You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Inference needs low latency. Most inference will actually happen on the edge, IMO. Less important for chat bots, but very important for a lot of other applications.
Think about a self driving car or humanoid robot, you don't want your car to have to make a request to the cloud and have to...
You can guess based on the size of the round.
Most companies raise for 2-3 years of burn. If you call it 2, you expect that they are burning $20b a year. If it's 3 you are talking about $15b a year in burn. I think it's probably somewhere in between those numbers.
Losses were $5b in 2024...
Looks like my table was edited out and I don't want to do my calculations again.
The numbers were around 35-40% of cash from operations being investing into AI for MSFT and META, 50% for AMZN, and almost 70% for BABA.
Cash from operations also needs to be investing in other areas, and is used...
To a limited extent, yes. EVs use fewer parts and more of those parts are standardized. But I don't see the foundry model emerging here.
What I'm seeing instead is companies are building EV platforms, which is basically the battery, drive train, and electrical architecture, and selling a...
10-20% of revenue may not seem like a lot, but really these numbers should be calculated as a percentage of operating cash flow, which is accounts for how much money a company actually has to spend and accounts for the different margins and financial structure between these companies. For...
I think there is a difference between the foundry model and contract manufacturing, that's what I'm trying to get at.
I think that difference is in the design space, how compex/varied of a design you can build on a relatively fixed production process.
TSMC, if they are building a 3nm chip for...
What is "the foundry model?" in your view?
Is it just the separation of design and manufacturing? Or is there something else involved?
I think the foundry model is more of a manufacturing platform. For it to work you need standardized design rules and IP blocks that designers can work with...
Not really true, China started off by having local companies partner with foreign entities. This helped Chinese local companies build expertise and competence.
I think the US needs to rebuild it's competences in manufacturing and also learn a lot as well from foreign companies that have now...
Yes China is still stealing US IP, but they are also producing a lot of their own important IP as well. 50 Years from now if/when China is the global technology leader no-one will talk about how China stole a lot of IP in the beginning of it's modernization, just like no-one talks about how the...
Yeah I think it's a bit backwards looking. By definition if you want to be one of the biggest value destroying companies... you need to start off as a very valuable company.
I have started predicting this over the last couple of years.
Chinese semi companies are catching up. They may not be able to produce leading edge chips just yet, but a lot of applications don't really need them, especially if we are moving into the world of chiplets. And the slow down in...
A lot of people will look at a spreadsheet of costs and determine that equipment is the highest capex item... however they are missing a key point.
It's equipment depreciation that's the largest expense.
Two companies can by the same piece of equipment with a useful life of 10 years. The...
Fabs could be given away for free and this would be a great deal for Intel.
I have been calling for an Intel foundry split for a very long time, looks like it could finally happen.
Microsoft's ARM instances are Ampere are they not?
https://azure.microsoft.com/en-us/blog/azure-virtual-machines-with-ampere-altra-arm-based-processors-generally-available/
I'm not really sure on the specifics of Amperes licensing arrangement with ARM.
In mobile, ARM would sell design licenses as well as an architecture license, and I think design licenses could be more lucrative since they would generate higher royalties. In datacenter I think ARM would prefer...
His statement that it’s better to buy a bad company for cheap vs a good company that looks expensive is usually wrong.
Amazon has looked expensive by most valuation metrics pretty much its entire life as a public company, but has been one of the best investments anyone could have possibly made...
DeepSeek will be sensitive to topics China is sensitive to, ChatGPT is also sensitive to certain topics. Either way you are consuming propaganda by using any of these models.
I agree with the idea that by making AI cheaper it will ultimately mean AI will become more widely deployed. The question is which businesses will benefit and which ones will be disrupted. When the Japanese started making more fuel efficient cars, gas demand may have rebounded... but that...
