Yes and no. They paid what everyone else pays for those gpus. NVIDIA make the profit and leaves crumbs for the rest. For other components they paid less, but since the gpus are the majority of the cost…
So recently I moved from a Anthropic model to a qwen 3.5 model running on my Mac to summarize ticket activity over 7 days. I used to do this manually with a colleague and it would take us a couple hours to go through. Opus took 58 seconds, and Qwen took 2.5 minutes. The quality of the qwen output was comparable, but the there was a 2.5x difference in time.
All that said I actually don’t think that matters much. I think we are dragging attention economy concepts in to ai responses, and it doesn’t matter. Both options saved me hours per week, and the difference between 3 and 1 minute may not be worth the additional cost.
Also there are times when the model output is much better with anthropic, but it’s not all the time. I think it becomes a question should we be using the best model for all questions?
It’s a very different company post the PwC purchase. They have around 1/3 of the revenue from consulting which tends to push the valuation down due to its relative low margin when compared to software. This also inflates the number of employees.
Just to level set here. I think its important to realize this is really focused on allowing things like search to operate on encrypted data. This technique allows you to perform an operation on the data without decrypting it. Think a row in a database with email, first, last, and mailing address. You want to search by email to retrieve the other data, but don't want that data unencrypted since it is PII.
In general, this solution would be expensive and targeted at data lakes, or areas where you want to run computation but not necessarily expose the data.
With regard to DRM, one key thing to remember is that it has to be cheap, and widely deployable. Part of the reason dvds were easily broken is that the algorithm chosen was inexpensive both computationally, so you can install it on as many clients as possible.
The large api/token providers, and large consumers are all investing in their own hardware. So, they are in an interesting position where the market is growing, and NVIDIA is taking the lion's share of enterprise, but is shrinking at the hyperscaler side (google is a good example as they shift more and more compute to TPU). So, they have a shrinking market share, but its not super visible.
Realistically groq is a great solution but has near impossible requirements for deployment. Just look at how many adapters you need to meet the memory requirements of a small llm. SRAM is fast but small.
I would guess their interconnect technology is what NVIDIA wants. You need something like 75 adapters for an 8b parameter model they had some really interesting tech to make the accelerator to accelerator communication work and scale. They were able to do that well before nvl 72 and they scale to hundreds of adapters since large models require more adapters still.
I would think this is for rental fleets or bike share. The weight and design would seem to make sense for that. Though the single speed seems like and odd choice for that.
The one vendor mentioned in the comments, AMI, is switching this code base to openbmc. Also it should be noted that often this software is system specific.
I think that the private carriers are more likely to be helped by this, since they will manage the paperwork.
It’s more likely a set of products that were shipping directly from factories disappears from the market. For example, the direct from factory Halloween costume.
It could end up being a step backwards in living standards and access to daily luxuries.
Gradual damage is consistent with over heating. I've seen racks of servers do the same thing.
Overall, there is a continued challenge with CPU temperatures that requires much tighter tolerances both in the thermal solution. The torque specs need to be followed and verified that they were met correctly in manufacturing.
It makes sense in any environment you have two workloads sharing compute from two parties, public clouds.
The protection here is to ensure the vms are isolated. Without doing this there is the potential you can leak data via speculative execution across guests.
I think it’s more pragmatic. We can eliminate hyperthreading to solve this, or increase memory safety at the cost of performance. One is a 50% hit in terms of vcpus, the other is now sub 50%.
More generally beats better. That’s the continual lesson from data intensive workloads. More compute, more data, more bandwidth.
The part that I’ve been scratching my head at is whether we see a retreat from aspects of this due to the high costs associated with it. For cpu based workloads this was a workable solution, since the price has been reducing. gpus have generally scaled pricing as a constant of available flops, and the current hardware approach equates to pouring in power to achieve better results.