@pequityresearch@burrytracker What the post doesn't explicitly mention is that over those years, Oracle also has to recognize $1.9B in Interest Expense on its income statement. So while top-line revenue and gross profit look higher at delivery, net income before delivery was reduced by the interest expense
@jukan05 The whole value proposition depends on yield and learning rate. TGV is the bottleneck. Neither the Korean or Chinese approach is better or worse. I wrote something about this earlier this year. Ultimately it will on a customer who is supply constrained taking a chance.
@jukan05 Historically suboptimal, but repriced memory $/mm2 could make it worth it at the right memory/logic area ratio. Has never worked in the past though
Think the AI capex surge will end in a classic bullwhip crash? Modern AI hardware operates by new rules. Custom silicon and balance sheet guarantees shift inventory risk away from stockrooms and directly onto corporate balance sheets. Our analysis below:
https://t.co/TBYfXzxh5B
@dwarkesh_sp Also, conflating AI performance with AI hardware utilization. Probably ok for coding. But does not extend uniformly across cognitive tasks.
Also conflating time regimes, productivity improvements and capacity additions can and probably will happen on different time scales.
@dwarkesh_sp If revenues grow 10x then total spend on training vs inference can go up without a change in percent. -% can go down and still be more than previously spent.
The rest of the conditions described only work in a hardware scarcity regime and that depends more on demand than supply
Some hard-tech failures are not killed by bad technology.
onsemi’s $20M NexGen in 2024 shows why stranded capability can matter: value proposition, market, and supply chain may catch up after.
It is scavenging, and it is possible to execute methodically.
https://t.co/Jc3T3rdu3X
@zephyr_z9 It is the same. You just have to look past the supply demand imbalance. The bottleneck in memory is the same as the bottleneck in power delivery. Same as the bottleneck in thermal management. Same as the bottleneck signal integrity.
Same themes; different walls.
@IanCutress I never could reconcile the hype about clearwater forest and glass earlier this year. Seamed like Intel was also playing up the ambiguity.
2030s seems bearish for glass commercialization given the substrate sizes on the roadmaps, but it could get scaled back like rubin ultra.
Nvidia’s move to assume financial risk for AI-era manufacturing changes traditional supply chain logic. Our latest blog, “The New Logic of Supply Chain Sovereignty,” outlines actionable steps for component suppliers. Read more:
https://t.co/MK2TsoMZb5
@firstadopter@jukan05 A better way to think about this is to ask if there was a fire at the datacenter and they lost all their A100s, would they ask Nvidia to make more of them to replace? Or would they just buy Blackwells? If the answer is buy Blackwell, then A100 should be fully depreciated already.
The Useful Life of AI Capital
Hyperscalers have imposed linear accounting on an exponential innovation cycle.
It is a decision that prioritizes earnings stability over the reality of the innovation cycle
https://t.co/ZkWMcHIkbr
#AIInfrastructure#Semiconductors#TechStrategy
@morganhousel
From Wild Minds blog post:
Naval once wrote:
The part of the person that we envy doesn’t exist without the rest of that person …
Who is Naval?