152 comments

[ 3.6 ms ] story [ 92.2 ms ] thread
The best piece of media advice I ever received was that “anything you see in a large media outlet, was put there in order to affect somebody’s stock price, somewhere”

This is double true for a Condé Nast publication.

> A key role enabling the AI gold rush thanks in large part to its recently ousted CEO Jensen Huang.

Well - I think the article kind of conflates the problems here.

First - I don't call it "AI gold rush" but instead "AI mafia rush". After all the RAM prices have exploded. Why do we have to pay more because of these gangster companies? Why are governments not doing anything about it? But this is a separate discussion really.

Second - I feel that the "Jensen Mess" is a strange expression, because ... which mess exactly? Does the article refer to NVIDIA having problems? Or does it refer to Jensen and his unstoppable affection for multiple ladies at the same time? Or is it not because of the female non-spy (though, do we know that) but the female spy (though, do we know that)? How many more ladies are part of the Jensen story here? Only two? That's not a very high count. I'd expect more to be revealed here. (Youtube already has tons of Jensen-memes, so he coined these, indirectly.)

To me this reminds me a little bit of the issues with regards to Kevin Spacey. Now - the situation is not the same, because personally I think Kevin is a bit creepier here, so this is not 1:1 comparable. But Kevin also is a great actor, he coined numerous iconic scenes (L. A. confidential, Usual Suspects, or my favourite being his performance in American Beauty); and Jensen also was highly successful with regards to NVIDIA. Or, we can also take Michael Jackson - I do not think anyone doubted his skills with regards to art, music and expression (dance), but many may remember the awkward connection he had with some children. These situations are all quite different, but what exactly does the article mean with "The Jensen Mess"? The sex life seems to always be of more interest than other things. That was also the case with Bill Clinton.

// Why do we have to pay more because of these gangster companies? Why are governments not doing anything about it?

Because they're being supplied with the means to realise the Panopticon.

GCHQ in particular are chomping at the bit to be able to utilise all that lovely mass-data gathering they pulled from their fibre taps into Europe.

https://en.wikipedia.org/wiki/Tempora

// what exactly does the article mean with "The Jensen Mess"?

He facilitated a massive circumvention of a strategic export ban to China and played a key role in their ability to distill IP and Advantage from the US AI Companies.

His major gaffe was allowing the public a view behind the curtain to see Oz as human and fallible - and thus the custodian of $2 Trillion of market worth as unsafe. Vitalik Buterin ain't ever falling for a honeypot by comparison.

> He facilitated a massive circumvention of a strategic export ban

This lacks an "allegedly".

For now, even if I also believe that it is very likely that he facilitated a circumvention of the export ban, because that would have been profitable for NVIDIA, no evidence whatsoever has been presented publicly to support this, only pure speculation.

> Why do we have to pay more because of these gangster companies?

This is the free market; demand is high and supply is limited so price goes up. Don't like it? Start making your own and benefit from the higher prices!

It’s very likely the limited supply has been exaggerated deliberately by NVIDIA. That’s the whole point of the circular financing scheme they’ve developed over the past 3+ years
No, it's wearing the free market's corpse as a face. The incentives and mechanisms are perverted and investigations into things like price fixing are being launched though.
> This is the free market; demand is high and supply is limited so price goes up.

Free markets can be abused. Demand, in NVIDIA's case, seems to be not all that Jensen, and the company, had been stating. If the core demand was AI companies that were being funded by NVIDIA itself - then it wasn't organic demand.

If "... at least $200 billion worth of NVIDIA AI GPUs are sitting idle in warehouses after being purchased by cloud giants that lacked either the electrical grid capacity or immediate end-user demand to power them on" [0] is true then the demand wasn't there. There's zero reason to pre-purchase a GPU a year, even 6 months, ahead of build. Sure, maybe Google / Amazon / SpaceX buy during build - but these companies actually build their DCs, they don't buy GPUs and store them in a warehouse for quarters on end unless something went wrong (maybe like in Finland [1]). Jensen may very well be in a lot deeper than anyone really knows. The reporting as of late has stated that commitments by NVIDIA to memory suppliers was coming in and there was likely not enough non-circular capital to cover. The reality is Jensen likely thought he was going to be the AI kingpin, having a piece of all the AI pie and that he'd have enough time to recoup his ponzi-like investments in the downstream by paying for his own products.

I, personally, don't think we've even seen the start of this story. It's probably a huge mess starting with Jensen and flowing down through his kids. I can't imagine everything was on the up and up given the scale of dollars involved. It almost feels like the affair was a scapegoat to get out of the hot seat. Ultimately time will tell.

But, TL;DR - if you think the demand for NVIDIA was completely organic, well then I have a crate of Founder's Edition 5090's to sell you for dirt cheap.

[0] https://wccftech.com/jensen-huang-financed-ai-startups-absor... [1] https://www.reuters.com/business/environment/finland-launche...

> at least $200 billion worth of NVIDIA AI GPUs are sitting idle in warehouses after being purchased by cloud giants that lacked either the electrical grid capacity

People I know with adequate space and power have also heard these rumors and have looked high and low for parties with these billions worth of idle GPUs-- either to purchase their hardware or offer them hosting-- and cannot find them.

I think people are just making this up. If I'm wrong and you know someone in need, let me know and I can connect them with someone with 2GW+ of capacity.

The thing is, if it is true (I don't doubt it since the numbers don't add up and because NVIDIA has conveniently been unable to account for GPUs exported) nobody is going to sell them outright to someone else. Why? Simply because they never paid for them in the first place. If NVIDIA funds the startup to buy the GPU it would be odd to think that there aren't legal restrictions in the deal financing that would inhibit resale. Maybe Jensen wanted to be the Ticketmaster of GPUs?

Sure, it sounds simple but again, I don't think we understand any of the complex web of moving parts that are involved here.

This mindset has its flaws now as it often does. It is a high tech, high barrier field, so it is unlikely for companies to just pop up. Because of the scope of creating something that could rival Nvidia, it is risky and thus you will also not get many investors.

So, you will not have money and you are going into a field with big competition where all you have to go on is "well, I surely could do better." Not impossible, but that takes existing establishment in the field. You are not going to see mom and pop chip companies. Not without some major breakthrough that sets them apart.

An awkward connection with some children? That's your gloss on pedophilia?

Disgusting.

MJ was acquitted on child sex abuse charges. That doesn't mean he was innocent, but it does mean one should be careful labeling him as a pedophile.
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.
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.

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.
Until the gold rush stops (as it did for crypto, as it will for LLMs).
Look for pick sellers who are well positioned to pivot to something else. What could manufacturers do if video cards (or SSDs, or RAM) are no longer needed for AI?
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.
You want to be selling picks, but you don’t want to be buying pick sellers, because they’re selling at overly inflated valuations because of their one-off business.
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.

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...

I think it lost over 10% in the month before the news hit.
In the long term EMT is roughly accurate. In the short term, it doesn’t describe the market, which is prone to panic and delusion. It overshoots in both directions.
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.
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.

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 shield 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.

It's in a bubble, prepare to have the market collapse.
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.
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.

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 emerges, it looks like they need him back to manage Asia in exchange for tweaking the org chart and coming up with a succession plan.

> 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.

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...

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?
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.
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.
It still won't make normal GPUs irrelevant like they did with crypto
> 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.

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'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.

TPUs aren’t as effective at training.

Inference has already been commoditised.

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.

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.

> 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?

Presumably manufacturing capacity, and the desire not to provide hardware to companies that compete with their cloud business.
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

"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
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.

And deals with memory manufacturers.
Various articles recently claimed most of the NVIDIA deals are "handshakes with Jensen".

This Huang scandal has so many dimensions that are clearly more impactful on the world outside than to NVIDIA itself -- the blast radius seems to extend to the US cabinet -- that it begs questions if Jensen Huang wasn't id'd as the essential piece to remove to affect a comprehensive restructing in the segment.

So if deals fall apart because Jensen and his handshake have left the room, that may unclog many memory and processing supply lines that are currently spoken for by deals with Jensen. That may have been a goal from the start.

So the GPUs were a cherry on top, the real target was removing Huang? And Jensen huang'd himself by making himself the brittle link in a chain (his handshake deals being a big one).
CUDA is useful if you want a human expert writing new experimental kernels, perhaps while testing different algorithms.

But once you do that, translating CUDA into one of many available vectorized backends is trivial.

We have a bunch of custom kernels for geometric operations (like merging polygons) and they were rather difficult to write a while ago. AI rewrote them in HIP during my lunch break.

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.
The special part is not paying NVIDIA markup.
What's the status of OpenAI Grasshopper?
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

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.
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.

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

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.
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.
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.

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.

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.
>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.

Check out Taalas, recently acquired by AMD.
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
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 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.
> 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.

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.

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.
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.
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?
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.
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.

All of the hyperscale data center companies have their own AI chips and their own CPUs as well. They don't want to be dependent on Nvidia (or AMD or Intel)

They still buy Nvidia and x86 chips but this is what the WSJ article from last week was saying. The hyperscalers are moving away and Nvidia is committed to TSMC wafer orders so they are funding smaller companies to buy Nvidia chips in exchange for equity.

Microsoft

https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-...

https://techcommunity.microsoft.com/blog/azureinfrastructure...

Google

https://cloud.google.com/tpu

https://cloud.google.com/products/axion

Amazon

https://en.wikipedia.org/wiki/AWS_Graviton

https://aws.amazon.com/ai/machine-learning/inferentia/

https://aws.amazon.com/ai/machine-learning/trainium/

(comment deleted)
Apples and oranges, really. SHA-256 is a fixed function with no memory. So it lends itself to being hardwired into pure logic. A CPU or GPU was absurdly wasteful for it. That's why ASICs saw like 100x gains over general purpose hardware. Most of the repeatable tasks in AI workloads, like matrix and vector math at reduced bit widths, has already been encoded into hardware in modern GPUs.
"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.

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?
I think anything that doesn't compete with HBM for wafers will get it's day in the sun somewhere.

I don't know if HBF in particular is going to be huge but it's just so hard to predict the future in this space.

My personal opinion is we should be paying attention to SerDes and base dies to keep compute flexible and able to adapt to whatever memory and storage solutions are not supply constrained.

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

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.

Apple was supporting their suppliers so that they could buy more supplies from them. Nvidia is funding their own sales to customers.
> Chip design is harder than that

Chip design is not NVIDIA's advantage, or at least, it wasn't the advantage that "won" it the AI race. NVIDIA's advantage is the quality of the software ecosystem that surrounds the hardware. That is (arguably) a less secure moat.

Funding _suppliers_ to expand their operations is not that unusual; while Apple's maybe unusual in how much it does it, it's hardly the only company doing so (and it's also not new; literally centuries old). Funding purchasers to buy your stuff, outside of certain consumer businesses (notably cars) _is_ very unusual.
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.
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.
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.

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".

I don't understand. What is your gripe with the reporting? A lot of people denigrate reporting and reporters and news but it is still relevant and influential. Ultimately, a man who proclaimed loudly that all knowledge work would be replaced very soon, was himself ousted by a reporter
>was himself ousted by a reporter

Outsted by reporter spreading lies and unconfirmed information as confirmed news.

This is incorrect. What about the original NYT article was lies?
(comment deleted)
I no longer have trust in these provocative stories. First it was a spy, then not really a spy. Then it was no chips allowed, then it was well-allowed-really-ask-Lutnick, then it was CIA leaking, then actually Rubio-of-State-Dept leaking because of hawks versus capitalists.

Why don't we have these drip drip stories about Tesla in trouble before Musk impregnated a direct subordinate?

I no longer have trust in these provocative stories. First it was a spy, then not really a spy. Then it was no chips allowed, then it was well-allowed-really-ask-Lutnick, then it was CIA leaking, then actually Rubio-of-State-Dept leaking because of hawks versus capitalists.

Why didn't we get these drip drip stories about Tesla being in trouble after Musk impregnated women in roughly the same time, including a direct subordinate?

> “Everything sold today was designed two-to-three years ago,” says the CEO of a large AI company, who spoke to WIRED on condition of anonymity to avoid hurting business relationships.

This part is true for every big chip company. I've been designing large complex chips for 30 years. This is how long it takes from architecture, design, manufacturing, testing, mass production. Even just fab production is 4 months at TSMC from GDS delivery to finished chip.

> Kress, or whoever replaces her permanently, will have to chart the course that the chip industry will take two years from now, the CEO says.

That's true of every company in this industry. If you want a product to sell 2 years from now you had to decide on what it is last year.

It can take 3-5 years or more between design phase and production of automobiles. That's not a bad thing.
(comment deleted)