What is Intel?

Anthropic's annualized revenue run rate exceeds that of Pfizer, Cisco, and Intel, and is $10B ahead of Uber and $15B ahead of Coca-Cola, despite the company being only a few years old.

“It's done more revenue than Fizer, Cisco, Kico Phillips, Intel, $10 billion ahead of Uber, $15 billion ahead of CocaCola.”

Intel is roadmapping process nodes at 1nm and 0.7nm, beyond the current 14A (1.4nm) node.

“We have the you know 14A is like 1.4 nanometer and we already starting to plan for 1 nanometer and 0.7 nanometer.”

Lip Bu Tan on AI compute demand being purely a supply problem.

“I don't see there anything to slow it down because the workload is increasing a lot and then I think the question mark is how >> we are supply constraint. >> We're supply constraint.”

Bittensor Subnet 51 achieved GPU hardware attestation without trusted execution environments by SSHing directly into GPUs, a novel alternative to Intel TEE-based verification worth tracking for decentralized compute security.

“Submit 51 figured out GPU testation you know trusted execution without trusted execution which is quite incredible. They don't even have tees which is what you know Intel does. They built the algorithms that could go in SSH into the GPU and check everything in the world to make sure that person has that computer and is not cheating.”

Lip Bu Tan explains that AI compute demand is not demand-constrained but supply-constrained, because inference and training workloads keep growing faster than capacity can be built.

“I don't see there anything to slow it down because the workload is increasing a lot and then I think the question mark is how >> we are supply constraint. >> We're supply constraint.”

The CPU-to-GPU ratio in AI training is expected to compress from 1:8 to as low as 1:4, implying rising CPU relevance in AI clusters.

“Versus one to eight in the training CPU to GPU now I can see one to four maybe one to.”

Kernel-level optimization yields only single-digit percentage gains on H100s (already heavily optimized) but 20 to 40 percent improvements on B40/B200, and 2x or more on Mac, AMD, and Intel hardware.

“If you start looking at think is like a B40, or a B200, you'll start seeing like much more significant gains to be had over there. you know, we have seen 20 30 40% improvements in performance in many cases.”

Lip Bu Tan on the US government becoming a significant Intel shareholder.

“I'm delighted you know US government become a big shareholder.”

CPUs outperform GPUs for reinforcement learning and multi-agent orchestration, countering the assumption that GPUs dominate all AI workloads.

“In term of reinforced learning in term of the speed of orchestrating all the agents and turn up the CPU is actually is better.”

Nvidia is testing Intel equipment for producing its next-generation Fineman chip set, with production targeted for 2028.

“Nvidia, meanwhile, is said to still be in the testing phase. Sources said that they're currently testing Intel equipment for producing the next generation Fineman chip set for production in 2028.”