ink contracts with other suppliers like MediaTek, but there're only so many semiconductor fabs in the world and TPUs need a fairly modern process.
Within Google itself the TPUs are definitely preferred for AI workloads, you just can't always get them.
Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not really the same markets. NVidia is selling chips that AI runs on, Google is selling services that use AI. Google's play is to make better services than the competition can because they control the full infrastructure stack, thus preventing their markets from being commoditized. NVidia's play is to commoditize the AI providers so that the number of services expands rapidly, with NVidia chips powering all of them. Arguably it's Google's strategy that is the market failure, as it's designed to keep competition out, but...
Google's used this strategy a few times before, and it frequently doesn't work out. They kept MapReduce and GFS proprietary so they could build better big-data services than everyone else, but Doug Cutting from Yahoo clean-roomed it, and then eventually you had interviewees going "MapReduce? That's like Hadoop, right?" They had a long string of internal ML frameworks well before the rest of the industry (I remember in the '00s and early '10s Google was by far the leader in large-scale machine-learning) but had to open-source TensorFlow in the face of external competition, and then lost in the marketplace to PyTorch because TensorFlow was filled with internal WTFs and footguns needed to support specific Google projects. They had a big lead in datacenter technology over everyone else in the industry in the early 2010s, but kept it internal and proprietary until they were forced to sell it as a service by the huge growth of AWS and Azure, then had to play catch-up as a distant 3rd because they lacked the customer relationships and developer mindshare.
> They're not really the same markets.
Both are fabless customers of TSMC, which is the choke point and center of the high-efficiency chip market.
Unlike gp, I do not believe its a market failures as much as the cutting-edge process node being a natural monopoly. Once TSMC processes got really good, TSMC started capturing most of revenue dollars and reinvesting them, which made TSMC products even better - a positive feedback loop that makes it impossible for Samsung or Texas Instruments to catch up. Intel rode on a similar wave for decades
Packaging is the choke point in AI. It's still TSMC but chip production is doing fine. It's the Chip-on-Wafer-on-Substrate build chain that is severly constrained.
https://www.insidedeeptech.com/tsmc-cowos-packaging-full-gui...
Why is no one going to Intel then for this kind of stuff ? Or even a legacy producer like IBM or Ti ? I would imagine they could buy some slightly older ASML litho machines and make something on a slightly older process (or package it using the older process for the other stuff and just keep the main processor latest gen) ?
From what I’ve read it’s that TSMC is just so much better at it… much higher yields and much more reliable than any of the other fabs. I wanna say something like 2-3 years ago, before Blackwell, NVidia tried to use Samsung for example but had nothing but issues. Intel hasn’t proven its fabs capable of anything really useful in nearly a decade.
I think even if you have the capital to acquire something from ASML… by the time it’s setup and tuned correctly TSMC will likely have enough capacity to eat your lunch. If it takes 2-3 years to get a fab online TSMC will likely have a new node and their old ones will be available.
TSMC has really proven over the past 5-10 years that the fab is its own beast and requires its own immense level of expertise and investment. When you have the worlds largest chip companies shoveling all their manufacturing dollars through one single org, that seemingly does the right thing, it’s going to be pretty much impossible to beat them.
By the time Hadoop came out, Google had already moved on to Big Table.
BigTable and MapReduce are differennt use-cases. The former is a database; the latter is a framework for running batch jobs. The (internal) successor to MapReduce is Flume (2010), which is quite a bit more capable, but didn't come out until Hadoop was already well-established. The external equivalent is Apache Spark (2009), but the gap between internal and external technologies had already shrunk to be nearly indistinguishable by then.
> Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not buying up "all the supply to build an inferior product". They're buying it up build a different, high-demand product.
TPUs are a systolic array. It's not comparable to the hardware architecture that backstops CUDA, nor do they serve the same markets.
The writing was on the wall for TSMC ever since the 5nm node; Apple proved that anyone with enough liquid cash could block competitors from fab access.
Modern NVIDIA GPUs contain tensor cores that are optimised for MMA operations as well though right? The standard CUDA execution path isn’t comparable, but the tensor core path very much is?
I'm no expert but I think we're dealing with two different ends of the spectrum here. Nvidia GPUs are likely cutting edge within the individual board / processor itself. Their main limiting factor right now seems in effectively networking them together (which is why they bought Mellanox or whatever awhile back to further develop Infiniband).
TPUs are significantly less flexible and I would guess at an individual level not as efficient or fast as Nvidia. But they're designed specifically to be bundled together in large arrays with similarly large data pipes to storage and interconnects to cache, ram and other compute. I would guess Google didn't bother trying to sell them until recently because you needed to be deploying them at large scale for it to be even worth also developing the software for.
What we really need is innovation in super fast (bandwidth and latency) and reliable interconnects. I suspect the days of slotted in GPU and RAM will be over soon and we'll see a move everywhere to pin based sockets (like we've always had for CPUs). Stuff like SXM and LPCAMM.
We've started to vertically stack ram out of necessity but have yet to really move on from the 2D plane for the motherboard itself. Slots allow more devices to connect to one flat board. But if the board itself had more surface area (by being maybe an L shape or box) you could slot in a lot more chips physically
It's logical that company like TSMC sells capacity to the highest bidder.
The main thing is that chip manufacturing is the most highly capitalized industry in the world. Each manufacturing step has gargantuan costs and so no company wants to manage (and pay for) all the steps - so TSMC buys chip manufacturing equipment from ASML and sets up a system where corporations buy capacity. It's essentially a very, very, very expensive print shop. The print shop doesn't make its own printers nor it print a newspaper.
And naturally, with very, very objects like leading-edge chips, "inferior" or "superior" is a ridiculous simplification. Nvidia chips are incredibly powerful and useful. Nvidia has been more committed than any other company to general purpose GPU programming.
But Nvidia also is doing it's best to sell at high, monopoly prices. That's why they buy a lot of capacity. The thing about the situation is these companies did spend a lot of money reaching this monopoly situation so from their viewpoint, this monopoly pricing might be fair. But I personally wouldn't care and hope some regulation stops it. But that kind of regulation would still have to get these companies significant profit on these investments, so they'd keep making them.
Interesting. What manufacturing process is Broadcom producing them at ? I'm guessing they're not newly as small (lithography wise) as the best of the best yet but probably good enough
Broadcom isn't a chip foundry so calling it Google's supplier doesn't make sense. TSMC has been the primary manufacturer of TPUs but Google's hoping to use Samsun and Intel also going forward (and these are the only leading edge foundries in the world). Broadcom is co-developer of the chip.
What node process do they need to be made on ?
I don't know details at this level but I know that TPUs like GPUs need to be made on the most advanced available nodes and TSMC, Samsung and Intel are exactly the only enterprises pursuing those advanced nodes. The reason is that this is
I mean, Google tell me it's TSMC's 3 and 2 mm process - exactly what you'd expect. But you can Google as well as I can.
Presenting broadcom as an alternative to tsmc is like wet roads cause rain. The average knowledge of computer hardware development cycle in hn comment sections is frustratingly low.
Broadcom is fabless I believe and therefore competing for that same TSMC space
Broadcom is just a chip designer or “mercenary” as I read the other day. TSMC is still the fab for most of the TPU, with Mediatek, and Intel used for auxiliary/supporting components.
google doesn't use them, they buy them to make them available for google cloud customers that require CUDA or other nvidia-only workloads. All training and inference at Google is done on TPUs.
> All training and inference at Google is done on TPUs.
No. Per Jeff dean.
> However the hardware isn't available to the public
It is available to Oracle.
https://www.oracle.com/news/announcement/ai-world-oracle-bri...
This may be, but look at Alphabet's share price:
> However the hardware isn't available to the public like Nvidia chips are.
Why? Don't they want to add another $3-5T of market cap?
You have basically managed to sum up why NVDA keeps dominating, no on can "just buy a few TPUs" and that has been the case since the first TPUs were announced years ago (yes yes GPU shortages and so on, but no one questions general availability of GPUs), and the current Tenstorrent model support page looks like an invitation to beta test HW that might (or might not) be able to run current and future models at some point.
But the next gen of Tenstorrent HW looks great on paper - problem is, I could have written that before the current gen was released.
All of the allegedly competing HW looks like that and has looked like that for years, either a lack of availability, capability, or both.
TPUs aren’t as effective at training.
Inference has already been commoditised.
Anthropic has probably paid around 12$ billion for training their models in total. Nvidia sales is > 200$ billion per year. So it's all for training?
Anthropic uses Trainium and Inferencia: https://www.youtube.com/watch?v=Bteba8KLeGc
https://d1.awsstatic.com/onedam/marketing-channels/website/a...
> Sure they suck now,
Some suck. Google's Tensor is an example of one that doesn't, that some consider superior to Nvidia's offering. There are likely others.
https://spectrum.ieee.org/inference-hardware-revolution
Relevent article from september 26. To me it’s obvious gpu are a « first response » solution. But I think nvidia can sell a lot before AI R&D slow down (because of the demand or the tech, infra, energy… or whatever reason it will be)
To be fair GPUs have already changed a lot in response to the AI demands. A lot of silicon these days are tensor cores which have very few applications outside of deep learning.
Inference, yes, see https://taalas.com/ and https://chatjimmy.ai/ As expected, efficiency boost is extreme. But you've got to pick the model you want to enshrine in silicon months or maybe years before you actually get that silicon (and maybe there are additional difficulties if the model is large. Maybe in a few years everybody will just want to use the same four or five models, with no tectonic shifts on the horizon. Then it'll be a no-brainer to make ASICs for those models.
Training? I'm not so sure. You need something flexible so that it can be trained. Burning the weights into the board is the opposite of that.
I read into Taalas when they first were putting out info and they also are wiring up massive CPU cache to the thing. It's not just the burned in models.
As much as I love the idea of Taalas, their chatjimmy demo is a hard-wired (and probably very quantized) 8b LLM producing garbage extremely fast (17k tokens/s). Extrapolating from the picture of their PCIe board in their demo, I would expect a frontier model to require an IC not smaller than my desk.
At these dimensions, you are better off producing general-purpose inference ICs that can run any model, instead of baking in a single fixed LLM.
So, Taalas (now AMD) will probably target small models with these (maybe something like qwen3.8-27b, which seems too large already), if at all.
The long term target is JEPA-like primitives so models become irrelevant. You only have to accelerate critical paths to see massive gains.
The difference is the algorithm needed to be embedded in an asic is static for Bitcoin. Models are changing too fast
It's very likely some 'good nuff' models will make it to silicon from some applications soon.
Something that could benefit from throughput/Wh but wouldn't benefit from flexibility and being cheaper than using bespoke chips feels like such a narrow set of uses though
It still won't make normal GPUs irrelevant like they did with crypto, there will always be demand for training and SOTA models while there is no demand for GPU mining
> there will always be demand for training and SOTA model
Some, but how much? Once you have a mass produced ASIC with a model that's "good enough" much (but obviously not all) of the demand for the cutting edge could evaporate quite rapidly I think.
Taalas is working on exactly that scenario.
NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto, but now it seems the will drop the entire customer graphics market long term they just haven’t committed themselves to it. They can’t go back to GPUs anyway there are not enough gamers to reach the revenue numbers they need now and soon gaming/graphics revenue will be irrelevant for them. Most US companies won’t maintain a boutique of products that don’t have synergy or are actively competing for internal resources long term they put all their resources where the margins and growth are best and for NVIDIA that’s not gaming anymore.
> NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto
But Bitcoin mining moved GPUs to ASICs many many many years ago and Ethereum switched from proof-on-work using GPUs to proof-of-stake: GPUs are pointless now for Ethereum. And Ethereum was the big one for GPU mining. All the others cryptocurrencies are too tiny to register.
Ethereum switching from proof-of-work to proof-of-stake has been the death blow to GPU mining at scale. I'm not saying some have not been profitable mining other less important coins: I'm saying the market for GPU mining shrank to next to nothing.
Nvidia has multiple B2B customers that require the consumer graphics featureset, like Nintendo. There's sufficient motivation in keeping the GPGPU/raster capabilities on the same card even if PC gaming dies as a market.
There are also the opposing architectural directions to extend toward:
one in the "Cerebras" direction of ever-larger scale general purpose HPC silicon running inference workloads. Cerebras may have skipped ahead to the last act of that scaling story, notwithstanding any further wafer-diameter enlargements past 300mm. Whether the ending is happy or not is still unclear. Lot of iterative room to grow into it further and they have a formidable patent moat waiting for conventional die-sized AI semi vendors when they inevitably run up against limitations of single-die designs.
the other is in the "Taalas" direction of freezing a specific model into something like high speed ROM / FPGA and getting orders of magnitude inference speed improvements - for that specific model. Like Cerebras, Taalas may also have jumped directly to the architectural telos of that philosophy, with only iterative improvements in cost per work unit left to grind, though like CBRS quite a lot of grinding left to do especially in getting cost per custom tapeout down to near-disposable (hopefully, recyclable) levels.
Both designs also massively increase processor cache and nobody talks about that.
But why wouldn't nvidia be expected to be a major successful player in that development?
It's Nvidia's 80% profit margin that's threatened.
Yes, agreed.
They didn't really get into it with Asics last time, at least. Are they looking into it this time?
I don't know. But my assumption is that they are. The crypto ASIC market was way smaller and less ambitious than this. It would have been just a relatively small cash grab for them to get into that market, whereas this market could represent an actual strategic direction.
Yeah, that's my thesis since the beginning of this madness. I see it as a temporary blip (albeit it's already much much longer than I would expect !) and once the bubble pops, Nvidia will be back to their normal cruising speed. Which is a good company, making great products, but their valuation will be 1/10 of what it is now. And if the boom continues, there is no way either that their clients won't develop a product of their own. It just doesn't make sense to buy chips by the 10s of billions $ on which Nvidia makes >50% margin.
So will people buy gaming GPUs fron Nvidia again after how if abused them this time ? Or will the AI obsession be their downfall & example for others what not to do and how not to alienate the customers who get you where you are now?
This is why I think their pivot/venture to general robotics was smart. However, I doesn't seem like they will become profitable on that front because they were outcompeted pretty hard in most measures I can conjure.
Yes. Half dozen companies are designing special chips to bypass/replace the GPU.
Bitcoin was lucky for Nvida, then AI was lucky. Third times the charm?
On other hand, luck favors the prepared. Beyond luck, they did make their chips accessible, easier to program than other GPU's. So bitcoin miners and then AI could utilize them.
There are still a lot of people that want to run regular liner algebra.
>Bitcoin was lucky for Nvida
Nvidia cards were the worst hashrate/purchase price of any cards. AMD were the lucky ones there if you consider anyone lucky in this situation. Consumers definitely weren't lucky because the GPU market hasn't recovered since the Etherium boom.
Interesting. I wonder if I'm accidentally swallowing some marketing spin?
I think things changed, in the early days of BTC AMD (ATI at the time) was king. Later GPU made a comeback with Nvidia rigs doing well during Ethereum era.
Maybe mojo will take off as an alternative to cuda?
Yeah. I wanted to mention Mojo, but that seems to rile people up. Not sure why exactly, but lot of Mojo haters around.
Mining ASICs are bound to a single cryptographic hashing algorithm, meaning they do one thing and they do it very fast. Something similar may make sense for running models locally, but I believe the more multifaceted Nvidia chips make more sense for training.
Taalas takes a similar approach but at a much higher level of complexity - instead of an array of nnn tiny specialized cryptographic cores running the same algo (or a mix of several with interlocking dependencies), it freezes a single entire LLM on to one die, running inference both far faster and cooler than any general purpose compute but with the same one trick pony obsolescence drawback. Question remains whether high volume fab tech can be wrangled into a form that allows this approach to drop unit costs near "disposable/recyclable" levels of cheapness per chip. This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
>This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
Yeah I was thinking this will probably be the case. We may eventually settle into a slower release schedule as well, where a new model only comes every year or two.
If you look at the sessions from 2026's "Hot Chips" there's like a dozen competitors and possibilities and new architectures and the like, yes. Some very compelling. And OpenAI and Google and others are also making their own chips.
The thing is, it's only NVIDIA really that has a proven product in high volumes pumping through TSMC with the existing software stack that the community has already proven. That's a huge difference. Everything else is riskier if you're building out in volume. For now.
Check out Taalas, recently acquired by AMD. They printed a model directly to silicon and it generates 16,000 tokens per second. You can even try a demo here:
https://chatjimmy.ai/
Whats amazing is that this is around 5x faster than the top end Cerebras wafer-scale silicon, at much less power/cooling cost per token. Of course Cerebras is a general purpose design, but perhaps Taalas could license CBRS wafer-scale patents to create a single monster chip that held either a single massive trillion parameter-scale model, or could an MoE design with a half dozen smaller specialized models on the same wafer with a shared SRAM pool.
I wonder if cooling such a beast, with very non-uniform hot spot distribution, would be too complicated to bother with though.
They were acquired for their ROM-like function block IP not ASIC stuff.
That function block was worth the cost of accusation.
You have any more info ? I was under the impression their use of processor cache like stuff was also their real advantage
won't
those Asics have a fixed algorithm
Cerebras already does that an need to make a new chip for each new model
that's why you couldn't make an ASIC for Ethereum
Cerebras don't need to customise their chips. To my knowledge they are crazy fast but struggle to fit larger models and it's not clear they're economically favourable.
Ethereum's work function was designed to be memory bandwidth limited, which is something gous are already quite well optimised for. An asic would be faster but not by enough to be worth the investment, at least while gpu margins were less crazy.
The main thing is the actual number crunching and memory interfaces for LLMs are not crazy hard to implement, and the potential scale is huge, so NVIDIA's moat is not likely to be big enough to make their current margins sustainable in the longer run.
Not to mention, many ASICs were outclassed by a few well-tuned GPUs in terms of power-efficiency. There were only a few coins that bucked that trend, usually just altcoin/shitcoin experiments that weren't worth mining in the first place.
I'm kinda bewildered, it's bizarre seeing people cheer on quickly-obsolete accelerators instead of demanding their OEM to adopt GPGPU. CUDA will be relevant for much longer than any ASIC can ever be, even (or especially) if the transformer gets replaced overnight with a better architecture.
That's because after bitcoin, most coins aimed for asic-resistance as a feature.
Yes, ASICs are application-specific e.g. "fixed". But, I think it's possible to target algorithm implementation that is inherently flexible, only limited by HW constraints (bit widths, RAM depth, etc) and thus cost. Yet, still fully pipelined and much higher performance than alternatives.
My background is ASIC design and more recently 3rd-party IP development for integration into ASICs. In our last design I made heavy use of resource sharing (RAM and holding regs for intermediary values, counters, LUTs, complex multipliers, adders, etc, even state machines).
That got the wheels turning. I believe such an approach can be generalized to support a wide range of computation and thus algorithm development. Even state machines can be configurable if designed that way. All controllable by programmable registers.
Just don't look for AI to take on such a task. Or go ahead and try, it already fails miserably when it comes to RTL design (verilog, vhdl). LOL
Yup. Pretty much. Any model can be a a hard wired circuit.
Post hoc rationalization. Nothing in here was unknown on 10/29 when the scandal broke, yet the stock market value was pretty durable before this and dropped a ton after.
It's common for a small story to break a perspective and cause re-evaluation.
That said, Nvidia is in a strong position.
Even in a slowdown, they probably would do ok.
Agreed, as the saying goes: In a gold rush you want to be selling picks. And they have the best publicly available picks.
Common saying. I always wonder how the stock market would have priced pick sellers? Would they have baked in 10 years of continued revenue growth from the sale of picks?
Look at Cisco. Only this year did its stock return to the price it reached at its peak in 2000. Since then the amount of people on the internet has grown ten-fold.
Until the gold rush stops (as it did for crypto, as it will for LLMs).
People will stop using LLMs like they stopped using the Internet in 1999.
It's peak delusion to think that AI is going away.
AI will not crash, but, a slowdown in AI could very well cause a general recession, precisely because it's he only thing holding the economy up.
That's a totally different thing though.
Caterpillar still makes lots of money selling shovels to gold mines. But they have a 15% profit margin, not Nvidia's 80% profit margin.
Excellent point. I would guess gold mining profit for Caterpillar is similar now than selling shovels to gold rush in 1849 though. Point being, Nvidia may grow revenue 5x during that margin compression.
Given that Nvidia currently has ~80% market share, that implies the whole market also has to grow 5x. And if Nvidia starts losing market share to their equivalent of Komatsu, Volvo, John Deere etc, then that means market growth of 20x or more. I don't see it. I don't think the world can absorb data center growth at that rate.
I don't think anyone here is claiming that LLMs will go away. However I do think there are probably a lot of opportunities for commoditization around LLMs, and if/when that happens it will signal an end to the "gold rush" days.
I think a lot of people do think somehow it will go away aka the individual I Was responding to.
I don't think demand for Frontier AI will go away either - even as in the near future 90% of AI will be 'commodity AI' - the Frontier foundries will still have competitive differentiation to offer.
Just like SaaS - technically almost all SaaS should be 'commodity' ... and yet it's not.
I think we can all agree that we are 'sick of the hype' but that doesn't mean the engine will diminish.
> It's peak delusion to think that AI is going away.
Even if it's not going away, it'll be a breath of fresh air when it's not shoved into every possible place as a marketing tactic, even when it's not necessary or beneficial. AI does have it's place, I just hope we realize soon how every place is not that place.
I think the ultimate role of AI is probably going to be "every place", just not... foregrounded.
It will follow the same pattern that technologies like phones, computers themselves, the internet, and (I guess) cloud storage now touch on nearly everything.
Example: even if you're an old-fashioned boutique leather artisan who makes handmade bespoke horse saddles in 2026 like it was 1783... your business also probably uses all that other stuff in some capacity. You probably have some computers for design and payroll and answering emails, a website that lives somewhere in the cloud, etc.
On a closely related note, if you're tired of hearing "AI, AI, AI" everywhere, there's reason to feel encouraged. Ten years ago, nobody would shut the fuck up about "the cloud." It didn't go away, it just matured and now it's how most people do many things.
I mean crypto is still here, but it's a shadow of its former self.
Just because LLMs are useful doesn't mean it will have a trillion dollar evaluation.
An economy built on a lie sooner or later crashes into the reality. The sooner the less damage it will cause.
I don't think LLMs are as world-changing as the internet, that is a very, very high bar.
They're useful certainly and won't be going away after any financial bubble bursts, whether they are so useful they need to be rammed into every facet of our lives remains to be seen.
One big difference between the internet and LLMS - people enthusiastically adopted the internet and found all sorts of uses for it. LLMs companies are mandating usage by employees and trying to force end users to use it, which seems really strange, why would they have to do that if it's so useful? They're giving it away for free at present.
The best usage I see so far (which people like, though there are problems) is to replace search, so it certainly could be profitable, though it would require google-like infrastructure as google isn't going to offer those search results for free to bots forever.
AI will have about 'on par' affect as the Internet.
Companies are 'mandating' use of AI because they see the potential in it.
In 1997, people were not sophisticated enough to grasp how broad the web would be - if they did - managers would have been pushing everyone to adopt the browser and email.
That said, in the 1990's they were actually pushing everyone to adopt 'Windows' etc..
We have a better grasp of how tech affects the economy, and so now there's a lot of hype. Probably too much.
AI is just in 'Phase 1'. The current Chat Bots are basically like the 'Commodore 64' - a 64K computer.
The hype is annoying - I use AI all day and find it useful and drives me crazy but at the same time 'it's real at the baseline'.
It will not 'explode' I think any time soon, it's going to take years, but in 5 years we will look back and see a huge amount of change.
In 5 years, up to 1/2 of AWS services will be AI based for example. All sorts of new features.
An example: RAG is being embedded into most DBs now, that's a big deal and taht's just 'embedding' not even really AI. In a few years, regular 'javascript jockeys' will be used to that kind of thing.
I'm by no means AI-pilled, and I'm skeptical of all these shenanigans ... but 'it's real'. This is happening on about the scale of the Web Browser.
I really don't think we can say what impact LLMS will have yet - AI is such a broad term and is much broader than the LLM craze, which is very specific and led to this dramatic demand for GPUs.
AI has already had a massive impact (machine learning to recognise pictures, spot patterns in data etc) and will continue to do so.
We're talking about LLMs here and the huge amount of money that has been pumped into them by companies like Nvidia in circular deals with their customers. Personally I think the jury is still out on whether LLMs will develop much further than they have. Comparing Opus 5.5 to Fable 5.2 this week I've seen little difference. We'll see in time though.
It does make me wonder if this is a concerted press attack by short sellers.
All the stuff about Jensen's affairs, NVidia sales to China, circular financing, increasing competition in the market, etc. was known months ago. Indeed, much of it came out in research notes by major short-selling funds. And NVDA is an extremely attractive target for shorts, having been bid up high by the AI bubble and then having a lot of retail interest that could easily be pushed the other way. Entirely possible that certain funds have taken a large short interest and are now trying to give it that push to profit.
I will agree it seems like a lot but
1) short sellers do not control the press. The best they can do is material leaks, like those 'Heisenberg' or whtever research companies, but you have to drop a real bomb for that
2) Trump is generally pro Jensen and most of the Cabinet and tech people are, and Dems don't have a specific fight with him
3) Jenses is not a likely face of 'tech evil' like Musk could be.
Most likely what you are witnessing is how ugly and nasty the press can be.
I've seen so, so many articles fro 'resepectable news outlets' with tawdry and slacious headlines - and almost no real information.
The 'timeline' of Mega Speed ? Had to go to AI to find it, and it's far more nuanced than what the headlines are saying.
The press loves waves and hype - it brings them revenue.
I think there is pent up anti-AI populism and the Jensen story was just the thing to pop first.
So he got it.
Also - the news just kept leaking aka 'yet another woman he slept with' so that kept the story alive.
There also might be things we don't know aka stuff that the Board has been made aware of that we don't know.
I think it's sad because I think Jensen is a reasonable and decent person that owes an apology to his wife, and not really anyone else.
I think it's an underreported detail in the english language press & internet that openai and anthropic are visibly lobbying to get the US government to ban open weight models or de facto make them difficult or impossible to use.
And Huang was by far the most powerful person outside of china taking the other side of that. I'm not going full conspiracy here, I genuinely don't think it was necessarily a coordinated takedown or anything like that. But if you look at whose public positions incentivize them to rile people up around this it is definitely worrying.
If you read amodei's remarks over the last few years he very clearly considers llm development to be a US vs china cold war 2 type deal that justifies dramatic action that would otherwise be morally abhorrent to him.
Now here's the conspiracy brain part. Say you're ideologically committed to impeding chinese labs' ability to develop open models, and willing to do anything short of literal bloodshed for it. We know tech companies consider manipulating public sentiment for their own purposes to be within the bounds of fair play. And in this hypothetical you have unlimited unrestricted access to claude mythos. What could you do with it?
> I'm not going full conspiracy here, I genuinely don't think it was necessarily a coordinated takedown or anything like that.
Unless it's some 4d chess tactic, the timeline doesn't have much sense, the ANT IPO is paying for this takedown.
And generally speaking, if you want to control you adversary you try to "leak" tech and information in order to control how much and what exactly he knows. Going full ban doesn't make any sense in the long term.
Epstein files, the American soldier who was arrested in UK for spying for Russia (https://www.bbc.com/news/articles/c6d7q77xgw0vo ), the blue wave, and I am sure I am missing a few others.
Nvidia's chip business seems certain to survive. They have also made themselves an "AI financing company" and the question is whether that enterprise could sink the Nvidia the corporation. Notably, Nvidia has "$530 billion in gross off-balance-sheet guarantees..." [1]. Nvidia total capitalization is 3.77 trillion and its yearly revenue is 200-300 billion. As long as things chugging along, the chip business can support the finance business but if you have a sudden shortfall, things may be tighter.
[1] https://finance.biggo.com/news/a10a2d57-9a11-43da-8600-4c5cf...
Which is very typical. It's one thing to recognize a bubble; it's another to predict when it's going to pop.
There were lots of traders who lost a lot of money betting against the housing bubble pop in 2005. The bubble was front page news on the Economist in 2005. Michael Burry made a lot of money because it took him a couple of years to assemble the money to be able to bet against the market and so he got lucky on the timing.
Or as the aphorism goes: the market can stay irrational longer than you can stay solvent.
Isn't there a bit of popular wisdom that goes like this? "If you want to time the market you have to get two bets almost perfectly right. When to get out and when to get back in"
I think it lost over 10% in the month before the news hit.
almost as if it's a bubble...
Or, the CEO who’s been at the helm during a long period of success has been fired, making future success uncertain. Perhaps?
The share price slide happened before the resignation. Uncertainty, sure, but if the only change were that Jensen had not yet resigned as of today, that uncertainty would still not be resolved.
Wasn’t the story being circulated in advance of that?
The story of the affair(s) and the Chinese spy were, yes. My point is that those stories and uncertainty seemed to have done most of the share price damage so far. The actual act of pushing out Jensen resolved some of that uncertainty. If everything had played out the same but the board had not pushed him to resign, I'm not sure things would be resolved any better currently or in the near future.
There were talks from the first hour whether Jensen would survive this. That he was fired came as no surprise to the market.
Many companies have continued being successful through CEO changes. Many others have not. Many successful CEOs have suddenly made some bad decisions.
Of course, you’re describing risk. Which is how players in the markets make investment decisions. Nvidia are now a more risky investment, until proven otherwise.
The markets saw Nvidia as an increasing risk before the actual decision to fire him, so that doesn't seem to be the single incident causing uncertainty. That risk and uncertainty already existed whether or not he resigned or was fired.
I never claimed it was a single incident. Risk profiles are an ongoing metric that take in myriad factors. Some weighted more than others. Losing the co-founder and CEO of 30+ years would have a heavy weighting, causing uncertainty until a new CEO has proven the ship is steady. Of course, the potential AI bubble is a factor too. But, losing your experienced CEO just before a potential popping of said bubble is also an extra factor to add to the risk metric.
Don't bother
Nvidia is down today on the rumor that Huang wants to come back! No future success is attributed to him.
People have claimed that competitors are catching up for a long time. But these journalists never understood why Nvidia was so dominant. And it seems they still don't. The Jensen story only made investors unhappy because of the high volatility and uncertainty. Noone who's well informed dumped Nvidia because of technological fears of competitors.
OTOH developers keep confusing R&D vs production. Nvidia is going to keep its dominance in R&D. But the vast majority of Nvidia chips are used for training and inference. Using Nvidia for that saves you 3 days of work porting from R&D to production.
Nvidia chips still have the best ratio of price to performance, so there’s not much motivation to buy anything else.
They have two years their head start is evaporating every day…
You can't buy Tensor or Cerebras as a consumer.
can’t buy nvidia chips that compete vs tensor/cerebras as a consumer either
It’s over if they don’t put a technocrat in charge a bean counter isn’t gonna keep them on top. See Apple when Jobs was fired or the clown lineup of bean counters at Intel Intel…
Apple seems to have done well under Tim Cook’s leadership, though.
Maybe they wouldn't be the laughing-stock of AI if someone better than Tim Cook was leading them. Siri is pretty bad, "Apple Intelligence" is a misnomer.
Apple is currently the most valuable company in the world and unlike many of the companies that aren't "laughing-stock of AI" it actually has a defensible moat for its products for the foreseeable future.
I'm generally not an Apple customer at all (I own an old Mac Mini to do occasional iOS development, but rarely use it), but think they are a better bet long term than any of the AI labs.
Then why has it been trading at such low levels, since even before the Jensen stuff?
It's forward P/E has been well below market average for a while now.
Fears of the bubble have reduced prices significantly across the sector for quite a while. Micron has a P/E of something like 5, Super micro as well.
Also for Nvidia, when a stock reaches number one and gets up to a few trillion market cap, there is hesitancy, same happened to Apple at times.
Nvidia biggest problem going forward they aren’t as versatile now essentially a one trick pony and now that they don’t have their CEO, they will be rudderless. That latest team up with Microsoft was a dud and a bean counter CEO isn’t gonna be able to fix that.
https://www.reddit.com/r/NvidiaStock/comments/1xvefbv/before...
All the more reason to find the board failed it's basic fiduciary duties.
To be fair, ram really is cyclical, and a commodity.
The Nvidia ecosystem… not so much. Most all of these pieces fail to recognise their already dominant position in the edge inference space - and most all of these pieces fall into the trap of AI=LLMs=ChatBots, and really, who wants to chat with a computer any more.
I don’t mind. This has been a great week to stock up on the equity, and there was enough blood in the water before the salacious spy sex story that I bought a bunch of very, very cheap $150 puts, as hysterical overreaction is the plat du jour, which I sold on Thursday.
RAM is not that cyclical, and booked solid for a couple of years at least.
"Nvidia did, however, acquire a startup making specialized chips, called Grok, in December of 2025."
I guess they mean Groq. Anyone know what nvidia did with groq btw?
Groq has been renamed to LPX and it's on the roadmap.
Disclaimer: I was at Groq and now work at NVIDIA on LPX.
Specifically: https://www.nvidia.com/en-us/data-center/lpx/
We ran a pure-LPU demo for HotChips that AA benchmarked: https://developer.nvidia.com/blog/how-nvidia-groq-3-lpx-unlo...
TLDR: LPUs are very efficient, quite a lot more efficient and faster tok/s/u than anything else at this years Hot Chips.
It's my personal opinion that with the way models change you don't want silicon that is too specialised but it needs to be specialised enough. LPUs are designed to strike this balance such that they remain flexible if model architectures change drastically.
In my personal opinion as well "specialized enough" chips are exactly the right way to go - though disclaimer, also At Nvidia, working on spec dec/external drafts :)
With things like https://developer.nvidia.com/blog/co-designing-ai-models-usi... you can specialize one chip for decode/drafting, and the other for prefill and draft verification.
Is LPX only going for rack size integration or is the plan to offer an LPX companion chip to something like a 6000 Blackwell ?
Could see a board with something like two SXM slots (maybe or similar) and one being for LPX optionally be cool
What's your opinion of high bandwidth flash? It seems ideal for storing and retrieving model weights. Do you see that dramatically changing inference chip design over the next couple of years?
Media are basically a reporting machine trying to get views. I have known that for quite some time, but this is another experience I will remember.
On the article, most of the tech discussions make zero sense as many of the comments already pointed out. Some of those challenges is also one of the reason why Nvidia has accelerated the whole roadmap and doing iteration of GPU every 18 months of so to stay competitive. Capitalising on leading edge node early and securing larger wafer quantities than even Apple. Large commitments so TSMC is also building Fabs faster than ever. The whole thing filtered out to all the downstream supply chain. Financing it as well making sure the capital is well positioned rather than forcing their partner to raise capital on additional capacity. Same playbook as old Apple. And in case anyone is wondering, it is Nvidia that is now pushing and funding the development of leading edge node along with all the other leading edge tech. Without them, TSMC A14 wouldn't happen as early. PCIe 6.0 would be a year later, Optical interconnect is 2 years early if not more.
It is that Datacenter are put on halt that is the biggest concern. And that didn't happen before "Jensen's mess".
This "trouble" is just silly, and mostly misdirected
(a) making chips is not making datacenters (esp. the memory fabric). Nvidia is working up the value chain (racks, datacenters), backwards (by supporting customers), and sideways (by lowering the cost of related technologies). No one else is close.
(b) Nvidia profit margins are so huge they can destroy competition
(c) The main cited competition is from their main customers, who are just trying to get a better deal from them.
However, the China trouble is real: Huawei is making chips for model-building and inference that the Chinese labs are happily using (to copy models?) that are good enough to dump models and chips to make a monopoly. Stirring up anti-tech, anti-datacenter concerns at the start of this race makes that a real concern, hardly mentioned in the article. It is not possible to face down a cheaper substitute with infinite resources.
Not good that people are stirring false worries when there are real concerns.
I don't agree with this.
One of Tim Apple's [sic] big innovations was to take Apple's giant pile of cash and use it for vendor financing. Building a new fab or process or factory is expensive. Borrowing money is expensive and risk. But Apple gave favorable terms in exchange for a period of exclusivity to whatever the vendor produced. This was huge for the iPhone for years.
NVidia is kind of in that position now. One can say that there is circular financing between NVidia and the big AI companies but is it really that different to Apple's approach? I'm not entirely convinced that there's an imminent domino effect collapse coming as a result of all of this. And, even if it does, I don't see it taking down NVidia.
Boeing (and Airbus) is the other company NVidia reminds me of. The business of selling planes is complicated. Boeing gets to throw its weight around and demand concessions. Launch partners for new planes get favorable delivery dates but that's also a chain. Qatar Airways famously went to war with Boeing and lost its A321 orders. That's the kind of power NVidia has now over the AI providers. It gets to decide who gets to buy and when.
As for competitors, there are always wannabees. Remember the decade of Google killers? I do. What happened to them all?
It surprises me on HN of all places how many people think you can vibecode a GPU competitor to NVidia, basically. I'm sorry but no. Chip design is harder than that. Managing relationships with suppliers, vendors, the fabs (particularly TSMC), etc is hard. NVidia has been doing it a long time, has a lot of built up expertise and has a lot of volume and weight it can throw around.
Jensen Huang was correct about China (IMHO). Restricting the chipmaking from ASML is one thing. Restricting the chips was a mistake in foreign policy terms. It has created captive domestic demand for Chinese chips. And it's too late to unring the bell. Xi Jinping realizes that chipmaking and AI is of national security interest to China. Yes there'll be some smuggling in the meantime but I fully expect China's EUV to be producing commercial chips by 2030. China simply doesn't want to be beholden to Western companies in the exact same way that most US tech companies were banned in China. I believe history will vindicate Xi's long-term strategic thinking in this area (yes, blocking the likes of Facebook and Uber predated Xi, I know).
I also think AI isn't going to be ASICed anytime soon like crypto mining was. Crypto had a fixed algorithm. AI does not, at least not yet. It's constantly changing. Even in recent years we've tried different quantization methods (eg fp4, int4). Model architecture is still evolving at a rapid rate. It's way too early to fix that into hardware.
It's in a bubble, prepare to have the market collapse.
Headline: Market goes down 20%
Internet: The end is near! The market is collapsing.
Investors: Well I was up 100% but now I'm up 80%.
> Well I was up 100% but now I'm up 80%.
That's bad math. If you had 100k invested, and you're up 100%, then you were at 200k. A 20% drop leaves you with 160k, and you're only up 60%.
You're right, I suck at math. Still up!
you still just lost $40k
12,000% here. And that’s after I bought a few houses.
But oh no, whatever shall I do. Oh, I’ll hold. That’s what I’ll do.
The press is hot trash. The financial press doubly so.
Polymarket says 7% and hasn't budged at all. Jensen looks like he will be returning as well, turns out it was just a coordinated attack from the media.
Trusting Polymarket on anything is a very unwise move.
It is a market, so I'm not really sure how this statement makes sense.
With all the money looking at it now, it's basically Goodhart's law in action: https://en.wikipedia.org/wiki/Goodhart%27s_law
There's a simpler explanation for the stock price drop here. The price had almost gone up 5x in five years, and bad news caused people to stop and take profits. You don't have to assume their business model is bullshit or that the AI bubble is popping or that there is fraud to justify a 10% pull back after a run up like that.
There's so much projection and wishful thinking about completely normal price movements in hyped stock.
> bad news caused people to stop and take profits
> You don't have to assume … that the AI bubble is popping
I’m not saying it is, but what is a bubble popping, if not a bunch of people saying “oh this might be the highest we go for now, I’m gonna take my profits”, which causes downward pressure for those who come slightly later who also want to take profits, etc?
The question is just if enough people stop to take profits to trigger that chain reaction. After all, it’s basically the same as saying that the seller thinks the current value in cash is more valuable than the future value of the company, otherwise people don’t really sell
Chanos calls this downside reflexivity. The feedback mechanisms which caused the rapid rise reverse and cause an accelerated downside phase.
Because people think the stock is worth holding 150 (or 140, or wherever it bounces from) but not at 180. It doesn't need to go _all the way down_. A normal overshoot and correction is not a "bubble".
> The price had almost gone up 5x in five years, and bad news caused people to stop and take profits
You just described a bubble...
A bubble would be if there's nothing backing it, it's not the same as the price temporarily overshooting, which happens _all the time_.
People who don't watch stocks desperately want to tell a story about what this stock price is telling them. If you watch tech stocks, this kind of thing is completely normal when a stock is on a bull run. Just look at the huge swoon nvidia had last year when it dropped from 140 down to 100.
Oh no help, I have one of the most in demand products in the world that everyone is paying a huge premium on. I am in trouble, please save me. Total horse shit. If you do a sensitivity analysis to a total collapse of data center demand it needs to be like a 75% reduction before its much of a problem to fundamentals.
The crypto bubble got saved by the AI bubble, but it appears there isn’t anything that’s going to save the AI bubble from crashing.
This whole thing just feels like different sides competing; no question there are "journalists" with short positions.
Not saying this author does. But senior leadership at Wired?
Everyone just using an obfuscated form of Elon's "GAMESTONK" strategy to boost their position.
Shorting is really really hard. And the end goal of shorting, sure for someone it's self enrichment, but why it's tolerated as a whole is to help identify and pop bubbles before they get out of hand.
Journalists plausibly have the same goal: speak truth to power to protect society. To be an investigative reporter you take on immense amounts of personal risk, and on very rare occasions get a big payoff in terms of recognition and career prestige.
So Wired is "short" Nvidia only in the sense that they benefit from pointing out its problems, and mostly only if they are right and get the timing right, otherwise they just lose status.
Wired is Soylent Green.
It's a legal fiction to insulate the real people from having their personal lives destroyed in court; they were just following orders type thing.
I never said shorting was easy; don't put words in others mouths.
Doesn't stop people from trying. This forum is naive af about a the world; there are billions of people who don't know you exist making choices that, like a spider-web, shake the various social webs of people and impact you.
They won't be fitting into your absurd reductionist take on human behavior.
Well, I don't have a subscription to Wired, so it would seem I can't read this, but the lesson I draw from this is that no one person should be so important to your company's value that them making human mistakes drags the stock value down that much. No single human is that smart or that valuable. It's the job of boards to prepare for eventualities like this. Nvidia was in a better position than most to be able to do something about Huang's conduct, but now they're having to make up a succession plan as they go along.
If you own shares in a company like that, you need to evaluate how that can be changed - either your ownership of the shares, or the corporate governance of the company.
Well before any of this happened one report that ranked boards of large publicly traded companies had NVIDIA’s board 609th out of 650 or so, by far the weakest out of peer companies. Their board tenure is more than twice as long as comparable companies (which is not a good thing), and the members do seem to cater towards Huang - such as one of his mentors being on it, who also owns a company being shoveled billions by NVIDIA.
I don’t think the report and rankings should be taken as gospel but it’s an interesting read to consider given the clear failures of the board.
https://jdrurypartners.com/news/weight-of-americas-boards-20...
> It's the job of boards to prepare for eventualities like this.
The board is the one who ousted him! I agree that they should have been better prepared, but from the outside this looks like a long-running power struggle. Jensen didn't accidentally wind up with 60 reports. The board finally found an opportunity to correct the issue.
To me, this smacks of having to cater to Huang's desire to be the "special boy". The board had actual power over him (a thing that many founders have unfortunately learned since to avoid) but at the same time, they never actually made him sit down and set up a succession plan, and never set up a board president/chairman until last week.
If they had forced him to do either of those things, it might have suggested that he was... replaceable.
They shouldn't have to force him and it shouldn't be adversarial. A $6t company shouldn't have a bus number of one.
It shouldn't, but if you look at the filings of both Meta and SpaceX, they both say that a CEO search would take an indeterminate amount of time and be unlikely to satisfy the value proposition of the business. Patrick Boyle just had a video about this on the YouTube.
What we have here are a bunch of boards that can't do anything, won't do anything, or both. SV founder culture is a bunch of people who believe themselves to be the special boy and being completely unwilling to deal with anyone who might tell them otherwise.
Corporate governance needs real reform.
And here I thought the lesson was control 51% of the vote like Zuckerberg, Karp, and Musk do, before letting your freaky flag fly.
But as this situation evolves, the spying allegations were BS which is why nothing moved on it for a year, and it looks like they need him back to manage Asia in exchange for tweaking the org chart and coming up with a succession plan. That org chart that led to his shenanigans is the same one that let him adapt faster to the moving target of AI than any of his peers. Good luck figuring out the best path forward here.
> And here I thought the lesson was control 51% of the vote like Zuckerberg, Karp, and Musk do, before letting your freaky flag fly.
That's the lesson lots of aspiring founders with untreated personality disorders will take from it, yes. The ones who don't are the ones who realize they're human, prone to mistakes, and will want to protect the legacy they've built with a system that constrains them.
Also, can't wait for the downvotes on this one. Every time I suggest that boards should have actual control over the c-suite, particularly founder CEOs, I get downvoted.
Why would I downvote you when I broadly agree with you in principle? It's just that reality gets in the way and you can either LARP the game you want to play or play the game by its rules.
Sorry, I should have clarified. I didn't mean to blame you for the downvotes.
Whenever I post something on here suggesting that perhaps we shouldn't give single individuals complete control over corporations worth trillions of dollars, regardless of whether or not they founded the company, I see downvotes. It makes sense; HN is the discussion forum for YC and my thoughts directly conflict with a sizable portion of the user base here, many of whom want to be the next Steve Jobs, just without the years in the wilderness.