Johnny Rose

B3 stopped being a gaming company in 2025 but we're still being incorrectly comped as one. B3IQ sold 8 figures of GPUs in its first two weeks. Since then, we've seen more than 9 figures in deployment demand for our modular data center. Here's a peak into what's changed and why we believe helping AI builders on their own AI infrastructure is at least a $100B opportunity. Through Andromeda Insights, we had a front-row seat to the GPU market. We made a bet that GPUs would become an emerging asset class, and that more AI builders would want to own the infrastructure their products run on. B3IQ is the business we built around that bet. The AI CapEx buildout is substantial. According to NVIDIA, the five largest hyperscalers are expected to spend ~2T over the next 2 years. Alongside them is a growing but underserved economy of AI builders running fine-tuning, inference, research and agents. They want to own smaller deployments of high-performance GPUs. Builders can rent compute or buy hardware, but bringing together ownership, financing, hosting and monetization still largely falls on them. B3IQ is built specifically for these builders. We bring procurement, financing, hosting and operations into one product. Builders own high-performance NVIDIA systems for their workloads. When those systems have idle capacity, we connect it to outside demand so owners can earn from the capacity they aren’t using. Each deployment also builds the supply for a larger product: a distributed compute network and matching engine. We are starting with the GPUs we bring online ourselves, then intend to open the network to other owners and facilities. Software will coordinate capacity, pricing, workloads and settlement across the network. Builders can be both customers and suppliers, using the compute they need and making the rest available to others. B3OS adds agents and automation that turn that compute into useful work. base:0xb3b32f9f8827d4634fe7d973fa1034ec9fddb3b3 remains the native token of the ecosystem. As B3IQ and B3OS scale, our goal is for base:0xb3b32f9f8827d4634fe7d973fa1034ec9fddb3b3 to become increasingly connected to the usage and economic activity these GPUs create. We are building B3 around AI builders: helping them own their AI infrastructure, earn from available capacity and participate in a broader compute economy.

How hard is it going to be to get compute in 2027? Harder than public markets are anticipating, according to @GavinSBaker on TBPN. He says securing a cluster as early as mid-to-late 2027 already means paying "vast amounts." B3IQ's customer-owned model is designed to get your compute online faster and put you back in control. You own the GPU servers. We handle provisioning, hosting and operations, so you don't have to run a data center. Instead of being beholden to a neocloud, be your own.

95% of the GPU capacity companies pay for sits idle. As per Cast AI's 2026 report, average GPU utilization across thousands of Kubernetes clusters on AWS, Azure, and GCP is just 5%. Most teams end up with more compute than they need, because: - GPUs are hard to get, so they hold on to what they find - Better rates require longer commitments, often with up to 30% paid upfront - When demand drops, they’re still locked in Meanwhile, many smaller teams are still waiting on GPUs. When you own the machine, you can rent out your idle hours to them.

Our engineer @0xaqt read 800+ sources and Anthropic’s brand-new IPO filing, then mapped every layer of the AI stack, from the chips up to the agents. Start with what runs out first in 2027.

Since 2019, US factory capacity for electrical equipment has fallen 9% while its prices rose 69%, per @a16z. Comms gear, semis, and computers all added capacity over the same stretch. a16z calls power one of the biggest bottlenecks in the AI buildout. Here’s how B3IQ brings capacity online against it: → About 6–8 weeks from order to a racked, running machine → About 1/4 the cost per MW → You own a dedicated NVIDIA machine, and we host and operate it b3iq.org

26 of 58 storage ideas, crowdsourced on b3iq, are about waiting for weights. the market asks to wait less, not to store more.

An H100 now rents for ~30% more than it did last fall. Newer GPUs are in the market, but demand is still ahead of supply. So even a chip launched before ChatGPT is holding its value, per @a16z. Compute is the new real estate. You can keep renting it, or own one and rent it out.

GPU rental prices went up 81 times in 60 days, per FastGPU's report. The biggest increases were on the chips used for frontier training and production inference: → B200: 7 increases (~17.7% each) → H200: 6 increases (~17.5% each) → H100: 8 increases (~14.3% each) → B300: 6 increases (~10.4% each) For a team renting 8 B200s around the clock, a single $1.20/GPU-hour increase adds roughly $7,000 to the monthly bill. And when supply is short, the provider controls the price, and you absorb the increase. It doesn’t have to work that way. With B3IQ, you own the machine and pay for it once. We build, host, and run it for you. And you can also rent out the hours you don’t need.

See how much each GPU can earn you and what it looks like down to its smallest component. Explore owning compute on our cloud at b3iq.org

If you have $500,000 invested and drive a $35,000 car, people think you’re broke. If you have $0 invested, $0 saved, and drive a $90,000 car, people think you’re rich. That pretty much sums up why so many people stay broke.

The B3IQ dashboard lets you easily manage, run and monetize your hardware remotely. Get bare-metal access for your workloads and even run models like Qwen3 directly from the dashboard.

In August, our cofounder & CTO @seangeng wrote about GPUs staying in service far longer than depreciation schedules assume. Six weeks later, the rental price of old chips went up. → H100 (2022) on-demand is up 16% vs a year ago → CoreWeave raised prices ~25% across every SKU in July and signed A100s (2020) through 2029 → Nebius raised H100 rates 17% on Oct 1, its second hike in three months → AWS raised reserved GPU prices 20% in July after 15% in January Most schedules write a GPU to zero in five or six years. The hourly rate on a four-year-old chip rose this year.

CME planned to launch futures on GPU rental prices on October 5, but the CFTC extended its review through November 9. The issue is that there still isn’t a trusted market price for GPU compute. Most pricing happens in private deals, making it difficult to build a reliable benchmark for buyers trying to hedge future costs. The fact that companies increasingly need financial products just to manage the price of compute says a lot about where the market is heading. For teams that need reliable access and predictable economics, there’s another option: own the underlying compute instead of continually buying exposure to the rental market. financefeeds.com/nvidia-gpu-…