3 ms·

They're booked out by the people making these bulk deals. Apple has significant weight to throw at TSMC, Micron, etc when it comes to negotiating both capacity

by 827a 9mo ago

They're booked out by the people making these bulk deals. Apple has significant weight to throw at TSMC, Micron, etc when it comes to negotiating both capacity and price; far more than the DIY manufacturers. The same laws apply to Nvidia, Quallcomm, and AMD. The sub-argument to this is that graphics cards will drive up fab prices for other packaged silicon products; this has probably been true for the past two years, but its very likely that we'll see this change in 2026. Even if theoretical demand stays high (which is debatable, but not for today): every major AI lab is sitting on warehouses of gigawatts of ready silicon, with nowhere to power them, so real demand will drop as the bottlenecks of data center construction and power delivery are solved. Those problems will take another few years. Even if these companies have the money and want to spend it: It makes zero sense to buy cards today, to have them sit in a warehouse for two years, when you can sit on the cash and buy newer-generation cards in a year. I would bet very real money that, sometime in 2026, we will see Nvidia reduce or cancel a committed fab order from TSMC.

bigyabai 9mo ago
I disagree with this for the most part. Nvidia's SOC designs have been industry-leading for almost a decade now, dGPUs are not their lifeline or their one-trick pony. They have a diverse base of customers that will schedule meetings with Jensen to re-allocate that fab capacity long before it gets cancelled. Apple doesn't, but Nvidia does. That's a real difference with the weight that Nvidia brings to the table, versus other customers. Not just the difference in capital, but the demand and diversity at play. There is Nvidia hardware in crop dusters and cruise missiles, datacenters and deep-sea SONAR. Apple kicked out their partners and doesn't even consider paying market-price when TSMC asks them to. They have the money to buy iPhone silicon at-cost, but the iPhone doesn't make enough money to compete with Raytheon's margins. Apple is smaller than you think, and all it takes is a team player to prove it. The performance side still won't add up in the favor of SOCs anyways. Distributed machines with high-speed interconnect run circles around the fastest Macs that an equivalent budget can buy. Benchmarks all show Nvidia being more power-efficient at raster and compute despite having the more complicated architecture. Nobody is ripping up their dGPU racks to install an SOC cluster right now, datacenters aren't putting their Nvidia cards on Ebay to buy more Ryzen blades. The opposite is happening really; SOCs are being fast-tracked into obsolescence to better handle heterogeneous workloads that homogenous SOCs can't do efficiently. Your initial claim ("they haven't made economic sense for a while") baffles me. SOCs would be cleaning up shop in the HPC niche, if they were any good at it.
- fc417fc802 9mo ago
  I think you've misunderstood what's being discussed here. AI demand for silicone has exceeded supply capacity and thus driven costs sky high. What's being debated is the projected impact on consumer electronics other than RAM and dGPUs - such as SoCs. GP suggests that AI hardware acquisition has outstripped the capacity to rack and power said hardware for the foreseeable future. If that's true we would expect orders to start getting postponed or even cancelled soon. It's an interesting hypothesis but I'm not sure I believe it.
- 827a 9mo ago
  The scale Apple operates at is sometimes hard for people to grasp. To present a small window into it: Apple did $73B in hardware revenue in the most recent quarter; Nvidia did $57B in total revenue. Now, consider that most of nvidia’s revenue comes from selling $30,000+ GPUs to businesses, with only a small number of TSMC-fabricated chips; whereas Apple sells $1000 phones. The scale of Nvidia’s contracts with TSMC, while increasing, doesn’t even begin to approach Apple’s. That’s why when TSMC began to open shop in the US, they did it with Apple. It’s true that Nvidia has a more diverse set of end-customers than Apple’s silicon. But, the customers you’re describing were the ones they had in 2020 when their revenue was $3B, versus the $57B they do today. The vast, vast majority of that revenue growth has came from less-than ten customers; the usual suspects. Revenue growth is a proxy for their contracts with TSMC. If the hyperscaler revenue growth takes a hit, the scale of that hit would outpace their ability to just shuffle fulfillment around (though, I have no doubt they will try) You’re also misunderstanding my initial claim, which might explain why you are baffled. I did not claim that SBCs make better economic sense than modular computers. I claimed that DIY computers rarely make economic sense over non-DIY computers today. Modular computers can be non-DIY; a quarter of every Best Buy is full of them, and it’s what most businesses in the HPC, CAD, etc spaces would buy their employees: a pre-build from Dell. Dell’s/etc relationships with suppliers grant them the same power laws that I explain fuel the SBC segment’s growth; power laws that the middlemen in the DIY space, like Corsair/etc, aren’t benefited from as strongly, and easily overwhelm with their own markup for sports car cooling fins and RGB.
  - bigyabai 9mo ago

by 827a 9mo ago

bigyabai 9mo ago

I disagree with this for the most part. Nvidia's SOC designs have been industry-leading for almost a decade now, dGPUs are not their lifeline or their one-trick pony. They have a diverse base of customers that will schedule meetings with Jensen to re-allocate that fab capacity long before it gets cancelled. Apple doesn't, but Nvidia does. That's a real difference with the weight that Nvidia brings to the table, versus other customers. Not just the difference in capital, but the demand and diversity at play. There is Nvidia hardware in crop dusters and cruise missiles, datacenters and deep-sea SONAR. Apple kicked out their partners and doesn't even consider paying market-price when TSMC asks them to. They have the money to buy iPhone silicon at-cost, but the iPhone doesn't make enough money to compete with Raytheon's margins. Apple is smaller than you think, and all it takes is a team player to prove it. The performance side still won't add up in the favor of SOCs anyways. Distributed machines with high-speed interconnect run circles around the fastest Macs that an equivalent budget can buy. Benchmarks all show Nvidia being more power-efficient at raster and compute despite having the more complicated architecture. Nobody is ripping up their dGPU racks to install an SOC cluster right now, datacenters aren't putting their Nvidia cards on Ebay to buy more Ryzen blades. The opposite is happening really; SOCs are being fast-tracked into obsolescence to better handle heterogeneous workloads that homogenous SOCs can't do efficiently. Your initial claim ("they haven't made economic sense for a while") baffles me. SOCs would be cleaning up shop in the HPC niche, if they were any good at it.

fc417fc802 9mo ago
I think you've misunderstood what's being discussed here. AI demand for silicone has exceeded supply capacity and thus driven costs sky high. What's being debated is the projected impact on consumer electronics other than RAM and dGPUs - such as SoCs. GP suggests that AI hardware acquisition has outstripped the capacity to rack and power said hardware for the foreseeable future. If that's true we would expect orders to start getting postponed or even cancelled soon. It's an interesting hypothesis but I'm not sure I believe it.
827a 9mo ago
The scale Apple operates at is sometimes hard for people to grasp. To present a small window into it: Apple did $73B in hardware revenue in the most recent quarter; Nvidia did $57B in total revenue. Now, consider that most of nvidia’s revenue comes from selling $30,000+ GPUs to businesses, with only a small number of TSMC-fabricated chips; whereas Apple sells $1000 phones. The scale of Nvidia’s contracts with TSMC, while increasing, doesn’t even begin to approach Apple’s. That’s why when TSMC began to open shop in the US, they did it with Apple. It’s true that Nvidia has a more diverse set of end-customers than Apple’s silicon. But, the customers you’re describing were the ones they had in 2020 when their revenue was $3B, versus the $57B they do today. The vast, vast majority of that revenue growth has came from less-than ten customers; the usual suspects. Revenue growth is a proxy for their contracts with TSMC. If the hyperscaler revenue growth takes a hit, the scale of that hit would outpace their ability to just shuffle fulfillment around (though, I have no doubt they will try) You’re also misunderstanding my initial claim, which might explain why you are baffled. I did not claim that SBCs make better economic sense than modular computers. I claimed that DIY computers rarely make economic sense over non-DIY computers today. Modular computers can be non-DIY; a quarter of every Best Buy is full of them, and it’s what most businesses in the HPC, CAD, etc spaces would buy their employees: a pre-build from Dell. Dell’s/etc relationships with suppliers grant them the same power laws that I explain fuel the SBC segment’s growth; power laws that the middlemen in the DIY space, like Corsair/etc, aren’t benefited from as strongly, and easily overwhelm with their own markup for sports car cooling fins and RGB.
- bigyabai 9mo ago