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!
This one is interesting because it highlights the chip company as the data center operator as well (just like Google TPU). It's also behind a paywall, but even the headlines are eye-opening.
OpenAI recently struck an unusual deal to lessen its dependence on Nvidia’s AI chips and potentially lower its computing expenses in the coming years.
OpenAI has agreed to pay Cerebras more than $20 billion to use servers powered by the firm’s chips over the next three years, double the figure that was previously associated with the deal, according to multiple people with knowledge of the deal.
As part of the agreement, OpenAI will receive warrants for a minority portion of Cerebras’ shares, and that ownership could increase as it spends more, according to two people with knowledge of the deal. OpenAI has also agreed to provide chip designer Cerebras with around $1 billion to fund the development of data centers that would run its AI products, according to three people with direct knowledge of the deal.
OpenAI recently struck an unusual deal to lessen its dependence on Nvidia’s AI chips and potentially lower its computing expenses in the coming years. OpenAI has agreed to pay Cerebras more than $20 billion to use servers powered by the firm’s chips over the next three years, double the figure ...
New Jersey datacenter expansion got $77m in tax breaks to create exactly one permanent job — JPMorgan's site already scored $35m and currently employs just 25 workers
Isnt this the issue , a lot of breaks to create very little for the local community.
New Jersey datacenter expansion got $77m in tax breaks to create exactly one permanent job — JPMorgan's site already scored $35m and currently employs just 25 workers
Isnt this the issue , a lot of breaks to create very little for the local community.
Maybe and maybe not - a lot depends on what kinds of tax breaks are offered and what the prognosis for the area would have been if the data centers hadn’t come. I’ve been watching the SE corner of Wisconsin where I grew up and the storyline is really interesting.
* Kenosha and Racine are former industrial areas that have experienced substantial de-industrialization. They are fortunate enough to be between Milwaukee and Chicago so other jobs and services have blossomed and manufacturing is smaller, and unskilled, but never left entirely.
* Racine became the site of the ill-fated scam next gen Foxconn LCD plant that was supposed to generate a $10 billion manufacturing investment anchored by 13,000 high-skill jobs for a mere $3B in tax incentives plus another $1B or so in local infrastructure improvements. Trump helped break ground via a 2018 photo op on what was to be the “eighth wonder of the world”. The deal collapsed in increments, finally downscaled to a a $672.8 million commitment for 1,454 jobs — with no manufacturing requirement attached, and only $80M in incentives (plus the bill for the already done infrastructure improvement). Foxconn also continues to pay property tax guarantees on the upvalued property.
* In the spirit of “build-it and someone else will come”, Microsoft has jumped in to build a series of next generation Fairwater class data centers on ready-improved land with great industrial power, water and road infrastructure. The projects come with far smaller long-term employee numbers, but larger final property values at around $15B for current and future. Microsoft does get sales tax breaks on all of that $15B investment but will pay full property taxes, which essentially doubles the property tax revenue of the county.
I’m sure some day someone produce economic analysis of this messy progression to deter in whether there was ROI for the local jurisdictions and the state. For now, the data centers are a far better alternative than the remains of the Foxconn deal.
Microsoft intends to expand in a place where it's been welcomed, months after backing down on a rezoning attempt in an adjacent village when residents objected.
It's funny that Cerebras don't say anything about efficiency of their product. Only speed is mentioned. I assume therefore that inference costs more using WSE than using GPUs.
Yes, low latency (TTFT), high token rate (TPS) tokens command a premium if you can guarantee service level, especially for coding and agents where you need fast turnaround. That’s why I find the SemiAnalysis InferenceX benchmarks to be fascinating because they really chart out the tradeoff curves (there are other dimensions as well - model/model size/input context/output size, etc).
I view inference data center token generation similarly to electricity generation - you have heterogeneous generation facilities with different delivery characteristics - base load plants that deliver only slightly variable capacity very inexpensively and need to be run close to their rated capacity, peaker plants that are much more flexible and can be spun up quickly, but cost much more to operate. There might even be a place for storage and time shifting (or in data center terms, saving up overnight regular jobs to be run when cheap capacity becomes available).
In this analogy, Cerebras runs a string of peaker power plants for tokens. And it sounds like OpenAI is willing to fund more peaker plants for its needs.
www.theinformation.com
