My read on the inference report is that AMD inference chips surprisingly do not suck. Competitive until you go against GB. Far behind if it was tested against GB300 most likely. Imo, the biggest factor for inference in power consumption per token. Going to 2 nm is important for that.
This is one of the sharpest takes on the AMD-OpenAI warrant deal I've seen. Your framing of warrants as 'sneaky discounts' that distort gross margins is brilliant - it makes the whole structure much clearer. The logic that OpenAI with 30-300 FTEs working on inference economics knows exactly what discount is needed to make AMD viable is devastating and elegant. I also appreciate your intellectual honesty here - you have massive financial skin in the NVDA game and still acknowledge the AMD warrant deal is genius. The observation that AMD will hit $600/share as that's where the last tranche vests is a clever catch. My only quesiton is whether the warrant dilution over time will be material enough to offset the revenue growth from actually winning these hyperscale deployments.
Yeah, i'm glad im not the only one who looked at the AMD OpenAI deal thought it was pricey but very clever. should solve AMD's chicken/egg problem in the AI DC.
I wonder if this partnership will be as critical to AMD as the Sony one and it's very funny to consider the amount of playstation in AMD's AI accelerator IP since presumably the future of the accelerators is gfx13. Maybe amd has already abandoned udna, who knows, but hard to imagine when sony has literally spoiled so much of the feature set
If anyone from AMD waited or had scheduled selling some of their shares or vested options, they made a very handsome profit. The way AMD structured the deal with OpenAI was also IMHO very smart. The last tranche of the sale of AMD shares for one Cent each only happens if AMD common stock hits $ 600 a share (about four times of the current price); if that would actually happen, I doubt any regular shareholder would mind OpenAI getting their shares cheap.
The AMD-OpenAI deal structure is really clever when you think about the incentive alignment. By tying the final tranche to $600/share, AMD essentially bets on their own success while getting critical engineering feedback from OpenAI's deployment. What's particularly intrsting is the inference performance data - MI300 being competitive until you hit GB200 suggests the gap isn't as insurmountable as many assumed. The power consumption angle will be huge going forward, especially as datacenters hit capacity constraints. If AMD can nail the 2nm transition and leverage what they learn from OpenAI, the competitive dynamics could shift meaningfully.
Any view on AMD 2nm chiplet appraoch next year to Nvidia competing 3nm monolithic chips ?
On a back of perplexitty calculation a datacentre with same power and cooling can pack in a third more AMD chips each running 10% faster so 45% performance gain ? Real life wont be like that but it is a good sales point for AMD when power and datacentres are supply constrained.
Total amateur here. I would argue it is also about if compute is really long term under that high demand. That is what everyone says and might be true. But is it so 100% clear that there is unlimited demand? OpenAI has a lot of incentive to make sure everyone believes this. If it is true, then probably nvidia can go up more indeed. Would be curious about the opinion of irrational analysis in that regard :)
I know little about AI scaling but understand HPC scaling well. What it seems to me is that network scaling is breaking. It is not going fast enough.
So, NV is announcing 3000W GPUs with four dies and both companies have gone to massive racks to keep things as connected as possible.
There are ways to scale the compute but per chip power may already be close to what's feasible with GB200. I think Nvidia has actual wizards if they can put four dies together (2000W) plus another 50%, starting with a 20% per die increase with Blackwell Ultra.
Maybe they can mitigate this somewhat with CPX and further optimization, reducing the need for so many high power chips in the rack. That is what it looks like.
600 at 2030 is 20% cagr and 1T market cap.. you think Nvidia can’t hit 10T+ in the next few years given what we are in right now?You still bothered by market cap?
Specs can turn out to be vaporware so it depends on execution incl if scale up in 400 actually works. Assuming execution, it also depends if they have to give other scalers the same deal. IrrAn is the first person that thought of the AMD the same way I did. Instead of riling on dilution, circular etc., I saw it as a discount as well, say close to 50% assuming they price 400 at full price for openai. The logic was simple. AMD is discounting the 300 series by 40%ish anyway. Why not go higher to 50% and benefit from free engineering from openai and scaling down COGS? Now the last assumption on charging full price is an unknown. In the end, that would depend on TCO per kWh and that is where the technical execution comes in.
My read on the inference report is that AMD inference chips surprisingly do not suck. Competitive until you go against GB. Far behind if it was tested against GB300 most likely. Imo, the biggest factor for inference in power consumption per token. Going to 2 nm is important for that.
This is one of the sharpest takes on the AMD-OpenAI warrant deal I've seen. Your framing of warrants as 'sneaky discounts' that distort gross margins is brilliant - it makes the whole structure much clearer. The logic that OpenAI with 30-300 FTEs working on inference economics knows exactly what discount is needed to make AMD viable is devastating and elegant. I also appreciate your intellectual honesty here - you have massive financial skin in the NVDA game and still acknowledge the AMD warrant deal is genius. The observation that AMD will hit $600/share as that's where the last tranche vests is a clever catch. My only quesiton is whether the warrant dilution over time will be material enough to offset the revenue growth from actually winning these hyperscale deployments.
Yeah, i'm glad im not the only one who looked at the AMD OpenAI deal thought it was pricey but very clever. should solve AMD's chicken/egg problem in the AI DC.
I wonder if this partnership will be as critical to AMD as the Sony one and it's very funny to consider the amount of playstation in AMD's AI accelerator IP since presumably the future of the accelerators is gfx13. Maybe amd has already abandoned udna, who knows, but hard to imagine when sony has literally spoiled so much of the feature set
How does all this change your views on AVGO?
Assume the article title was supposed to be “inferenceMAX”?
If anyone from AMD waited or had scheduled selling some of their shares or vested options, they made a very handsome profit. The way AMD structured the deal with OpenAI was also IMHO very smart. The last tranche of the sale of AMD shares for one Cent each only happens if AMD common stock hits $ 600 a share (about four times of the current price); if that would actually happen, I doubt any regular shareholder would mind OpenAI getting their shares cheap.
The AMD-OpenAI deal structure is really clever when you think about the incentive alignment. By tying the final tranche to $600/share, AMD essentially bets on their own success while getting critical engineering feedback from OpenAI's deployment. What's particularly intrsting is the inference performance data - MI300 being competitive until you hit GB200 suggests the gap isn't as insurmountable as many assumed. The power consumption angle will be huge going forward, especially as datacenters hit capacity constraints. If AMD can nail the 2nm transition and leverage what they learn from OpenAI, the competitive dynamics could shift meaningfully.
Thoughts about Google's TPU?
Any view on AMD 2nm chiplet appraoch next year to Nvidia competing 3nm monolithic chips ?
On a back of perplexitty calculation a datacentre with same power and cooling can pack in a third more AMD chips each running 10% faster so 45% performance gain ? Real life wont be like that but it is a good sales point for AMD when power and datacentres are supply constrained.
MI355 3nm compare to B200 5nm, any performance gain?
What's the actual cluster utilization rate in inference for GB200?
Total amateur here. I would argue it is also about if compute is really long term under that high demand. That is what everyone says and might be true. But is it so 100% clear that there is unlimited demand? OpenAI has a lot of incentive to make sure everyone believes this. If it is true, then probably nvidia can go up more indeed. Would be curious about the opinion of irrational analysis in that regard :)
I know little about AI scaling but understand HPC scaling well. What it seems to me is that network scaling is breaking. It is not going fast enough.
So, NV is announcing 3000W GPUs with four dies and both companies have gone to massive racks to keep things as connected as possible.
There are ways to scale the compute but per chip power may already be close to what's feasible with GB200. I think Nvidia has actual wizards if they can put four dies together (2000W) plus another 50%, starting with a 20% per die increase with Blackwell Ultra.
Maybe they can mitigate this somewhat with CPX and further optimization, reducing the need for so many high power chips in the rack. That is what it looks like.
Is it reasonable to assume that AMD’s stock will rise faster and more sharply than NVIDIA’s, given that AMD is a smaller company?
600 at 2030 is 20% cagr and 1T market cap.. you think Nvidia can’t hit 10T+ in the next few years given what we are in right now?You still bothered by market cap?
Specs can turn out to be vaporware so it depends on execution incl if scale up in 400 actually works. Assuming execution, it also depends if they have to give other scalers the same deal. IrrAn is the first person that thought of the AMD the same way I did. Instead of riling on dilution, circular etc., I saw it as a discount as well, say close to 50% assuming they price 400 at full price for openai. The logic was simple. AMD is discounting the 300 series by 40%ish anyway. Why not go higher to 50% and benefit from free engineering from openai and scaling down COGS? Now the last assumption on charging full price is an unknown. In the end, that would depend on TCO per kWh and that is where the technical execution comes in.