Ben Blaine

General Compute’s raise shows that a meaningful amount of capital is now willing to fund specialized inference hardware instead of treating general-purpose GPUs as the default. Their premise is that demand for fast inference will be substantially larger than demand for slower inference and is set to grow quickly. That is the same underlying condition that makes custom cores shaped around a specific model more valuable: once operators know what they are running and how often they are running it, the efficiency gap between general-purpose hardware and hardware built for that workload becomes larger and more worth closing. The round also demonstrates that buyers and investors are prepared to accept hardware that is not Nvidia and that is optimized for inference rather than training. That lowers the barrier for designs that go further in the same direction... cores that keep memory next to the circuits that use it, that only activate what is needed for a given model, and that can be rebalanced for power, speed, or context as the model evolves. Those characteristics only matter if there is sustained demand for high-volume, low-latency inference and a willingness to deploy non-standard silicon. Both are more clearly present after a $100 million raise at an $800 million valuation on that exact thesis. In practical terms, the more the market treats specialized inference capacity as a category worth scaling, the larger the set of customers who will eventually look for the next step in efficiency rather than simply more of the same chips. That is the step where embedding a model into purpose-built silicon, rather than running it on existing specialized chips, becomes relevant. The raise does not create that demand by itself, but it provides independent evidence that the demand exists and is being financed at scale.

Two of the most powerful people in chips sat on a stage this week and said, more or less, that AI is underwater. Hock Tan of Broadcom put the numbers on it. The world now spends roughly $200B a year generating tokens, and those tokens bring in roughly $150B of value. Most companies, he said, can't find any return at all. They're spending tokens the way a kid spends arcade money, and they're learning about as much. I don't think that's a model problem. The models are extraordinary. It's a cost problem, and underneath the cost problem is a power problem. Lip-Bu Tan of Intel said it plainly. GPUs are power hungry, which is why AI needs so much energy. He also said the next frontier is neuromorphic: designs that work more like a brain and don't need anywhere near as much power. Then Hock added the part I keep thinking about. A standard GPU is one size fits all, and one size fits all isn't the most efficient. Hardware built for one workload can beat a general GPU while drawing a fraction of the power. Put those together and the picture gets pretty clear. General-purpose hardware made sense while everyone was experimenting. But more and more companies now own their models. They've fine-tuned them on their own data and built products around them. Those models do one job, millions of times a day. Running that on hardware designed to do everything means paying, every second, for capability you never use. Most of the power doesn't even go on the maths. It goes on moving data back and forth. The brain doesn't work like that. It only fires what it needs, and it keeps memory and compute in the same place. That's why it runs on about 20 watts. I don't think the answer to "AI is underwater" is spending less on AI. I think the answer is hardware that's shaped around the model, instead of models squeezed onto hardware that was never built for them. The demand isn't slowing down. Lip-Bu said as much. The real question is who can afford to keep up with it. This is the problem I spend my days on at Aercrist, and I think the next few years will be decided by efficiency, not scale. So here's my question: If running your model cost 100x less, what would you build that you can't afford to build today?

Wow