Logically, they probably resolved the issue with HBM yields on the nerfing. Now they need to assemble the GPUs with the newly qualified HBM and that is causing the delay.
It seems like they held on as long as they could to try and get that last contract signed with Elon before the de-spec. Then a month to make it seem like they didn’t know.
Real slimy, dirtbag operation, get the contract signed then break out the rug pull.
It will be hilarious to watch these chips fail at unprecedented rates in space. It seems like an amazing waste of money to launch a rack, then watch the GPUs gradually go dark. They have enough trouble with reliability on earth, now you are going to launch overclocked, fragile, warping, HBM yield issues into space?
teapaw.org/0xmitsurii/status/2001…
I guess the nerfing Nvidia did to Rubin didn’t do the trick.
They are really in for some shit now. How low can they go?
AMD doesn’t seem to have yield issues at 7.6 Gbps, Nvidia could try that. That will give them 15.5 TB/s bandwidth to 23.3 TB/s for MI455X.
Basically 66% performance on memory bound workloads, so inference. The future of compute.
Is anyone paying attention out there?
$AMD
I guess the nerfing Nvidia did to Rubin didn’t do the trick.
They are really in for some shit now. How low can they go?
AMD doesn’t seem to have yield issues at 7.6 Gbps, Nvidia could try that. That will give them 15.5 TB/s bandwidth to 23.3 TB/s for MI455X.
Basically 66% performance on memory bound workloads, so inference. The future of compute.
Is anyone paying attention out there?
$AMD
Taiwan’s KYEC lowered its Q4 guidance due to delays affecting a major GPU, believed to be NVIDIA’s Rubin, caused by HBM4 yield issues at a memory supplier.
finance.technews.tw/2026/11/…
AMD has its own proprietary packaging technology that only they will be able to deploy. EFB
This allows them to package more compute into a single GPU.
This leads to more compute per watt.
If something isn’t getting installed, it is an ASIC.
Limitations would only make AMD’s compute more valuable.
$AMD
Every so often Wallstreet wants to shuffle money around.
So they clobber the best ideas and prop up the shittiest ideas in an attempt to move money from the shittiest to the best.
AI is going away and we are all going back to using software, got it.
Some congressman said something. Good luck.
Okay AI is totally finished, now what?
Stupid never lasts, there are many more sunny days ahead.
$AMD
Everyone seems convinced that there will be less datacenters in the future. I guess Democrats in congress will be able to limit international datacenters too.
So now we should only install the most efficient GPUs on the market. Anything else will not provide the compute per limited watts.
$AMD
Nvidia gets roughly $500,000 per CoWoS wafer in annual revenue.
If you believe AMD gets half that at $250,000 then 530,000 wafers would be $132B in revenues for 2027.
Today we get our first report with Helios shipping.
Every quarter afterthis quarter we get more Helios.
$AMD
I am going to look back with fond memories on the shares I bought today when it is July 2027 and Helios is banging out billions in revenues while MI555X starts its ramp.
$AMD
It has been 508 days since $AMD introduced Helios at Advancing AI in 2025.
Tomorrow we see the first revenue hit the books.
Then every quarter after this day we see the revenue go up at an amazing rate for years.
We have come a long way, but we have only now reached the starting line.
Time to run.
What do hyperscalers think about Nvidia dropping the performance on them by 17% at the last hour?
Are they in love with Nvidia? They waited until after all of them had no alternative but to move forward, and then they dropped the specs.
They had to know this was going to happen, months ago. People might have made different choices months ago.
Nvidia is a piece of shit company.
They are really trying to piss the customers off intentionally now. Let’s see how that works out for them.
Everyone questions why Nvidia is so cheap. $NVDA forward PE is 16.
Investors are not as stupid as many think, how sustainable is growth expansion when you don’t have the performance lead?
Nvidia can’t just buy all their own chips, at some point customers get a say in which rack gets ordered.
A slow down in client / consumer demand for the market will provide infinite wafer availability to $AMD via TSMC.
Roughly 50-60% of TSMC’s 2 and 3 nm wafers are directed towards client and consumer electronics.
AMD will have any amount of wafers they need for server CPUs.
Reduced demand for client products will also free up memory capacity for AI servers.
I missed the point where $AMD lost market share here? Where AMD revenues fell?
$INTC sucks because they suck.
Did demand suddenly go down? No, demand is way up. You can’t get memory because demand is so freaking high.
Intel is crashing because they suck at making chips. You know who doesn’t suck at making chips, TSMC.
Intel’s incompetence is an opportunity for AMD. The more those fucktards bumble fuck around for another year, the more opportunities AMD will have to fill the demand they can’t execute on.
Aletheia is the best research house. Their industry intelligence is truly impressive.
Aletheia: Anthropic has increased its AMD GPU orders for 2028 to 3 GW—three times its original deployment target of 1 GW.
Aletheia: OpenAI has increased its total planned AMD GPU deployments over 2026–2028 from 6 GW to 8–9 GW.
I don’t know when $AMD is going to double, but if I had to guess, it would be sooner rather than later.
Not financial advice.
Perhaps the bigger fraud on this spec sheet is 3.6 exaflops.
No one uses sparce compute numbers on their spec sheet.
It is truly the act of a charlatan to try and pull the wool over the eyes of the smartest people on earth with the dumbest scam.
teapaw.org/0xmitsurii/status/2001…
$AMD Helios now clears Rubin NVL72 by a wide margin.
AMD trades at $0.854T while Nvidia trades at $5.18T.
It is fast becoming time to ask whether or not AMD will sell more AI racks than Nvidia in the not too distant future.
Helios provides vastly more HBM capacity, meaning more model instances over the same number of racks.
Now they will provide more throughput by a substantial margin as Nvidia continues to nerf their racks.
Why would hyperscalers continue to spend more on Nvidia to get less?
This is how we end up with Taalas based fixed weight models brought to you by $AMD.
Frontier models are now too powerful for general release. There are no new models coming folks.
We might spend the next few months making the existing frontier models more efficient to inference or higher RL capabilities, but more powerful is done.
Next we take the best models and etch them into silicon with Taalas based chips. This will improve model inference efficiency 50x.
Helios can handle prefill and Taalas can handle decode. This will smoke anything on the market and send AMD into the stratosphere.
two weeks ago Sam Altman defended open source AI..
now he says his best model is too dangerous to release..
then..
society will have to accept severe cyber incidents..
as the price of open source liberty..
now..
OpenAI can't responsibly open source its best internal model..
not until cyber defense and bio safety catch up..
to be fair..
he's backing open source in general..
just not at the very frontier..
but then he admits the scariest part..
the gap to an open model just as powerful..
is "not super" big..
this same week..
hackers used AI agents to attack hundreds of organizations..
the closed labs think they're buying the world time..
the real question is how much time is left
One thing not accounted for with Anthropic is their access to Superhuman Intelligence.
They could do something like double the efficiency of solar cells with new materials and get a 10% cut of every solar cell produced worldwide.
We won’t see that AI, it is too powerful for public release, but it doesn’t stop them from licensing who knows how many breakthroughs.