What is really happening in AI
1/ The real bottleneck in AI isn’t just GPUs.
It’s power, cooling, and data center architecture.
Hyperscalers own the entire stack.
Neo clouds mostly rent space + GPUs.
We’re building the layer in-between:
GPU-as-a-Service + vertically integrated
Exabits
2,048 posts
The Mothership of AI Compute: The Base Layer Everyone Builds On. Powered by Exascale Labs. 💪 linktr.ee/exabits
- Not all AI compute is urgent. Instead of "compute as fast as possible," we're asking: How do we meet grid constraints while still delivering on performance SLAs? Result? Power capping, intelligent workload shifting, and data centers acting as flexible reserves for the grid. AI
00:00Link to full TedAi video: youtube.com/watch?v=i3CA9o… - In the last 10 years, Moore's Law gave us approximately 100x compute. We scaled 1,000,000x. We'll keep doing it. Energy efficiency directly drives factory revenues and token costs. Hardware prices rise, but token generation gets so much faster- token costs drop an order of
00:00 - Why do HVDC + Solid-State Transformers (SSTs) matter for next-gen AI? Everyone focuses on GPUs, but as racks race toward 1MW with Blackwell & Vera Rubin, the real bottleneck is delivering massive, clean power efficiently. NVIDIA’s 800 VDC architecture boosts efficiency up to 5%,
- F.03, the first humanoid robot to enter the White House alongside Melania Trump, just created history. Serious AI compute powers advanced robotics from research to reality. Elite compute is needed to train, optimize, and operate these intelligent systems. Exabits can power the
00:00 - Power grid is overbuilt for rare peak days. 99% of the time it has massive excess capacity sitting idle. We should design data centers to flex: throttle down or shift load when the grid needs it most. Use the idle power instead of forcing endless new buildouts. Why isn’t this
00:00
02:25:58Here's my conversation with Jensen Huang, CEO of NVIDIA, the most valuable & one of the most influential companies in the history of human civilization. It is the engine powering the AI revolution. This was a fascinating & inspiring conversation, in parts super-technical on - NVIDIA's Isaac GR00T & Cosmos models redefining robotics dexterity at GTC2026. Simulate, train, deploy faster with Exabits' global GPU marketplace. Accessible compute, lower costs, and no vendor lock-in. Ready to build the next humanoid? Visit
- Live from #GTC2026 in San Jose, Exabits Co-Founder Dr. Hoansoo Lee <@hoansoo>, CIO Jonathan Jaranilla <@JJcosino>, and CMO Mark Fidelman <@markfidelman> on the ground witnessing the next AI era unfold- agentic systems, physical AI, massive Blackwell/Rubin scale, DLSS 5,
- Where's the power for exploding AI data centers coming from? ✅Solar/wind only half the time + batteries? Sketchy & pricey for 24/7 gigawatt loads. ✅Small modular nuclear: clean, steady, reliable-seems like the winner. @markfidelman asks- any OTHER clean solution (no gas/oil)
00:00 - Next week, don’t treat NVIDIA GTC (March 16-19) like “just another AI event.” That’s crazy. Mark <@markfidelman>, Exabits CMO will be there- look him up! Key themes already signaling: inference, agentic AI, physical AI (robotics), open models & AI factories. Predictions: 1.
00:00 - George sums it up perfectly: Organizations remain stuck until the entire factory is redesigned for Institutional Intelligence, but individual AI is like switching from steam to electric motors- huge personal gains. However, without massive, dependable compute at the base layer,
- Big Tech clouds charge more because they're racing to build the next ChatGPT themselves. Exabits? We're just in the business of selling compute- no distractions, no self-dealing. We're different: pure compute providers. No internal AI races. No scarcity games. Just high-end
00:00 - Kevin O'Leary on AI right now: If he were 25 today, he'd go ALL IN on implementation/execution for small businesses desperate to adopt AI (not just consulting) + data center development. Demand is insane- only 5GW under construction vs 30GW needed! At Exabits, we're building the
00:00









